text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'The detailed abundance patterns of the stars within galaxies provide a unique window into the history of star formation (SF) at early times. Two widely used ‘chronometers’ include the $\alpha$ and iron-peak elements, which are created on short and long timescales, respectively. These two clocks have been widely used to estimate SF timescales from moderate-resolution spectra of early-type galaxies. Elements formed via $s$-process neutron captures (e.g., Sr and Ba) comprise a third type of chronometer, as the site of the main $s$-process is believed to be intermediate and low-mass asymptotic giant branch stars. The \[$\alpha$/Ba\] ratio in particular should provide a powerful new constraint on the SF histories of galaxies, in part because it is insensitive to the uncertain distribution of Type Ia SNe detonation times and the overall Ia rate. Here we present new measurements of the abundance of Sr and Ba in nearby early-type galaxies by applying stellar population synthesis tools to high S/N optical spectra. We find a strong anti-correlation between \[Mg/Fe\] and \[Ba/Fe\], and a strong positive correlation between \[Mg/Ba\] and galaxy velocity dispersion. These trends are consistent with the idea that more massive galaxies formed their stars on shorter timescales compared to less massive galaxies, and rule out several other proposed explanations for the observed super-solar \[Mg/Fe\] values in massive galaxies. In contrast, \[Sr/Fe\]$\sim0$, with no strong variation across the sample. It is difficult to interpret the Sr trends without detailed chemical evolution models owing to the multiplicity of proposed nucleosynthetic sites for Sr.'
author:
- 'Charlie Conroy, Pieter G. van Dokkum , and Genevieve J. Graves'
title: 'Strontium and Barium In Early-Type Galaxies'
---
Introduction {#s:intro}
============
While it is clear that the majority of massive early-type galaxies formed the bulk of their stars at $z>1$, the precise nature and duration of star formation in these galaxies has proven difficult to constrain. The difficulty arises from the fact that the main sequence turnoff point varies little with age at late times, implying that the integrated spectral energy distribution varies little with age at late times. For old stellar systems such as early-type galaxies, it has become common to instead rely on elemental abundance ratios to probe the star formation histories at early times [e.g., @Worthey92; @Thomas05]. The classic example is the ratio of $\alpha$ to iron-peak elements, \[$\alpha$/Fe\][^1]. The $\alpha$ elements form mostly in massive stars, which evolve on short timescales, while iron-peak elements form mostly in Type Ia supernovae (SNe), which occur on longer timescales. The \[$\alpha$/Fe\] ratio is thus a sensitive probe of SF on $\sim10^{8-9}$ yr timescales [e.g., @Tinsley79; @Thomas98].
One of the principal difficulties in using \[$\alpha$/Fe\] to measure a SF timescale is the unknown Type Ia SNe delay time distribution. Other difficulties include the possibility that the overall Ia rate varies in some systematic way with galaxy properties, the possibility of selective mass-loss such that Fe is preferentially lost from the system, and potential variation in the initial mass function [@Worthey92; @Thomas99; @Trager00]. A promising alternative chronometer is Ba, which is believed to form predominately within the envelopes of asymptotic giant branch stars via $s$-process neutron captures [@Burbidge57; @Busso99; @Herwig05]. The precise mass-dependent yields of Ba are presently not well-known, but there is hope that progress on this topic can be made in the near-term (in contrast to the delay time distribution for Ia’s, which appears to be governed largely by the initial conditions of stellar binarity). The prospects for measuring Ba in the integrated light spectra of galaxies are favorable owing to the strong transitions of [Ba$\;$[<span style="font-variant:small-caps;">[ii]{}</span>]{}]{} in the blue spectral region.
Sr is another neutron capture element with strong transitions in the blue. Like Ba, Sr is predominantly produced by the $s$-process, at least in the solar system. These two elements probe two of the three $s$-process peaks, with Sr belonging to the first (along with Y and Zr), and Ba belonging to the second (along with La, Ce, Pr, and Nd). While the nucleosynthetic origin of Ba is relatively secure, the same cannot be said for the elements in the first $s$-process peak [@Couch74; @Woosley92; @Raiteri93; @Sneden08]. Indeed, current chemical evolution models of the Galaxy are unable to reproduce the observed behavior of Sr-Y-Zr at low metallicity without appealing to exotic and/or ad hoc nucleosynthetic sites [e.g., @Travaglio04; @Qian08]. Perhaps the \[Sr/Ba\] ratios in metal-rich massive galaxies will provide some insight into this problem.
In this Letter we employ our new stellar population synthesis (SPS) model to interpret high quality optical spectra of early-type galaxies in order to constrain the abundances of Sr and Ba. In Sections \[s:model\] and \[s:data\] we describe the model and data, and in Section \[s:res\] we present our results.
Model {#s:model}
=====
The SPS model used herein was developed in @Conroy12a [CvD12]. The model adopts standard SPS techniques, including constructing libraries of isochrones and stellar spectra. Empirical spectra form the core of the model. The empirical stars are of approximately solar metallicity and have solar abundance patterns. We computed a large grid of model stellar atmospheres and spectra in order to construct response functions (i.e., the relative change in the spectrum of a star due to a change in the abundance of a single element). The model atmospheres and spectra were computed with the ATLAS model atmosphere and spectrum synthesis package [@Kurucz70; @Kurucz93], ported to Linux by @Sbordone04. Specifically, we use the ATLAS12 code, computing new atmospheres for each change in abundance pattern. The line list was provided by R. Kurucz[^2], including linelists for TiO and H$_2$O, amongst other molecules. The spectral response functions were applied to the empirical stellar spectra in order to create models with arbitrary abundance patterns. The model allows for arbitrary variation in the initial mass function (IMF) and spans ages from $3-13.5$ Gyr. See CvD12 for further details regarding the model.
We follow @Conroy12b in fitting the model to data. In its present form the model contains 27 free parameters, including the redshift and velocity dispersion, a two-part power-law IMF, two population ages, four nuisance parameters, and the abundances of C, N, Na, Mg, Si, Ca, Ti, V, Cr, Mn, Fe, Co, Ni, Sr, Y, and Ba, and O,Ne,S are varied in lock-step. These parameters are fit to the data via a Markov Chain Monte Carlo fitting technique. The data and models are split into four wavelength intervals (defined in the following section) and, within each interval the spectra are normalized by a high-order polynomial [with degree $n$ where $n\equiv
(\lambda_{\rm max}-\lambda_{\rm min})/100$Å; see @Conroy12b for details]. We have masked the spectral regions surrounding the H$\alpha$, \[[N$\;$[<span style="font-variant:small-caps;">[ii]{}</span>]{}]{}\], \[[S$\;$[<span style="font-variant:small-caps;">[ii]{}</span>]{}]{}\], H$\beta$, \[[O$\;$[<span style="font-variant:small-caps;">[iii]{}</span>]{}]{}\], and \[[N$\;$[<span style="font-variant:small-caps;">[i]{}</span>]{}]{}\] emission lines.
The effective resolution of the data varies from $R\approx400-1200$ owing to intrinsic Doppler broadening. Both because of this low resolution and the fact that in our model the synthetic spectra are only used differentially, several uncertainties associated with abundance analysis at high resolution (including the adopted microturbulent velocity, treatment of convection, non-LTE effects, and hyperfine splitting of certain levels) are mitigated to some extent herein. Moreover, @Bergemann12 has demonstrated that the strong [Sr$\;$[<span style="font-variant:small-caps;">[ii]{}</span>]{}]{} line at 4077Å suffers from negligible non-LTE corrections at \[Fe/H\]$\approx0.0$. Of course when working at low resolution one is more dependent on the fidelity of the model than the case where equivalent widths of individual unsaturated lines are directly measured from the data.
As an example of what can be measured from low resolution spectra, in Figure \[fig:response\] we show the change in the model spectrum due to a 0.3 dex increase in the \[Sr/Fe\] and \[Ba/Fe\] abundance ratios (holding \[Fe/H\] fixed to the solar value). The reference model is for an age of 13 Gyr, with a Galactic [@Kroupa01] IMF, and for solar metallicity. The model was convolved with a velocity dispersion of $350\kms$. Notice that even at this low resolution, at least two [Sr$\;$[<span style="font-variant:small-caps;">[ii]{}</span>]{}]{} lines and two [Ba$\;$[<span style="font-variant:small-caps;">[ii]{}</span>]{}]{} lines produce a change in the spectrum of $>0.3$% per 0.3 dex change in abundance. As we will demonstrate in Section \[s:res\], the existence of multiple, well-separated Sr and Ba lines allows us to test the robustness of our results in a relatively model-independent fashion.
Data {#s:data}
====
In this Letter we consider two samples of early-type galaxies. The first sample of 34 early-type galaxies (and the nuclear bulge of M31) was presented in @vanDokkum12. The sample was selected from the SAURON survey of nearby galaxies [@Bacon01]. High S/N spectra of these objects were obtained with the Low Resolution Imaging Spectrometer (LRIS) on the Keck I telescope. The wavelength coverage spans the range $3500$Å$-10,400$Å with a gap at $5600$Å$-7100$Å. Four wavelength intervals were defined, $0.4\mu
m-0.46\mu m$, $0.46\mu m-0.55\mu m$, $0.80\mu m-0.89\mu m$, and $0.96\mu m-1.02\mu m$ for fitting models to data. The wavelength choices reflect the limitations of the model and data wavelength coverage and regions of severe telluric corrections. The S/N at 5000Å varies from $\approx150$ Å$^{-1}$ to $\approx450$ Å$^{-1}$. Spectra were extracted within $R<R_e/8$, with a weighting intended to mimic a circular aperture, where $R_e$ is the effective radius. These data have already been compared to the CvD12 models in @Conroy12b [see the Appendix of that paper for a comparison of the best-fit models and data for each of the 35 LRIS galaxies]. In that earlier work, we did not consider variation in Sr and Ba, and only briefly commented on the abundances of the other elements considered in the fit. We found that the ages, \[Fe/H\] and \[Mg/Fe\] values derived with our full spectrum fitting technique agreed with conventional Lick index-based techniques.
The second source of data is provided by the Sloan Digital Sky Survey [SDSS; @York00]. Following the methodology outlined in @Graves09a, we have selected a sample of galaxies within a narrow redshift interval of $0.02-0.06$, with no detected emission in H$\alpha$ nor in \[[O$\;$[<span style="font-variant:small-caps;">[ii]{}</span>]{}]{}\], and with concentrated, de Vaucouleurs-like light profiles. The SDSS obtains spectra with fibers that have 3” diameters. The typical S/N of SDSS spectra in our sample is modest, of order 20 Å$^{-1}$, and so we have chosen to stack the spectra in seven bins of velocity dispersion with mean dispersions in each bin of 88, 112, 138, 167, 203, 246, and 300$\kms$. For the smallest sigma bin the SDSS fiber samples the inner $0.8\,R_e$, while for the most massive bin the fiber samples the inner $0.4\,R_e$. We have verified that our results do not change if we select galaxies within each bin such that the fiber samples the same fraction of $R_e$. Each spectrum was continuum-normalized and convolved to an effective dispersion of $350\kms$ before stacking, and each spectrum contributed equally to the stack. The resulting S/N of the stacked spectra at 5000Å ranges from $\approx500$ Å$^{-1}$ to $\approx1800$ Å$^{-1}$. As with the LRIS data, four wavelength intervals were defined for the SDSS stacks, $0.4\mu m-0.48\mu m$, $0.48\mu m-0.58\mu m$, $0.58\mu m-0.64\mu m$, and $0.80\mu m-0.88\mu
m$ for fitting models to data. These stacked data will be the focus of future work aimed at measuring the detailed abundance patterns as a function of galaxy properties. In this work we limit our attention to the neutron capture elements Sr and Ba.
As a first look at the data, in Figure \[fig:indx\] we show spectral indices for both the SDSS stacked spectra and the individual LRIS galaxies. The Mg[*b*]{} index is sensitive to Mg, while the other indices are sensitive to Ba and Sr lines (with central wavelengths indicated by the index name). The Br and Sr indices are defined in @Serven05, while the Mg[*b*]{} index is defined in @Worthey94b [no effort was made to place the latter on the Lick index scale]. For this figure, the LRIS data were convolved to a common dispersion of $350\kms$ in order to afford a direct comparison with the SDSS stacked spectra. The LRIS data do not cover the 6142Å spectral region. Overall the agreement between the LRIS and SDSS samples is encouraging, given the very different observational setups and reduction techniques. It is however difficult to interpret spectral indices because an index is in general sensitive not only to the feature of interest but also to features contributing to the sidebands (i.e., the index measures a feature strength relative to a pseudocontinuum). This is especially true when the feature of interest is weak compared to other features. Full spectrum fitting such as the kind performed in the present work does not suffer from this drawback.
Results {#s:res}
=======
Before showing the derived abundance trends, in Figure \[fig:spec\] we explore the quality of the fit around the [Ba$\;$[<span style="font-variant:small-caps;">[ii]{}</span>]{}]{} lines for the stacked SDSS spectrum in the second highest $\sigma$ bin. In the figure we compare two model fits, one in which \[Ba/Fe\]$=+0.0$ and another in which the Ba abundance is allowed to vary in the fit (with a resulting best-fit value of \[Ba/Fe\]$=-0.4$). We also mark the locations of the strongest [Ba$\;$[<span style="font-variant:small-caps;">[ii]{}</span>]{}]{} lines in this spectral region. It is noteworthy that the residuals decrease in [*all*]{} spectral regions sensitive to Ba. This is an important test of the results to follow because it is also evident from this figure that there are numerous places where the residuals exceed the formal S/N limits of the data (demarcated by the grey band). If there were only one strong Ba line then it would have been difficult to argue that a region of high residual required non-solar \[Ba/Fe\] values. We return to this point later in this section.
Our main result is shown in Figure \[fig:basr\]. In this figure we show the best-fit \[Ba/Fe\] and \[Sr/Fe\] abundance ratios as a function of \[Mg/Fe\]. We include both the stacked SDSS spectra and the LRIS spectra of individual galaxies. The formal statistical errors from the SDSS stacks are very small owing to the very high S/N of the spectra. It is encouraging that the LRIS data and SDSS stacks follow the same general trends of decreasing \[Ba/Fe\] and constant \[Sr/Fe\] with increasing \[Mg/Fe\]. Such low values of \[Ba/Fe\] are not uncommon amongst metal-poor halo stars in the Galaxy [@Sneden08]. Indeed, below \[Fe/H\]$\approx-3$ essentially all halo stars have \[Ba/Fe\]$<0.0$, with many stars having Ba abundances as low as \[Ba/Fe\]$\approx-2.0$ [@Francois07]. Such stars are also $\alpha$ enhanced, and have \[Sr/Ba\]$>0.0$, again broadly consistent with the mean abundance ratios of the massive early-type galaxies (although with very different overall metallicities).
Taken at face value, the \[Ba/Fe\] versus \[Mg/Fe\] trends are consistent with the standard interpretation of \[Mg/Fe\] probing the SF timescale in these galaxies. A shorter timescale should lead to higher \[Mg/Fe\] and lower \[Ba/Fe\] values, as observed. Although not shown, the SDSS data also display a strong correlation between \[Mg/Ba\] and $\sigma$. The derived \[Fe/H\] values range from $-0.1$ to $0.0$ over the SDSS early-type sample, and so the trends observed in Figure \[fig:basr\] are unlikely to be due to metallicity-dependent yields. Other explanations for the \[Mg/Fe\]$-\sigma$ relation have been proposed, including a varying IMF, varying SN Ia rate, and selective mass-loss of SN Ia ejecta [@Worthey92; @Thomas99; @Trager00]. The \[Mg/Ba\] data allow us to rule out the explanations that are particular to the Ia’s, since the Ia’s influence neither Mg nor Ba, but it cannot discriminate between a varying SF timescale and a varying high-mass IMF for the observed variation in \[Mg/Fe\] with $\sigma$.
The near constancy of \[Sr/Fe\] is puzzling. It has been suggested that Sr may have multiple nucleosynthetic origins, including the $r$-process, both the weak and main $s$-processes, and perhaps even hypernovae [@Raiteri93; @Travaglio04; @Qian08]. As \[Mg/Fe\] (and presumably the SF timescale) varies, then perhaps the main nucleosynthetic site varies in such a way as to keep a rough constancy in \[Sr/Fe\]. With detailed chemical evolution models constrained to match the observed \[Mg/Fe\] and \[Ba/Fe\] trends, it may be possible to place interesting constraints on the origin(s) of Sr in these massive galaxies.
We turn now to several tests that have been performed to assess the robustness of these results. In Figure \[fig:sys\] we show the same abundance trends as before, for only the SDSS stacked spectra, and for a variety of permutations to the fiducial model. We consider a model in which the IMF is held fixed to the Galactic [@Kroupa01] form, a model in which only the blue ($\lambda<5800$Å) spectral region is fit, and a model in which the SDSS stacks are not smoothed to a common resolution of $\sigma=350\kms$. The latter is a particularly important test because much more information is available in the low dispersion bins, which effectively have higher spectral resolution, compared to the high dispersion bins. Relatedly, when the data are not broadened to $\sigma=350\kms$, the associated response functions (Figure \[fig:response\]) become much stronger. The derived abundance ratios vary between these model permutations by $\sim0.1$ dex, but the overall qualitative trends are clearly robust to these details. We have also masked each of the two strongest [Ba$\;$[<span style="font-variant:small-caps;">[ii]{}</span>]{}]{} lines (one at a time) and refit the models to the data. The resulting \[Ba/Fe\] versus \[Mg/Fe\] trends are unaltered, although in the case of masking the [Ba$\;$[<span style="font-variant:small-caps;">[ii]{}</span>]{}]{} 4555Å line the overall \[Ba/Fe\] abundances shift lower by $\approx0.2$ dex. This highlights that while trends appear to be robust in our analysis, there are zero-point uncertainties at the $0.1-0.2$ dex level.
A concern is that the features we are trying to measure are much weaker than other metal lines. In fact, as shown in Figure \[fig:spec\] the expected Ba line strengths are similar to the typical systematic residual in the spectra [*after*]{} subtraction of our best-fit model! A robust test of our result is to randomly shuffle the central wavelengths of the strongest [Ba$\;$[<span style="font-variant:small-caps;">[ii]{}</span>]{}]{} lines and recompute $\chi^2$ around the shuffled lines. In practice, this is achieved by altering the response function for Ba (see Figure 1) by assigning random central wavelengths in the interval $4000$Å$-6000$Å while keeping the relative strengths and widths of the lines the same. We have performed this test $10^3$ times for each of the SDSS stacked spectra. As an example of the results for the highest $\sigma$ bin, the residuals decrease around the strongest shuffled [Ba$\;$[<span style="font-variant:small-caps;">[ii]{}</span>]{}]{} line in only $0.1$% of the shuffled models when using the best-fit \[Ba/Fe\] value shown in Figure \[fig:basr\]. Moreover, in [*none*]{} of the shuffled models do the residuals decrease in both of the two strongest shuffled [Ba$\;$[<span style="font-variant:small-caps;">[ii]{}</span>]{}]{} lines. In the four highest $\sigma$ bins, where the best-fit \[Ba/Fe\] ratios are $<-0.1$, the shuffled models are never able to reduce the residuals in all three of the strongest [Ba$\;$[<span style="font-variant:small-caps;">[ii]{}</span>]{}]{} lines. It is therefore highly unlikely that the derived \[Ba/Fe\] abundances are the result of a chance alignment between the strong [Ba$\;$[<span style="font-variant:small-caps;">[ii]{}</span>]{}]{} lines and spectral regions where the residuals happen to be high.
The elemental abundances derived herein should provide novel constraints on the duration of star formation in early-type galaxies once chemical evolution models that include yields from the $s$- and $r$-processes are employed. The \[Mg/Ba\] ratio is particularly promising because the derived timescale will not depend on the uncertain delay time distribution or overall rate of Type Ia SNe, in contrast to timescales derived from the \[Mg/Fe\] ratio. In future work, we will present detailed abundance patterns for $\approx17$ elements, which promise to provide further insight into the early formation histories of these now dormant galaxies.
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. 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/. The SDSS is managed by the Astrophysical Research Consortium for the Participating Institutions. The Participating Institutions are the American Museum of Natural History, Astrophysical Institute Potsdam, University of Basel, University of Cambridge, Case Western Reserve University, University of Chicago, Drexel University, Fermilab, the Institute for Advanced Study, the Japan Participation Group, Johns Hopkins University, the Joint Institute for Nuclear Astrophysics, the Kavli Institute for Particle Astrophysics and Cosmology, the Korean Scientist Group, the Chinese Academy of Sciences (LAMOST), Los Alamos National Laboratory, the Max-Planck-Institute for Astronomy (MPIA), the Max-Planck-Institute for Astrophysics (MPA), New Mexico State University, Ohio State University, University of Pittsburgh, University of Portsmouth, Princeton University, the United States Naval Observatory, and the University of Washington.
[29]{} natexlab\#1[\#1]{}
, R. [et al.]{} 2001, , 326, 23
, M., [Hansen]{}, C. J., [Bautista]{}, M., & [Ruchti]{}, G. 2012, , 546, A90
, E. M., [Burbidge]{}, G. R., [Fowler]{}, W. A., & [Hoyle]{}, F. 1957, Reviews of Modern Physics, 29, 547
, M., [Gallino]{}, R., & [Wasserburg]{}, G. J. 1999, , 37, 239
, C. & [van Dokkum]{}, P. 2012, , 747, 69
, C. & [van Dokkum]{}, P. G. 2012, , 760, 71
, R. G., [Schmiedekamp]{}, A. B., & [Arnett]{}, W. D. 1974, , 190, 95
, P. [et al.]{} 2007, , 476, 935
, G. J., [Faber]{}, S. M., & [Schiavon]{}, R. P. 2009, , 693, 486
, F. 2005, , 43, 435
, P. 2001, , 322, 231
, R. L. 1970, SAO Special Report, 309
—. 1993, [SYNTHE spectrum synthesis programs and line data]{}, ed. [Kurucz, R. L.]{}
, Y.-Z. & [Wasserburg]{}, G. J. 2008, , 687, 272
, C. M., [Gallino]{}, R., [Busso]{}, M., [Neuberger]{}, D., & [Kaeppeler]{}, F. 1993, , 419, 207
, L., [Bonifacio]{}, P., [Castelli]{}, F., & [Kurucz]{}, R. L. 2004, Memorie della Societa Astronomica Italiana Supplementi, 5, 93
, J., [Worthey]{}, G., & [Briley]{}, M. M. 2005, , 627, 754
, C., [Cowan]{}, J. J., & [Gallino]{}, R. 2008, , 46, 241
, D., [Greggio]{}, L., & [Bender]{}, R. 1998, , 296, 119
—. 1999, , 302, 537
, D., [Maraston]{}, C., [Bender]{}, R., & [Mendes de Oliveira]{}, C. 2005, , 621, 673
, B. M. 1979, , 229, 1046
, S. C., [Faber]{}, S. M., [Worthey]{}, G., & [Gonz[á]{}lez]{}, J. J. 2000, , 120, 165
, C., [Gallino]{}, R., [Arnone]{}, E., [Cowan]{}, J., [Jordan]{}, F., & [Sneden]{}, C. 2004, , 601, 864
, P. G. & [Conroy]{}, C. 2012, , 760, 70
, S. E. & [Hoffman]{}, R. D. 1992, , 395, 202
, G., [Faber]{}, S. M., & [Gonzalez]{}, J. J. 1992, , 398, 69
, G., [Faber]{}, S. M., [Gonzalez]{}, J. J., & [Burstein]{}, D. 1994, , 94, 687
, D. G. [et al.]{} 2000, , 120, 1579
[^1]: \[X/Fe\]$\equiv$log(X/Fe)$-$log(X/Fe)$_\odot$
[^2]: `kurucz.harvard.edu/`
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The SIDIS transverse spin asymmetries weighted with powers of $P_T$, the hadron transverse momentum in the $\gamma N$ reference system, have been introduced already twenty years ago and are considered quite interesting. While the amplitudes of the modulations in the azimuthal distribution of the hadrons are the ratios of convolutions over transverse momenta of the transverse–momentum dependent parton distributions and of the corresponding fragmentation functions, and can be evaluated analytically only making assumptions on the transverse–momentum dependence of these functions, the weighted asymmetries allow to solve the convolution integrals over transverse–momenta without those assumptions. Using the high statistics data collected in 2010 on transversely polarized proton target COMPASS has evaluated in x-bins the $P_T$ weighted Sivers asymmetry which is proportional to the product of the first transverse moment of the Sivers function and of the fragmentation function. The results are compared to the standard unweighted Sivers asymmetry.'
author:
- Franco Bradamante
title: 'COMPASS measurement of the $P_T$ weighted Sivers asymmetry'
---
In the generalized parton model, admitting a finite intrinsic transverse momentum $\vec{k}_T$ for the quarks, a total of eight Transverse Momentum Dependent (TMD) Parton Distribution Functions (PDFs) are needed for a full description of the structure of the nucleon at leading twist [@Kotzinian:1994dv; @Mulders:1995dh; @Bacchetta:2006tn]. These functions lead to asymmetries in the azimuthal distributions of hadrons produced in Semi-Inclusive measurements of Deeply Inelastic Scattering (SIDIS) processes off polarised and unpolarised nucleons and can be disentangled by measuring the amplitudes of the different angular modulations. Among the eight TMD PDFs, the T-odd Sivers function [@Sivers:1989cc] is of particular interest. This function arises from a correlation between the transverse momentum of an unpolarised quark in a transversely polarised nucleon and the nucleon spin. It is responsible for the Sivers asymmetry, $A_{Siv}$, which is proportional to the convolution of the Sivers function $f_{1T}^{\perp }$ and of the unpolarised fragmentation function $D_1$. This asymmetry consists in a sin $\Phi_{Siv}$ modulation in the number of the produced hadrons, where $\Phi_{Siv} = \phi_h-\phi_S$ is the difference of the azimuthal angle of the hadron and of the nucleon spin. The azimuthal angles are defined in a reference system in which the z axis is the virtual photon direction and the xz plane is the lepton scattering plane.
The Sivers effect was experimentally observed in SIDIS on transversely polarised proton targets, first by the HERMES Collaboration [@Airapetian:2004tw; @Airapetian:2009ae] and then by the COMPASS Collaboration [@Alekseev:2010rw; @Adolph:2012sp; @Adolph:2014zba]. A small effect was also observed at JLab for positive hadrons produced on a $^3$He target [@Qian:2011py], while the COMPASS measurements on a transversely polarised deuteron target [@Alexakhin:2005iw; @Ageev:2006da; @Alekseev:2008aa] gave asymmetries compatible with zero. Combined analysis of the proton and deuteron data soon allowed for first extractions of the Sivers function for u- and d-quarks [@Efremov:2003tf; @Vogelsang:2005cs; @Collins:2005ie; @Anselmino:2012aa], which turned out to be different from zero, with similar strength and opposite in sign, a most important result in TMD physics. In the standard “Amsterdam” notation the Sivers asymmetry can be written as $$\begin{aligned}
A_{Siv}(x,z) &=&
\frac{\sum_q e_q^2 x f_{1T}^{\perp \, q}(x) \otimes D_1^q(z)}
{\sum_q e_q^2 x f_{1}^q(x) \cdot D_1^q(z)}
\label{eq:sa}\end{aligned}$$ where $x$ is the Bjorken variable, $z$ is the fraction of the available energy carried by the hadron, and $\otimes$ indicates a convolution over the transverse momenta of the Sivers function $f_{1T}^{\perp}$ and the fragmentation function $D_1$. In all those analyses, due to the presence of the convolution, some functional form had to be assumed for the transverse momentum dependence of both the quark distribution functions and of the fragmentation functions. In most of the analyses these functions were assumed to be Gaussian, and as a result the first transverse moment of the Sivers function $$\begin{aligned}
f_{1T}^{\perp (1)}(x) = \int d^2 \vec{k}_T
\, \frac{k_T^2}{2 M^2} \,
\, f_{1T}^{\perp}(x, k_T^2)
\label{k_T_moment}\end{aligned}$$ could be determined from the asymmetry data.
Already twenty years ago an alternative method was proposed to determine $f_{1T}^{\perp (1)}$ without making any assumption on the functional form neither of the distribution functions nor of the fragmentation functions. The method consists in measuring asymmetries weighted by the transverse momentum of the hadron $P_T$ [@Kotzinian:1995cz; @Kotzinian:1997wt; @Boer:1997nt] as described in the following. For some reasons the method was not pursued: the only results (still preliminary) came from HERMES [@Gregor:2005qv] and have already been used to extract the transverse moment of the Sivers function [@Collins:2005ie]. Recently, much interest has been dedicated again to the weighted asymmetries (see, f.i. [@Kang:2012ns]). In this contribution the first COMPASS results [@Sbrizzai2016] on the weighted Sivers asymmetries from the data collected in 2010 on a transversely polarized proton target are presented.\
Keeping only the relevant terms, the SIDIS cross-section can be written as $$\begin{aligned}
\frac{d\sigma}{dx dy dz d\Phi_{Siv}} &= &
\frac{\alpha^2}{s}\,
\frac{1-y+ y^2/2}{x^2 y^2} \nonumber \\
& \times & \left[ F_{UU} + {S}_{T} \,F_{UT}^{\sin \Phi_{Siv}} \, \sin \Phi_{Siv}+...
\right] \; \; \; \;
\label{eq:sidiscs}\end{aligned}$$ where $\alpha$ is the fine structure constant, $s=Q^2/xy$ is the center-of-mass energy squared, and terms of the order of $\gamma^2 = (2Mx/Q)^2$, where $M$ is the nucleon mass, have been neglected. The quantities $y$ and $Q^2$ are the fraction of lepton energy carried away by the virtual photon and the photon virtuality respectively.
The Sivers asymmetry $A_{Siv}$ is the ratio of the structure functions $$\begin{aligned}
&F_{UT}^{sin \Phi_{Siv}} &=\int d^2\vec{P}_T \, P_T F(P_T^{\, 2}) \nonumber \\
&F_{UU} \; \; \; \; \; \; &= \int d^2\vec{P}_T
\int d^2\vec{k}_T \int d^2\vec{p}_{\perp} \delta^2 \, (\vec{P}_T-\vec{p}_{\perp}-z\vec{k}_T) \, \nonumber \\
&& \; \; \; \; \; \; \times \; f_{1}(k_T^2) D_1(p_{\perp}^2)\end{aligned}$$ where $$\begin{aligned}
F(P_T^{\, 2}) &=&
\int d^2\vec{k}_T \int d^2\vec{p}_{\perp} \, \delta^2 (\vec{P}_T-\vec{p}_{\perp}-z\vec{k}_T) \nonumber \\
&& \; \; \; \; \; \; \times \; \frac{\vec{P}_T \cdot k_{T}}{M P_T^{\, 2}}
f_{1T}^{\perp}(k_T^2) D_1(p_{\perp}^2) \end{aligned}$$ and $\vec{p}_{\perp}$ is the hadron transverse momentum relative to the fragmenting quark. Although it has not been explicitly indicated the quark distribution functions depend also on $x$ and $Q^2$ and the fragmentation functions on $z$ and $Q^2$. Also, the sum over the quark flavor is not indicated.
The convolution which appears in $F_{UU}$ can be easily calculated (the usual relation among the transverse momenta of the quark and of the hadron is guaranteed by the delta function), but in general this is not the case for $F_{UT}^{sin \Phi_{Siv}}$. It is possible to calculate it by assuming specific functional forms for the transverse momentum dependence of $f_{1T}^{\perp}$ and $D_1$, and a common choice is the Gaussian “ansatz” $$\begin{aligned}
f_{1T}^{\perp}(x, k_T^2) &=& f_{1T}^{\perp}(x) \, \frac{e^{- k_T^2/\langle k_T^2 \rangle_S}}{\pi
\langle k_T^2 \rangle_S}\,, \; \; \; \;
\nonumber \\
D_1(z, p_{\perp}^2) &=& D_1(z) \, \frac{e^{- p_{\perp}^2/\langle p_{\perp}^2 \rangle}}{\pi
\langle p_{\perp}^2 \rangle}\,,
\label{tmd_gauss3} \end{aligned}$$ which gives $$\begin{aligned}
A_{Siv,G}(x,z) &=& a_G
\frac{\sum_q e_q^2 x f_{1T}^{\perp \, (1) \, q}(x) \cdot D_1^q(z)}
{\sum_q e_q^2 x f_{1}^q(x) \cdot D_1^q(z)} \, .
\label{eq:gsa}\end{aligned}$$ where $a_G = \sqrt{\pi}M/\sqrt{\langle k_T^2 \rangle _S +
\langle p^2_{\perp} \rangle /z^2}$. From eq. (\[eq:gsa\]) it is clear that in the Gaussian ansatz the first moment of the Sivers function $f_{1T}^{\perp \, (1)}$ can be evaluated directly from the measured Sivers asymmetries using some assumption for the transverse momenta which appear in $a_G$.
As stated above, the first moment of the Sivers function can be accessed without making any hypothesis on the specific functional forms for $f_{1T}^{\perp}$ and $D_1$ by measuring the $P_T$ weighted Sivers asymmetries. These asymmetries are obtained by weighting the events by the measured transverse momentum of the hadron. Only the spin-dependent part of the cross-section has to be weighted, leaving unweighted the unpolarised cross-section. In this work the weighting is done with $P_T/zM$, where $M$ is the nucleon mass. After some algebra one gets the simple result $$\begin{aligned}
F_{UT}^{sin \Phi_{Siv}, \, w}=\int d^2\vec{P}_T \frac{P_T^{\, 2}}{zM}\, F(P_T^{\, 2})
=2\,f_{1T}^{\perp (1)} \, D_1\end{aligned}$$ namely the convolution becomes the product of the first transverse moment of the Sivers function and the fragmentation function $D_1$, so that the weighted Sivers asymmetry is $$\begin{aligned}
A_{Siv}^{w}(x,z) &=&
2\frac{\sum_q e_q^2 x f_{1T}^{\perp \, (1) \, q}(x) \, D_1^q(z)}
{\sum_q e_q^2 x f_{1}^q(x) \, D_1^q(z)} \,.
\label{eq:wsa}\end{aligned}$$ An interesting remark is that, knowing the unpolarised distributions and fragmentation functions, both $f_{1T}^{\perp \, (1) \, u}$ and $f_{1T}^{\perp \, (1) \, d}$ could easily be obtained following the procedure of Ref. [@baronespin]. Also, assuming u-dominance for positive hadrons produced on a proton target, the fragmentation function cancels out and the asymmetry simply becomes $$\begin{aligned}
A_{Siv}^{w}(x,z) &\simeq&
2\frac{f_{1T}^{\perp \, (1) \, u}(x)}
{f_{1}^u(x) } \,.
\label{eq:wsag}\end{aligned}$$
The data we used for this analysis are the data collected in 2010 with the transversely polarised proton target. Since the comparison with the “standard” Sivers asymmetry is also important, the data production and all the cuts to select the muons and the hadrons are the same as for the published data [@Adolph:2012sp]. In particular, the selected phase space is defined by $0.004 < x < 0.7$, $Q^2 > 1$ GeV/c$^2$, $0.1<y < 0.9$, $W > 5$ GeV/c$^2$, $P_T > 0.1$ GeV/c, and $z > 0.2$. Presently, the asymmetry has been measured only as a function of $x$ and the results have been extracted in the nine $x$-bins of ref. [@Adolph:2012sp].
The distributions of the weight factor $P_T/zM$ in the different $x$ bins are shown in Fig. \[fig:weights\] for the positive hadrons.
{width="98.00000%"}
The corresponding distributions for the negative hadrons are very similar. The $P_T/z$ acceptance of the spectrometer is about 60% and rather flat.
{width="49.00000%"} {width="49.00000%"}
To extract $A_{Siv}^{w}$ we used an ad-hoc double-ratio method which utilizes the information coming from the different cells in which our target system is divided and insures cancellation of the target acceptance and of the beam flux. The method used in so far had to be modified since only the counts in the numerator of the expression of $A_{Siv}^{w}$ are weighted, while the counts at the denominator are unweighted. Three different estimators have been used, and it has been checked that the results are essentially identical.\
The results in the case of positive hadrons are given in Fig. \[fig:results\] left, while Fig. \[fig:results\] right gives the results for negative hadrons. In both cases the “standard” Sivers asymmetries published in [@Adolph:2012sp] are also plotted for comparison. As expected, the trend of the asymmetries is similar both for positive and negative hadrons. Assuming u-dominance, the results for positive hadrons which are clearly different from zero in particular at large $x$, where $\langle Q^2 \rangle$ reaches $\sim 20$ GeV$^2$, constitute the first direct measurement of $f_{1T}^{\perp \, (1) \, u}(x) / f_{1}^u(x)$.
Given the similar trend of the weighted asymmetries $A_{Siv}^{w}$ and of the standard asymmetries $A_{Siv}$, we have evaluated in each $x$ bin their ratios $R^w=A_{Siv}^{w}/A_{Siv}$. The values of $R^w$ are given in Fig. \[fig:ratio\_pos\] for positive hadrons, which exhibit a large Sivers asymmetry.
![The ratios $R^w$ between $A_{Siv}^{w}$ and $A_{Siv}^{w}$ in the nine $x$-bins for positive hadrons.[]{data-label="fig:ratio_pos"}](ratio_pos.eps){width="49.00000%"}
In spite of the large statistical uncertainties the ratios are compatible with a constant value, which is rather well determined ($1.6 \pm 0.1$). Also, the values in the different $x$ bins agree rather well with those of $\langle P_T / z M \rangle$, as expected.
To gain insight into the physics of the ratio $R^w$ it is useful to take for $A_{Siv}$ the expression given in eq. (\[eq:gsa\]) obtained in the Gaussian model which from the past phenomenological analyses of the Sivers effect is known to work rather well. By comparison of eq. (\[eq:gsa\]) and eq. (\[eq:wsa\]) it is clear that their ratio provides some information on the quantity $a_G$. In particular $$\begin{aligned}
R^w_G(x)&=&\frac{A_{Siv}^{w}(x)}{A_{Siv,G}(x)} \\
&=&\frac{2 }{\sqrt{\pi} M}
\frac{\sum_q e_q^2 x f_{1T}^{\perp \, (1) \, q}(x) \cdot \int dz D_1^q(z)}
{\sum_q e_q^2 x f_{1T}^{\perp \, (1) \, q}(x) \cdot \int dz z D_1^q(z) / \sqrt{ \langle P_T^2 \rangle _S} } \nonumber\end{aligned}$$ can be further simplified by making the (reasonable) assumption that $\sqrt{\langle P_T^2 \rangle _S}
=\sqrt{z^2 \langle k_T^2 \rangle _S + \langle p^2_T \rangle}$ can be regarded as a constant since the dependence on $z$ of $\sqrt{\langle P_T^2 \rangle}$ is known to be much weaker than that of $D_1(z)$. Further simplifications on the ratios $R^w_G$ can be obtained by weighting the events by $P_T/M$ rather than $P_T/zM$.\
As a conclusion, COMPASS is extracting the weighted Sivers asymmetry in the SIDIS process of 160 GeV muons on transversely polarized protons extending the standard measurements already performed. Preliminary results from the data collected in 2010 have already been derived, using as weight $P_T/zM$ and are presented in this paper as a function of $x$. No major experimental problems have been encountered and the results look very promising in view of more precise extractions of the Sivers function, and further measurements are foreseen.
[99]{}
A. Kotzinian, Nucl. Phys. B [**441**]{} (1995) 234.
P. J. Mulders and R. D. Tangerman, Nucl. Phys. B [**461**]{} (1996) 197 Erratum: \[Nucl. Phys. B [**484**]{} (1997) 538\].
A. Bacchetta [*et al.*]{}, JHEP [**0702**]{} (2007) 093.
D. W. Sivers, Phys. Rev. D [**41**]{} (1990) 83.
A. Airapetian [*et al.*]{} \[HERMES Collaboration\], Phys. Rev. Lett. [**94**]{} (2005) 012002.
A. Airapetian [*et al.*]{} \[HERMES Collaboration\], Phys. Rev. Lett. [**103**]{} (2009) 152002.
M. G. Alekseev [*et al.*]{} \[COMPASS Collaboration\], Phys. Lett. B [**692**]{} (2010) 240.
C. Adolph [*et al.*]{} \[COMPASS Collaboration\], Phys. Lett. B [**717**]{} (2012) 383.
C. Adolph [*et al.*]{} \[COMPASS Collaboration\], Phys. Lett. B [**744**]{} (2015) 250.
X. Qian [*et al.*]{} \[Jefferson Lab Hall A Collaboration\], Phys. Rev. Lett. [**107**]{} (2011) 072003.
V. Y. Alexakhin [*et al.*]{} \[COMPASS Collaboration\], Phys. Rev. Lett. [**94**]{} (2005) 202002.
E. S. Ageev [*et al.*]{} \[COMPASS Collaboration\], Nucl. Phys. B [**765**]{} (2007) 31.
M. Alekseev [*et al.*]{} \[COMPASS Collaboration\], Phys. Lett. B [**673**]{} (2009) 127.
A. V. Efremov, K. Goeke and P. Schweitzer, Phys. Lett. B [**568**]{} (2003) 63.
W. Vogelsang and F. Yuan, Phys. Rev. D [**72**]{} (2005) 054028.
J. C. Collins [*et al.*]{}, Phys. Rev. D [**73**]{} (2006) 014021.
M. Anselmino, M. Boglione and S. Melis, Phys. Rev. D [**86**]{} (2012) 014028.
A. M. Kotzinian and P. J. Mulders, Phys. Rev. D [**54**]{} (1996) 1229.
A. M. Kotzinian and P. J. Mulders, Phys. Lett. B [**406**]{} (1997) 373.
D. Boer and P. J. Mulders, Phys. Rev. D [**57**]{} (1998) 5780.
I. M. Gregor \[HERMES Collaboration\], Acta Phys. Polon. B [**36**]{} (2005) 209.
Z. B. Kang, I. Vitev and H. Xing, Phys. Rev. D [**87**]{} (2013) no.3, 034024.
G. Sbrizzai for the COMPASS Collaboration, “TMD measurements at COMPASS” QCD-N – 4rd Workshop on the QCD Structure of the Nucleon, Getxo, Bilbao, Spain, 11-15 July, 2016.
A. Martin, F. Bradamante and V. Barone, these proceedings, and arXiv:1701.08283 \[hep-ph\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Static cylindrical shells composed of massive particles arising from matching of two different Levi-Civita space-times are studied for the shell satisfying either isotropic or anisotropic equation of state. We find that these solutions satisfy the energy conditions for certain ranges of the parameters.'
author:
- 'Metin Arik[^1] and Özgür Delice[^2]'
title: Static Cylindrical Matter Shells
---
KEY WORDS: Levi-Civita spacetime, Cylindrical shells.
Introduction
============
The cylindrically symmetric static vacuum space-time is given by the Levi-Civita metric [@levicivita], which can be written in the form $$ds^{2}=-q^{4s }dt^{2}+ q^{4s (2s -1)}(dq ^{2}+ dz^{2}) +\alpha^2
q^{2(1-2s)}d\phi^{2} \label{Lc}$$ where $t$, $q$, $z$ and $\phi$ are the time, the radial, the axial, and the angular coordinates with the ranges $-\infty <t< \infty$, $0 \leq q < \infty$, $-\infty<z<\infty$ and $0<\phi \leq 2\pi$. This metric has two real parameters which cannot be removed by a coordinate transformation. Cylinders [@Bonnor; @philbin] and cylindrical shells [@Stachel]-[@gleiser] were studied as a physically acceptable sources of this metric in order to understand the meaning and the behavior of these parameters. For recent reviews on cylinders see [@Bonnor]. The first parameter, $s$, is related to the energy density of the source whereas the second one, $\alpha$, is related to the global conicity of space-time. When $s=0$ or $1/2$ the space-time is flat and corresponds to Minkowski space-time. For $0<s<1/2$ this metric is the exterior metric of an infinite line mass (line singularity). For $s=0$ the metric is flat and if $\alpha=1$ it is regular whereas if $\alpha
\neq 1$ it becomes the well known exterior cosmic string metric [@Vilenkin]. For $s=1/2$ this metric correspond to an infinite plane [@dasilva]. In this paper we limit our analysis to $-1/2\leq s \leq 1/2$.
Our aim is to study infinitely thin cylindrical shells composed of matter particles having isotropic or anisotropic equation of state and satisfying certain energy conditions as a source of exterior Levi-Civita metric. Actually, it is known [@zouzou]- [@gleiser],[@bicak] that the matching conditions yield the conditions for the shell to be matter shell or shell composed of massless particles. The shells composed of photons is studied recently in detail. Some properties of the photonic shells around a regular interior or a cosmic string [@bicak] or a line singularity [@AD1] have been discussed and generalized to the multiple shell case [@AD2]. The particle motion about a photonic shell with flat interior and exterior is discussed in [@Zofka]. In this paper, we will focus on non-photonic shells-matter shells- having isotropic or anisotropic equation of state. Imposing several equation of states for the shell one can find admissible values of the metric parameters satisfying some energy conditions. We will compare the results with the previous works on this topic, especially with [@bicak] which studied shells made of various types of matter as a source of the Levi-civita metric. They relate the interior properties of the shell,especially the mass per unit lenght of the shell with the two parameters of the exterior metric. They limited their analysis with the case where the interior region is flat or locally flat and a cosmic string can be present on the symmetry axis. We also refer [@bicak] for a review of previous works on this topic since a good comparison of their results with the previous works has been given. Our main difference is that we will allow the inside of the shell not flat. This may not seem physically attractive, however this is not the case, since the shell can surround another shell or cylinder satisfying (or not satisfying) physical requirements. Specifically, the shell can be one member of a multiple shell system[@gleiser; @AD2]. Thus, the interior region of the shell is not necessarily flat in the most general case.
Matching conditions
===================
We can write the Levi-Civita metric in normal form by transforming the radius $q$ into a proper radius $r$ by defining $dr=q^{2s(2s
-1)}dq$ which results in $$ds^{2}=-R^{4s /N}dt^{2}+dr^{2}+ R^{4s (2s -1)/N}dz^{2} +\alpha^2
R^{2(1-2s )/N}d\phi^{2}, \label{Lc1}$$ where $$\
q =R^{1/N},\ \ R=Nr, \ \ N=4s ^{2}-2s +1.$$
Let us denote the interior and the exterior radial coordinates as $r_\pm$ and the location of the shell at $(r_{\pm}=r_{0\pm})$. We want both the interior and the exterior metrics to be continuous on the shell. After rescaling the metrics, the interior and the exterior metrics can be written as:
$$\begin{aligned}
&&ds^2_-=-r_-^{\frac{4s}{N}}dt^2+dr_-^2+r_-^{\frac{4s(2s-1)}{N}}dz^2+\alpha_-^2
r_-^{\frac{2(1-2s)}{N}}d\phi^2,\quad (r_-\leq r_{0-}), \\
&&ds^2_+=-r_+^{\frac{4s'}{N'}}dt'^2+dr_+^2+r_+^{\frac{4s'(2s'-1)}{N'}}dz'^2+\alpha_+^2
r_+^{\frac{2(1-2s')}{N'}}d\phi^2,\ \ (r_+\geq r_{0+}).
\phantom{aa}\end{aligned}$$
To achieve the continuity at $r_-=r_{0-}$, $r_+=r_{0+}$, we define $t'=Kt$, $z'=Lz$, then we find: $$K=\frac{r_{0-}^{2s/N}}{r_{0+}^{2s'/N'}}, \quad
L=\frac{r_{0-}^{2s(2s-1)/N}}{r_{0+}^{2s'(2s'-1)/N'}}, \quad
\alpha_+=\frac{\alpha_- r_{0-}^{(1-2s)/N}}{r_{0+}^{(1-2s')/N'}}.$$
After a straightforward calculation using Israel’s thin shell formalism [@israel], the Energy-Momentum tensor of the shell, whose form is anisotropic in general with $S^i_{\phantom{i}j}=diag\left(-\rho,p_z,p_{\phi} \right)$, can be written as (we take $c=G=1$):
$$\begin{aligned}
\label{emt}
%8\pi S^0_{\phantom{0}0}&=&
-8\pi \rho&=&\frac{N'-2s'}{N'r_{0+}}-\frac{N-2s}{Nr_{0-}}%\nonumber \\%=\frac{1-a'}{r_0+B}-\frac{1-a}{r_0+A}
=8\pi(p_z+p_\phi)+\frac{4s}{Nr_{0-}}-\frac{4s'}{N'r_{0+}},\phantom{AAA}\\
%8\pi S^z_{\phantom{z}z}&=&
8\pi
p_z&=&\frac{1}{N'r_{0+}}-\frac{1}{Nr_{0-}},%\label{emtpz}\\ %=\frac{1-b'}{r_0+B}-\frac{1-b}{r_0+A}
%8 \pi S^{\phi}_{\phantom{\phi}\phi}&=&
\quad \quad 8 \pi p_{\phi}=\frac{4s'^2}{N'r_{0+}}-\frac{4s^2}{Nr_{0-}}.\label{emtpp}%=\frac{1-c'}{r_0+B}-\frac{1-c}{r_0+A}\end{aligned}$$
Matter Shells
=============
The trace of the energy momentum tensor of the shell (\[emt\]-\[emtpp\]) is $$4 \pi\, S=\frac{1}{r_{0+}}-\frac{1}{r_{0-}}.$$
Thus, the vanishing of the trace of energy-momentum tensor of the shell requires $r_{0+}=r_{0-}.$ Equivalently, if we perform a coordinate transformation on the inner or the outer independent radial coordinates $r_-$ or $r_+$ to express one of them in terms of the other, than the condition can be expressed as follows: If there is no relative shift in the values of the interior and the exterior radial coordinates on the shell, than the shell is composed of massless particles. If there is a shift, than the shell is composed of massive particles.
Isotropic Shells
================
In this section we study shells with $p_z=p_\phi=p.$ Also we impose an equation of state $p=\gamma \rho.$ where $\gamma$ is a constant. These constraints give $$\label{isotropicgamma}
r_{0+}=\frac{(1-4s^2)N'\,r_{0-}}{(1-4s'^2)N},\quad
\gamma=\frac{s+s'}{(2s-1)(2s'-1)}.$$ From these equations, we can find the exterior metric parameters for given $\gamma$ and interior metric parameters or vice versa. The energy density and the pressure become: $$\label{isotropicpress}
\rho=\frac{(s-s')(2s-1)}{2 \pi(1+2s')N r_{0-}},\quad
p=\frac{s^2-s'^2}{2\pi(4s'^2-1)N r_{0-}}.$$ The values of these quantities for flat interior can be found by taking $s=0.$
For a shell with the stress-energy tensor of the form $S_{ij}=diag(\rho,p_z,p_\phi)$ we have the weak energy condition (WEC) with $\rho\geq 0$, $\rho+p_z\geq 0$, $\rho+p_\phi \geq 0$; the dominant energy condition (DEC) with $\rho \geq 0$, $\rho\geq
p_z\geq - \rho$, $\rho\geq p_\phi\geq - \rho$; and the strong energy condition (SEC) with $\rho+p_z+p_\phi \geq 0$ together with the conditions for WEC [@HawkingEllis]. For $s=0$ this shell satisfies WEC and SEC for $0\leq s' \leq 1/2$, DEC for $0\leq s' \leq 1/3$. Notice that when the interior metric parameter $s$ is getting bigger, the range of $s'$ when the shell satisfies the energy conditions decreases. For $s=1/2$, $\rho$ vanishes. And for $s>1/2$ there is no solution satisfying the energy conditions except WEC since $\rho$ is always smaller than $p$ for this case. These results are in agreement with the previous work [@bicak] who studied this case for $s=0$. The changing of the energy density and pressure with $s'$ is presented in Figure \[perfect\] for the case when the interior metric parameter $s=0$. As previous studies of the cylindrical shells suggest [@bicak], the energy density of the shell is bounded from above.
Some examples on isotropic shells
---------------------------------
If we choose $\gamma=0$ in (\[isotropicgamma\]) we find a solution with shell composed of pressureless dust. We have $$s'=-s,\quad r_{0+}=\frac{1+2s'+4s'^2}{N'}r_{0-}, \quad
\rho=\frac{s'}{\pi N r_{0-}}.$$
We know that Levi-Civita metrics with negative parameter are isomorphic to the positive parameter cases and correspond to a repulsive (negative energy) line mass. In this case, the energy density of this shell is positive when $s$ is negative and vice versa. Thus , when the interior metric parameter $s$ is negative, the shell satisfies all the energy conditions. The energy density of the shell is bounded from above and reaches its maximum value for $s'=1/2$. The variation of the energy density for a given $s'$ is presented in Figure \[dust\].
Domain walls are topological defects which might arise during the phase transitions of the early universe [@Vilenkin]. They have the equation of state $\gamma =-1$. Applying our shell to this equation of state we find that, $$s'=\frac{s-1}{4s-1},\quad r_{0+}=-r_{0-}\quad \rho=-p=\frac{1}{6
\pi r_{0-}}.$$
Thus, we can have a domain wall with positive energy density provided that the outer radius of the shell is negative of the inner radius of the shell. Thus, this solution is not physically attractive.
We can have a shell with isotropically moving photons in $z$ and $\phi$ directions with $\gamma=1/2$. Then we find $s'=(s-1)/(4s-1)$ and $r_{0+}=r_{0-}$, thus inner and outer radii agree for this case as expected.
Anisotropic shells
==================
Here we will consider the case where the shell is composed of anisotropic massive particles having the equation of state $p_z=\kappa\, \rho$, $p_\phi=\lambda\, \rho$ where $\kappa$ and $\lambda$ are constants. From (\[emt\]-\[emtpp\]) we find $$\begin{aligned}
\kappa=\frac{N r_{0-}-N' r_{0+}}{(1-2s)^2N'r_{0+}-(1-2s')^2 N
r_{0-}},\\
\lambda=\frac{4s'^2Nr_{0-}-4s^2N'r_{0+}}{(1-2s)^2N'r_{0+}-(1-2s'^2)N
r_{0-}}.\end{aligned}$$
This result is general and for any given $\kappa$ and $\lambda$ we can determine the outer metric parameters in terms of the inner metric parameters or vice versa. Then using (\[emt\]-\[emtpp\]) we can determine their energy density and pressure. Let us present some examples.
Some examples on anisotropic shells
-----------------------------------
We first present dust-like solutions where $\kappa+\lambda=0$. Applying these constraints on the field equations, we find $$\begin{aligned}
&&r_{0+}=\frac{(1+4s'^2)N r_{0-}}{(1+4s^2)N'},\quad
\rho=\frac{s'(1+4s^2)-s(1+4s'^2)}{2\pi(1+4s'^2) N\,r_{0-}}, \\
&& p_z=-p_\phi=\frac{(s^2-s'^2)}{2\pi(1+4s'^2) N r_{0-}}.\end{aligned}$$
The changing of the energy density and the pressures for the interior metric parameter $s=0$ is presented in figure \[dustlike\] for $r_{0-}=1$. For this case the shell satisfies WEC, DEC and SEC for $0\leq s'\leq 1/2$.
Next, let us discuss an anisotropic shell solution which corresponds to a hollow cylinder with infinitely long straight cosmic strings [@Vilenkin] with the equation of state $\rho=-p_z$. The exterior metric should be the same of the exterior metric of infinitely long static cosmic string metric [@Vilenkin]. Indeed, only for $s'=0$ we have a solution of this form with $s=0$ or $s=1$ which gives $$8 \pi\rho=\frac{1}{r_{0-}}-\frac{1}{r_{0+}},\quad (s=0);\quad 8
\pi \rho= \frac{1}{3r_{0-}}-\frac{1}{r_{0+}},\quad (s=1).$$
These solutions correspond to a cosmic string shell and in the first one the interior region is flat [@zouzou; @Tsoubelis]. Same results are also presented in [@bicak].
Let us now discuss a massive shell with only counter rotating particles or counter moving particles in $z$ direction. For a shell with counter rotating particles ($p_z=0$) we need $\kappa=0$, then we find: $$r_{0+}=\frac{Nr_{0-}}{N'},\quad \rho=\frac{s'(1-s')-s(1-s)}{2\pi
Nr_{0-}},\quad p_\phi=\frac{s'^2-s^2}{2\pi Nr_{0-}}.$$ For flat interior, ($s=0$) this shell has positive energy density for $0\le s'\le 1$, and satisfies WEC and DEC for this range whereas SEC for $0\le s'\le 1/2$.
For a shell with counter moving particles in $z$ direction ($p_\phi=0$) we need $\lambda=0$ which yields: $$r_{0+}=\frac{s'^2Nr_{0-}}{s^2N'},\quad
\rho=\frac{s'^2(1-4s)-s^2(1-4s')}{8\pi s'^2Nr_{0-}},\quad
p_z=\frac{s^2-s'^2}{8\pi s'^2Nr_{0-}}.$$ For this shell for flat interior there is no solution. For negative $s$, we can find solutions satisfying all the energy conditions, and for $0<s<1$ there is no solutions satisfying all the energy conditions. Only for the interval $s > 1/2$ $\cap$ $s/
(4 s-1) < s' < s$ WEC and SEC is satisfied.
Two-fluid shell
===============
It is interesting to note that the shell also admits a two-fluid solution with $S_{ij}=S_{ij}^{(p)}+S_{ij}^{(m)}$ with $S_{j}^{i(p)}=diag(-\rho^{(p)},p_z^{(p)},p_\phi^{(p)})$ and $S_{j}^{i(m)}=diag(\rho^{(m)},p_z^{(m)},p_\phi^{(m)})$. We first make a coordinate transformation to the outer radial coordinate as $r_{+}=r_{-}+a$. Then we find: $$\begin{aligned}
&&\rho^{(p)}=\frac{s'N-sN'}{4\pi NN'(r_{0-}+a)},\quad
p_z^{(p)}=\frac{N-N'}{8 \pi NN'(r_{0-}+a)} \\
&&p_{\phi}^{(p)}=\frac{s'^2N-s^2N'}{2 \pi NN'(r_{0-}+a)},\quad
\rho^{(m)}=\frac{a(N-2s)}{8\pi N r_{0-}(r_{0-}+a)},\\
&&p_{z}^{(m)}=\frac{-a}{8\pi N r_{0-}(r_{0-}+a)},\quad
p_{\phi}^{(m)}=\frac{-s^2a}{2\pi N r_{0-}(r_{0-}+a)}.
%&&\rho^{(p)}=\frac{2(s'-s-4s'^2s+4s^2s')}{NN'(r_{0-}+a)},\quad
%p_z^{(p)}=\frac{2(s'(1-2s')-s(1-2s))}{NN'(r_{0-}+a)} \phantom{aaaa}\\
%&&p_{\phi}^{(p)}=\frac{}{NN'(r_{0-}+a)},\quad
%\rho^{(m)}=\frac{a(1-2s^2)}{Nr_{0-}(r_{0-}+a)}, \\
%&&p_z^{(m)}=\frac{-a}{Nr_{0-}(r_{0-}+a)} ,\quad
%p_\phi^{(m)}=\frac{4s^2N'a}{Nr_{0-}(r_{0-}+a)}\end{aligned}$$ It is easy to see that $S_{ij}^{(p)}$ is traceless but $S_{ij}^{(m)}$ is not. Thus, this solution represents a two-fluid shell composed of a mixture of photons and matter. When $s=s'$ the photonic part vanishes whereas for $a=0$ the matter part vanishes. Thus the existence of photons in the shell deviates the outer metric parameter $s'$ from the inner metric parameter $s$ and the existence of matter particles results a shift in the values of the inner and the outer radius of the shell. When $a=0$ we get photonic shells discussed recently in great detail [@bicak; @AD1]. For the matter part, the pressures of the shell in $z$ and $\phi$ directions are negative. When the interior is locally flat, the matter part satisfies $\rho=-p_z$ and this two-fluid represents a shell composed of both cosmic strings and photons [@Wang; @gleiser].
Conclusions
===========
In this work, we studied some exact solutions of the Einstein equations corresponding to static cylindrical shells composed of massive particles. The shells constructed from matching of two different Levi-Civita space-times can be classified as matter or photonic shells whether their radial coordinates have relative shifts or not. In this paper we have focused on to the matter shells. The solutions we have presented can be seen as sources for exterior Levi-Civita vacuum space-time. Our difference is that we have chosen the interior space-time non-flat and it is described by a Levi-Civita metric and contains a line singularity at the symmetry axis in general. However we can replace this singularity with a shell or a cylinder with finite radius. Choosing the interior not flat, we can consider the possibility that the shell can be a part of a multiple shell system. The generalization of these results to multiple shell system is straightforward. Choosing the interior of the innermost shell regular, we can avoid having a singularity at the axis. Two concentric shells [@gleiser] or multiple shells composed of photons [@AD2] is presented recently. Thus considering previous works and our results, one can find multiple shells having different equation of states and satisfying certain energy conditions.
We have imposed isotropic or anisotropic equations of state for the shell. Some solutions satisfy energy conditions. For example, the perfect fluid, dustlike shell, shells with $p_z=0$, satisfies the weak, the strong and the dominant energy conditions for some ranges of the parameters. For a fixed $s$, one can find these ranges which are only depend on $s'$ and the interior and the exterior radial coordinates on the shell. The shells can have any radius. Their energy density and mass per unit length are bounded from above. For the pressureless dust shell, it satisfies all of the energy conditions with the expense of negative line mass at the axis. This is also true for a shell with particles moving only in the axial direction. Similar behavior has been observed for a static cylindrical shell with counter moving photons in $z$ direction [@AD1]. This might be an expected result since for these cases there is no pressure in the fluid against collapse ($p_\phi=0$) and we need an repelling force to keep these shells static.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Ahmet Baykal for reading the manuscript and help with the figures.
[99]{} Levi-Civita, T. (1919). Rend. Acc. Lincei [**28**]{}, 101. Bonnor, W. B. (1999). The static cylinder in general relativity [*On Einstein’s Path*]{} ed A Harvey (New York: Springer) p 113;\
Bicak, J. Ledvinka, T. Schmidt, B. G. and Zofka, M. (2004). Class. Quant. Grav. [**21**]{}, 1583. Marder, L. (1958). Proc. R. Soc. A [**244**]{}, 524;\
Bonnor, W. B. (1979). J Phys. A: Math. Gen. [**12**]{}, 847;\
Lathrop, J. D. and Orsense, M. S.(1980). J. Math. Phys. [**21**]{}, 152;\
Bonnor, W. B. and Davidson, W. (1992). Class. Quantum Grav. [**9**]{}, 2065;\
da Silva, M. F. A., Herrera,L., Paiva, F. M. and Santos, N. O. (1995). J. Math. Phys. [**36**]{}, 3625;\
Haggag, S. and Desokey, F. (1996). Class Quantum Grav. [**13**]{}, 3221;\
Philbin, T. G. (1996). Class. Quantum Grav. [**13**]{}, 1217. Stachel, J. (1983). J. Math. Phys. [**25**]{}, 338. Cl´ement,G., and Zouzou,I. (1994). Phys. Rev. D [**50**]{}, 7271. Wang, A. Z., da Silva,M. F. A., and Santos,N. O. (1997). Class. Quantum Grav. [**14**]{}, 2417. Herrera, L., Santos, N. O., Teixeira, A. F. F. and Wang, A. Z. (2001). Class. Quantum Grav. [**18**]{}, 3847. Gleiser, R. J., and Tiglio, M. H. (2000). Phys Rev D [**61**]{}, 104006. Vilenkin, A. and Shellard, E. P. S. (1994). [*Cosmic Strings and other Topological Defects*]{}, Cambridge University Press, Cambridge. da Silva, M. F. A., Wang, A. Z., and Santos, N. O. (1998). Phys. Lett. A [**244**]{}, 462. Bicak, J. and Zofka, M. (2002). Class. Quantum Grav. [**19**]{}, 3653. Arik, M. and Delice, O.(2003). Int. J Mod. Phys. D [**12**]{}, 1095. Arik, M. and Delice, O. (2003). Gen. Relat. Gravit. [**35**]{}, 1285. Zofka, M. (2004). Class. Quantum Grav. [**21**]{}, 465. Israel,W. (1966) Nuovo Cimento [**B44**]{}, 1; Israel, W. (1967). [*ibid*]{} [**B48**]{}, 463(E). Hawking, S. W. and Ellis, G. F. R. (1973). [*The Large Scale Structure of Spacetime*]{} p 88, Cambridge University Press, Cambridge. Tsoubelis, D. (1989). Class. Quantum Grav. [**6**]{}, 101.
[^1]: [Department of Physics, Bogazici University, Bebek, Istanbul, Turkey; e-mail: [email protected]]{}
[^2]: Department of Physics, Bogazici University, Bebek, Istanbul, Turkey; e-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study by photoacoustic spectroscopy the band–gap shift effect of CdS films. The CdS films were grown by chemical bath deposition and exposed to different annealing atmospheres over a range of temperature in which the sample structure is observed to change. We show the band–gap evolution as a function of temperature of thermal annealing and determine the process which produces the best combination of high band–gap energy and low resistivity. It allows us to know a possible procedure to obtain low–resistivity CdS/CdTe solar cells with high–quantum efficiency.'
author:
- |
S.A. Tomás, O. Vigil$^*$, J.J. Alvarado-Gil, R. Lozada-Morales,\
O. Zelaya-Angel\
Departamento de Física\
H. Vargas$^{\dag}$\
Programa Multidisciplinario en Ciencias Aplicadas y\
Tecnología Avanzada.\
Centro de Investigación y de Estudios Avanzados del IPN\
Apdo. Postal 14-740, México D.F. 07000, México\
A. Ferreira da Silva\
Instituto Nacional de Pesquisas Espaciais-INPE\
Laboratorio Associado de Sensores e Materiais-LAS\
CP 515,12201-970 Sao Jose dos Campos, S.P., Brazil
date:
title: 'Study of the band–gap shift in CdS films: Influence of thermal annealing in different atmospheres.'
---
1em
$^*$ Permanent address: Physics Faculty, University of La Havana, La Havana, Cuba.\
$^{\dag}$ On leave of absence from the Universidade Estadual de Campinas, S.P., Brazil.\
PACS N. 43.85, 62.65, 78.20.D, 78.65, and 81.40.E
24.1pt [Introduction]{}\
The study of heterojunctions and their use in a large variety of technological applications has become one of the principal subjects in solid state research. Among other applications, we mention the use of heterojunctions in transistors and diodes, as well as in other optoelectronic devices. Likewise, it is well known the importance represented by heterojunctions in solar cells development. In particular, CdS/CdTe heterojunctions, represent one of the principal candidates for low cost and high conversion-efficiency solar cells [@Ferekides]. In this type of solar cells, the obtention of CdS films with low resistivity plays an important role because it helps to diminish the dipositive sheet resistance and obtain the space charge region in the active zone, i.e., in the CdTe film.\
It is also known that CdS samples present two different structural phases, namely, the highly stable hexagonal phase and the metastable cubic phase. Even though it is possible to know by means of reflectivity measurements at room–temperature that these two phases have a band–gap energy differing less than 0.1eV, it is not possible, however, to infer any other conclusion about them [@Cardona].\
CdS films present a band–gap shift (BGS) effect when exposed to certain experimental conditions. This effect was reported by Balkanski [*et al.*]{} [@Balkanski] for Si films as a function of temperature. Later on, because of its technological importance in optoelectronic design, many researchers have investigated the BGS of intrinsic and extrinsic semiconductors [@Berggren]-[@Wagner]. Recently, Zelaya–Angel [*et al.*]{} [@Zelaya] have reported the first investigation of BGS as a function of temperature of thermal annealing (TTA) for CdS films. Nevertheless, it is worth saying that such a BGS effect, as well as the growth procedures to obtain CdS films in either of the phases, are not well understood up to now.\
In this work we investigate by photoacoustic spectroscopy (PAS) the band–gap shift effect of CdS films as a function of TTA in different annealing atmospheres. By extension, we show the interesting and technologically important combination of low resistivity CdS/CdTe solar cells with high energy-gap. The annealing atmospheres investigated are Ar, Ar+S$_2$, H$_2$ + In, H$_2$, and air, each of them fixed at certain temperatures within a range from 200$^o$C to 450$^o$C. The purpose of studying it by PAS is that this technique allows us to obtain spectra which clearly show the BGS effect [@Zelaya]-[@Vargas].\
[Experimental.]{}\
The approximately 0.25$\mu$m thickness cadmium sulfide films were deposited by the chemical bath deposition method (CBD). This deposition occurs due to an ionic precipitation of salts dissolved in an aqueous solution. The salts used were CdCl$_2$, K0H, NH$_4$NO$_3$, and CS\[NH$_2$\]$_2$ (thiourea). The solution was kept at (80$\pm$3)$^o$C and continuously stirred to guarantee an homogeneous distribution of the chemical compounds, among other effects [@Kaur].\
The postdeposition thermal treatments given to the samples were performed in a hot wall furnace with a quartz tube. Taking advantage of the furnace temperature profile, CdS samples were simultaneously placed within a range of temperature from 200$^o$C to 450$^o$C. For Ar gas and Ar gas + S$_2$ vapor flux annealings, CdS samples were placed at 208, 260, 353, 398, and 447$^o$C (419$^o$C the last for Ar+S$_2$) during 28h. The S$_2$ vapor flux was obtained by placing a boat containing sulfur at 190$^o$C. For H$_2$ gas, H$_2$ gas + In vapor flux, and air annealings, CdS samples were placed at 200, 250, 300, 350, 400, and 450$^o$C during 1h. In the H$_2$+In case, a boat containing In was placed at 630$^o$C, whereas for the H$_2$ annealing, an H$_2$ flow was kept for 30 min. before placing the samples. The stimated pressure for all the treatments was about 1 atm.\
The photoacoustic absorption spectra of the films were obtained in the region 400–700nm by use of a standard photoacoustic spectrometer. This spectrometer is fitted with a 1000W Xenon lamp (Oriel), whose radiation is focused onto a variable–frequency light chopper set at 17 Hz. The photoacoustic signal is preamplified before providing the input to the signal channel lock–in amplifier. The resultant photoacoustic spectra are recorded in a computer, which simultaneously display the wavelength–dependent signal intensity.\
[Results and discussion.]{}\
In Fig.1 we illustrate the PA spectra corresponding to both the as–grown CdS sample and the whole set of CdS films annealed in Ar+S$_2$ atmosphere. We assume that the value of the band–gap energy is given by the energy which matches the spectrum inflexion point. With this assumption, we observe the BGS effect as a function of TTA. The values of the band–gap energy (BGE) for the samples annealed in Ar and Ar+S$_2$ atmosphere are shown in Table I, whereas the ones for H$_2$, H$_2$+In, and air are shown in Table II. Focusing us on the Ar+S$_2$ annealing we obtain the BGE, as shown in Fig.2. We observe that the BGE starts decreasing its value as a consequence of the thermal annealing. This effect carries down the BGE from its 2.42 eV initial value at 80$^o$C to 2.28 eV at roughly 350$^o$C, where a minimum is reached. Above this temperature the BGE presents an opposite behavior, increasing its value until 2.35eV at 418$^o$C. This narrowing–widening–like behavior presented by CBD/CdS films annealed in Ar+S$_2$ atmosphere has been interpreted as a cubic to hexagonal phase transition, which occurs at the temperature corresponding to the minimum BGE [@Zelaya]. The TTA coefficients of the BGE have also been approximately found in the linear part of the data [@Burger] with values of about $\mp 10^{-4}$ eV/TTA($^oK$), corresponding respectively to the narrowing and widening region. The magnitude of this coefficient agrees generally with those of other reported materials, e.g., Si, Ge, GaAs, and MgI$_2$ [@Burger],[@Sze].\
With regard to the other annealing atmospheres, we show in Tables I and II that all of them start with the same behavior as the Ar+S$_2$ case, that is, the samples present a narrowing–like behavior. Some of these annealings decrease the BGS faster of slower and reach a higher or lower minimum value than Ar+S$_2$. Nevertheless, it is very important to point out that none presents a clear widening–like behavior as does Ar+S$_2$. On the contrary, all of them show a stable tendency above the temperature corresponding to the minimum BGE. These results are illustrated in Fig.3 where the BGS, given by
$$\Delta Eg(T) = Eg(80^oC) - Eg(T),$$
is plotted versus the TTA. Here, $Eg(80^o$C) is the band–gap energy of the as–grown samples. We notice that, among the investigated annealings, the highest $\Delta Eg$ values are produced by the air, the maximum $\Delta Eg$ = 0.17 eV taking place at roughly 350$^o$C. Moreover, only the Ar+S$_2$ annealing produces a considerable diminution of the BGE after having reached its maximum value.\
Owing to our interest in finding the thermal process which produces CBD/CdS films with the best combination of high band-gap energy and low resistivity, we deal now with the sample electrical resistivity as a function of TTA. As all the annealing atmospheres produce similar resistivity effects in the samples, we take advantage of it and describe qualitatively their general behavior. Electrical measurements were performed by the two probes method for high resistivity films and the four probes method for low resistivity films at room temperature. The four probes technique was not satisfactory for high resistivity films because of the same order of magnitude of the film resistance and the electrometer input impedance. Resistivity of the as–grown samples was measured to be $\rho = 2.3\times 10^7 \Omega$ cm. Subsequently, as the annealings were performed, the resistivity decreased with increasing TTA until reaching a minimum value, then increasing again. As can be seen in Table III, $\rho(T)$ takes different minimum values for each different annealing atmosphere.\
The different minimum values of $\rho$ are analysed below. The Ar+S$_2$ annealing diminishes the resistivity by increasing the grain size, therefore decreasing the number of grain boundaries in the films. On the other hand, the Ar annealing, besides of in creasing the grain size, originates a high amount of sulfur vacancies as a consequence of non-equilibrium conditions. In the air annealing, the presence of oxygen creates both CdO and SO$_4$Cd layers on the surface, increasing slightly the resistivity of the films respect to the Ar case [@Kolhe]. The thermal treatments in H$_2$ and H$_2$+In present the advantage of containing H$_2$, which is a strong agent for grain boundary passivation by oxygen chemiabsorption [@Melo],[@Chopra], decreasing the resistivity by 4 and 5 or ders of maagnitude, respectively, respect to the former cases.\
Now, in order to consider simultaneously the band–gap energy and the resistivity, we define
$$\xi(T) = \frac{\Delta\rho (T)}{\Delta Eg(T)} \ \ = \ \ \frac{\rho (80^oC)-
\rho(T)}{Eg(80^oC)-Eg(T)} \quad .$$
\
The above parameter contains the information we are interested in. In effect, high band–gap energy $Eg(T)$ is equivalent to low $\Delta Eg(T)$, and similarly, low resistivity $\rho(T)$ implies high $\Delta\rho(T)$. Thus, the higher $\xi(T)$ value, the better combination of high $Eg(T)$ and low $\rho(T)$. According with it, we have found a procedure to find out the TTA which produces samples with most favorable characteristics. We should observe that $\xi(T)$ is basically governed by $\Delta\rho(T)$ because of its high values compared to the $\Delta Eg(T)$ ones. It gives us an idea of what TTA should be in advance investigated, namely, those with high $\Delta\rho(T)$ values.\
We have calculated $\xi(T)$ for all the annealing atmospheres. With their highest values, which are written in Table IV, we have plotted the figure of merit shown in Fig.4. This figure allows us to know that, among the analysed annealings, the best effects are produced by H$_2$ and H$_2$+In atmospheres. In particular, H$_2$+In annealing is lightly better than H$_2$ annealing, its best value $\xi= 4.6 \times 10^6 \Omega$ cm/eV taking place at 250$^o$C. It is worth to point out that the measured resistivity of this sample is $\rho= 5 \times 10^{-2} \Omega$ cm. A similar value has been reported [@Toyotomi] but for the crystal grown by the melt, using the Tamman method in Ar gas of about 110 atm pressure. Moreover, for the Ar+S$_2$ atmosphere, the minimum resistivity is $\rho= 1.9 \times 10^3 \Omega$ cm. This resistivity is roughly reproduced $(2.3 \times 10^3 \Omega$ cm) for a normal CdS crystal but at 4.2$^oK$[@Toyotomi].\
The X-Ray diffraction patterns show that all the annealed samples present a light interplanar–distance increase compared to the as–grown samples, but only the samples exposed to Ar and Ar+S$_2$ annealings undergo a cubic to hexagonal phase transition. On the contrary, the samples annealed in H$_2$, H$_2$+In and air atmospheres preserve the same structural phase along the complete thermal process [@Melo]. The Mechanism that explains this microestructural behavior is under study at present.\
Photoluminiscence spectra of CdS samples annealed in Ar and Ar+S$_2$ atmospheres present a shift of the luminescent peaks as a function of TTA. This shift has been interpreted as an evidence of evolution from S$_2$ vacancies to interstitial S$_2$ formation. For these two annealings, where a cubic to hexagonal phase transition is completed, interstitial S$_2$ becomes part of the hexagonal structure itself. PL spectra of CBD/CdS samples annealed in H$_2$, H$_2$+In, and air atmospheres has not been performed up to now.\
In summary we have found that CBD/CdS films present optimum conditions of low resistivity and high band–gap energy when exposed to H$_2$+In annealing atmosphere at 250$^o$C. This result allows us to know a possible procedure to obtain low resistivity CdS/CdTe solar cells with high quantum efficiency, therefore a real candidate for high conversion-efficiency solar cells.\
[Acknowledgements]{}\
This work was partially supported by the Mexican agency CONACYT. H. V. and A. F. S. acknowledge the Brazilian agency CNPq/MCT for financial support. A. F. S. acknowledges CINVESTAV-IPN for its hospitality.
[99]{} J. Britt and C. Ferekides, Appl. Phys. Lett. [**62**]{}, 2851 (1993) M. Cardona, M. Weinstein, and G. A. Wolff, Phys. Rev. [**140**]{}, A633 (1965) M. Balkanski, A. Aziza and E. Amzallag, Phys. Status Solidi [**31**]{}, 323 (1969) K.-F. Berggren and B.E. Sernelius, Phys. Rev. B[**24**]{}, 1971 (1981) W.P. Dumke, Appl. Phys. Lett. [**42**]{}, 196 (1983) K.E. Newman, A. Lastras-Martínez, B. Kramer, S.A. Barnett, M.A. Ray, J.D. Dow, J.E. Green, and P.M. Raccach, Phys. Rev. Lett. [**50**]{}, 1466 (1983) B.E. Sernelius, K.-F. Berggren, Z.C. Jin, I. Hamberg, and C.G. Granquist, Phys. Rev. B[**37**]{}, 10244 (1988) J. Wagner and J.A. del Alamo, J. Appl. Phys. [**63**]{}, 425 (1988) O. Zelaya–Angel, J.J. Alvarado-Gil, R. Lozada-Morales, H. Vargas, and A. Ferreira da Silva, Appl. Phys. Lett [**64**]{}, 291 (1994) L. Eaves, H. Vargas, and P.J. Williams, Appl. Phys. Lett. [**38**]{}, 768 (1981) A. Mandelis, [*Photoacoustic and Thermal Wave Phenomena in Semiconductores*]{}, (North Holland, New York, 1987) H. Vargas and L.C.M. Miranda, Phys. Rep. [**161**]{}, 43 (1988) I. Kaur, D.K. Pandya, and K.L. Chopra, J. Electrochem. Soc. [**127**]{}, 943 (1980) A. Burger and D. Nason, J. Appl. Phys. [**71**]{}, 2717 (1992) S.M. Sze, [*Physics of Semiconductor Devices*]{}, (New York, John Wiley, 1981), Pg. 15 S. Kolhe, S.K. Kulkarni, A.S. Nigavekar, and S.K. Sharma, Sol. Energy Mat. [**10**]{}, 47 (1984) O. de Melo, L. Hernández, O. Zelaya-Angel, R. Lozada-Morales, M. Becerril, and E. Vasco, Appl. Phys. Lett [**65**]{}, 1278 (1994) K.L. Chopra, R.C. Kaaintla, D.K. Pandya, and A.P. Thakoor, Phys. Thin Films [**12**]{}, 167 (1987) S. Toyotomi and K. Morigaki, J. Phys. Soc. Jpn. [**25**]{}, 807 (1968)
[TABLE I]{}. Values of the band–gap energy, $E_g$, and band–gap shift, $\Delta E_g$, as a function of temperature of thermal annealing for Ar and Ar+S$_2$ annealings.
--------------- ---------- ----------------- ---------- -----------------
$Eg$(eV) $\Delta Eg$(eV) $Eg$(eV) $\Delta Eg$(eV)
80 (as–grown) 2.42 – 2.42 –
208 2.40 0.02 2.40 0.02
260 2.36 0.06 2.37 0.05
296 2.35 0.07 2.29 0.13
353 2.30 0.12 2.28 0.14
398 2.29 0.13 2.33 0.09
418 2.34 0.08
447 2.31 0.11
--------------- ---------- ----------------- ---------- -----------------
[TABLE II.]{} Values of the band–gap energy, $E_g$, and band–gap shift, $\Delta E_g$, as a function of temperature of thermal annealing for H$_2$, H$_2$+In and air annealings.
-------------- ---------- ----------------- ---------- ----------------- ---------- ---------
$Eg$(eV) $\Delta Eg$(eV) $Eg$(eV) $\Delta Eg$(eV) $Eg$(eV) $\Delta
Eg$(eV)
80(as–grown) 2.42 – 2.42 – 2.42 –
200 2.40 0.02 2.40 0.02 2.37 0.05
250 2.39 0.03 2.37 0.05 2.34 0.08
300 2.37 0.05 2.34 0.08 2.27 0.15
350 2.31 0.11 2.28 0.14 2.25 0.17
400 2.27 0.15 2.27 0.15 2.26 0.16
450 2.27 0.15 2.28 0.14 2.26 0.16
-------------- ---------- ----------------- ---------- ----------------- ---------- ---------
[TABLE III.]{} Resistivity minimum values and corresponding $E_g$ as a function of temperature in different annealing atmospheres.
----------- ------------------ ------- -------------
annealing minimum $\rho$ $E_g$ temperature
($\Omega$ cm) (eV) ($^o$C)
Ar 2.5$\times 10^2$ 2.30 353
Ar+S$_2$ 1.9$\times 10^3$ 2.29 296
H$_2$ 0.15 2.37 300
H$_2$+In 0.05 2.37 250
air 1.0$\times 10^3$ 2.27 300
----------- ------------------ ------- -------------
[TABLE IV.]{} $\xi$(T) highest values in different annealing atmospheres.
----------- ----------------------- -------------
annealing highest $\xi$ temperature
(10$^8 \Omega$ cm/eV) ($^o$C)
air 1.53 300
Ar+S$_2$ 1.76 296
Ar 2.09 447
H$_2$ 4.59 300
H$_2$+In 4.60 250
----------- ----------------------- -------------
[Figure Captions.]{}\
1. PA Normal Intensity of CdS films annealed in Ar+S$_2$ atmosphere for different TTA as a function of energy.
2. Band–gap energy dependence on TTA for CdS annealed in Ar+S$_2$ atmosphere.
3. Band–gap shift dependence on TTA for CdS annealed in different atmospheres.
4. $\xi(T)$ highest value for each different annealing atmosphere.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study Barut’s covariant equations describing the electromagnetic interactions between $N$ spin-$1/2$ particles. In the covariant formulation each particle is described by a Dirac spinor. It is assumed that the interactions between the particles are not mediated by a bosonic field (direct interactions). Within this formulation, using the Lagrangian formalism, we derive the approximate (semirelativistic) Breit equation for two interacting spin-$1/2$ particles.'
---
[**A derivation of the Breit equation from Barut’s covariant formulation of electrodynamics in terms of direct interactions**]{} .6in
Domingo J. Louis-Martinez .2in
Department of Physics and Astronomy,\
University of British Columbia\
Vancouver, Canada, V6T 1Z1
[email protected]
PCAS No. 03.30.+p, 03.50.-z, 03.65.Pm, 11.10.Ef.
It is well known that an isolated system of $N$ particles interacting electromagnetically may be described, to terms of second order, by the Darwin Lagrangian [@darwin],[@landau].
At the quantum level one can describe a system of two spin-$1/2$ particles interacting electromagnetically, including second order terms, by the Breit equation [@breit; @dirac1]. Perturbative solutions of the Breit equation give results for the energy spectrum of relativistic bound systems in agrement with experiments [@bethe; @kroli; @fushchich].
A derivation of the Breit equation from standard quantum field theory was given in [@bethe]. A more recent derivation in terms of direct interactions between fermions was given in [@duviryak]. The derivation presented in [@duviryak] uses elements of quantum field theory and requires employing an unconventional vacuum state.
In this paper we work within the theory proposed by Barut [@barut1; @barut2]. This theory is an alternative approach to quantum field theory and does not require second quantization. It is a theory of direct interaction between spin-$1/2$ particles. Each particle is described by a Dirac spinor. It is assumed that the electromagnetic interactions between the particles are not mediated by a bosonic field (direct interactions).
Our goal in this paper is to derive the Breit equation directly from Barut’s covariant equations describing the electromagnetic interactions between two spin-$1/2$ particles.
The Darwin Lagrangian [@darwin],[@landau], describing a system of $N$ particles interacting electromagnetically, to terms of second order, can be written as:
$$\begin{aligned}
L_{Darwin} & = & - c^{2}\sum\limits^{}_{a} m_{a}
+ \frac{1}{2} \sum\limits^{}_{a} m_{a} v_{a}^{2}
+ \frac{1}{8 c^{2}}\sum\limits^{}_{a} m_{a} v_{a}^{4}
- \frac{1}{2} \sum\limits^{}_{a} \sum\limits^{}_{b \neq a}
\frac{e_{a} e_{b}}{r_{ab}}
\nonumber\\
& & + \frac{1}{4 c^{2}}
\sum\limits^{}_{a} \sum\limits^{}_{b \neq a}
\frac{e_{a} e_{b}}{r_{ab}}
\left((\vec{v}_{a} \vec{v}_{b})
+(\vec{n}_{ab} \vec{v}_{a})(\vec{n}_{ab} \vec{v}_{b})\right),
\label{P1}\end{aligned}$$
where $c$ is the speed of light, $m_{a}$ and $e_{a}$ are the mass and electric charge of particle $a$ ($a = 1,2,...,N$), $\vec{v}_{a}$ its velocity, and $\vec{r}_{ab} = \vec{r}_{a} - \vec{r}_{b}$ the relative position of particle $a$ with respect to particle $b$. In (\[P1\]), $\vec{n}_{ab} \equiv \frac{\vec{r}_{ab}}{r_{ab}}$.
The Darwin Lagrangian can be derived either from Faraday-Maxwell’s field theory of electrodynamics [@landau] or from the relativistic action-at-a-distance theory of Wheeler and Feynman [@anderson; @domingo3].
To terms of fourth order ($\frac{v^4}{c^4}$), ignoring the radiation effects (which for Faraday-Maxwell electrodynamics appear at third order) the Lagrangian for an isolated system of N particles interacting electromagnetically was obtained in [@smoro], [@ohta].
At the quantum level a system of two spin-$1/2$ particles interacting electromagnetically, including second order terms, can be described by the Breit equation [@breit; @dirac1]:
$$\begin{aligned}
& & \left(i\hbar \frac{\partial}{\partial t} + i\hbar c \vec{\alpha}_{1}\frac{\partial}{\partial \vec{r}_{1}} +
i\hbar c \vec{\alpha}_{2}\frac{\partial}{\partial \vec{r}_{2}} - m_{1}c^{2}\beta_{1} - m_{2}c^{2}\beta_{2}\right.\nonumber\\
& & \left. - \frac{e_{1} e_{2}}{|\vec{r}_{1} -\vec{r}_{2}|} \left(1 - \frac{1}{2}\left((\vec{\alpha}_{1}\vec{\alpha}_{2})
+ \frac{(\vec{\alpha}_{1}(\vec{r}_{1} - \vec{r}_{2}))}{|\vec{r}_{1}
-\vec{r}_{2}|}\frac{(\vec{\alpha}_{2}(\vec{r}_{1} - \vec{r}_{2}))}{|\vec{r}_{1} -\vec{r}_{2}|}\right)\right)\right)
\Psi(t, \vec{r}_{1}, \vec{r}_{2}) = 0.
\label{P2}\end{aligned}$$
where $\Psi(t, \vec{r}_{1}, \vec{r}_{2})$ is a two-body bilocal field defined as follows:
$$\Psi(t,\vec{r}_{1},\vec{r}_{2}) = \psi_{1}\left(t, \vec{r}_{1}\right) \otimes \psi_{2}\left(t, \vec{r}_{2}\right).
\label{P3}$$
and $\vec{\alpha}$ and $\beta$ are Dirac’s matrices[@bjorkendrell], $\otimes$ is used to denote a tensor product, and $\hbar$ is Planck’s constant.
Equation (\[P2\]) is consistent with the Darwin Lagrangian (\[P1\]) for $N = 2$ [@breit]. It was obtained by following the simple recipe [@breit; @2bDirac] of substituting the free terms in (\[P1\]) by the left hand sides of the free Dirac equation for each particle, and by substituting the velocities $\vec{v}_{a}$ by $c \vec{\alpha}_{a}$ in the interaction terms of (\[P1\]).
Let us consider a system of $N$ distinguishable spin-$1/2$ particles interacting electromagnetically. In Barut’s formulation each particle is described by a Dirac spinor $\psi_{a}(x)$ ($a=1,2,...,N$). It is assumed that the interactions between particles are not mediated by a bosonic field. The action functional for each individual particle can be written as[@barut1; @barut2]:
$$S^{(a)} = \int d^{4}x \left(i\hbar \bar{\psi}_{a}(x) \gamma^{\mu} \partial_{\mu} \psi_{a}(x) - m_{a}c \bar{\psi}_{a}(x)\psi_{a}(x)\right)
- \frac{e_{a}}{c} \int d^{4}x A^{(a)}_{\mu}(x) \bar{\psi}_{a}(x) \gamma^{\mu} \psi_{a}(x),
\label{1}$$
where,
$$A^{(a)}_{\mu}(x) = \sum\limits_{b \ne a}^{} e_{b} \int d^{4}y \delta\left((x - y)^{2}\right) \bar{\psi}_{b}(y) \gamma_{\mu} \psi_{b}(y).
\label{2}$$
The argument of the Dirac delta function in (\[2\]) is $(x - y)^{2} = \eta_{\mu\nu} (x^{\mu} - y^{\mu})(x^{\nu} - y^{\nu})$, where the metric tensor $\eta_{\mu\nu} = diag(+1, -1, -1, -1)$. The Dirac matrices $\gamma^{\mu}$ are defined as usual[@bjorkendrell]: $\gamma^{0}=\beta$, $\vec{\gamma}=\beta\vec{\alpha}$. In (\[1\],\[2\]) $\bar{\psi}_{a}(x) = \psi^{\dagger}_{a}(x) \gamma^{0}$.
The action functional for the whole system of $N$ interacting distinguishable spin-$1/2$ particles can be given in the form[@barut2]:
$$\begin{aligned}
S & = & \sum\limits_{a}^{}\int d^{4}x \left(i\hbar \bar{\psi}_{a}(x) \gamma^{\mu} \partial_{\mu} \psi_{a}(x) - m_{a}c \bar{\psi}_{a}(x)\psi_{a}(x)\right)\nonumber\\
& & - \frac{1}{2c} \sum\limits_{a}^{} \sum\limits_{b \ne a}^{} e_{a} e_{b}\int\int d^{4}x d^{4}y \delta\left((x - y)^{2}\right)
\bar{\psi}_{a}(x) \gamma^{\mu} \psi_{a}(x) \bar{\psi}_{b}(y) \gamma_{\mu} \psi_{b}(y).
\label{3}\end{aligned}$$
The action (\[3\]) is invariant under Lorentz transformations. At the classical level, (\[3\]) corresponds to the action functional of Wheeler-Feynman action-at-a-distance theory of electrodynamics[@wheeler; @anderson; @hoyle; @barut]:
$$S = - \sum\limits^{}_{a} m_{a} c \int d s_{a} \left(\dot{z}_{a}^{2}\right)^{\frac{1}{2}}
- \frac{1}{2 c}\sum\limits^{}_{a} \sum\limits^{}_{b\neq a}
e_{a} e_{b}
\int \int d s_{a} d s_{b}
\delta\left(\left(z_{a} - z_{b}\right)^{2}\right) \left(\dot{z}_{a} \dot{z}_{b}\right).
\label{4}$$
In (\[4\]), $s_{a}=c\tau_{a}$, where $\tau_{a}$ is the proper time of particle $a$, $z_{a}^{\mu}(s_{a})$ is its world line in Minkowski spacetime, and the 4-vector velocity $\dot{z}_{a}^{\mu} = \frac{d z_{a}^{\mu}}{d s_{a}}$. The Dirac delta function in (\[4\]) accounts for the interactions propagating at the speed of light forward and backward in time.
The fully relativistic equations of motion corresponding to the action functional (\[4\]) admit exact circular solutions for any number of particles [@schild; @domingo].
From the action functional (\[3\]) the covariant equations for the spin-$1/2$ particles can be obtained as follows[@barut2]:
$$i\hbar \gamma^{\mu} \partial_{\mu} \psi_{a}(x) - m_{a}c \psi_{a}(x)
- \frac{e_{a}}{c} \sum\limits_{b \ne a}^{} e_{b}\int d^{4}y \delta\left((x - y)^{2}\right)
\bar{\psi}_{b}(y) \gamma_{\mu} \psi_{b}(y) \gamma^{\mu} \psi_{a}(x) = 0,
\label{5}$$
$$i\hbar \partial_{\mu} \bar{\psi}_{a}(x) \gamma^{\mu} + m_{a}c \bar{\psi}_{a}(x)
+ \frac{e_{a}}{c} \sum\limits_{b \ne a}^{} e_{b}\int d^{4}y \delta\left((x - y)^{2}\right)
\bar{\psi}_{b}(y) \gamma_{\mu} \psi_{b}(y) \bar{\psi}_{a}(x)\gamma^{\mu} = 0.
\label{6}$$
From (\[5\], \[6\]) it follows that:
$$\partial_{\mu}\left(\bar{\psi}_{a}(x) \gamma^{\mu} \psi_{a}(x)\right) = 0,
(a=1,...,N).
\label{7}$$
These are the well known continuity equations[@bjorkendrell].
We use the normalization conditions:
$$\int d^{3}\vec{x} \psi^{\dagger}_{a}(t,\vec{x})\psi_{a}(t,\vec{x}) = 1.
\label{8}$$
In [@barut1; @barut2] from (\[5\], \[6\]) an exact covariant equation for the two-body bilocal field:
$$\Phi(t,\vec{x},\vec{y}) = \psi_{1}\left(t, \vec{x}\right) \otimes \psi_{2}\left(t - \frac{|\vec{x}-\vec{y}|}{c}, \vec{y}\right)$$
was obtained.
In this paper we are interested in obtaining an approximate (semirelativistic) equation for the two-body bilocal field (\[P3\]):
$$\Psi(t,\vec{x},\vec{y}) = \psi_{1}\left(t, \vec{x}\right) \otimes \psi_{2}\left(t, \vec{y}\right).
\label{9}$$
We will now derive the approximate (semirelativistic) Breit equation for the two-body bilocal field (\[9\]) directly from Barut’s covariant equations describing the interactions between two spin-$1/2$ particles.
In order to obtain the Breit equation for $\Psi(t,\vec{x},\vec{y})$ (\[9\]) in the case of electrodynamics, we return to the action functional (\[3\]) for $N=2$:
$$\begin{aligned}
S & = & \int d^{4}x \left(i\hbar \bar{\psi}_{1}(x) \gamma^{\mu} \partial_{\mu} \psi_{1}(x) - m_{1}c \bar{\psi}_{1}(x)\psi_{1}(x)\right)\nonumber\\
& + & \int d^{4}y \left(i\hbar \bar{\psi}_{2}(y) \gamma^{\mu} \partial_{\mu} \psi_{2}(y) - m_{2}c \bar{\psi}_{2}(y)\psi_{2}(y)\right)\nonumber\\
& & - \frac{e_{1} e_{2}}{c} \int\int d^{4}x d^{4}y \delta\left((x - y)^{2}\right)
\bar{\psi}_{1}(x) \gamma^{\mu} \psi_{1}(x) \bar{\psi}_{2}(y) \gamma_{\mu} \psi_{2}(y).
\label{10}\end{aligned}$$
In (\[10\]) we write the Dirac delta function:
$$\delta\left((x - y)^{2}\right) = \frac{1}{2|\vec{x}-\vec{y}|} \left(\delta(y^{0} - x^{0} + |\vec{x}-\vec{y}|) +
\delta(y^{0} - x^{0} - |\vec{x}-\vec{y}|)\right).
\label{11}$$
Therefore, we can write the third term in (\[10\]) as:
$$\begin{aligned}
& & - \frac{e_{1} e_{2}}{c} \int\int d^{4}x d^{4}y \delta\left((x - y)^{2}\right)
\bar{\psi}_{1}(x) \gamma^{\mu} \psi_{1}(x) \bar{\psi}_{2}(y) \gamma_{\mu} \psi_{2}(y) = \nonumber\\
& & - \frac{e_{1} e_{2}}{2} \int dt \int d^{3}\vec{x} \int \frac{d^{3}\vec{y}}{|\vec{x}-\vec{y}|}
\left(\bar{\psi}_{2}\left(t - \frac{|\vec{x}-\vec{y}|}{c},\vec{y}\right) \gamma_{\mu} \psi_{2}\left(t- \frac{|\vec{x}-\vec{y}|}{c},\vec{x}\right) +\right.\nonumber\\
& &\left. \bar{\psi}_{2}\left(t + \frac{|\vec{x}-\vec{y}|}{c},\vec{y}\right) \gamma_{\mu} \psi_{2}\left(t + \frac{|\vec{x}-\vec{y}|}{c},\vec{x}\right)\right)
\bar{\psi}_{1}(t,\vec{x}) \gamma^{\mu} \psi_{1}(t,\vec{x})
\label{12}\end{aligned}$$
Using the approximate expressions:
$$\begin{aligned}
\bar{\psi}_{2}\left(t - \frac{|\vec{x}-\vec{y}|}{c},\vec{y}\right) \gamma_{\mu} \psi_{2}\left(t- \frac{|\vec{x}-\vec{y}|}{c},\vec{x}\right)
& \approx & \bar{\psi}_{2}(t,\vec{y}) \gamma_{\mu} \psi_{2}(t,\vec{y})
- \frac{|\vec{x}-\vec{y}|}{c}\frac{\partial}{\partial t}\left(\bar{\psi}_{2}(t,\vec{y}) \gamma_{\mu} \psi_{2}(t,\vec{y})\right)\nonumber\\
& & +\frac{|\vec{x}-\vec{y}|^{2}}{2c^{2}}\frac{\partial^{2}}{\partial t^{2}}\left(\bar{\psi}_{2}(t,\vec{y}) \gamma_{\mu} \psi_{2}(t,\vec{y})\right)
\label{13}\end{aligned}$$
$$\begin{aligned}
\bar{\psi}_{2}\left(t + \frac{|\vec{x}-\vec{y}|}{c},\vec{y}\right) \gamma_{\mu} \psi_{2}\left(t + \frac{|\vec{x}-\vec{y}|}{c},\vec{x}\right)
& \approx & \bar{\psi}_{2}(t,\vec{y}) \gamma_{\mu} \psi_{2}(t,\vec{y})
+ \frac{|\vec{x}-\vec{y}|}{c}\frac{\partial}{\partial t}\left(\bar{\psi}_{2}(t,\vec{y}) \gamma_{\mu} \psi_{2}(t,\vec{y})\right)\nonumber\\
& & +\frac{|\vec{x}-\vec{y}|^{2}}{2c^{2}}\frac{\partial^{2}}{\partial t^{2}}\left(\bar{\psi}_{2}(t,\vec{y}) \gamma_{\mu} \psi_{2}(t,\vec{y})\right)
\label{14}\end{aligned}$$
and substituting (\[13\],\[14\]) in (\[12\]) we obtain:
$$\begin{aligned}
& & - \frac{e_{1} e_{2}}{c} \int\int d^{4}x d^{4}y \delta\left((x - y)^{2}\right)
\bar{\psi}_{1}(x) \gamma^{\mu} \psi_{1}(x) \bar{\psi}_{2}(y) \gamma_{\mu} \psi_{2}(y) \approx \nonumber\\
& & - e_{1} e_{2} \int dt \int d^{3}\vec{x} \int \frac{d^{3}\vec{y}}{|\vec{x}-\vec{y}|}
\bar{\psi}_{1}(t,\vec{x}) \gamma^{\mu} \psi_{1}(t,\vec{x}) \bar{\psi}_{2}(t,\vec{y}) \gamma_{\mu} \psi_{2}(t,\vec{y})\nonumber\\
& & -\frac{e_{1}e_{2}}{2c^{2}} \int dt \int d^{3}\vec{x} \int d^{3}\vec{y} |\vec{x}-\vec{y}|
\bar{\psi}_{1}(t,\vec{x}) \gamma^{\mu} \psi_{1}(t,\vec{x}) \frac{\partial^{2}}{\partial t^{2}}\left(\bar{\psi}_{2}(t,\vec{y}) \gamma_{\mu} \psi_{2}(t,\vec{y})\right)
\label{15}\end{aligned}$$
Integrating the second term in (\[15\]) by parts and recalling that $(\gamma^{0})^{2} = I, \gamma^{0}\vec{\gamma} = \vec{\alpha}$ we find:
$$\begin{aligned}
& & - \frac{e_{1} e_{2}}{c} \int\int d^{4}x d^{4}y \delta\left((x - y)^{2}\right)
\bar{\psi}_{1}(x) \gamma^{\mu} \psi_{1}(x) \bar{\psi}_{2}(y) \gamma_{\mu} \psi_{2}(y) \approx \nonumber\\
& & - e_{1} e_{2} \int dt \int d^{3}\vec{x} \int \frac{d^{3}\vec{y}}{|\vec{x}-\vec{y}|}
\psi^{\dagger}_{1}(t,\vec{x}) \psi_{1}(t,\vec{x}) \psi^{\dagger}_{2}(t,\vec{y}) \psi_{2}(t,\vec{y})\nonumber\\
& & + e_{1} e_{2} \int dt \int d^{3}\vec{x} \int \frac{d^{3}\vec{y}}{|\vec{x}-\vec{y}|}
\psi^{\dagger}_{1}(t,\vec{x}) \vec{\alpha} \psi_{1}(t,\vec{x}) \psi^{\dagger}_{2}(t,\vec{y}) \vec{\alpha}\psi_{2}(t,\vec{y})\nonumber\\
& & +\frac{e_{1}e_{2}}{2c^{2}} \int dt \int d^{3}\vec{x} \int d^{3}\vec{y} |\vec{x}-\vec{y}|
\frac{\partial}{\partial t}\left(\psi^{\dagger}_{1}(t,\vec{x}) \psi_{1}(t,\vec{x})\right)
\frac{\partial}{\partial t}\left(\psi^{\dagger}_{2}(t,\vec{y}) \psi_{2}(t,\vec{y})\right)\nonumber\\
& & -\frac{e_{1}e_{2}}{2c^{2}} \int dt \int d^{3}\vec{x} \int d^{3}\vec{y} |\vec{x}-\vec{y}|
\frac{\partial}{\partial t}\left(\psi^{\dagger}_{1}(t,\vec{x}) \vec{\alpha} \psi_{1}(t,\vec{x})\right)
\frac{\partial}{\partial t}\left(\psi^{\dagger}_{2}(t,\vec{y}) \vec{\alpha} \psi_{2}(t,\vec{y})\right).
\label{16}\end{aligned}$$
Using the continuity equations (\[7\]), the third term in (\[16\]) can be rewritten as:
$$\begin{aligned}
& & \frac{e_{1}e_{2}}{2c^{2}} \int dt \int d^{3}\vec{x} \int d^{3}\vec{y} |\vec{x}-\vec{y}|
\frac{\partial}{\partial t}\left(\psi^{\dagger}_{1}(t,\vec{x}) \psi_{1}(t,\vec{x})\right)
\frac{\partial}{\partial t}\left(\psi^{\dagger}_{2}(t,\vec{y}) \psi_{2}(t,\vec{y})\right)\nonumber\\
& \approx & \frac{e_{1}e_{2}}{2} \int dt \int d^{3}\vec{x} \int d^{3}\vec{y} |\vec{x}-\vec{y}|
\frac{\partial}{\partial \vec{x}}\left(\psi^{\dagger}_{1}(t,\vec{x}) \vec{\alpha}\psi_{1}(t,\vec{x})\right)
\frac{\partial}{\partial \vec{y}}\left(\psi^{\dagger}_{2}(t,\vec{y}) \vec{\alpha}\psi_{2}(t,\vec{y})\right).
\label{17}\end{aligned}$$
Integrating by parts and taking into account that
$$\frac{\partial}{\partial \vec{y}}\left(|\vec{x} - \vec{y}|\right) = - \frac{\vec{x} - \vec{y}}{|\vec{x} - \vec{y}|}
\label{18}$$
we obtain:
$$\begin{aligned}
& & \frac{e_{1}e_{2}}{2c^{2}} \int dt \int d^{3}\vec{x} \int d^{3}\vec{y} |\vec{x}-\vec{y}|
\frac{\partial}{\partial t}\left(\psi^{\dagger}_{1}(t,\vec{x}) \psi_{1}(t,\vec{x})\right)
\frac{\partial}{\partial t}\left(\psi^{\dagger}_{2}(t,\vec{y}) \psi_{2}(t,\vec{y})\right)\nonumber\\
& \approx & \frac{e_{1}e_{2}}{2} \int dt \int d^{3}\vec{x} \int d^{3}\vec{y}
\frac{\partial}{\partial \vec{x}}\left(\psi^{\dagger}_{1}(t,\vec{x}) \vec{\alpha}\psi_{1}(t,\vec{x})\right)
\psi^{\dagger}_{2}(t,\vec{y}) \frac{\left(\vec{\alpha}(\vec{x} -\vec{y})\right)}{|\vec{x}-\vec{y}|}\psi_{2}(t,\vec{y}).
\label{19}\end{aligned}$$
Performing a second integration by parts and taking into account that:
$$\frac{\partial}{\partial \vec{x}}\left(\frac{1}{|\vec{x} - \vec{y}|}\right) = - \frac{\vec{x} - \vec{y}}{|\vec{x} - \vec{y}|^{3}},
\label{20}$$
we obtain:
$$\begin{aligned}
& & \frac{e_{1}e_{2}}{2c^{2}} \int dt \int d^{3}\vec{x} \int d^{3}\vec{y} |\vec{x}-\vec{y}|
\frac{\partial}{\partial t}\left(\psi^{\dagger}_{1}(t,\vec{x}) \psi_{1}(t,\vec{x})\right)
\frac{\partial}{\partial t}\left(\psi^{\dagger}_{2}(t,\vec{y}) \psi_{2}(t,\vec{y})\right) \approx \nonumber\\
& & - \frac{e_{1}e_{2}}{2} \int dt \int d^{3}\vec{x} \int \frac{d^{3}\vec{y}}{|\vec{x} - \vec{y}|}
\psi^{\dagger}_{1}(t,\vec{x}) \vec{\alpha}\psi_{1}(t,\vec{x})
\psi^{\dagger}_{2}(t,\vec{y}) \vec{\alpha}\psi_{2}(t,\vec{y})\nonumber\\
& & + \frac{e_{1}e_{2}}{2} \int dt \int d^{3}\vec{x} \int \frac{d^{3}\vec{y}}{|\vec{x} - \vec{y}|}
\psi^{\dagger}_{1}(t,\vec{x}) \frac{\left(\vec{\alpha}(\vec{x} -\vec{y})\right)}{|\vec{x}-\vec{y}|}\psi_{1}(t,\vec{x})
\psi^{\dagger}_{2}(t,\vec{y}) \frac{\left(\vec{\alpha}(\vec{x} -\vec{y})\right)}{|\vec{x}-\vec{y}|}\psi_{2}(t,\vec{y}).
\label{21}\end{aligned}$$
Substituting (\[21\]) into (\[16\]) we finally obtain:
$$\begin{aligned}
& & - \frac{e_{1} e_{2}}{c} \int\int d^{4}x d^{4}y \delta\left((x - y)^{2}\right)
\bar{\psi}_{1}(x) \gamma^{\mu} \psi_{1}(x) \bar{\psi}_{2}(y) \gamma_{\mu} \psi_{2}(y) \approx \nonumber\\
& & - e_{1} e_{2} \int dt \int d^{3}\vec{x} \int \frac{d^{3}\vec{y}}{|\vec{x}-\vec{y}|}
\psi^{\dagger}_{1}(t,\vec{x}) \psi_{1}(t,\vec{x}) \psi^{\dagger}_{2}(t,\vec{y}) \psi_{2}(t,\vec{y})\nonumber\\
& & + \frac{e_{1} e_{2}}{2} \int dt \int d^{3}\vec{x} \int \frac{d^{3}\vec{y}}{|\vec{x}-\vec{y}|}
\psi^{\dagger}_{1}(t,\vec{x}) \vec{\alpha} \psi_{1}(t,\vec{x}) \psi^{\dagger}_{2}(t,\vec{y}) \vec{\alpha}\psi_{2}(t,\vec{y})\nonumber\\
& & +\frac{e_{1}e_{2}}{2} \int dt \int d^{3}\vec{x} \int \frac{d^{3}\vec{y}}{|\vec{x} - \vec{y}|}
\psi^{\dagger}_{1}(t,\vec{x}) \frac{\left(\vec{\alpha}(\vec{x} -\vec{y})\right)}{|\vec{x}-\vec{y}|}\psi_{1}(t,\vec{x})
\psi^{\dagger}_{2}(t,\vec{y}) \frac{\left(\vec{\alpha}(\vec{x} -\vec{y})\right)}{|\vec{x}-\vec{y}|}\psi_{2}(t,\vec{y})\nonumber\\
& & -\frac{e_{1}e_{2}}{2c^{2}} \int dt \int d^{3}\vec{x} \int d^{3}\vec{y} |\vec{x}-\vec{y}|
\frac{\partial}{\partial t}\left(\psi^{\dagger}_{1}(t,\vec{x}) \vec{\alpha} \psi_{1}(t,\vec{x})\right)
\frac{\partial}{\partial t}\left(\psi^{\dagger}_{2}(t,\vec{y}) \vec{\alpha} \psi_{2}(t,\vec{y})\right).
\label{22}\end{aligned}$$
If we assume that the last term in (\[22\]) can be neglected, then, substituting (\[22\]) into (\[10\]) we obtain the approximate action functional for two spin-$1/2$ particles in the form:
$$\begin{aligned}
& & S \approx\nonumber\\
& & \int dt\int d^{3}\vec{x}\int d^{3}\vec{y} \psi^{\dagger}_{2}(t, \vec{y})\psi_{2}(t, \vec{y}) \psi^{\dagger}_{1}(t, \vec{x})
\left(i\hbar \frac{\partial}{\partial t} \psi_{1}(t, \vec{x})
+ i\hbar c \vec{\alpha}\frac{\partial}{\partial \vec{x}}\psi_{1}(t, \vec{x})
- m_{1}c^{2} \beta\psi_{1}(t, \vec{x})\right) \nonumber\\
& & + \int dt \int d^{3}\vec{x} \int d^{3}\vec{y}\psi^{\dagger}_{1}(t, \vec{x})\psi_{1}(t, \vec{x})
\psi^{\dagger}_{2}(t, \vec{y})\left(i\hbar \frac{\partial}{\partial t} \psi_{2}(t, \vec{y})
+ i\hbar c \vec{\alpha}\frac{\partial}{\partial \vec{y}}\psi_{2}(t, \vec{y})
- m_{2}c^{2} \beta\psi_{2}(t, \vec{y})\right) \nonumber\\
& & - e_{1} e_{2} \int dt \int d^{3}\vec{x} \int \frac{d^{3}\vec{y}}{|\vec{x}-\vec{y}|}
\psi^{\dagger}_{1}(t,\vec{x}) \psi_{1}(t,\vec{x}) \psi^{\dagger}_{2}(t,\vec{y}) \psi_{2}(t,\vec{y})\nonumber\\
& & + \frac{e_{1} e_{2}}{2} \int dt \int d^{3}\vec{x} \int \frac{d^{3}\vec{y}}{|\vec{x}-\vec{y}|}
\psi^{\dagger}_{1}(t,\vec{x}) \vec{\alpha} \psi_{1}(t,\vec{x}) \psi^{\dagger}_{2}(t,\vec{y}) \vec{\alpha}\psi_{2}(t,\vec{y})\nonumber\\
& & +\frac{e_{1}e_{2}}{2} \int dt \int d^{3}\vec{x} \int \frac{d^{3}\vec{y}}{|\vec{x} - \vec{y}|}
\psi^{\dagger}_{1}(t,\vec{x}) \frac{\left(\vec{\alpha}(\vec{x} -\vec{y})\right)}{|\vec{x}-\vec{y}|}\psi_{1}(t,\vec{x})
\psi^{\dagger}_{2}(t,\vec{y}) \frac{\left(\vec{\alpha}(\vec{x} -\vec{y})\right)}{|\vec{x}-\vec{y}|}\psi_{2}(t,\vec{y}).
\label{23}\end{aligned}$$
Following [@barut1; @barut2], we have multiplied the free terms in the action functional by the normalization conditions (\[8\]). This can of course be done without altering these terms (it is equivalent to multiplying each term by one).
In terms of the two-body bilocal spinor (\[9\]) the action (\[23\]) can simply be written as:
$$\begin{aligned}
S & \approx &
\int dt\int d^{3}\vec{x}\int d^{3}\vec{y} \Psi^{\dagger}(t, \vec{x}, \vec{y})
\left(i\hbar \frac{\partial}{\partial t} + i\hbar c \vec{\alpha}_{1}\frac{\partial}{\partial \vec{x}}
+ i\hbar c \vec{\alpha}_{2}\frac{\partial}{\partial \vec{y}} - m_{1}c^{2}\beta_{1} - m_{2}c^{2}\beta_{2}\right.\nonumber\\
& & \left. - \frac{e_{1} e_{2}}{|\vec{x} -\vec{y}|} \left(1 - \frac{1}{2}\left((\vec{\alpha}_{1}\vec{\alpha}_{2})
+ \frac{(\vec{\alpha}_{1}(\vec{x} - \vec{y}))}{|\vec{x} -\vec{y}|}\frac{(\vec{\alpha}_{2}(\vec{x} - \vec{y}))}{|\vec{x} -\vec{y}|}\right)\right)\right)
\Psi(t, \vec{x}, \vec{y})
\label{24}\end{aligned}$$
From this action functional, one can immediately obtain Breit’s equation for electrodynamics in the form:
$$\begin{aligned}
& & \left(i\hbar \frac{\partial}{\partial t} + i\hbar c \vec{\alpha}_{1}\frac{\partial}{\partial \vec{x}} +
i\hbar c \vec{\alpha}_{2}\frac{\partial}{\partial \vec{y}} - m_{1}c^{2}\beta_{1} - m_{2}c^{2}\beta_{2}\right.\nonumber\\
& & \left. - \frac{e_{1} e_{2}}{|\vec{x} -\vec{y}|} \left(1 - \frac{1}{2}\left((\vec{\alpha}_{1}\vec{\alpha}_{2})
+ \frac{(\vec{\alpha}_{1}(\vec{x} - \vec{y}))}{|\vec{x} -\vec{y}|}\frac{(\vec{\alpha}_{2}(\vec{x} - \vec{y}))}{|\vec{x} -\vec{y}|}\right)\right)\right)
\Psi(t, \vec{x}, \vec{y}) = 0.
\label{25}\end{aligned}$$
which coincides with (\[P2\]).
We have presented a very simple derivation of the Breit equation from Barut’s formulation of electrodynamics in terms of direct interactions between fermions. In this formulation the spinors describing the spin-$1/2$ particles are not operators. This is reflected at every step of our derivation, which proceeds completely within the realm of the Lagrangian formalism of relativistic wave equations.
Since it is known that the Breit equation gives results, up to terms of order $\alpha^{4}$, for the energy spectrum of relativistic bound systems (such as positronium) in agreement with experiments[@kroli; @fushchich], our proof that the Breit equation follows from Barut’s formulation of electrodynamics (in terms of direct interactions) lends further support to Barut’s approach.
[99]{}
C.G. Darwin, Phil. Mag. [**39**]{}, 537 (1920). L. D. Landau and E. M. Lifshitz, [*The Classical Theory of Fields*]{} (4th revised English edition), Butterworth- Heinenann, Oxford (1996). G. Breit, Phys. Rev. [**34**]{}, 553 (1929). P.A.M. Dirac, R. Peierls and M.H.L. Pryce, Proc. Cambridge Phil. Soc. [**35**]{}, 186 (1939). H.A. Bethe and E.E. Salpeter, [*Quantum Mechanics of One and Two Electron Atoms*]{} (Springer, Berlin, 1957). W. Krolikowski, Acta Phys. Pol. [**B12**]{}, 891 (1981). V.I. Fushchich and A.G. Nikitin, [*Symmetries of Equations of Quantum Mechanics*]{} (Allerton Press, 1994). A. Duviryak and J.W. Darewych, CEJP, 3(4), 467 (2005). A.O. Barut, Lecture Notes in Physics, Vol. [**180**]{}, p. 332 (Springer-Verlag, Berlin, 1983). A.O. Barut, J. Math. Phys. [**32**]{}, 1091 (1991). J.L. Anderson, [*Principles of Relativity Physics*]{} (Academic Press, New York, 1967). D.J. Louis-Martinez, Phys. Lett. A[**364**]{}, 93 (2007). V.N. Golubenkov and Ya. A. Smorodinskii, Zh. Eksp. Teor. Fiz. [**31**]{}, 330 (1956) \[Sov. Phys. JETP [**4**]{}, 442 (1957)\]. T. Ohta and T. Kimura, J. Math. Phys. [**34**]{}, 4655 (1993). J.D. Bjorken and S.D. Drell, [*Relativistic Quantum Mechanics*]{} (McGraw-Hill, New York, 1964). P. Van Alstine and H. Crater, Found. Phys. [**27**]{}, 67 (1997). J. A. Wheeler and R. P. Feynman, Rev. Mod. Phys. [**17**]{}, 157 (1945); [**21**]{}, 425 (1949). F. Hoyle and J.V. Narlikar, [*Lectures on Cosmology and Action at a Distance Electrodynamics*]{}, (World Scientific, 1996). A. O. Barut, [*Electrodynamics and Classical Theory of Fields and Particles*]{}, (Dover Publications Inc, New York, 1980). A. Schild, Phys. Rev. [**131**]{}, 2762 (1963). D.J. Louis-Martinez, Phys. Lett. A[**320**]{}, 103 (2003).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'When the planning horizon is long, and the safe asset grows indefinitely, isoelastic portfolios are nearly optimal for investors who are close to isoelastic for high wealth, and not too risk averse for low wealth. We prove this result in a general arbitrage-free, frictionless, semimartingale model. As a consequence, optimal portfolios are robust to the perturbations in preferences induced by common option compensation schemes, and such incentives are weaker when their horizon is longer. Robust option incentives are possible, but require several, arbitrarily large exercise prices, and are not always convex.'
address:
- 'Boston University, Department of Mathematics and Statistics, 111 Cummington Street Boston, MA 02215, USA, and Dublin City University, School of Mathematical Sciences, Glasnevin, Dublin 9, Ireland.'
- 'ETH Zürich, Departement für Mathematik, Rämistrasse 101, CH-8092, Zürich, Switzerland, and Swiss Finance Institute'
- 'London School of Economics and Political Science, Department of Statistics, 10 Houghton St, London WC2A 2AE, UK'
author:
- Paolo Guasoni
- 'Johannes Muhle-Karbe'
- Hao Xing
bibliography:
- 'biblio.bib'
title: 'Robust Portfolios and Weak Incentives in Long-Run Investments'
---
[^1]
Introduction
============
Investors pursue long-term goals both by managing their portfolios and by designing incentives, such as stock and option grants, to align managers’ actions with their interests. This paper explores the implications of long-term investment for both portfolio choice and incentive contracts, overcoming their traditional separation in the literature, and some puzzling results that this separation has generated.
The main message in the literature for long-term portfolio choice comes from turnpike theorems:[^2] when the horizon is distant, optimal portfolios depend only on preferences at high levels of wealth, hence a generic investor should invest like an isoelastic investor with the same asymptotic risk aversion.[^3] These theorems are welcome news for long-term investors who seek simple portfolio allocation strategies, as they imply that local differences in preferences are irrelevant at long horizons, and allow to focus on the one-parameter family of isoelastic strategies.
However, turnpike theorems also have unsettling implications for managers’ compensation. When investors offer incentive stock options to a manager, the incentive is a perturbation of the manager’s preferences. If turnpike theorems hold, they imply that incentive schemes lose their strength as their horizon (i.e., vesting period) increases. The central goal of this paper is to understand the extent and the limits of this effect – and its potential remedies.
Our main theorem provides conditions on preferences under which, for a long horizon, an isoelastic portfolio is approximately optimal for a generic investor. This is not another turnpike theorem: while turnpikes show that generic portfolios are close to isoelastic portfolios *when the horizon is distant*, we show that the welfare loss to an investor who uses an isoelastic portfolio *throughout the entire period* is negligible – in *relative monetary terms*. That is, the fraction of wealth lost by using a simple isoelastic portfolio rather than the exact, but more complex, optimal portfolio, declines to zero as the planning horizon increases. The theorem holds in a general semimartingale market without arbitrage opportunities, in which the safe asset grows indefinitely.
To appreciate the strength of this statement, compare this welfare effect with that of a small reduction in the interest rate. Compounded over a long horizon, a lower interest rate translates into a large relative decline in wealth, as the ratio between investments growing at different rates diverges. From this viewpoint, isoelastic portfolios are surprisingly robust – and incentives remarkably weak.
The message of this result is as positive for portfolio choice as it is negative for typical incentive schemes (those based on options with strike price near the money): if, at long horizons, the optimal strategy for an investor is robust to local perturbations in preferences, then, by the same token, a manager’s policy is insensitive to common stock and option grants, which modify preferences only locally. Thus, to be effective at long horizons, an incentive contract must modify preferences at levels much higher than current wealth.
We argue that incentive contracts based on options with several, arbitrarily high strike prices are robust to the horizon. A simple example is a (multiple of a) contract that pays one option for each strike price in a regularly spaced grid: its payoff is approximately the underlying assets’s squared price, and the value of the contract is finite in most common models.[^4]
Our main result is a natural complement to the extant turnpike literature. The latter compares, as the horizon increases, optimal portfolios for generic investors to optimal portfolios of isoelastic investors with the same horizon, and establishes joint restrictions on markets and preferences, under which these portfolios are close. Put differently, the question is whether the optimizers of the two maximization problems are similar at the early stages of the investment process. A natural practical question (but, somewhat unnaturally, not addressed in the literature) is whether these maximizers are good substitutes: if a generic investor chooses to use an isoelastic portfolio over the entire investment horizon, then under which conditions will utility be close to optimal, and in which sense? Although such a result is plausible, there are two potential pitfalls.
First, turnpike theorems only assert that optimizers are close at the beginning, not necessarily at the end, of the period. Thus, switching a portfolio with another may have dreadful results if they grow apart as time passes – a concrete possibility, as even optimal isoelastic portfolios shift between the beginning and the end of the period.[^5] Second, the generic investor may have nearly isoelastic preferences at high wealth levels, but may be far more risk averse for very low wealth, and it is unclear whether such bad states have a negligible effect, even in the long run. This concern is especially relevant when the isoelastic portfolio is very risky.
The main result of the present study clarifies these issues in a general setting. It also helps to reconcile some of the puzzling, or at least counterintuitive results in the literature on executive compensation. @jensen1976theory already recognized that *managers of large publicly held corporations seem to behave in a risk averse way to the detriment of the equity holders*, as projects with positive net present value may be foregone to avoid additional risk. Stock-based compensation attempts to align managers’ interests with those of shareholders, but may also backfire, leading managers to engage in hedging activities, as predicted by @amihud1981risk as well as @smith1985determinants, and confirmed empirically by @may1995managerial. As a remedy, option-based incentives aim at rewarding managerial risk-taking by introducing convexity in their payoffs. The intuition, which finds its roots in the familiar notion that prices of options with convex payoffs increase with volatility, is that *the asymmetric payoffs of call options make it more attractive for managers to undertake risky projects.* [@defusco1990effect]. However, [@Carpenter] and [@ross2004compensation] show that this intuition is misleading: with a nontradeable option, the manager will focus on its private reservation value (that is, the certainty equivalent) rather than on its risk-neutral value, and the effect of convex incentives is ambiguous in general. In a full-information model with risk-sharing, @cadenillas2007optimal characterize optimal contracts, and also find that they may be either convex or concave. The numerical results of @larsen2005optimal also lead to a similar conclusion.
Adding to the debate, [@hall2000optimal] argue that the standard practice of awarding stock options with exercise price equal to the stock price at grant date, potentially motivated by a favorable accounting treatment, is suboptimal, and find that optimal exercise prices are generally higher. @cadenillas2004leverage also find that optimal exercise prices increase with manager performance and firm size. @edmans2012dynamic note the dynamic problems with option compensation: *securities given to incentivize the CEO may lose their power over time: if firm value declines, options may fall out of the money and bear little sensitivity to the stock price*. In fact, this issue motivates the common industry practice of “repricing”, that is resetting the exercise price after a sharp drop in the stock price [@acharya2000optimality; @chen2004executive], a custom that has attracted criticism for rewarding poor performance and weakening original incentives.
Our results clarify the interplay between the exercise price and the horizon in option compensation. If the exercise price is held constant, the option loses its effect as the horizon increases, in that its certainty equivalent becomes arbitrarily small compared to the manager’s wealth. By contrast, an incentive contract with multiple exercise prices remains effective even after large changes in the stock price, and for long horizons.
Multiple exercise prices also shed light on the ambiguous effects of option grants, and convex incentives in general, on managerial risk-taking. The intuition is that [@carr2001optimal] a portfolio of call and put options can recreate any regular function of the underlying, including incentive contracts of power type $x^\alpha$, $\alpha>0$. If the manager’s risk aversion is high, the incentives contracts that reduce it correspond to $0<\alpha<1$, hence are *concave*, not convex. By contrast, a manager with low risk aversion is motivated to take risks by a package of call options with all strikes, and this convex incentive is robust to changes in stock prices and to long horizons.
Finally, note that this paper focuses on the effects of compensation contracts based on combinations of cash, stock, and options, which are prevalent in practice, but it does not investigate their optimality for a possible principal. (See @bolton2005contract and @Cvitanic-Zhang for recent surveys in contract theory.) On the other hand, our results allow considerable flexibility in both investment opportunities and preferences, shedding new light on the effect of the horizon in typical compensation contracts.
The rest of the paper is organized as follows: Section 2 introduces the model, and presents the main result and its implications. Section 3 lays the groundwork for the proof of the main result, recalling the general duality results of [@Bouchard-Touzi-Zeghal] and some auxiliary results that will be used repeatedly in the sequel. The main result is proved for power utilities ($p \neq 0$) in Section 4, and for logarithmic utility ($p=0$) in Section 5. Finally, Section 6 contains a counterexample illustrating the necessity of our assumptions.
Main Result
===========
Model and Main Result
---------------------
We focus on an agent who invests in assets $S$, thereby affecting total wealth $X$, so as to maximize expected utility from terminal wealth $X_T$ at time $T$: $$\label{eq: exp U}
\max_{X \text{ admissible}} {\mathbb{E}}[U(X_T)] \ .$$ To encompass the applications below, the utility function $U$ is strictly increasing and concave, but not necessarily continuously differentiable or strictly concave. The model has the usual portfolio choice interpretation, in which the agent is an investor, $S$ represents the financial asset(s), $X$ the portfolio value, and $U$ the investor’s utility function. A second interpretation is that of the agent as a corporate manager, $S$ a real investment opportunity, and $X$ the firm’s value. In this case, the function $U$ combines the manager’s preferences and incentives: for example, if the manager receives a compensation equal to $F(X_T)$ and has utility function $u$, then $U(x) = u (F(x))$. A hybrid interpretation (cf. @Carpenter) is that of a fund manager, who invests in financial assets $S$, so as to maximize the expected payoff of some function $U$ of the terminal fund value.
Formally, there are $d+1$ assets available. A safe asset, with price denoted by $S^0$, and $d$ risky assets, with prices $S=(S^1,\ldots,S^d)$. These assets are traded continuously, without frictions, and no arbitrage opportunities are available. Let $(\Omega,\mathcal F,(\mathcal F_t)_{t\ge 0},P)$ be a filtered probability space satisfying the usual conditions of right-continuity and saturatedness.
\[ass: growth\] The safe asset $S^0: [0,\infty) \mapsto {\mathbb R}$ is a deterministic, strictly positive function satisfying $S^0_t \uparrow \infty$ as $t\uparrow\infty$.[^6]
Moreover, the discounted prices $S/S^0$ of the risky assets are local martingales under some probability $Q$ equivalent to $P$.
The above assumption $\lim_{t\rightarrow \infty} S^0_t = \infty$ is satisfied, for example, in models with positive interest rates bounded away from zero. The existence of a martingale measure ensures that the market is free of arbitrage opportunities (cf. @DS98).
The agent’s objective is described by a utility function $U$, which incorporates the combined effect of preferences and incentives. Henceforth, fix $p<1$, and denote by $\tilde U(x)$ the isoelastic utility function, defined by $\tilde{U}(x)=x^p/p$, $0 \neq p<1$, resp. $\tilde{U}(x) = \log(x)$ for $p=0$. Compared to the benchmark $\tilde{U}$, the generic utility function $U$ satisfies the following restrictions, which yield the main result.
\[ass: utility\]
i) $U(x): (0,\infty)\rightarrow {\mathbb R}$ is strictly increasing, concave, not necessarily differentiable or strictly concave for low wealth levels, but differentiable and strictly concave for large enough wealth.
ii) As wealth increases ($x\uparrow\infty$), the utility $U$ becomes similar to the isoelastic utility $\tilde{U}$, in that their marginal utilities are asymptotically equivalent: $$\label{ass: conv}
\lim_{x\uparrow \infty} \frac{U'(x)}{\tilde{U}'(x)} =1.$$
iii) The utility $U$ satisfies additional conditions at low wealth levels, depending on the sign of $p$ in $\tilde{U}(x) = x^p/p$:
a) For $0<p<1$, $U$ is bounded from below;
b) For $p=0$, i.e., $\tilde{U}(x)=\log(x)$, $$\label{ass: U/tU'}
\liminf_{x\downarrow 0} \frac{U(x)}{\tilde{U}'(x)} >-\infty.$$
c) For $p<0$, $\lim_{x\uparrow \infty} U(x) =0$ and is satisfied.
Condition $i)$ implies that the agent is risk averse when wealth is high. Condition $ii)$ requires that, when the agent is rich, the utility (either by preferences or by incentives) is close to isoelastic, which is the central assumption in turnpike theorems. In particular, implies that $U$ satisfies the Inada condition at infinity, i.e., $\lim_{x\uparrow \infty} U'(x) =0$. However, the Inada condition may not be satisfied at zero.
Condition $iii)$ is new, and requires that $U$ is not too risk averse compared to $\tilde U$ when wealth is low. For example, if $U(x) = x^{p^*}/p^*$ for $x$ small, where $p^*<1$, the condition in boils down to $p^*\geq p-1$, that is, the risk aversion $1-p^*$ of $U$ should not be greater than one plus the risk aversion $1-p$ of $\tilde U$ at low wealth. In general, means that the ratio of utilities $U(x)/\tilde{U}(x)$ does not diverge faster than $x^{-1}$ for $p\neq 0$ or $(x\log x)^{-1}$ for $p=0$, as the wealth $x$ tends to zero. These conditions are satisfied, in particular, if the ratio between $U$ and $\tilde U$ remains bounded near zero. The example outlined in Section \[sec:example\] and analyzed in Section \[sec:analysis\] below shows that if is dropped, the main result can fail even in the Black-Scholes model.
The agent invests in the assets subject to the usual budget constraint: if $x$ denotes the initial capital, and $(\varphi^i_t)^{1\le i\le d}_{0\le t\le T}$ the number of shares of the $i$-th asset at time $t$, the corresponding wealth $X^\varphi_t$ equals $$X^\varphi_t = S^0_t \left( x + \int_0^t \varphi_s d(S_s/S^0_s) \right).$$ To simplify notation, without loss of generality we set $x=1$, which amounts to scaling the numeraire by a factor of $x$. An ${\mathbb R}^d$-valued process $\varphi$ is an admissible strategy if it is predictable, $S$-integrable, and the corresponding wealth process satisfies $X^\varphi_t \ge 0$ a.s. for all $t\ge 0$. The class of admissible wealth processes is denoted by $$\mathcal{X}:= \{X: X_t\geq 0, {\mathbb{P}}-a.s.\ \text{ for all } t\geq 0\}.$$ Thus, the value functions for the utility maximization problems for the generic utility $U$ and its isoelastic counterpart $\tilde U$ are: $$\label{eq: primary}
u^T(x) = \sup_{X\in \mathcal X}{\mathbb{E}}[U(X)]
\qquad\text{and}\qquad
{\tilde u}^T(x) = \sup_{X\in \mathcal X}{\mathbb{E}}[\tilde U(X)].$$
The final assumption is that the isoelastic utility maximization problem is well posed. This assumption is necessary only for $p\ge 0$ because it is always satisfied if the utility function is bounded from above for $p<0$:
\[ass: wellposedness\] If $0\leq p<1$, let $$\tilde{V}(y) = \begin{cases} -y^q/q &\text{for $p \in (0,1)$ and $q=p/(p-1)$},\\ -\log(y)-1 &\text{for $p=0$},\end{cases}$$ be the convex dual of the isoelastic utility $\tilde{U}$, and assume that $$\label{eq:dualfinite}
\inf_{Y\in \mathcal{Y}} {\mathbb{E}}[\tilde{V}(Y_T)] <\infty,$$ where $\mathcal{Y}$ is the set of *stochastic discount factors*: $$\begin{aligned}
\mathcal{Y} := \{Y = \overline{Y}/S^0: \overline{Y}>0 \text{ with } \overline{Y}_0=1 \text{ such that } X\overline{Y} &\text{ is a supermartingale}\\
&\qquad \quad \text{for all $X \in \mathcal{X}$}\}.\end{aligned}$$
Condition ensures that the isoelastic dual (and in turn primal) problem is well posed. For $p \neq 0$, this requires the existence of the $q$-th moment for *some* stochastic discount factor, which is satisfied, for example, if the asset price follows an Itô process, and the market price of risk is bounded. Indeed, in this case one can choose the density process of the minimal martingale measure, and follows from Novikov’s condition. The argument for $p=0$ is similar.
With the above notation and assumptions, the main result reads as follows.
\[thm:mt\] Let Assumptions \[ass: growth\] - \[ass: wellposedness\] hold. Then for any horizon $T>0$ there exist an optimal payoff $X^T_T$ for the generic utility $U$ and $\tilde{X}_T^T$ for the isoelastic utility $\tilde{U}$.[^7] They satisfy $$\label{eq: limit ce}
\lim_{T\rightarrow \infty} \frac{U^{-1}({\mathbb{E}}[U(\tilde{X}^T_T)])}{U^{-1}({\mathbb{E}}[U(X^T_T)])}=1.$$ That is, in the long run, the certainty equivalent of the isoelastic portfolio is arbitrarily close in relative terms to that of the optimizer.
The strength of this result is that the certainty equivalents converge – not just their growth rates. In contrast, the long-run portfolio choice literature based on risk-sensitive control[^8] and large deviations[^9] focuses on the maximization of the equivalent safe rate, defined as: $$\liminf_{T\rightarrow\infty} \frac1T \log U^{-1}({\mathbb{E}}[U(X_T)]).$$ It is clear that if two processes $(X_t)_{t\ge 0}$ and $(\tilde X_t)_{t\ge 0}$ satisfy , then they share the same equivalent safe rate (if it exists). It is also clear, however, that is a much stronger property. As a trivial example, even with a risk-neutral utility ($U(x) = x$), consider the two processes $\tilde X_t = e^{(\mu-\sigma^2/2)t + \sigma W_t}$, $X_t = \frac12 e^{r t} + \frac12 e^{(\mu-\sigma^2/2)t + \sigma W_t}$, which correspond to a full stock investment, and to a half-stock, half-bond investment (without rebalancing) in a safe asset earning a constant interest rate $r$ and a stock following geometric Brownian motion with expected return $\mu>r$ and volatility $\sigma$. Then, both wealth processes have the same growth rate $\mu$, but the ratio of the corresponding certainty equivalents converges to 2. In fact, examples are also available, in which such a ratio diverges while the equivalent safe rate remains equal. Thus, even if two investment policies share the same equivalent safe rate, they may have very different certainty equivalents, which means that the agent may value one policy much higher than the other.
When holds, however, the optimal policy can only be marginally better than its isoelastic counterpart for long horizons, in that the gain from choosing the superior policy is smaller than any fraction of the value of the inferior one. It is precisely this property that makes the theorem relevant for incentive schemes, and the next section explores in detail the theorem’s implications in this area.
Incentives
----------
### Weakness of incentives with one exercise price {#weakness-of-incentives-with-one-exercise-price .unnumbered}
If a manager has isoelastic preferences ($\tilde U(x)=x^p/p$ with $0\neq p<1$), and compensation that includes a cash component $c_1\geq 0$ and a fraction $c_2>0$ of the equity $X_t$, the objective function is $$\label{eq:private}
{\mathbb{E}}[\tilde{U}(c_1 + c_2 X_T)],$$ where $X$ runs through the class of admissible wealth processes. Suppose now that shareholders are concerned that the manager’s high equity exposure is likely to discourage investment in projects with positive expected value, and are contemplating to grant $c_3>0$ call options with exercise price $K$, as an incentive to take risks. Such executive stock options typically have a vesting period of ten years, so that our focus on long horizons is relevant. Including the option grant, the manager maximizes the objective $$\label{eq:manager}
{\mathbb{E}}[\tilde{U}(c_1+c_2 X_T +c_3 (X_T-K)^+)].$$ Both the optimizer and the certainty equivalent of this problem are the same as for $$\label{eq:manager2}
{\mathbb{E}}[{\bar U}(X_T)] \ ,$$ where $\bar U(x) = \tilde{U}(c_1+c_2 x + c_3 (x-K)^+)/(c_2+c_3)^p$. In other words, awarding the option grant is equivalent to replacing the individual utility $\tilde U$ with the *effective* utility $\bar U$. This utility $\bar U$ is strictly increasing, differentiable on $(K,\infty)$, and satisfies $\bar U(\infty)=0$ if $p<0$ as well as $$\label{eq:lim1}
\lim_{x \uparrow \infty} \frac{\bar U'(x)}{\tilde{U}'(x)}=1,
\qquad
\lim_{x \downarrow 0} \frac{\bar U(x)}{\tilde{U}'(x)}=0.$$ Thus, the effective utility $\bar U$ satisfies Assumption 2 for any fixed compensation $c_1\ge 0$, with the exception that $\bar U$ is no longer concave in a neighborhood of the exercise price $K$, which is illustrated in Figure \[fig:convex\]. Indeed, creating such a region is the main purpose of option incentives.
![Effective utility function $\bar{U}$ (solid) and corresponding concave envelope $U$ (dashed), for sample parameter values ($\gamma=1/2$, $c_1=1$, $c_2=2$, $c_3=3$, $K=4$). []{data-label="fig:convex"}](envelope.pdf){width="80.00000%"}
Under different sets of assumptions[^10] @Carpenter, @cuoco2011equilibrium, and @Bichuch-Sturm show that maximizing the expected utility $\bar U$ is actually equivalent to maximizing its concave envelope $U$, that is, the minimal concave function that dominates $\bar U$: $$\label{eq:manager3}
{\mathbb{E}}[{U}(X_T)],
\qquad
\text{where } U (x) = \inf \{ g(x) : g \text{ concave}, g\ge \bar U\} .$$ In the present setting, the concave envelope coincides with $\bar U$ for sufficiently large or small wealth levels, therefore it preserves properties of $\bar U$. As it is also concave, Theorem \[thm:mt\] applies to $U$, and yields the following result:
\[thm:incentive\] If the manager’s problem and the concavified problem are well posed and have the same solution, and the safe asset grows indefinitely, then, as the horizon increases, the certainty equivalents of the optimal strategies with incentives and without incentives are asymptotically equivalent. In particular, the manager’s private value of an option grant with fixed exercise price becomes relatively negligible as the horizon increases.
With hindsight, this result looks natural, and indeed follows from the the local effect on preferences of a single option. Yet, the option pricing intuition, which shapes much of the literature on incentives, points to the opposite conclusion. As in the Black-Scholes model (and also in more sophisticated extensions) the arbitrage-free price of a call option increases with volatility, it is plausible to conclude that a manager is encouraged to increase its value by taking more risk. A longer horizon also increases the option’s value, suggesting a stronger, not weaker incentive effect.
However, option pricing heuristics are misleading in this context because the manager, who can neither sell nor hedge the option grant,[^11] does not focus on the hypothetical risk-neutral value of the option, but rather on its private value, for which risk-aversion is central. As a single option affects risk-aversion only locally, long horizons combined with wealth growth make their impact vanish.
In summary, Theorem \[thm:incentive\] supports the broad observation that the size and the exercise price of an option grant need to be chosen carefully, depending on the horizon. A small quantity of options with a low strike may result, similar to a large stock position, in discouraging risk-taking. A large number of options with a high strike may also be ineffective, as the risk necessary to make the options profitable may be too much for the manager to bear. And even an exercise price chosen optimally at the time of award may soon become inadequate after large changes in the asset price.
### Robustness of several strike prices, and power(ful) incentives {#robustness-of-several-strike-prices-and-powerful-incentives .unnumbered}
We argue that option grants that include several (in theory, infinitely many) exercise prices retain their incentive effects after large price changes, and are robust to long horizons. Once again, the intuition comes from option theory: @carr2001optimal show that a European option with (smooth) payoff $f(S_T)$ admits the following representation as a portfolio of call and put options of all strikes:[^12] $$\begin{split}
f(S_T) =\ &f(\bar K) + f'(\bar K)(S_T-\bar K)\\
&+\int_0^{\bar K} f''(k)(k-S_T)^+dk + \int_{\bar K}^\infty f''(k)(S_T-k)^+dk.
\end{split}$$ In this representation, the term $f(\bar K)$ represents a cash amount, $f'(\bar K)(S_T-\bar K)$ a position in a forward contract, and the two integrals correspond to static portfolios in puts and calls, respectively. The threshold $\bar K$ is arbitrary, and determines the strike above which calls rather than puts are used.
For example, setting $\bar K=0$, this decomposition shows that a portfolio consisting of an equal number of call options at all strikes leads to a payoff of the form $f(x) = c x^2$ for some $c>0$. In general, a payoff of power type $f(x) = x^\alpha$ is replicated by a portfolio of options with weights $f''(x) = \alpha (\alpha -1) x^{\alpha - 2}$. We restrict to $\alpha>0$, as otherwise the incentive does not reward higher asset values.
Consider now the effect of such a power incentive on a manager with isoelastic utility $\tilde U(x) = x^p/p$. With fixed compensation set to zero for simplicity, the utility function including incentives becomes $U(x) = x^{\alpha p}/p$, which means that the incentive replaces the manager’s risk aversion $\gamma = 1-p$ with the *effective* risk aversion:[^13] $$\label{eq:gammastar}
\gamma^* = \alpha \gamma + (1-\alpha) .$$ This formula has several implications. First, the incentive scheme reduces risk aversion ($\gamma^*<\gamma$) if and only if $(1-\alpha)(1-\gamma)<0$. In particular, a convex payoff ($\alpha>1$) does not reduce risk aversion if the latter is originally higher than logarithmic ($\gamma>1$), as it typically is. In general, a concave payoff ($\alpha<1$) makes a manager closer to logarithmic $\gamma=1$, and a convex payoff does the opposite.
Thus, for a manager who has an already low risk aversion ($\gamma<1$), an options grant that includes calls with several exercise prices is an incentive scheme that remains robust to the horizon, and to large movements in the stock price, preventing the need for future repricing.
The payoff of a concave incentive is qualitatively similar to that of a combination of covered-call positions, but devising concave incentives may not be practical, because they would imply large stock holdings, combined with short positions in options of all maturities. The unusual nature of such arrangements, combined with the resulting difficulties for tax and accounting purposes, may make such schemes hard to implement.
In spite of these difficulties, note that the above formula shows that such concave power incentive are implicit in the high-water mark provisions in the compensation of hedge fund managers. Indeed, @guasoni2011incentives find that a hedge fund manager with risk aversion $\gamma$, who receives as performance fee a fraction $1-\alpha$ of profits, invests the fund’s assets like an owner-investor with the same effective risk aversion as in . This observation shows that concave incentives, although virtually absent in corporate compensation, are in fact implicitly present in the hedge-fund industry, and that the typical performance fees of $20\%$ correspond to a power incentive with $\alpha=0.8$.
Counterexample {#sec:example}
--------------
This section outlines a counterexample, which shows that Theorem \[thm:mt\] can fail even in the usual Black-Scholes model if condition is not satisfied, i.e., if the generic utility $U$ is too risk averse compared to the reference isoelastic utility $\tilde U$ at low wealth levels. The detailed calculations are presented in Section \[sec:analysis\] below.
Suppose the safe asset earns a constant interest rate $r>0$, and there is a single risky asset following geometric Brownian motion: $$dS_t/S_t=(\mu+r)dt+\sigma dW_t,$$ for a standard Brownian motion $W_t$ and constants $\mu, \sigma>0$. Let $\tilde{U}(x)=x^p/p$, $p<0$ be a reference isoelastic utility with risk aversion $1-p>1$ and consider the generic utility function given by $U(x)=x^p/p$ for sufficiently large $x$ and by $U(x)=x^{p^*}/p^*$, $p^* < p-1$, for $x \leq 1$, with smooth interpolation in between. Then, at high wealth levels, the generic utility $U$ has the same risk aversion $1-p$ as its isoelastic counterpart $\tilde{U}$, but the corresponding risk aversion is bigger (by more than 1) at low wealth levels. In particular, a simple calculation shows that the generic utility $U$ does not satisfy .
In this setting, the optimal isoelastic portfolio $\widetilde{X}^T_T$ can be too risky for the generic utility at low wealth levels, leading to a diminishing ratio of certainty equivalents compared to the optimizer $X^T_T$ for $U$ in the long run: $$\label{eq: ce conv 0}
\lim_{T\rightarrow \infty} \frac{U^{-1}({\mathbb{E}}[U(\tilde{X}^T_T)])}{U^{-1}({\mathbb{E}}[U(X^T_T)])}=0.$$ Indeed, we show in Section \[sec:analysis\] that this result holds when the risky asset is attractive enough, relative to the safe asset: $$\label{eq: para rest}
\left(\frac{\mu}{\sigma^2 (1-p)}\right)^2 > 2\max \left\{1,\frac{1}{p-p^*-1}\right\} \frac r{\sigma^2}.$$ To understand this parameter restriction, recall that in this model $\mu/(\sigma^2 (1-p))$ is the optimal portfolio weight in the risky asset for the isoelastic utility. Thus, the lower bound ensures that the risky asset is sufficiently attractive, compared to the safe asset, to lead to a sufficiently large risky investment. In particular, as the difference $p-p^*$ between risk aversions at low wealth levels declines to one, the isoelastic risky weight rises, leading to a sufficiently large probability of reaching low wealth levels. Indeed, the lower bound in tends to infinity, as the difference in risk aversions approaches one ($p^* \uparrow p-1$), in line with our main result, which holds if this difference is less than or equal to one.
The result in may seem puzzling, because it implies that the properties of the generic utility function at low wealths are important in the long run, even though the safe asset has a positive growth rate, so that any initial safe investment grows arbitrarily, thereby avoiding low wealth with certainty. However, the optimal isoelastic portfolio keeps a constant *fraction* of wealth in the risky asset. Hence, if the stock price drops and the risky position declines, the safe *position* is reduced as well, so that wealth can decrease even further.[^14] Such a portfolio may be unacceptable for another investor, who is substantially more risk averse at low wealth levels.
Notation and Preliminaries {#sec:dual}
==========================
We begin by recalling the basic notions and duality results for the present non-smooth setting, which have been established by [@Bouchard-Touzi-Zeghal].
Recall the value functions $u^T$ and $\tilde{u}^T$ for the generic utility $U$ and its isoelastic counterpart $\tilde{U}$ from . Let $$V(y) := \sup_{x>0} (U(x) - xy) \quad \text{and} \quad\tilde{V}(y) := \sup_{x>0} (\tilde{U}(x) - xy)$$ be the dual functions of the generic utility $U$ and the isoelastic utility $\tilde{U}$, respectively. Define the domain of $V$ as $\text{dom}(V) := \{y>0: |V(y)|<\infty\}$ and let $\overline{\text{dom}(V)}$ be its closure. Consider the dual problems associated to : $$\label{eq: dual}
v^T(y) = \inf_{Y\in \mathcal{Y}} {\mathbb{E}}[V(y Y_T)] \quad \text{ and } \quad \tilde{v}^T(y) = \inf_{Y\in \mathcal{Y}} {\mathbb{E}}[\tilde{V}(yY_T)],$$ where $\mathcal{Y}$ denotes the set of stochastic discount factors: $$\begin{aligned}
\mathcal{Y} := \{Y = \overline{Y}/S^0: \overline{Y}>0 \text{ with } \overline{Y}_0=1 \text{ such that } X\overline{Y} &\text{ is a supermartingale}\\
&\qquad \quad\text{for all $X \in \mathcal{X}$}\}.\end{aligned}$$
\[thm:duality BTZ\] Suppose the following holds:
a) $U:(0,\infty)\rightarrow {\mathbb R}$ is nonconstant, nondecreasing and concave;
b) $\overline{\text{dom}(V)} = [0,\infty)$;
c) $V$ satisfies the dual asymptotic elasticity condition $$AE_0(V) := \limsup_{y\downarrow 0} \sup_{x\in -\partial V(y)} \frac{yx}{V(y)}<\infty;$$
d) There is $y>0$ such that $v^T(y)$ defined in is finite.
Then, optimal solutions $X^T \in \mathcal{X}$ for $u^T$ and $Y^T \in \mathcal{Y}$ for $v^T$ exist such that $X^T Y^T$ is a uniformly integrable martingale and $$\label{eq: X^T resp}
X^T_T \in -\partial V(y^T Y^T_T) \quad \text{for some $y^T>0$}.$$
For isoelastic utilities $\tilde{U}$, the same statements hold (cf. [@Kramkov-Schachermayer-99]). We denote the corresponding optimal solutions of $\tilde{u}^T$ and $\tilde{v}^T$ by $\tilde{X}^T$ and $\tilde{Y}^T$, respectively. Both of them are unique. Moreover, because $\tilde{V}$ is differentiable, simplifies to $\tilde{X}^T_T = - \tilde{V}'(\tilde{y}^T \tilde{Y}^T_T)$ for some $\tilde{y}^T>0$.
\[rem: stoch interest\] If the generic utility $U$ is differentiable and strictly concave, the aforementioned duality results also hold for a nondeterministic safe asset $S^0$, assuming that the latter is bounded from above and below by two deterministic, positive processes $\underline{S}$ and $\overline{S}$ (cf. [@karatzas.zitkovic.03 Theorem 3.10]).
In what follows, we will verify all prerequisites of Theorem \[thm:duality BTZ\] in our setting. We begin by deriving some consequences of the convergence of the ratio of marginal utilities.
\[lem:rv\] Suppose holds and let $p\neq 0$. Then, for large wealth levels, the generic utility $U$ lies between arbitrarily close multiples of its isoelastic counterpart $\tilde{U}$ up to an additive constant. That is, for any $\epsilon>0$ there exists some sufficiently large $M_\epsilon>0$ such that:
1. If $p\in (0,1)$, then for some constants $A_\epsilon$ and $B_\epsilon$, $$\label{eq: U bdd p>0}
(1-\epsilon) x^p/p + B_\epsilon \leq U(x) \leq (1+\epsilon) x^p/p + A_\epsilon, \quad \mbox{for $x\geq M_\epsilon$};$$
2. If $p<0$, then $$\label{intest2}
(1-\epsilon) x^p/p \geq U(x) \geq (1+\epsilon) x^p/p, \quad \mbox{for $x\geq M_\epsilon$.}$$
Moreover, $U$ is regularly varying at infinity in both cases.
For any $\epsilon\in (0,1)$, the convergence of the ratio of marginal utilities yields the existence of some sufficiently large $M_\epsilon$ such that $$\label{eq:intest}
(1-\epsilon) x^{p-1} \leq U'(x) \leq (1+\epsilon) x^{p-1}, \quad \mbox{for $x\geq M_\epsilon$}.$$ Up to a larger $M_\epsilon$, we can assume that the generic utility $U$ is differentiable for $x \geq M_\epsilon$. When $p\in (0, 1)$, integrating on $(M_\epsilon, x)$ for $x\geq M_\epsilon$ gives the estimates in i). When $p<0$, integrating on $(x, \infty)$ for $x\geq M_\epsilon$ and using $U(\infty)=0$ in Assumption \[ass: utility\] iii) c) yields the estimates in ii). As for the regular variation of $U$, the bounds in i) give $$\limsup_{x\uparrow \infty} \frac{U(cx)}{U(x)} \leq \limsup_{x\uparrow \infty} \frac{(1+\epsilon) (cx)^p/p + A_\epsilon}{(1-\epsilon) x^p/p + B_\epsilon} = \frac{1+\epsilon}{1-\epsilon} \,c^p, \quad \mbox{for any $c>0$.}$$ As $\epsilon$ is chosen arbitrarily, it follows that $\limsup_{x\rightarrow \infty} U(cx)/U(x) \leq c^p$. The converse statement $\liminf_{x\rightarrow \infty} U(cx)/U(x) \geq c^p$ is obtained analogously. Hence, $U$ is regularly varying at infinity. The regular variation of $U$ for the case $p<0$ is obtained along the same lines.
As $U$ is differentiable and strictly concave at sufficiently large wealth levels (cf. Assumption \[ass: utility\] i)), the subdifferential $-\partial V(y)$ is single-valued and equals $(U')^{-1}(y)$ when $y\in (0, y_0)$ for some sufficiently small $y_0>0$. Set $$I(y):= -\partial V(y) = (U')^{-1}(y), \quad \text{for $y\in (0,y_0)$}.$$
Suppose holds. Then: $$\label{eq: I conv}
\lim_{y\downarrow 0} I(y) y^{\frac{1}{1-p}} =1.$$
Set $x= I(y)$, which tends to infinity as $y\downarrow 0$. Then, follows from $$\frac{I(y)}{y^{\frac{1}{p-1}}} = \frac{I(U'(x))}{(U'(x))^{\frac{1}{p-1}}} = \frac{x}{(U'(x))^{\frac{1}{p-1}}} = {\left(\frac{x^{p-1}}{U'(x)}\right)}^{\frac{1}{p-1}} \rightarrow 1, \quad \text{as $y\downarrow 0$},$$ where the convergence holds due to .
Let us now verify the prerequisites of Theorem \[thm:duality BTZ\].
\[lem: dom V\] Let Assumption \[ass: utility\] hold. Then $\overline{\text{dom}(V)} = [0,\infty)$.
We prove the statement for the case $p<0$; the proofs for $0<p<1$ and $p=0$ are similar. As $U$ is increasing with $\lim_{x\uparrow \infty} U(x) =0$, we have $U(x) - xy \leq 0$ for any $x, y>0$. Therefore, $V(y)\leq 0$ for any $y>0$. To obtain a lower bound for $V$, recall the second inequality in , which gives a lower bound for $U$ for $x\geq M_{\epsilon}$. For the same $M_\epsilon$, implies the existence of $C_M<0$ such that $U(x)\geq C_M x^{p-1}$ for $x< M_\epsilon$. Hence, when $y$ is small, $V(y)$ is bounded from below by the convex dual of $(1+\epsilon) x^p/p$, which is $-(1+\epsilon)^{\frac{1}{1-p}} \frac{p-1}{p} y^{\frac{p}{p-1}}$; when $y$ is large, $V(y)$ is bounded from below by the convex dual of $C_M x^{p-1}$, which is $-\frac{p-1}{p-2} ((p-1) C_M)^{-\frac{1}{p-2}} y^{\frac{p-1}{p-2}}$. Both convex dual functions are larger than negative infinity when $y\in (0,\infty)$. Therefore, $V(y)>-\infty$ for $y\in (0,\infty)$ confirming that $\text{dom}(V) =(0,\infty)$.
The convex dual $V$ of the generic utility $U$ satisfies the dual asymptotic elasticity condition.
As $-\partial V(y) = I(y)$ for sufficiently small $y$, the statement is equivalent to $\limsup_{y\downarrow 0} \frac{y I(y)}{V(y)}<\infty$. For any $\epsilon>0$, yields $(1-\epsilon) y^{\frac{1}{p-1}} \leq I(y) \leq (1+\epsilon) y^{\frac{1}{p-1}}$ for sufficiently small $y$, say $y\leq y_\epsilon$ for some $y_\epsilon$. Integrating the first inequality on $(y, y_\epsilon)$ yields $V(y) \geq (1-\epsilon) \tilde{V}(y) + D_\epsilon$, for $y\leq y_\epsilon$ and some $D_\epsilon$. Then, $\limsup_{y\downarrow 0} \frac{y I(y)}{V(y)}<\infty$ follows from this lower bound for $V(y)$ and the upper bound for $I(y)$.
\[lem: finite\] The value functions $u^T, {\tilde u}^T$ in , and $v^T, {\tilde v}^T$ in are finite.
If $p<0$, yields $I(y)\leq (1+\epsilon)y^{\frac{1}{p-1}}$ for $y\leq y_\epsilon$. Integrating this inequality on $(y, y_\epsilon)$ yields $V(y)\leq (1+\epsilon) \tilde{V}(y) + \tilde{D}_\epsilon$, $y\leq y_\epsilon$, for some $\tilde{D}_\epsilon$. As $\tilde{V}(y)<0$ and $V(y)$ is decreasing, this upper bound implies that $V(y)$ is uniformly bounded from above on $(0,\infty)$, so that $v^T(y)<\infty$ for any $y>0$. On the other hand, it follows from $\tilde{U}(x) \leq 0$ and the first inequality in that $U(x)$ is uniformly bounded from above. Hence, $u^T <\infty$ for any $T>0$ as well. When $0\leq p<1$, the same argument as in the case $p<0$ gives $V(y) \leq (1+\epsilon) \tilde{V}(y) + \tilde{D}_\epsilon$, $y\leq y_\epsilon$, for some $\tilde{D}_\epsilon$. As $V(y)$ is decreasing, $V(y)$ is also bounded from above for $y\geq y_\epsilon$. Combining this upper bound for $V(y)$ with Assumption \[ass: wellposedness\] gives $v^T(y)<\infty$, for any $T>0$ and $y>0$. Hence $u^T$ is also finite, due to the duality relation $u^T \leq \inf_{y>0}(v^T(y) + y)$.
The previous three lemmata verify all assumptions in Theorem \[thm:duality BTZ\], hence the statements therein hold. For the isoelastic problem $\tilde{u}^T$, we recall from [@Guasoni-Robertson Lemma 5] the following duality result, which will be frequently used below.
\[lem: power dual\] Let $X,Y$ be ${\mathcal{F}}_T$-measurable random variables such that $X,Y>0$ almost surely and $\mathbb{E}[XY] \leq 1$. Then, for any $0\neq p<1$, $$\frac{1}{p} \mathbb{E}[X^p] \leq \frac{1}{p}\mathbb{E}[Y^q]^{1-p}, \quad \text{for $q=\frac{p}{p-1}$},$$ and equality holds if and only if $\mathbb{E}[XY]=1$ and $X^{p-1}=\alpha Y$ for some $\alpha>0$.
Finally, we note that $\lim_{T\rightarrow \infty} S^0_T=\infty$ in Assumption \[ass: growth\] implies the following long-run property for stochastic discount factors: $$\label{eq: expec Y conv}
{\lim_{T\rightarrow \infty}}\sup_{Y\in \mathcal{Y}}{\mathbb{E}}[Y_T] =0.$$ Indeed, a full safe investment is admissible, so that the supermartingale property of $S^0Y$ yields $S^0_0 \geq S^0_T{\mathbb{E}}[Y_T]$ for any $Y\in \mathcal{Y}$. Assertion in turn follows from Assumption \[ass: growth\].
Proof of the main result for $p\neq 0$
======================================
Different arguments are needed to establish the main result in the cases where the corresponding isoelastic utilities are of power type ($p\neq 0$) or logarithmic ($p=0$), respectively.
First, consider the case $p \neq 0$. Recall from Lemma \[lem:rv\] that the generic utility $U$ is regularly varying at infinity. As a result, the convergence of the ratio of certainty equivalents follows from the convergence of the ratio of the respective utilities:
\[lem:equiv\] Let Assumptions \[ass: growth\] - \[ass: wellposedness\] hold. Then holds provided that $$\label{eq: value conv 1}
{\lim_{T\rightarrow \infty}}\frac{{\mathbb{E}}[U(X^T_T )]}{{\mathbb{E}}[U(\tilde{X}^T_T )]} =1.$$
To simplify notation, define $a_T = U^{-1}({\mathbb{E}}[U(X^T_T)])$ as well as $b_T = U^{-1}({\mathbb{E}}[U(\tilde{X}^T_T)])$, and note that $a_T, b_T\geq 0$ and $a_T\geq b_T$.
Observe that $\lim_{T\rightarrow \infty} a_T = \lim_{T\rightarrow \infty} b_T = \infty$. Indeed, for $p\in (0,1)$, ${\lim_{T\rightarrow \infty}}{\mathbb{E}}[U(X_T^T )] \geq {\lim_{T\rightarrow \infty}}{\mathbb{E}}[U(S^0_T )] = \infty$ follows from $\lim_{x\uparrow \infty} U(x) = \infty$ (cf. the first inequality in ) and Assumption \[ass: growth\]. In view of , we also have $\lim_{T\rightarrow \infty} {\mathbb{E}}[U(\tilde{X}^T_T)] =\infty$. As $U^{-1}(\infty)=\infty$, this confirms that $\lim_{T\rightarrow \infty} a_T = \lim_{T\rightarrow \infty} b_T = \infty$. When $p<0$, the argument is analogous, with the difference that $U(\infty)=0$, whence both expected utilities converge to zero, and therefore certainty equivalents are also become infinite.
Now, prove by contradiction. Suppose that, for any $\epsilon>0$, there exists a subsequence $\{T_n\}_{n\geq 0}$ with $\lim_{n\rightarrow \infty} T_n=\infty$ such that $a_{T_n} / b_{T_n} \geq 1+\epsilon$. For $p\in (0,1)$, the regular variation of $U$ at infinity implies that $$\liminf_{T_n\rightarrow \infty} \frac{U(a_{T_n})}{U(b_{T_n})} \geq \liminf_{T_n\rightarrow \infty} \frac{U((1+\epsilon) b_{T_n})}{U(b_{T_n})} = (1+\epsilon)^p >1.$$ For $p<0$, the inequality $U(b_{T_n}) <0$ and the regular variation of $U$ at infinity imply that $$\limsup_{T_n\rightarrow \infty} \frac{U(a_{T_n})}{U(b_{T_n})} \leq \limsup_{T_n\rightarrow \infty}\frac{U((1+\epsilon) b_{T_n})}{U(b_{T_n})} = (1+\epsilon)^p <1,$$ and both inequalities above contradict .
To prove Theorem \[thm:mt\], it therefore remains to show the following:
\[prop: limit value\] Let Assumptions \[ass: growth\] - \[ass: wellposedness\] hold. Then $$\label{eq: value conv}
{\lim_{T\rightarrow \infty}}\frac{{\mathbb{E}}[U(X^T_T )]}{{\mathbb{E}}[U(\tilde{X}^T_T )]} =1.$$
In order to compare the utilities ${\mathbb{E}}[U(X^T_T)]$ and ${\mathbb{E}}[U(\tilde{X}^T_T)]$ as $T\rightarrow \infty$, compare both of them to the maximal isoelastic utilities ${\mathbb{E}}[(\tilde{X}^T_T)^p/p]$. Henceforth, we prove Proposition \[prop: limit value\] separately for $0<p<1$ and $p<0$, as the contribution of low wealth levels requires a different treatment in each case.
Proof of Proposition \[prop: limit value\] for $0<p<1$
------------------------------------------------------
We first focus on the long-run performance of the isoelastic portfolio $\tilde{X}^T_T$.
\[lem: U-p-tx p>0\] Let Assumptions \[ass: growth\] - \[ass: wellposedness\] hold. Then: $${\lim_{T\rightarrow \infty}}\frac{{\mathbb{E}}[U(\tilde{X}^T_T)]}{{\mathbb{E}}[(\tilde{X}^T_T)^p/p]} =1.$$
To simplify notation, the superscript $T$ in $\tilde{X}^T_T$ is omitted throughout this proof. By , $$\begin{split}
&\frac{{\mathbb{E}}{\left[U(\tilde{X}_T ) \, {\mathbb{I}}_{\{\tilde{X}_T < M_\epsilon\}}\right]}}{{\mathbb{E}}[\tilde{X}^p_T/p]} + B_\epsilon \frac{{\mathbb{P}}(\tilde{X}_T \geq M_\epsilon)}{{\mathbb{E}}[\tilde{X}^p_T/p]} + (1-\epsilon) \frac{{\mathbb{E}}{\left[(\tilde{X}_T )^p \, {\mathbb{I}}_{\{\tilde{X}_T \geq M_\epsilon\}}\right]}}{{\mathbb{E}}[\tilde{X}^p_T]} \\
\leq\ & \frac{{\mathbb{E}}[U(\tilde{X}_T )]}{{\mathbb{E}}[\tilde{X}_T^p/p]} \label{eq: same X ub p>0}\\
\leq\ & \frac{{\mathbb{E}}{\left[U(\tilde{X}_T ) \, {\mathbb{I}}_{\{\tilde{X}_T < M_\epsilon\}}\right]}}{{\mathbb{E}}[\tilde{X}^p_T/p]} + A_\epsilon \frac{{\mathbb{P}}(\tilde{X}_T \geq M_\epsilon)}{{\mathbb{E}}[\tilde{X}^p_T/p]} + (1+\epsilon) \frac{{\mathbb{E}}{\left[(\tilde{X}_T )^p \, {\mathbb{I}}_{\{\tilde{X}_T \geq M_\epsilon\}}\right]}}{{\mathbb{E}}[\tilde{X}^p_T]}.
\end{split}$$ Note that $U(\tilde{X}_T )$ is uniformly bounded from below because this is assumed for the generic utility $U$, cf. Assumption \[ass: utility\] iii)-a). As ${\lim_{T\rightarrow \infty}}{\mathbb{E}}[\tilde{X}^p_T/p] \geq {\lim_{T\rightarrow \infty}}{\mathbb{E}}[(S^0_T)^p/p] =\infty$ by Assumption \[ass: growth\], this implies that the first two terms on both sides of converge to zero as $T\rightarrow \infty$.
Now, focus on the third terms. Define an auxiliary probability measure ${\mathbb{P}}^T$ on ${\mathcal{F}}_T$ via $$\label{eq: def P^T}
\frac{d{\mathbb{P}}^T}{d{\mathbb{P}}} = \frac{\tilde{X}^p_T}{{\mathbb{E}}[\tilde{X}^p_T]} ,$$ where the expectation is finite by Lemma \[lem: finite\]. Recall from Section \[sec:dual\] that $\tilde{y}^T \tilde{Y}^T_T = \tilde{X}_T^{p-1}$ and, moreover, $1={\mathbb{E}}[\tilde{Y}^T_T\tilde{X}_T]\geq {\mathbb{E}}[\tilde{Y}^T_T X_T]$ and hence $$\label{eq: numeraire}
0\geq {\mathbb{E}}{\left[\tilde{y}^T \tilde{Y}^T_T (X_T - \tilde{X}_T)\right]} = {\mathbb{E}}{\left[\tilde{X}^{p-1}_T (X_T - \tilde{X}_T)\right]} = {\mathbb{E}}{\left[\tilde{X}_T^p\right]}{\mathbb{E}}^{{\mathbb{P}}^T}{\left[\frac{X_T}{\tilde{X}_T}-1\right]},$$ for any admissible wealth process $X$. This inequality yields that $$\label{eq: tX ubb}
{\lim_{T\rightarrow \infty}}{\mathbb{P}}^T(\tilde{X}_T \geq N) =1, \quad \text{ for any } N>0.$$ Indeed, choosing $X_T=S^0_T$ in , it follows that $$1\geq {\mathbb{E}}^{{\mathbb{P}}^T}[S^0_T/ \tilde{X}_T] \geq {\mathbb{E}}^{{\mathbb{P}}^T}{\left[S^0_T/\tilde{X}_T \, {\mathbb{I}}_{\{S^0_T\geq L, \tilde{X}_T\leq N\}}\right]} \geq \frac{L}{N} \, {\mathbb{P}}^T(S^0_T\geq L, \tilde{X}_T\leq N),$$ for any positive $L$ and $N$. Combined with Assumption \[ass: growth\], this yields $$\begin{aligned}
\limsup_{T\rightarrow \infty} {\mathbb{P}}^T(\tilde{X}_T\leq N) &\leq \limsup_{T\rightarrow \infty} {\mathbb{P}}^T(S^0_T\geq L, \tilde{X}_T \leq N) + \limsup_{T\rightarrow \infty} {\mathbb{P}}^T(S^0_T <L)\\
&\leq N/L,
\end{aligned}$$ which confirms because $L$ was arbitrary.
Now, coming back to the third term of the upper and lower bounds in , $$\begin{aligned}
\frac{{\mathbb{E}}{\left[(\tilde{X}_T )^p \, {\mathbb{I}}_{\{\tilde{X}_T \geq M_\epsilon\}}\right]}}{{\mathbb{E}}[\tilde{X}^p_T]} = {\mathbb{P}}^T(\tilde{X}_T \geq M_\epsilon) \rightarrow 1, \quad \text{as } T\rightarrow \infty,
\end{aligned}$$ where the convergence follows from . Therefore, the above estimates on both sides of yield $$1-\epsilon \leq \liminf_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(\tilde{X}_T )]}{{\mathbb{E}}[\tilde{X}^p_T]} \leq \limsup_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(\tilde{X}_T )]}{{\mathbb{E}}[\tilde{X}^p_T]} \leq 1+\epsilon.$$ This proves the assertion because $\epsilon$ was chosen arbitrarily.
The proof of Proposition \[prop: limit value\] (and hence the main result) now proceeds as follows. The lower bound in yields $${\mathbb{E}}[U(\tilde{X}^T_T) {\mathbb{I}}_{\{\tilde{X}^T_T \geq M_\epsilon\}}] \geq (1-\epsilon) \,{\mathbb{E}}[(\tilde{X}^T_T)^p/p \, {\mathbb{I}}_{\{\tilde{X}^T_T \geq M_\epsilon\}}] + B_\epsilon {\mathbb{P}}(\tilde{X}^T_T \geq M_\epsilon),$$ which converges to $\infty$ as $T\rightarrow \infty$, because ${\lim_{T\rightarrow \infty}}{\mathbb{E}}[(\tilde{X}^T_T)^p/p] \geq {\lim_{T\rightarrow \infty}}{\mathbb{E}}[(S^0_T)^p/p] =\infty$ follows from Assumption \[ass: growth\]. Therefore, given that $U$ is bounded from below, ${\lim_{T\rightarrow \infty}}{\mathbb{E}}[U(\tilde{X}^T_T)] =\infty$. On the other hand, the optimality of $X^T_T$ for the generic utility $U$ yields ${\mathbb{E}}[U(\tilde{X}^T_T )] \leq {\mathbb{E}}[U(X^T_T )]$, so that $$\liminf_{T\rightarrow \infty}\frac{{\mathbb{E}}[U(X^T_T )]}{{\mathbb{E}}[U(\tilde{X}^T_T )]} \geq 1.$$ Therefore, Proposition \[prop: limit value\] follows by showing $$\limsup_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(X^T_T )]}{{\mathbb{E}}[U(\tilde{X}^T_T )]} \leq 1.$$ In view of Lemma \[lem: U-p-tx p>0\], it suffices to prove that $$\limsup_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(X^T_T)]}{{\mathbb{E}}[(\tilde{X}^T_T)^p/p]} \leq 1.$$ To establish this inequality, we begin with the following auxiliary result:
\[lem: ratio Y p>0\] Suppose Assumptions \[ass: growth\] - \[ass: wellposedness\] hold and let $p\in (0,1)$. Then: $$1={\lim_{T\rightarrow \infty}}\frac{{\mathbb{E}}[(Y^T_T)^q]^{1-p}}{{\mathbb{E}}[(\tilde{Y}^T_T)^q]^{1-p}} = {\lim_{T\rightarrow \infty}}\frac{y^T}{\tilde{y}^T}.$$
To simplify notation, we omit the superscript $T$ in $X^T$, $\tilde{X}^T$, $Y^T$, and $\tilde{Y}^T$ throughout this proof. For the $M_\epsilon$ in , the martingale property of $XY$ implies $$\label{eq: XY primary p>0}
\begin{split}
1 = {\mathbb{E}}[X_T Y_T] &= \frac{1}{y^T}\,{\mathbb{E}}{\left[U'(X_T ) X_T \, {\mathbb{I}}_{\{X_T \geq M_\epsilon\}}\right]} + {\mathbb{E}}[X_T Y_T \, {\mathbb{I}}_{\{X_T <M_\epsilon\}}]\\
&\leq \frac{1+\epsilon}{y^T} \,{\mathbb{E}}{\left[X_T^{p-1} X_T \, {\mathbb{I}}_{\{X_T \geq M_\epsilon\}}\right]} + {\mathbb{E}}[X_T Y_T \, {\mathbb{I}}_{\{X_T <M_\epsilon\}}].
\end{split}$$ Here, the second identity uses the first-order condition $U'(X_T) = y^T Y_T$ when $X_T \geq M_\epsilon$, and the inequality follows from the second inequality in . The second term on the right-hand side of vanishes as $T\rightarrow \infty$ due to . For the first term, note that ${\mathbb{E}}[X_T \hat{Y}_T] \leq 1$ for any $\hat{Y}\in \mathcal{Y}$. Hence, it follows from Lemma \[lem: power dual\] that $$\label{eq: GR-duality}
\frac{1}{p} {\mathbb{E}}[X_T^p] \leq \frac{1}{p} {\mathbb{E}}[\hat{Y}_T^q]^{1-p}, \quad \text{for $q=\frac{p}{p-1}$}.$$ The previous inequality and $p>0$ imply that $$\liminf_{T\rightarrow \infty} {\mathbb{E}}{\left[X_T^p \, {\mathbb{I}}_{\{X_T \geq M_\epsilon\}}\right]} \leq \liminf_{T\rightarrow \infty} {\mathbb{E}}{\left[X_T^p\right]} \leq \liminf_{T\rightarrow \infty} {\mathbb{E}}[\hat{Y}^q_T]^{1-p},\quad \mbox{ for any } \hat{Y}\in \mathcal{Y}.$$ Coming back to , the above estimates for the two terms on the right-hand side yield $$\label{eq: Y^q est p>0}
\frac{1}{1+\epsilon} \leq \liminf_{T\rightarrow \infty} \frac{1}{y^T} {\mathbb{E}}[\hat{Y}^q_T]^{1-p}, \quad \mbox{ for any } \hat{Y}\in \mathcal{Y}.$$
For any $\epsilon>0$, shows that there exists a sufficiently small $\delta_\epsilon<y_0$, so that $$\label{eq:estI p>0}
(1-\epsilon) \, y^{\frac{1}{p-1}} \leq I(y) \leq (1+\epsilon) \, y^{\frac{1}{p-1}}, \quad \mbox{for } y<\delta_\epsilon.$$ Fix such a $\delta_\epsilon$; the martingale property of $X Y$ then implies $$\label{eq: XY dual p>0}
1= {\mathbb{E}}[X_T Y_T] = {\mathbb{E}}{\left[Y_T I(y^TY_T) \, {\mathbb{I}}_{\{y^T Y_T\leq \delta_\epsilon\}}\right]} + {\mathbb{E}}{\left[Y_T X_T \, {\mathbb{I}}_{\{y^T Y_T> \delta_\epsilon\}}\right]},$$ where the second identity follows from $X_T = I(y^TY_T)$ for $y^T Y_T \leq \delta_\epsilon$. Continue by estimating the second term on the right-hand side. Let $V'_-$ be the increasing left derivative of the convex function $V$. Then $-V'_-(y)$ dominates all other elements of the subdifferential $-\partial V$ at $y$, i.e., $- V'_-(y) \geq x$ for any $x\in -\partial V(y)$. Moreover, again as $-V'_-$ is nonincreasing, there exists $C_\delta$ such that $x\leq C_\delta$ for any $x\in -\partial V(y)$ and $y> \delta_\epsilon$. In view of , it therefore follows that $X_T \leq C_\delta$ when $y^T Y_T >\delta_\epsilon$. As a result, gives $$\label{eq: YX large y}
{\mathbb{E}}{\left[Y_T X_T \, {\mathbb{I}}_{\{y^T Y_T > \delta_\epsilon\}}\right]} \leq C_\delta {\mathbb{E}}{\left[Y_T \,{\mathbb{I}}_{\{y^T Y_T >\delta_\epsilon\}}\right]} \rightarrow 0, \quad \text{as} \quad T\rightarrow \infty.$$ Turning to the first term on the right-hand side of , it follows from the first inequality in that $$\label{eq: YI est p>0}
\begin{split}
&{\mathbb{E}}{\left[Y_T I(y^T Y_T) \, {\mathbb{I}}_{\{y^T Y_T \leq \delta_\epsilon\}}\right]}\\
\geq\ & (1-\epsilon) \frac{{\mathbb{E}}[Y_T^q \, {\mathbb{I}}_{\{y^T Y_T\leq \delta_\epsilon\}}]}{(y^T)^{1/(1-p)}}\\
=\ & \frac{1-\epsilon}{(y^T)^{1/(1-p)}} {\mathbb{E}}[Y^q_T] - \frac{1-\epsilon}{(y^T)^{1/(1-p)}} {\mathbb{E}}[Y^q_T \, {\mathbb{I}}_{\{y^T Y_T >\delta_\epsilon\}}].
\end{split}$$ Here, the second term tends to zero as the horizon grows because $$\frac{1}{(y^T)^{1/(1-p)}} {\mathbb{E}}[Y^q_T \, {\mathbb{I}}_{\{y^T Y_T >\delta_\epsilon\}}] = {\mathbb{E}}[Y_T (y^T Y_T)^{\frac{1}{p-1}} \, {\mathbb{I}}_{\{y^T Y_T >\delta_\epsilon\}}] \leq \delta_\epsilon^{\frac{1}{p-1}} {\mathbb{E}}[Y_T]\rightarrow 0,$$ due to . These estimates together with and imply that $$\frac{1}{1-\epsilon} \geq \limsup_{T\rightarrow \infty} \frac{1}{(y^T)^{1/(1-p)}} {\mathbb{E}}[Y^q_T].$$ Raising both sides to the power $(1-p)$, and using the optimality of $\tilde{Y}$ for the dual problem $\tilde{v}^T$ in , we obtain $${\left(\frac{1}{1-\epsilon}\right)}^{1-p} \geq \limsup_{T\rightarrow \infty} \frac{1}{y^T} {\mathbb{E}}[Y_T^q]^{1-p} \geq \limsup_{T\rightarrow \infty} \frac{1}{y^T} {\mathbb{E}}[\tilde{Y}^q_T]^{1-p}.$$ Together with , and recalling that $\epsilon$ was chosen arbitrarily, it follows that $$1= {\lim_{T\rightarrow \infty}}\frac{1}{y^T} {\mathbb{E}}[Y^q_T]^{1-p} = {\lim_{T\rightarrow \infty}}\frac{1}{y^T} {\mathbb{E}}[\tilde{Y}^q_T]^{1-p}.$$ Together with ${\mathbb{E}}[\tilde{Y}_T^q]^{1-p} = {\mathbb{E}}[\tilde{X}^p_T] = \tilde{y}^T$ (cf. Lemma \[lem: power dual\]) this yields the assertion.
We are now ready to complete the proof for Proposition \[prop: limit value\], and hence the main Theorem \[thm:mt\], for the case $0<p<1$.
We continue to omit the superscript $T$ to ease notation. As discussed before Lemma \[lem: ratio Y p>0\], it suffices to show $$\label{eq: UX/ptX p>0}
\limsup_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(X_T)]}{{\mathbb{E}}[\tilde{X}_T^p/p]} \leq 1.$$ The second inequality in implies $$\label{eq: est UX/ptX p>0}
\begin{split}
&\frac{{\mathbb{E}}[U(X_T )]}{{\mathbb{E}}[\tilde{X}_T^p/p]}\\
=\ & \frac{{\mathbb{E}}[U(X_T ) \,{\mathbb{I}}_{\{X_T <M_\epsilon\}}]}{{\mathbb{E}}[\tilde{X}_T^p/p]} + \frac{{\mathbb{E}}[U(X_T )\, {\mathbb{I}}_{\{X_T \geq M_\epsilon\}}]}{{\mathbb{E}}[\tilde{X}_T^p/p]}\\
\leq\ & \frac{{\mathbb{E}}{\left[U(X_T ) \, {\mathbb{I}}_{\{X_T < M_\epsilon\}}\right]}}{{\mathbb{E}}[\tilde{X}^p_T/p]} + A_\epsilon \frac{{\mathbb{P}}(X_T \geq M_\epsilon)}{{\mathbb{E}}[\tilde{X}^p_T/p]} + (1+\epsilon) \frac{{\mathbb{E}}{\left[X_T^p \, {\mathbb{I}}_{\{X_T \geq M_\epsilon\}}\right]}}{{\mathbb{E}}[\tilde{X}^p_T]}\\
\leq\ & \frac{{\mathbb{E}}{\left[U(X_T ) \, {\mathbb{I}}_{\{X_T < M_\epsilon\}}\right]}}{{\mathbb{E}}[\tilde{X}^p_T/p]} + A_\epsilon \frac{{\mathbb{P}}(X_T \geq M_\epsilon)}{{\mathbb{E}}[\tilde{X}^p_T/p]} + (1+\epsilon) \frac{{\mathbb{E}}{\left[X_T^p\right]}}{{\mathbb{E}}[\tilde{X}^p_T]}.
\end{split}$$ As we have seen in the proof of Lemma \[lem: U-p-tx p>0\], the first two terms on the right-hand side vanish as $T\rightarrow \infty$. For the third, Lemma \[lem: power dual\] implies $
(1/p){\mathbb{E}}[X_T^p] \leq (1/p) {\mathbb{E}}[Y_T^q]^{1-p}
$ so that $\limsup_{T\rightarrow \infty} {\mathbb{E}}[X_T^p] \leq \limsup_{T\rightarrow \infty} {\mathbb{E}}[Y_T^q]^{1-p}$ for $q=p/(p-1)$ because $p>0$. On the other hand, ${\mathbb{E}}[\tilde{X}_T^p] = {\mathbb{E}}[\tilde{Y}_T^q]^{1-p}$ (again cf. Lemma \[lem: power dual\]). Therefore the previous estimates together with imply $$\limsup_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(X_T )]}{{\mathbb{E}}[\tilde{X}_T^p/p]} \leq (1+\epsilon) \limsup_{T\rightarrow \infty} \frac{{\mathbb{E}}[Y_T^q]^{1-p}}{{\mathbb{E}}[\tilde{Y}_T^q]^{1-p}} = 1+\epsilon,$$ where the last identity follows from Lemma \[lem: ratio Y p>0\]. Hence, Assertion is confirmed because $\epsilon$ was chosen arbitrarily.
Proof of Proposition \[prop: limit value\] for $p<0$
----------------------------------------------------
The overall strategy is similar to the case $0<p<1$. However, the contribution of low wealth levels to the total expected utility is more delicate, as utilities may be unbounded from below near zero and the value functions $u^T$ and $\tilde{u}^T$ converge to zero rather than infinity as $T\rightarrow \infty$.[^15] Therefore, the additional Assumptions on $U$ are needed to ensure that the contribution of low wealth levels is still negligible in the long run. We start with the following analogue of Lemma \[lem: U-p-tx p>0\].
\[lem: U-p-tx\] Let Assumptions \[ass: growth\] and \[ass: utility\] hold. Then: $${\lim_{T\rightarrow \infty}}\frac{{\mathbb{E}}[U(\tilde{X}^T_T)]}{{\mathbb{E}}[(\tilde{X}^T_T)^p/p]} =1.$$
To simplify notation, the superscript $T$ in $\tilde{X}^T$ and $\tilde{Y}^T$ is omitted throughout this proof. In view of , we have $$\label{eq: same X ub}
\begin{split}
&\frac{{\mathbb{E}}{\left[U(\tilde{X}_T ) \, {\mathbb{I}}_{\{\tilde{X}_T < M_\epsilon\}}\right]}}{{\mathbb{E}}[\tilde{X}^p_T/p]} + (1-\epsilon) \frac{{\mathbb{E}}{\left[\tilde{X}_T^p \, {\mathbb{I}}_{\{\tilde{X}_T \geq M_\epsilon\}}\right]}}{{\mathbb{E}}[\tilde{X}^p_T]} \\
\leq\ & \frac{{\mathbb{E}}[U(\tilde{X}_T )]}{{\mathbb{E}}[\tilde{X}_T^p/p]}\\
\leq\ & \frac{{\mathbb{E}}{\left[U(\tilde{X}_T ) \, {\mathbb{I}}_{\{\tilde{X}_T < M_\epsilon\}}\right]}}{{\mathbb{E}}[\tilde{X}^p_T/p]} + (1+\epsilon) \frac{{\mathbb{E}}{\left[\tilde{X}_T^p \, {\mathbb{I}}_{\{\tilde{X}_T \geq M_\epsilon\}}\right]}}{{\mathbb{E}}[\tilde{X}^p_T]}.
\end{split}$$ Let us estimate separately the two terms in the upper and lower bounds. For the second term, similar arguments as in the proof of Lemma \[lem: U-p-tx p>0\], utilizing the auxiliary measure ${\mathbb{P}}^T$, yield $$\label{eq: same X ub 2}
\lim_{T\rightarrow \infty} \frac{{\mathbb{E}}{\left[\tilde{X}_T^p \, {\mathbb{I}}_{\{\tilde{X}_T \geq M_\epsilon\}}\right]}}{{\mathbb{E}}[\tilde{X}^p_T]} = 1.$$
Compared to the proof of Lemma \[lem: U-p-tx p>0\], the case $p<0$ differs with respect to the estimation of the first term in the upper and lower bounds in . Here, the first-order condition $\tilde{X}_T^{p-1}=\tilde{y}^T \tilde{Y}_T$ and the martingale property of $\tilde{Y} \tilde{X}$ imply ${\mathbb{E}}[\tilde{X}_T^p]=\tilde{y}^T$. On the other hand, it follows from that $pU(x) \leq C_M x^{p-1}$ for some constant $C_M>0$ and any $x<M_\epsilon$. As a result: $$\label{eq: Utx/txp 2}
\begin{split}
\frac{{\mathbb{E}}[U(\tilde{X}_T ) \, {\mathbb{I}}_{\{\tilde{X}_T <M_\epsilon\}}]}{{\mathbb{E}}[\tilde{X}_T^p/p]} &=\frac{{\mathbb{E}}[p\,U(\tilde{X}_T ) \, {\mathbb{I}}_{\{\tilde{X}_T < M_\epsilon\}}]}{\tilde{y}^T} \\
&\leq C_M \frac{{\mathbb{E}}[\tilde{X}_T^{p-1} \, {\mathbb{I}}_{\{\tilde{X}_T <M_\epsilon\}}]}{\tilde{y}^T}\\
&= C_M {\mathbb{E}}[\tilde{Y}^T_T {\mathbb{I}}_{\{\tilde{X}_T < M_\epsilon\}}] \rightarrow 0, \quad \text{as $T\rightarrow \infty$}.
\end{split}$$ Therefore, the first terms of the upper and lower bounds in vanish as the horizon grows. Together with and , it follows that $$1-\epsilon\leq \liminf_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(\tilde{X}_T)]}{{\mathbb{E}}[\tilde{X}^p_T/p]} \leq \limsup_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(\tilde{X}_T )]}{{\mathbb{E}}[\tilde{X}^p_T/p]} \leq 1+\epsilon.$$ This yields the assertion because $\varepsilon$ was arbitrary.
\[rem: Utx/txp\] The calculation of the contribution from low wealth levels relative to the expected power utility in will be used in the counterexample in Section \[sec:analysis\]. Therefore, for future reference, we summarize it here: for any $M$ and $U$ satisfying , $$\lim_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(\tilde{X}^T_T) \, {\mathbb{I}}_{\{\tilde{X}^T_T< M\}}]}{{\mathbb{E}}[(\tilde{X}^T_T)^p/p]} =0.$$ A careful examination of shows that $U$ does not need to be concave to ensure the above convergence.
Similarly as in the case $0<p<1$, the proof of Proposition \[prop: limit value\] (and hence the main result) now proceeds as follows. The optimality of $X^T_T$ for the utility $U$ yields ${\mathbb{E}}[U(\tilde{X}^T_T)] \leq {\mathbb{E}}[U(X^T_T)] <0$, so that $$\frac{{\mathbb{E}}[U(X^T_T)]}{{\mathbb{E}}[U(\tilde{X}^T_T)]} \leq 1, \quad \mbox{for any $T>0$.}$$ Therefore, Proposition \[prop: limit value\] follows by showing $$\liminf_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(X^T_T)]}{{\mathbb{E}}[U(\tilde{X}^T_T)]} \geq 1.$$ Due to Lemma \[lem: U-p-tx\], it suffices to prove $$\liminf_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(X^T_T)]}{{\mathbb{E}}[(\tilde{X}^T_T)^p/p]} \geq 1,$$ which will be established in the sequel. The following auxiliary result is the analogue of Lemma \[lem: ratio Y p>0\]:
\[lem: ratio Y\] Suppose Assumptions \[ass: growth\] and \[ass: utility\] hold for $p<0$. Then, even without assuming : $$1={\lim_{T\rightarrow \infty}}\frac{{\mathbb{E}}[(Y^T_T)^q]^{1-p}}{{\mathbb{E}}[(\tilde{Y}^T_T)^q]^{1-p}} = {\lim_{T\rightarrow \infty}}\frac{y^T}{\tilde{y}^T}.$$
To simplify notation, we once more omit the superscript $T$ in $X^T$, $\tilde{X}^T$, $Y^T$, and $\tilde{Y}^T$ throughout this proof. Coming back to and using the second inequality in Equation , $$\label{eq: dual p<0 lb}
\begin{split}
1&\leq (1+\epsilon) \frac{{\mathbb{E}}[Y_T^q \, {\mathbb{I}}_{\{y^T Y_T\leq \delta_\epsilon\}}]}{(y^T)^{1/(1-p)}} + {\mathbb{E}}{\left[Y_T X_T \, {\mathbb{I}}_{\{y^T Y_T> \delta_\epsilon\}}\right]}\\
&\leq (1+\epsilon) \frac{{\mathbb{E}}[Y_T^q]}{(y^T)^{1/(1-p)}} + {\mathbb{E}}{\left[Y_T X_T \, {\mathbb{I}}_{\{y^T Y_T> \delta_\epsilon\}}\right]},
\end{split}$$ where $q=p/(p-1)$. Now, repeating the argument after , the second term on the right-hand side vanishes as $T\rightarrow \infty$. Then, raise both sides of to the $(1-p)$-th power, obtaining $$\label{eq: Y<tY}
{\left(\frac{1}{1+\epsilon}\right)}^{1-p} \leq \liminf_{T\rightarrow \infty} \frac{1}{y^T} {\mathbb{E}}[Y^q_T]^{1-p} \leq \liminf_{T\rightarrow \infty} \frac{1}{y^T} {\mathbb{E}}[\tilde{Y}^q_T]^{1-p},$$ where the second inequality follows from the optimality of $\tilde{Y}$ for the dual problem $\tilde{v}^T$ in .
On the other hand, for any fixed $a>0$, it follows from that $$\label{eq: conv a}
\lim_{x\uparrow \infty} \frac{U'(x)}{(x+a)^{p-1}} =1.$$ Hence, for any $\epsilon>0$, there is $M_{\epsilon, a}>0$ such that $U'(x)$ exists and $$\label{eq:intest a}
(1-\epsilon)(x+a)^{p-1} \leq U'(x) \leq (1+\epsilon) (x+a)^{p-1}, \quad \mbox{for $x\geq M_{\epsilon, a}$}.$$ The martingale property of $XY$, the first-order condition $y^T Y_T=U'(X_T)$ for large $X_T$, and the first inequality in in turn yield $$\label{eq: XY primary}
\begin{split}
1&= {\mathbb{E}}[X_T Y_T] \\
&= {\mathbb{E}}{\left[X_T Y_T \, {\mathbb{I}}_{\{X_T < M_{\epsilon, a}\}}\right]} - \frac{a}{y^T} {\mathbb{E}}{\left[U'(X_T) \, {\mathbb{I}}_{\{X_T\geq M_{\epsilon, a}\}}\right]}\\
&\qquad \qquad \qquad \qquad \qquad \quad+ \frac{1}{y^T} {\mathbb{E}}{\left[U'(X_T) (X_T +a) \, {\mathbb{I}}_{\{X_T\geq M_{\epsilon, a}\}}\right]}\\
& \geq {\mathbb{E}}{\left[X_T Y_T \, {\mathbb{I}}_{\{X_T < M_{\epsilon, a}\}}\right]} - a {\mathbb{E}}{\left[Y_T \, {\mathbb{I}}_{\{X_T \geq M_{\epsilon, a}\}}\right]}\\
& \qquad \qquad \qquad \qquad \qquad \quad+ (1-\epsilon)\frac{1}{y^T}{\mathbb{E}}{\left[(X_T+a)^{p} \, {\mathbb{I}}_{\{X_T \geq M_{\epsilon, a}\}}\right]}.
\end{split}$$ By , the first two terms on the right-hand side go to zero as $T\rightarrow \infty$. Therefore $$\label{eq: exp X/y}
\begin{split}
\frac{1}{1-\epsilon} &\geq \limsup_{T\rightarrow \infty} \frac{1}{y^T} {\mathbb{E}}[(X_T+a)^p \, {\mathbb{I}}_{\{X_T \geq M_{\epsilon, a}\}}] \\
&\geq \limsup_{T\rightarrow \infty} \frac{1}{y^T} {\mathbb{E}}[(X_T+a)^p] - \limsup_{T\rightarrow \infty} \frac{1}{y^T} {\mathbb{E}}[(X_T+a)^p \, {\mathbb{I}}_{\{X_T< M_{\epsilon, a}\}}].
\end{split}$$ Let us estimate the second term on the right-hand side below. To this end, $p<0$ and $X_T\geq 0$ imply that $$\label{eq: p/y 1}
\begin{split}
\frac{1}{y^T} {\mathbb{E}}[(X_T+ a)^p \, {\mathbb{I}}_{\{X_T < M_{\epsilon, a}\}}] &\leq \frac{a^p}{y^T} {\mathbb{E}}[{\mathbb{I}}_{\{X_T < M_{\epsilon, a}\}}] \\
&\leq \frac{a^p}{y^T} {\mathbb{E}}{\left[\frac{y^TY_T}{\delta_M} {\mathbb{I}}_{\{X_T < M_{\epsilon, a}\}}\right]} \\
&= \frac{a^p}{\delta_M} {\mathbb{E}}[Y_T {\mathbb{I}}_{\{X_T < M_{\epsilon, a}\}}] \rightarrow 0, \quad \text{as $T\rightarrow \infty$},
\end{split}$$ where the last step is once again a consequence of . Here, $\delta_M$ is a positive constant such that $y^T Y_T \geq \delta_M$ on $\{X_T <M_{\epsilon, a}\}$. The reason for the existence of such a constant is the following. Notice that $X_T \in -\partial V(y^T Y_T)$ and that every element $x$ of $-\partial V(y)$ dominates $-V'_+(y)$, where $V'_+(y)$ is the right derivative of $V$ at $y$. As $V(y)$ is strictly convex when $y$ is close to zero, $-V'_+$ is strictly decreasing. Together with $-V'_+(0)=\infty$, this implies the existence of $\delta_M>0$ such that $-V'_+(y) \geq M_{\epsilon, a}$ for $y<\delta_M$, and in turn $y^T Y_T \geq \delta_M$ on $\{X_T <M_{\epsilon, a}\}$.
In view of , the second term on the right-hand side of vanishes as the horizon grows, so that $$\label{eq: X/y ub}
\frac{1}{1-\epsilon} \geq\limsup_{T\rightarrow \infty} \frac{1}{y^T} {\mathbb{E}}[(X_T+ a)^p].$$ Now, note that ${\mathbb{E}}{\left[(X_T + a)\frac{\hat{Y}_T}{a{\mathbb{E}}[\hat{Y}_T] + 1}\right]}\leq 1$ for any $\hat{Y}\in \mathcal{Y}$. Hence, Lemma \[lem: power dual\] shows $$\label{eq: GR-duality-a}
\frac{1}{p} {\mathbb{E}}[(X_T+a)^p] \leq \frac{1}{p} \frac{{\mathbb{E}}[\hat{Y}_T^q]^{1-p}}{(a{\mathbb{E}}[ \hat{Y}_T]+1)^{-p}}, \quad \text{for $q=\frac{p}{p-1}$}.$$ Taking into account that ${\lim_{T\rightarrow \infty}}{\mathbb{E}}[\hat{Y}_T]=0$ and $p<0$, the previous two inequalities imply $$\label{eq: Y/y ub}
\frac{1}{1-\epsilon} \geq \limsup_{T\rightarrow \infty} \frac{1}{y^T} {\mathbb{E}}[(X_T+ a)^p] \geq \limsup_{T\rightarrow \infty} \frac{1}{y^T}{\mathbb{E}}[\hat{Y}^q_T]^{1-p}, \quad \text{ for any } \hat{Y}\in \mathcal{Y}.$$ In particular, this holds for $Y_T$ and $\tilde{Y}_T$. Combining this inequality with and recalling that $\epsilon$ was chosen arbitrarily, it follows that $$1= {\lim_{T\rightarrow \infty}}\frac{1}{y^T} {\mathbb{E}}[Y^q_T]^{1-p} = {\lim_{T\rightarrow \infty}}\frac{1}{y^T} {\mathbb{E}}[\tilde{Y}^q_T]^{1-p}.$$ Together with ${\mathbb{E}}[\tilde{Y}_T^q]^{1-p} = {\mathbb{E}}[\tilde{X}^p_T] = \tilde{y}^T$ (cf. Lemma \[lem: power dual\]), this yields the assertion.
We are now ready to complete the proof of Proposition \[prop: limit value\] and hence Theorem \[thm:mt\] in the case $p<0$.
As discussed before Lemma \[lem: ratio Y\], it suffices to show that $$\label{eq: liminf U/p}
\liminf_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(X_T)]}{{\mathbb{E}}[\tilde{X}_T^p/p]} \geq 1.$$ Fix $a>0$. Recall that $\lim_{x\uparrow \infty} U'(x)/(x+a)^{p-1} =1$ by . As in , it follows that $$\label{eq: U bdd p<0 a}
(1-\epsilon) (x+a)^p/p \geq U(x) \geq (1+\epsilon) (x+a)^p/p, \quad \mbox{for } x\geq M_{\epsilon, a}.$$ Hence, $$\begin{aligned}
\frac{{\mathbb{E}}[U(X_T)]}{{\mathbb{E}}[\tilde{X}_T^p/p]} &= \frac{{\mathbb{E}}[U(X_T)\,{\mathbb{I}}_{\{X_T<M_{\epsilon, a}\}}]}{{\mathbb{E}}[\tilde{X}_T^p/p]} + \frac{{\mathbb{E}}[U(X_T)\,{\mathbb{I}}_{\{X_T\geq M_{\epsilon, a}\}}]}{{\mathbb{E}}[\tilde{X}_T^p/p]} \\
&\geq (1-\epsilon) \frac{{\mathbb{E}}[(X_T+a)^p {\mathbb{I}}_{\{X_T
\geq M_{\epsilon, a}\}}]}{{\mathbb{E}}[\tilde{X}^p_T]}\\
&= (1-\epsilon) \frac{{\mathbb{E}}[(X_T+a)^p]}{{\mathbb{E}}[\tilde{X}^p_T]} - (1-\epsilon) \frac{{\mathbb{E}}[(X_T+a)^p {\mathbb{I}}_{\{X_T
< M_{\epsilon, a}\}}]}{{\mathbb{E}}[\tilde{X}^p_T]},\end{aligned}$$ where the inequality follows from the first inequality in , multiplied by $p<0$ on both sides. We have seen in that $${\lim_{T\rightarrow \infty}}\frac{1}{y^T}{\mathbb{E}}[(X_T+a)^p {\mathbb{I}}_{\{X_T
< M_{\epsilon, a}\}}] =0.$$ As ${\mathbb{E}}[\tilde{X}^p_T] = \tilde{y}^T$ and ${\lim_{T\rightarrow \infty}}y^T/\tilde{y}^T =1$ by Lemma \[lem: ratio Y\], it follows that $${\lim_{T\rightarrow \infty}}\frac{{\mathbb{E}}[(X_T+a)^p {\mathbb{I}}_{\{X_T
< M_{\epsilon, a}\}}]}{{\mathbb{E}}[\tilde{X}^p_T]} =0,$$ and in turn $$\label{eq: UX/ptX}
\liminf_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(X_T)]}{{\mathbb{E}}[\tilde{X}_T^p/p]} \geq (1-\epsilon) \liminf_{T\rightarrow \infty}\frac{{\mathbb{E}}[(X_T+a)^p]}{{\mathbb{E}}[\tilde{X}^p_T]}.$$ Now take $\hat{Y}$ in to be $Y$. Then, as ${\lim_{T\rightarrow \infty}}{\mathbb{E}}[Y_T]=0$, and ${\mathbb{E}}[\tilde{X}^p_T] = {\mathbb{E}}[\tilde{Y}_T^q]^{1-p}$ by another application of Lemma \[lem: power dual\], it follows from and Lemma \[lem: ratio Y\] that $$\liminf_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(X_T)]}{{\mathbb{E}}[\tilde{X}_T^p/p]} \geq (1-\epsilon) \liminf_{T\rightarrow \infty}\frac{{\mathbb{E}}[(X_T+a)^p]}{{\mathbb{E}}[\tilde{X}^p_T]}\geq (1-\epsilon) \liminf_{T\rightarrow \infty} \frac{{\mathbb{E}}[Y^q_T]^{1-p}}{{\mathbb{E}}[\tilde{Y}_T^q]^{1-p}} = 1-\epsilon.$$ As $\epsilon$ was arbitrary, this proves and in turn Proposition \[prop: limit value\].
Proof of the main result for $p=0$
==================================
If $p=0$, the isoelastic utility $\tilde{U}$ is logarithmic and – compared to the power case $p \neq 0$ – different arguments are needed to establish the main result. In this case, the convergence of the ratio of marginal utilities implies the following:
\[lem: inv U\] Suppose holds. Then, for any $\epsilon>0$, there exists a sufficiently large $M_\epsilon$ such that $$(1-\epsilon) (a-b) \leq \log \frac{U^{-1}(a)}{U^{-1}(b)} \leq (1+\epsilon) (a-b), \quad \mbox{for $a\geq b\geq M_\epsilon$}.$$
For any $\epsilon \in (0,1)$, the same argument as in Item $i)$ of Lemma \[lem:rv\] yields the existence of $M_\epsilon$ such that $$\label{eq: U bd log}
(1-\epsilon) \log x + B_\epsilon \leq U(x) \leq (1+\epsilon) \log x +A_\epsilon, \quad \mbox{for } x\geq M_\epsilon$$ and constants $A_\epsilon, B_\epsilon$. The lower bound implies $\lim_{x\uparrow \infty} U(x) =\infty$. Set $y= U(x)$; then, $x\uparrow \infty$ as $y\uparrow \infty$. Combined with the convergence of ratio of marginal utilities , this yields $$\frac{(U^{-1})'(y)}{U^{-1}(y)} = \frac{1}{x U'(x)} \rightarrow 1, \quad \mbox{as $y\uparrow \infty$}.$$ Therefore, $1-\epsilon \leq (\log U^{-1}(y))' \leq 1+\epsilon$ for $y\geq M_\epsilon$, after enlarging $M_\epsilon$ if necessary. The assertion then follows from integrating these inequalities over the interval $(b, a)$ for $a\geq b\geq M_\epsilon$.
In the long run, the generic expected utility diverges, both for the corresponding optimal portfolio and for its isoelastic counterpart:
\[lem: log value\] Let Assumptions \[ass: growth\] - \[ass: wellposedness\] hold. Then $${\lim_{T\rightarrow \infty}}{\mathbb{E}}[U(X^T_T)] = {\lim_{T\rightarrow \infty}}{\mathbb{E}}[U(\tilde{X}^T_T)] =\infty.$$
Once again, the superscript $T$ for $X^T$, $\tilde{X}^T$, and $\tilde{Y}^T$ is omitted throughout this proof. As ${\mathbb{E}}[U(X_T)]\geq {\mathbb{E}}[U(\tilde{X}_T)]$ by optimality of $X_T$ for the generic utility $U$, the first convergence is implied by the second one. To prove the second convergence, first note that the lower bound in yields $$\label{eq: Utx log}
\lim_{T\rightarrow\infty}
{\mathbb{E}}[U(\tilde{X}_T) \, {\mathbb{I}}_{\{\tilde{X}_T \geq M_\epsilon\}}] \geq (1-\epsilon) \,{\mathbb{E}}[\log(\tilde{X}_T) \,{\mathbb{I}}_{\{\tilde{X}_T \geq M_\epsilon\}}] + B_\epsilon {\mathbb{P}}(\tilde{X}_T \geq M_\epsilon) = \infty
.$$ Here, the last convergence holds because the optimality of $\tilde{X}_T$ for the logarithmic utility implies ${\lim_{T\rightarrow \infty}}{\mathbb{E}}[\log(\tilde{X}_T)] \geq {\lim_{T\rightarrow \infty}}{\mathbb{E}}[\log(S^0_T)] =\infty$ by $\lim_{T\rightarrow \infty} S^0_T = \infty$. Now, turn to the contribution of the low wealth levels $\{\tilde{X}_T <M_\epsilon\}$. To this end, Assumption guarantees the existence of $C_M$ such that $U(x) \geq C_M x^{-1}$ for all $x<M_\epsilon$. Then, $${\mathbb{E}}[U(\tilde{X}_T ) \, {\mathbb{I}}_{\{\tilde{X}_T <M_\epsilon\}}] \geq C_M \,{\mathbb{E}}[(\tilde{X}_T)^{-1} \, {\mathbb{I}}_{\{\tilde{X}_T <M_\epsilon\}}].$$ Recall from Section \[sec:dual\] that $\tilde{X}_T^{-1} = \tilde{y}^T\tilde{Y}_T$ and $\tilde{y}^T \equiv 1$.[^16] In view of , the term on the right-hand side therefore converges to zero as $T\rightarrow \infty$, so that $\liminf_{T\rightarrow \infty}[U(\tilde{X}_T ) \, {\mathbb{I}}_{\{\tilde{X}_T < M_\epsilon\}}]\geq 0$. Combined with , this confirms the second (and in turn the first) convergence in the assertion.
Set $a_T = {\mathbb{E}}[U(X^T_T)]$ and $b_T = {\mathbb{E}}[U(\tilde{X}^T_T)]$. As we have seen above, $a_T \geq b_T$ and both utilities tend to infinity as $T\rightarrow \infty$. In view of Lemma \[lem: inv U\], the convergence $U^{-1}(a_T)/U^{-1}(b_T) \to 1$ of the ratio of corresponding certainty equivalents is thereby equivalent to the *difference* $a_T-b_T$ of utilities vanishing in the long run. This is in contrast to the power case $p \neq 0$, where the convergence of the ratio of certainty equivalents was found to be equivalent to the convergence of the *ratio* $a_T/b_T$ of utilities in Lemma \[lem:equiv\]. As a result, different estimates are needed in the case $p=0$. More specifically, the proof of the main result is based on the following long-run asymptotics in this case, whose technical proof is deferred to Section \[sec:proofp0\].
\[prop: ratio X\] Let $p=0$ and suppose Assumptions \[ass: growth\] - \[ass: wellposedness\] hold. Then:
1. ${\mathbb{P}}-{\lim_{T\rightarrow \infty}}\tilde{X}^T_T =\infty$;
2. ${\mathbb{P}}-{\lim_{T\rightarrow \infty}}X^T_T =\infty$;
3. ${\mathbb{P}}-{\lim_{T\rightarrow \infty}}X^T_T /\tilde{X}^T_T =1$.
Here, ${\mathbb{P}}-{\lim_{T\rightarrow \infty}}$ denotes convergence in ${\mathbb{P}}$-probability.
With Proposition \[prop: ratio X\] at hand, we can now complete the proof of the main result also in the case $p=0$:
The superscript $T$ for $X^T$, $\tilde{X}^T$, and $\tilde{Y}^T$ is again omitted throughout this proof. As discussed above, due to Lemmas \[lem: inv U\] and \[lem: log value\], it suffices to prove $${\lim_{T\rightarrow \infty}}{\mathbb{E}}[U(X_T ) - U(\tilde{X}_T )] =0.$$ As ${\mathbb{E}}[U(X_T)] \geq {\mathbb{E}}[U(\tilde{X}_T )]$, we only need to show $$\label{eq: Ux-Utx}
\limsup_{T\rightarrow \infty} {\mathbb{E}}[U(X_T ) - U(\tilde{X}_T )]\leq 0.$$ For any $\epsilon>0$, the convergence of the ratio of marginal utilities implies that there exists a sufficiently large $M_\epsilon$ such that $x U'(x) \leq 1+\epsilon$ for $x\geq M_\epsilon$. As a result: $$U(x) - U(\tilde{x}) = \int_{\tilde{x}}^x \frac{1}{y} y U'(y)\, dy \leq (1+\epsilon) (\log x - \log\tilde{x}), \quad \mbox{for } x\geq \tilde{x}\geq M_\epsilon.$$ Choosing $x= X_T$ and $\tilde{x}= {\widetilde{X}}_T$ in the previous inequality leads to $$\label{eq: diff U est>M}
\begin{split}
&{\mathbb{E}}{\left[{\left(U(X_T) - U(\tilde{X}_T)\right)} \,{\mathbb{I}}_{\{\tilde{X}_T \geq M_\epsilon, X_T \geq M_\epsilon\}}\right]} \\
\leq\ &{\mathbb{E}}{\left[{\left(U(X_T) - U(\tilde{X}_T)\right)} \,{\mathbb{I}}_{\{X_T \geq \tilde{X}_T \geq M_\epsilon\}}\right]}\\
\leq\ &(1+\epsilon)\,{\mathbb{E}}{\left[\log \frac{X_T}{\tilde{X}_T} \,{\mathbb{I}}_{\{X_T \geq \tilde{X}_T\geq M_\epsilon\}}\right]},
\end{split}$$ where the first inequality holds because $U(X_T ) < U(\tilde{X}_T )$ when $X_T < \tilde{X}_T $. Now, observe that $$0 \leq {\mathbb{E}}{\left[\log \frac{X_T }{\tilde{X}_T } \, {\mathbb{I}}_{\{X_T > \tilde{X}_T \geq M_\epsilon\}}\right]} \leq {\mathbb{E}}{\left[\log \frac{X_T }{\tilde{X}_T } \, {\mathbb{I}}_{\{X_T > \tilde{X}_T\}}\right]} = {\mathbb{E}}{\left[\log {\left(\frac{X_T }{\tilde{X}_T } \vee 1\right)}\right]},$$ and, moreover, $$\begin{aligned}
{\mathbb{E}}{\left[\exp \log {\left(\frac{X_T }{\tilde{X}_T } \vee 1\right)}\right]} = {\mathbb{E}}{\left[\frac{X_T }{\tilde{X}_T } \vee 1\right]} \leq {\mathbb{E}}{\left[\frac{X_T }{\tilde{X}_T }\right]} + 1 \leq 2, \quad \mbox{for all } T>0,
\end{aligned}$$ where the last inequality holds due to the numeraire property of $\tilde{X}_T$, that is, ${\mathbb{E}}[\hat{X}_T/\tilde{X}_T] \leq 1$ for any admissible $\hat{X}$ (cf. with $p=0$). As a result, de la Vallee-Poussin’s criterion as in [@Shiryaev Lemma 3] implies that the family $\log{\left(\frac{X_T}{\tilde{X}_T } \vee 1\right)}$ is uniformly integrable in $T$. Together with Proposition \[prop: ratio X\] iii), this yields $$\label{eq: lim log ratio}
{\lim_{T\rightarrow \infty}}{\mathbb{E}}{\left[\log \frac{X_T }{\tilde{X}_T } \, {\mathbb{I}}_{\{X_T > \tilde{X}_T \geq M_\epsilon\}}\right]} =0.$$ In view of , it follows that $$\limsup_{T\rightarrow \infty} {\mathbb{E}}{\left[{\left(U(X_T ) - U(\tilde{X}_T)\right)} \, {\mathbb{I}}_{\{\tilde{X}_T \geq M_\epsilon, X_T \geq M_\epsilon\}}\right]} \leq 0.$$ In the next two paragraphs, we will show $$\begin{aligned}
& \limsup_{T\rightarrow \infty} {\mathbb{E}}{\left[U(X_T) \,{\mathbb{I}}_{\{X_T \leq M_\epsilon \text{ or } {\widetilde{X}}_T \leq M_\epsilon\}}\right]} \leq 0, \label{eq: diff U est 1}\\
& \liminf_{T\rightarrow \infty} {\mathbb{E}}{\left[U(\tilde{X}_T)\,{\mathbb{I}}_{\{X_T \leq M_\epsilon \text{ or } \tilde{X}_T \leq M_\epsilon\}}\right]} \geq 0. \label{eq: diff U est 2}
\end{aligned}$$ Combining these three inequalities then yields , and in turn completes the proof of Theorem \[thm:mt\] in the case $p=0$.
To establish , note that Proposition \[prop: ratio X\] ii) gives $$\limsup_{T\rightarrow \infty} {\mathbb{E}}{\left[U(X_T) \,{\mathbb{I}}_{\{X_T \leq M_\epsilon\}}\right]}\leq U(M_\epsilon) \limsup_{T\rightarrow \infty} {\mathbb{P}}(X_T \leq M_\epsilon) =0.$$ On the other hand, the upper bound on the generic utility $U$ in implies $$\begin{aligned}
& {\mathbb{E}}[U(X_T ) \, {\mathbb{I}}_{\{X_T > M_\epsilon, \tilde{X}_T \leq M_\epsilon\}}]\\
\leq\ & (1+\epsilon) \,{\mathbb{E}}[\log(X_T ) \, {\mathbb{I}}_{\{X_T > M_\epsilon, \tilde{X}_T \leq M_\epsilon\}}] + A_\epsilon {\mathbb{P}}(\tilde{X}_T \leq M_\epsilon)\\
\leq\ & (1+\epsilon)\, {\mathbb{E}}{\left[\log \frac{X_T }{\tilde{X}_T } \, {\mathbb{I}}_{\{X_T > \tilde{X}_T\}}\right]} + (1+\epsilon)\,{\mathbb{E}}[\log(\tilde{X}_T ) \, {\mathbb{I}}_{\{X \geq M_\epsilon, \tilde{X}_T \leq\ M_\epsilon\}}]\\
&\qquad + A_\epsilon {\mathbb{P}}(\tilde{X}_T \leq M_\epsilon)\\
\leq\ & (1+\epsilon)\, {\mathbb{E}}{\left[\log \frac{X_T }{\tilde{X}_T } \, {\mathbb{I}}_{\{X_T > \tilde{X}_T\}}\right]} + ((1+\epsilon) \log M_\epsilon + A_\epsilon) \,{\mathbb{P}}(\tilde{X}_T \leq M_\epsilon)\\
=\ & (1+\epsilon) {\mathbb{E}}{\left[\log \frac{X_T }{\tilde{X}_T } \vee 1\right]} + ((1+\epsilon) \log M_\epsilon + A_\epsilon) \,{\mathbb{P}}(\tilde{X}_T \leq M_\epsilon).
\end{aligned}$$ Now, Proposition \[prop: ratio X\] iii) and the uniform integrability established in the derivation of show that the first term on the right-hand side converges to zero as $T \to \infty$. Likewise, by Proposition \[prop: ratio X\] i), the second term also tends to zero as the horizon grows, confirming .
To prove , note that $$\liminf_{T\rightarrow \infty} {\mathbb{E}}[U(\tilde{X}_T ) \, {\mathbb{I}}_{\{\tilde{X}_T >M_\epsilon, X_T \leq M_\epsilon\}}] \geq U(M_\epsilon) {\lim_{T\rightarrow \infty}}{\mathbb{P}}(\tilde{X}_T > M_\epsilon, X_T \leq M_\epsilon) =0$$ by Proposition \[prop: ratio X\] ii). It therefore suffices to show $$\label{eq: Utx log <M}
\liminf_{T\rightarrow \infty} {\mathbb{E}}[U(\tilde{X}_T ) \, {\mathbb{I}}_{\{\tilde{X}_T \leq M_\epsilon\}}]\geq 0.$$ To this end, yields the existence of $C_M$ such that $U(x) \geq C_M x^{-1}$ for $x\leq M_\epsilon$. Together with the first-order condition $\tilde{X}_T^{-1}=\tilde{y}^T\tilde{Y}_T$ with $\tilde{y}^T\equiv 1$ and , it follows that $$\begin{aligned}
{\mathbb{E}}[U(\tilde{X}_T ) \, {\mathbb{I}}_{\{\tilde{X}_T \leq M_\epsilon\}}] \geq C_M \, {\mathbb{E}}[ (\tilde{X}_T)^{-1} \, {\mathbb{I}}_{\{\tilde{X}_T \leq M_\epsilon\}}] = C_M {\mathbb{E}}[\tilde{Y}_T \, {\mathbb{I}}_{\{\tilde{X}_T \leq M_\epsilon\}}] \rightarrow 0,
\end{aligned}$$ as $T\rightarrow \infty$. Therefore and in turn is confirmed in both cases, completing the proof of Theorem \[thm:mt\].
Proof of Proposition \[prop: ratio X\] {#sec:proofp0}
--------------------------------------
This section concludes the proof of Theorem \[thm:mt\] in the case $p=0$, by establishing the auxiliary Proposition \[prop: ratio X\]. In all the following proofs, the superscript $T$ in $\tilde{X}^T$, $X^T$, $Y^T$, and $\tilde{Y}^T$ is omitted to ease notation. We begin with the analogue of Lemma \[lem: ratio Y p>0\]:
\[eq: ratio y p=0\] Let $p=0$ and suppose Assumptions \[ass: growth\] - \[ass: wellposedness\] hold. Then $$\lim_{T\rightarrow \infty} \frac{y^T}{\tilde{y}^T} =1,$$ where $\tilde{y}^T\equiv 1$ as we have seen before (cf. Footnote 15).
The proof follows the argument in Lemma \[lem: ratio Y p>0\], where many estimates are simplified when $p=0$. Indeed, on the one hand, the same estimate as in yields $$1 \leq \frac{1+\epsilon}{y^T} \,{\mathbb{E}}{\left[\frac{X_T}{X_T} \, {\mathbb{I}}_{\{X_T \geq M_\epsilon\}}\right]} + {\mathbb{E}}[X_T Y_T \, {\mathbb{I}}_{\{X_T <M_\epsilon\}}] \leq \frac{1+\epsilon}{y^T} + {\mathbb{E}}[X_T Y_T \, {\mathbb{I}}_{\{X_T <M_\epsilon\}}].$$ Here, the term ${\mathbb{E}}[X_T Y_T \, {\mathbb{I}}_{\{X_T <M_\epsilon\}}]$ vanishes as $T\rightarrow \infty$ due to , so that $$\label{eq: ratio y p=0 liminf}
\frac{1}{1+\epsilon} \leq \liminf_{T\rightarrow \infty} \frac{1}{y^T}.$$ On the other hand, the same argument that leads to also yields $$1= {\mathbb{E}}{\left[Y_T I(y^TY_T)\, {\mathbb{I}}_{\{y^T Y_T\leq \delta_\epsilon\}}\right]} + {\mathbb{E}}{\left[Y_T X_T \, {\mathbb{I}}_{\{y^T Y_T> \delta_\epsilon\}}\right]},$$ where the second term on the right-hand side vanishes as $T\rightarrow \infty$ by the same reasoning as after . Concerning the first term, for $p=0$ shows that $I(y)\geq (1-\epsilon)y^{-1}$ for any $y<\delta_\epsilon$ and some sufficiently small $\delta_\epsilon$. Therefore: $${\mathbb{E}}{\left[Y_T I(y^T Y_T) \, {\mathbb{I}}_{\{y^T Y_T \leq \delta_\epsilon\}}\right]}
\geq (1-\epsilon) \frac{{\mathbb{P}}(y^T Y_T\leq \delta_\epsilon)}{y^T}
= \frac{1-\epsilon}{y^T} - \frac{1-\epsilon}{y^T} {\mathbb{P}}(y^T Y_T >\delta_\epsilon).$$ Here the second term also tends to zero because $$\frac{1}{y^T}{\mathbb{P}}(y^TY_T >\delta_\epsilon) = {\mathbb{E}}{\left[\frac{Y_T}{y^T Y_T} {\mathbb{I}}_{\{y^T Y_T >\delta_\epsilon\}}\right]} \leq \delta_{\epsilon}^{-1} {\mathbb{E}}[Y_T]\rightarrow 0,$$ due to . Combining the above estimates, we obtain $$\label{eq: ratio y p=0 limsup}
\frac{1}{1-\epsilon}\geq \limsup_{T\rightarrow \infty} \frac{1}{y^T}.$$ Thus, the assertion follows by combining and because $\varepsilon$ was arbitrary.
Using the previous results, we can now verify the first two items of Proposition \[prop: ratio X\].
\[lem: X tX ub\] Let $p=0$ and suppose Assumptions \[ass: growth\] - \[ass: wellposedness\] hold. Then: $${\mathbb{P}}-{\lim_{T\rightarrow \infty}}\tilde{X}^T_T = \infty \quad \mbox{and} \quad {\mathbb{P}}-{\lim_{T\rightarrow \infty}}X^T_T = \infty.$$
Recall the numeraire property of the log-optimal portfolio $\tilde{X}$: ${\mathbb{E}}[\hat{X}_T / \tilde{X}_T] \leq 1$ for any $T$ and any admissible payoff $\hat{X}$. The first part of the assertion in turn follows verbatim as in the derivation of , where ${\mathbb{P}}^T$ is replaced by ${\mathbb{P}}$. As for the second part of the assertion, for any $N>0$, there exists $y_N$ such that $I(y) > N$ for $y\leq y_N$. Here, $y_N$ is chosen sufficiently small so that $-\partial V(y)$ is single valued and is denoted by $I(y)$ for $y\leq y_N$. For the chosen $N$, $$\begin{aligned}
{\mathbb{P}}(X_T \leq N) \leq {\mathbb{P}}(X_T \leq N, y^T Y_T \leq y_N) + {\mathbb{P}}(y^T Y_T >y_N).
\end{aligned}$$ The second term on the right-hand side vanishes as $T\rightarrow \infty$, due to , ${\lim_{T\rightarrow \infty}}y^T/\tilde{y}^T = 1$, and $\tilde{y}^T \equiv 1$ (cf. Lemma \[eq: ratio y p=0\]). The first term is identically zero, because $X_T = I(y^T Y_T) >N$ on $\{y^T Y_T \leq y_N\}$. Therefore, the second part of the assertion follows.
To complete the proof of Proposition \[prop: ratio X\], it remains to verify Item iii). To this end, some auxiliary results are established first.
\[lem: lim XY\] Suppose Assumptions \[ass: growth\] - \[ass: wellposedness\] hold for $p=0$. Then: $${\mathbb{P}}-{\lim_{T\rightarrow \infty}}X^T_T Y^T_T =1.$$
For any $\epsilon>0$, the convergence of the ratio of marginal utilities implies the existence of $M_\epsilon$ such that $U(x)$ is differentiable beyond $M_\epsilon$ and, moreover, $xU'(x) \geq 1-\epsilon/2$ for $x\geq M_\epsilon$. For such a $M_\epsilon$, $$\label{eq: p XY<1-e}
{\mathbb{P}}(X_TY_T \leq 1-\epsilon) \leq {\mathbb{P}}(X_T Y_T \leq 1-\epsilon, X_T \geq M_\epsilon) + {\mathbb{P}}(X_T < M_\epsilon).$$ In view of Lemma \[lem: X tX ub\], the second term on the right-hand side vanishes as $T\rightarrow \infty$. To estimate the first term, recall that $Y_T = U'(X_T ) /y^T$ on $\{X_T \geq M_\epsilon\}$. As a result, for $T \geq T_*$, $$\begin{aligned}
{\mathbb{P}}(X_T Y_T \leq 1-\epsilon, X_T \geq M_\epsilon) &= {\mathbb{P}}{\left(X_T U'(X_T ) \leq y^T (1-\epsilon), X_T \geq M_\epsilon\right)}\\
&\leq {\mathbb{P}}(1-\epsilon/2\leq y^T (1-\epsilon)),
\end{aligned}$$ where the inequality holds because $X_T U'(X_T ) \geq 1-\epsilon/2$ on $\{X_T \geq M_\epsilon\}$. As ${\lim_{T\rightarrow \infty}}y^T=1$ by Lemma \[eq: ratio y p=0\], the above estimates imply $${\lim_{T\rightarrow \infty}}{\mathbb{P}}(X_T Y_T \leq 1-\epsilon, X_T \geq M_\epsilon)=0.$$ Coming back to , this gives ${\lim_{T\rightarrow \infty}}{\mathbb{P}}(X_T Y_T \leq 1-\epsilon)=0$. Along the same lines, ${\lim_{T\rightarrow \infty}}{\mathbb{P}}(X_T Y_T \geq 1+\epsilon) =0$ follows, completing the proof.
\[cor: XY conv\] Let $p=0$ and suppose Assumptions \[ass: growth\] - \[ass: wellposedness\] hold. Then: $${\lim_{T\rightarrow \infty}}{\mathbb{E}}[|X^T_T Y^T_T-1|] =0.$$
By the convergence in probability established in the previous lemma and ${\mathbb{E}}[X^T_TY^T_T]=1$, this follows from Scheffe’s lemma [@williams.91 5.5.10].
Set $r_T = X^T_T /\tilde{X}^T_T$. The following estimate is a key to prove Proposition \[prop: ratio X\] iii).
\[lem: r est\] Let $p=0$ and suppose Assumptions \[ass: growth\] - \[ass: wellposedness\] hold. Then: $${\lim_{T\rightarrow \infty}}{\mathbb{E}}{\left[\left|1-\frac{X^T_T Y^T_T}{r_T}\right| |r_T-1|\right]}=0.$$
Recall the first-order condition $\tilde{y}^T \tilde{Y}_T = \tilde{X}^{-1}_T$. As $1= {\mathbb{E}}[\tilde{Y}_T \tilde{X}_T] \geq {\mathbb{E}}[\tilde{Y}_T X_T]$, it follows that $${\mathbb{E}}[\tilde{X}^{-1}_T (X_T - \tilde{X}_T)]\leq 0.$$ Similarly, $1={\mathbb{E}}[Y_T X_T] \geq {\mathbb{E}}[Y_T \tilde{X}_T]$ yields $${\mathbb{E}}[Y_T (\tilde{X}_T - X_T)]\leq 0.$$ Summing up the previous two inequalities and using the definition $r_T = X^T_T /\tilde{X}^T_T$ gives $$\label{eq: r^T-1 exp}
0\geq {\mathbb{E}}[(\tilde{X}^{-1}_T - Y_T) (X_T - \tilde{X}_T)] = {\mathbb{E}}{\left[{\left(1-\frac{X_T Y_T}{r_T}\right)}(r_T-1)\right]}.$$ Observe that $(1- X_T Y_T /r_T)(r_T-1)\leq 0$ when $X_T Y_T \leq r_T \leq 1$ or $1\leq r_T \leq X_T Y_T$. Therefore, $$\begin{aligned}
& {\mathbb{E}}{\left[{\left({\left(1-\frac{X_T Y_T}{r_T}\right)}(r_T-1)\right)}_-\right]}\\
\leq\ & {\mathbb{E}}{\left[{\left(1-\frac{X_T Y_T}{r_T}\right)}(1-r_T) \, {\mathbb{I}}_{\{X_T Y_T \leq r_T \leq 1 \, \text{or} \, 1\leq r_T \leq X_T Y_T\}}\right]}.\end{aligned}$$ As ${\mathbb{E}}[((1-X_TY_T/r_T)(r_T-1))_+] \leq {\mathbb{E}}[((1-X_TY_T/r_T)(r_T-1))_-]$ by , it follows that $$\label{eq: abs XY}
\begin{split}
&{\mathbb{E}}{\left[\left|1-\frac{X_T Y_T}{r_T}\right| |r_T-1|\right]}\\
\leq\ & 2 \, {\mathbb{E}}{\left[{\left(1-\frac{X_T Y_T}{r_T}\right)}(1-r_T) \, {\mathbb{I}}_{\{X_T Y_T \leq r_T \leq 1 \, \text{or} \, 1 \leq r_T \leq X_T Y_T\}}\right]}.
\end{split}$$ Now, estimate the expectation on the right-hand side. On the set $\{X_T Y_T\leq r_T \leq 1\}$ we have $(1-X_T Y_T/r_T)(1-r_T) \leq (1- X_T Y_T)^2$, so that $$\begin{split}
&{\mathbb{E}}{\left[{\left(1-\frac{X_TY_T}{r_T}\right)}(1-r_T)\, {\mathbb{I}}_{\{X_T Y_T \leq r_T \leq 1\}}\right]}\\
\leq\ &{\mathbb{E}}{\left[(1- X_T Y_T)^2 \, {\mathbb{I}}_{\{X_T Y_T \leq 1\}}\right]}\\
\leq\ &{\mathbb{E}}[(1-X_T Y_T)^2 \, {\mathbb{I}}_{\{X_T Y_T\leq 1-\epsilon\}}] + {\mathbb{E}}[(1-X_T Y_T)^2 \, {\mathbb{I}}_{\{1-\epsilon < X_T Y_T \leq 1\}}]\\
\leq\ &{\mathbb{P}}(X_T Y_T \leq 1-\epsilon) + \epsilon^2, \quad \mbox{for any $\epsilon>0$}.
\end{split}$$ Lemma \[lem: lim XY\] shows that the first term on the right vanishes as $T\rightarrow \infty$. As $\epsilon$ was chosen arbitrarily, this yields $$\label{eq: est XY<r<1}
{\lim_{T\rightarrow \infty}}{\mathbb{E}}{\left[{\left(1-\frac{X_TY_T}{r_T}\right)}(1-r_T)\, {\mathbb{I}}_{\{X_T Y_T \leq r_T \leq 1\}}\right]} =0.$$ On $\{1\leq r_T \leq X_T Y_T\}$ we have $X_T Y_T/r_T + r_T \geq 2$ and in turn $$(1- X_T Y_T/ r_T)(1-r_T)\leq X_T Y_T -1.$$ As a consequence: $$\begin{aligned}
&{\mathbb{E}}{\left[{\left(1-\frac{X_T Y_T}{r_T}\right)}(1-r_T) \, {\mathbb{I}}_{\{1\leq r_T \leq X_T Y_T\}}\right]}\\
=\ &{\mathbb{E}}{\left[{\left(1-\frac{X_T Y_T}{r_T}\right)}(1-r_T)\, {\mathbb{I}}_{\{1\leq r_T \leq X_T Y_T, X_T Y_T \leq 1+\epsilon\}}\right]}\\
&+ {\mathbb{E}}{\left[{\left(1-\frac{X_T Y_T}{r_T}\right)}(1-r_T)\, {\mathbb{I}}_{\{1\leq r_T \leq X_T Y_T, X_T Y_T > 1+\epsilon\}}\right]}\\
\leq\ &{\mathbb{E}}{\left[(1-X_T Y_T)^2\, {\mathbb{I}}_{\{1\leq r_T \leq X_T Y_T, X_T Y_T \leq 1+\epsilon\}}\right]} + {\mathbb{E}}{\left[(X_T Y_T -1) \,{\mathbb{I}}_{\{X_T Y_T > 1+\epsilon\}}\right]}\\
\leq\ &\epsilon^2 + {\mathbb{E}}[|X_T Y_T-1|].\end{aligned}$$ As $\epsilon$ was chosen arbitrarily and the second term on the right-hand side converges to $0$ by Corollary \[cor: XY conv\], we obtain $$\label{eq: est 1<r<XY}
{\lim_{T\rightarrow \infty}}{\mathbb{E}}{\left[{\left(1-\frac{X_T Y_T}{r_T}\right)}(1-r_T) \, {\mathbb{I}}_{\{1\leq r_T \leq X_T Y_T\}}\right]} =0.$$ Together, , and yield the assertion.
We are now ready to complete the proof of Proposition \[prop: ratio X\] iii).
Lemma \[lem: r est\] implies $${\mathbb{P}}-{\lim_{T\rightarrow \infty}}(1- X_T Y_T/r_T) (1-r_T) =0.$$ Combined with Lemma \[lem: lim XY\], this yields the assertion ${\mathbb{P}}-{\lim_{T\rightarrow \infty}}r_T=1$.
Analysis of the Counterexample {#sec:analysis}
==============================
In this section, we provide a detailed analysis of the counterexample from Section \[sec:example\]. Recall that $U(x) = x^p/p$, $p<0$ for sufficiently large $x$ and $U(x) = x^{p^*}/{p^*}$, $p^*<p-1$, for $x\leq 1$. In what follows, we will show that if is satisfied, then $$\label{eq: ratio exp}
\lim_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(\tilde{X}_T)]}{{\mathbb{E}}[\tilde{U}(\tilde{X}_T)]} =\infty \quad \text{and} \quad \limsup_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(X_T)]}{{\mathbb{E}}[\tilde{U}(\tilde{X}_T)]}\leq 1.$$ Hence $$\label{eq: ratio conv 0}
\lim_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(\tilde{X}_T)]}{{\mathbb{E}}[U(X_T)]} =\infty,$$ which confirms as $U$ is regularly varying at infinity (cf. Lemma 3.1). Indeed, set $a_T = U^{-1}({\mathbb{E}}[U(\tilde{X}_T)])$ and $b_T = U^{-1}({\mathbb{E}}[U(X_T)])$. If $\limsup_{T\rightarrow \infty} {\mathbb{E}}[U(\tilde{X}_T)]$ is bounded away from zero, then $\limsup_{T\rightarrow \infty}a_T<\infty$. However, $U(\infty) =0$ and $\lim_{T\rightarrow \infty} S^0_T=\infty$ yield $0\geq {\mathbb{E}}[U(X_T)] \geq {\mathbb{E}}[U(S^0_T)]\rightarrow 0$ as $T\rightarrow \infty$, hence $\lim_{T\rightarrow \infty}b_T =\infty$. Therefore $\lim_{T\rightarrow \infty} a_T/b_T =0$ holds. When $\lim_{T\rightarrow \infty} {\mathbb{E}}[U(\tilde{X}_T)] =0$, then $\lim_{T\rightarrow \infty} a_T=\infty$. If there exists $\delta>0$ such that $\liminf_{T\rightarrow \infty}a_T/b_T \geq \delta$, then $$\limsup_{T\rightarrow \infty}\frac{U(a_T)}{U(b_T)} \leq \limsup_{T\rightarrow \infty}\frac{U(\delta b_T )}{U(b_T)}= \delta^p<\infty$$ because $U \leq 0$, which contradicts .
To prove the first convergence in , it suffices to show $$\label{eq: tXp*/tXp}
\lim_{T\rightarrow \infty} \frac{{\mathbb{E}}[\tilde{X}^{p^*}_T \,{\mathbb{I}}_{\{\tilde{X}_T\leq 1\}}]}{{\mathbb{E}}[\tilde{X}^p_T]} = \infty.$$ In the Black-Scholes model, the optimal risky weight for power utility $\tilde{U}(x)=x^{p}/p$ is $\tilde{\pi}= \frac{1}{1-p} \frac{\mu}{\sigma^2}$. The associated wealth process starting from unit initial capital is $$\label{eq: tX}
\tilde{X}_T = \exp{\left({\left(r+ \frac{1-2p}{2(1-p)^2} \frac{\mu^2}{\sigma^2}\right)}T + \frac{1}{1-p} \frac{\mu}{\sigma} W_T\right)}.$$ Straightforward calculations show $$\label{eq: txp*}
\frac{\tilde{X}^{p^*}_T}{{\mathbb{E}}[\tilde{X}_T^p]} = \exp{\left((p^*-p) rT + {\left(\frac{p^*(1-2p+p^*)}{2(1-p)^2} + \frac12 q\right)} \frac{\mu^2}{\sigma^2} T\right)} \mathcal{E}{\left(\frac{p^*}{1-p} \frac{\mu}{\sigma} W_T\right)},$$ where $q=p/(p-1)$. Define a new probability measure ${\mathbb{Q}}^*$ via $$\frac{d{\mathbb{Q}}^*}{d{\mathbb{P}}}|_{{\mathcal{F}}_T} = \mathcal{E}{\left(\frac{p^*}{1-p} \frac{\mu}{\sigma} W_T\right)}.$$ Then, $W^*_t = W_t - \frac{p^*}{1-p}\frac{\mu}{\sigma} t$ is a ${\mathbb{Q}}^*$-Brownian motion on $[0,T]$. As a result, $$\begin{aligned}
&{\mathbb{E}}^{{\mathbb{P}}}{\left[\mathcal{E}{\left(\frac{p^*}{1-p} \frac{\mu}{\sigma} W_T\right)} \,{\mathbb{I}}_{\{\tilde{X}_T\leq 1\}}\right]}\\
=\ & {\mathbb{Q}}^*{\left(\frac{1}{1-p}\frac{\mu}{\sigma} W_T \leq - {\left(r+ \frac{1-2p}{2(1-p)^2} \frac{\mu^2}{\sigma^2}\right)} T\right)}\\
=\ & {\mathbb{Q}}^*{\left(\frac{1}{1-p}\frac{\mu}{\sigma} W^*_T \leq -rT - \frac{1-2p+2p^*}{2(1-p)^2} \frac{\mu^2}{\sigma^2}T\right)}\\
= \ & {\mathbb{Q}}^*{\left(\frac{1}{1-p}\frac{\mu}{\sigma} \frac{W^*_T}{\sqrt{T}} \leq -r\sqrt{T} - \frac{1-2p+2p^*}{2(1-p)^2} \frac{\mu^2}{\sigma^2}\sqrt{T}\right)}.\end{aligned}$$ Observe that $$-r-\frac{1-2p+2p^*}{2(1-p)^2}\frac{\mu^2}{\sigma^2} > -r + \frac{1}{2(1-p)^2} \frac{\mu^2}{\sigma^2} >0,$$ where the first inequality follows from $p^*<p-1$ and the second inequality holds due to $\mu^2/\sigma^2 > 2(1-p)^2 r$ in . As a consequence: $$\label{eq: Q* lim}
\begin{split}
&\lim_{T\rightarrow \infty} {\mathbb{E}}^{{\mathbb{P}}}{\left[\mathcal{E}{\left(\frac{p^*}{1-p} \frac{\mu}{\sigma} W_T\right)} \,{\mathbb{I}}_{\tilde{X}_T\leq 1}\right]}\\
=\ & \lim_{T\rightarrow \infty} {\mathbb{Q}}^*{\left(\frac{1}{1-p}\frac{\mu}{\sigma} \frac{W^*_T}{\sqrt{T}} \leq -r\sqrt{T} - \frac{1-2p+2p^*}{2(1-p)^2} \frac{\mu^2}{\sigma^2}\sqrt{T}\right)}= 1.
\end{split}$$ For the exponential factor on the right-hand side of , note that $$\begin{aligned}
\frac{p^*(1-2p + p^*)}{2(1-p)^2} + \frac12 q >0\end{aligned}$$ because $p^*(1-2p+p^*)$ is strictly decreasing in $p^*$ when $p^*<p-1$. If is satisfied, it follows that $$\label{cond2}
(p^*-p) r + {\left(\frac{p^*(1-2p+p^*)}{2(1-p)^2} + \frac12 q\right)}\frac{\mu^2}{\sigma^2} = (p^*-p) {\left(r+\frac{p^*+1-p}{2(1-p)^2} \frac{\mu^2}{\sigma^2}\right)}>0,$$ so that the exponential term on the right-hand side of diverges as $T\rightarrow \infty$. Therefore, is obtained after taking into account .
Now, consider the second convergence in . Let us first prove $$\label{eq: ratio exp small x}
\lim_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(X_T) \, {\mathbb{I}}_{\{X_T\leq M\}}]}{{\mathbb{E}}[\tilde{X}^p/p]} =0.$$ As the market is complete, there exists a common stochastic discount factor $Y$ and $y^T, \tilde{y}^T>0$ such that $U'(X_T) = y^T Y_T$ and $\tilde{X}^{p-1}_T = \tilde{y}^T Y_T$. Hence, $X_T = I{\left(\frac{y^T}{\tilde{y}^T} \tilde{X}_T^{p-1}\right)}$, where $I= (U')^{-1}$. Define $U^*(T,x) := U{\left(I{\left(\frac{y^T}{\tilde{y}^T} x^{p-1}\right)}\right)}$. Recall that $U(x) = x^{p^*}/p^*$ for small $x$. Therefore, $$U^*(T, x)= \frac{1}{p^*} {\left(\frac{y^T}{\tilde{y}^T}\right)}^{\frac{p^*}{p^*-1}} x^{\frac{p-1}{p^*-1}p^*}, \quad \mbox{for small } x.$$ On the other hand, there exists $T_0$ such that $1/2\leq y^T/\tilde{y}^T \leq 2$ for any $T\geq T_0$ due to Lemma \[lem: ratio Y\]. As a result, is satisfied when $U$ is replaced by $U^*$, i.e. $$\liminf_{x\downarrow 0}\frac{U^*(T, x)}{x^{p-1}} = \liminf_{x\downarrow 0} \frac{1}{p^*} {\left(\frac{y^T}{\tilde{y}^T}\right)}^{\frac{p^*}{p^*-1}} x^{\frac{p-1}{p^*-1}} = 0, \quad \mbox{for } T>T_0,$$ where the second convergence holds because $p^*<p-1<0$. It then follows from Remark \[rem: Utx/txp\] that $$\begin{aligned}
\lim_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(X_T) \, {\mathbb{I}}_{\{X_T <M\}}]}{{\mathbb{E}}[\tilde{X}^p_T/p]} &= \lim_{T\rightarrow \infty} \frac{{\mathbb{E}}{\left[U^*(T, \tilde{X}_T)\, {\mathbb{I}}_{\{\tilde{X}_T \leq (U'(M) \tilde{y}^T/y^T)^{1/(p-1)}\}}\right]}}{{\mathbb{E}}[\tilde{X}_T^p/p]}\notag\\
&=0, \label{eq: Utx conv}
\end{aligned}$$ for any $M>0$. On the other hand, fix $a>0$. For any $\epsilon>0$ there is $M_{a,\epsilon}$ such that $1-\epsilon \leq U'(x)/(a+x)^{p-1} \leq 1+\epsilon$ for $x\geq M_{a,\epsilon}$. Then follows, and the second inequality therein yields $$\label{eq: Ux/tx}
\frac{{\mathbb{E}}[U(X_T)]}{{\mathbb{E}}[\tilde{X}^p_T /p]} \leq \frac{{\mathbb{E}}[U(X_T) \, {\mathbb{I}}_{\{X_T <M_{a,\epsilon}\}}]}{{\mathbb{E}}[\tilde{X}^p_T/p]} + (1+\epsilon) \frac{{\mathbb{E}}{\left[(a+X_T)^p\,{\mathbb{I}}_{\{X_T \geq M_{a,\epsilon}\}}\right]}}{{\mathbb{E}}[\tilde{X}^p_T]},$$ where the first term on the right-hand side vanishes as $T\rightarrow \infty$ due to . For the second term, use with $\lim_{T\rightarrow \infty} y^T/\tilde{y}^T=1$ and $\tilde{y}^T = {\mathbb{E}}[\tilde{X}_T^p]$, to obtain $$\limsup_{T\rightarrow \infty} \frac{{\mathbb{E}}{\left[(a+X_T)^p\,{\mathbb{I}}_{\{X_T \geq M_\epsilon\}}\right]}}{{\mathbb{E}}[\tilde{X}^p_T]} \leq \frac{1}{1-\epsilon}.$$ In summary, the estimates for the two terms on the right side of yield $$\limsup_{T\rightarrow \infty} \frac{{\mathbb{E}}[U(X_T)]}{{\mathbb{E}}[\tilde{X}^p_T/p]} \leq \frac{1+\epsilon}{1-\epsilon}.$$ which confirms the second convergence in as $\epsilon$ was chosen arbitrarily.
[^1]: Mathematics Subject Classification: (2010) 91G10, 91G50. JEL Classification: G11, J33.The authors are grateful to Jak[š]{}a Cvitani[ć]{}, Ren Liu, and an anonymous referee for their useful comments. Paolo Guasoni is partially supported by the ERC (278295), NSF (DMS-1109047), SFI (07/MI/008, 07/SK/M1189, 08/SRC/FMC1389), and FP7 (RG-248896). Hao Xing is supported in part by STICERD at London School of Economics. Johannes Muhle-Karbe gratefully acknowledges partial support by the National Centre of Competence in Research “Financial Valuation and Risk Management” (NCCR FINRISK), Project D1 (Mathematical Methods in Financial Risk Management), of the Swiss National Science Foundation (SNF)
[^2]: See [@mossin1968optimal], [@leland1972turnpike], [@ross1974portfolio], [@hakansson1974convergence], [@MR736053], [@cox1992continuous], [@MR1629559], [@MR1805320], @dybvig1999portfolio, @detemple2010dynamic, and @guasoni.al.11.
[^3]: Here and henceforth, “isoelastic” refers to preferences with constant relative risk aversion.
[^4]: In Black and Scholes’ model, the price of the contract $e^{-\sigma^2 T} S_T^2$ is $S_0^2$, i.e., the square of the current stock price. This contract is replicated by a portfolio with equal weights in call options of all strikes.
[^5]: [@Guasoni-Robertson] show examples in which the long-run optimal portfolio has utility equal to negative infinity, if used over the entire period.
[^6]: If the utility function is strictly concave and continuously differentiable, the safe asset can be stochastic as long as it is bounded from below and above by two deterministic processes $\underline{S}$ and $\overline{S}$ such that $\lim_{t\rightarrow \infty} \underline{S}_t =\infty$; cf. Remark \[rem: stoch interest\].
[^7]: Here, the superscript $T$ indicates the optimal wealth process for horizon $T$, whereas the subscript $T$ refers to its evaluation at time $T$.
[^8]: See @MR1358100, @MR1675114 [-@MR1675114; -@MR1790132], and several others.
[^9]: @MR1968944, @follmer:aaa
[^10]: Which include a complete market where the unique equivalent martingale measure has no atoms, as well as other settings.
[^11]: In practice, incentive stock options grants include clauses that prevent a manager from taking offsetting positions even with private accounts.
[^12]: The formula follows using the fundamental theorem of calculus twice, and then integrating by parts.
[^13]: Here we assume that the resulting optimization problem is well-posed in the sense of Assumption \[ass: wellposedness\]; in particular, the effective risk aversion should be positive.
[^14]: As an extreme case, recall that, if risk aversion is small enough, the wealth of the optimal Merton portfolio converges to zero almost surely, even though its expected return is high.
[^15]: This is because $0\geq u^T\geq {\mathbb{E}}[U(S^0_T)]$ and ${\lim_{T\rightarrow \infty}}{\mathbb{E}}[U(S^0_T)]=0$ due to $\lim_{T\rightarrow \infty} S^0_T=\infty$ and $\lim_{x\uparrow \infty} U(x) =0$. The same statement holds for the isoelastic utility $\tilde{U}$.
[^16]: The identity $\tilde{y}^T \equiv 1$ follows from $1= {\mathbb{E}}[\tilde{X}_T/\tilde{X}_T] = {\mathbb{E}}[\tilde{X}_T \tilde{y}^T \tilde{Y}_T] = \tilde{y}^T$, where ${\mathbb{E}}[\tilde{X}_T \tilde{Y}_T] =1$ is used to obtain the third identity.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Weakly interacting massive particle (WIMP) is the leading paradigm of dark matter (DM). However, in this paradigm, the particle mass of DM, $m_\chi$, can not be determined by investigating its relic abundance. In order to obtain more information to constrain $m_\chi$, this work models the phase of DM production for WIMP paradigm, and investigate the suppression of tensor-to-scalar ratio of metric perturbation $r$ caused by the pair productions of DM particles. It yields a new relation between $m_\chi$ and $r$, which also contains the reheating parameters. As $r$ will be stringently constrained in the future searches of primordial gravitational wave, this relation can be used to constrain $m_\chi$ and the reheating parameters.'
author:
- Changhong Li
title: Shedding light on Dark Matter and Cosmic Reheating with Primordial Gravitational Wave
---
[*Introduction.–*]{} The particle mass of dark matter (DM), $m_\chi$, is one of the most important parameters of the Universe. However, current astrophysical evidence is insufficient to determine it as the evidence only reflects the properties of global and/or local DM abundance [@Bertone:2016nfn]. As pointed out by the latest research [@Li:2019cjp], to constrain $m_\chi$, we can turn our attention to DM production. As illustrated in FIG.\[fig:feyman.pdf\], each pair production of DM particles can generate a small local fluctuation $\Delta\Phi_i$, which encodes $m_\chi$, in the trace of metric [^1]. During DM production, these fluctuations accumulate to drive a resonance between DM density perturbation and scalar modes of metric perturbation [@Li:2019cjp], and amplify them. By investigating such amplification of scalar modes of metric perturbations, $\Delta\Phi=\sum_i \Delta\Phi_i+\Delta\Phi_{b}$, we can thus determine $m_\chi$, where $\Delta\Phi_{b}$ is the back-reaction from DM density perturbation.
![A pair production of DM particles, $\chi$, in the curved and nonrigid spacetime. A small local scalar fluctuation of spacetime $\Delta\Phi_i$ can be generated in each pair production ([*c.f.*]{} FIG. 1(b) in [@Li:2019cjp]). []{data-label="fig:feyman.pdf"}](feyman.pdf){width="48.00000%"}
As weakly interacting massive particle (WIMP) is the leading paradigm of DM, we are motived to apply this strategy to determine $m_\chi$ for this paradigm. However, in existing WIMP paradigm, the process of DM production is neglected as it is irrelevant to the relic abundance of DM [@Bertone:2004pz]. To make this strategy available for WIMP paradigm, a phase of DM production should be introduced to existing WIMP paradigm.
This work models a phase of DM production embedded in WIMP paradigm (sometimes we call it new WIMP paradigm to avoid confusion), and computes $\Delta\Phi$ to determine $m_\chi$. For new WIMP paradigm, $\Delta\Phi$ increases fastly during DM production, and its amplitude is frozen when DM attains thermal equilibrium. Since the tensor modes are unaffected by the pair productions of DM particles, such amplification results in a suppression of tensor-to-scalar ratio of metric perturbation, $r$. More specifically, the magnitude of the suppression of $r$ not only depends on $m_\chi$ but also relies on how Universe is reheated. By implementing a typical realization of reheating [@Bassett:2005xm; @Allahverdi:2010xz; @Amin:2014eta], we obtain a new relation between $m_\chi$ and $r$, which also contains the reheating parameters. As $r$ will be further constrained in the coming round of primordial gravitational wave (PGW) searches [@Abazajian:2016yjj], this relation can be, therefore, used to constrain $m_\chi$ and the reheating parameters.
[*Modeling DM production for WIMP Paradigm –* ]{} To model the phase of DM production for WIMP paradigm, we consider that DM particles $\chi$ are produced by the pair annihilations of scalar particles $\phi$ with the minimal coupling $\mathcal{L}=\lambda \phi^2\chi^2$. In FIG.\[fig:capab.pdf\], we illustrate the new WIMP paradigm. In standard cosmology [@Dodelson:2003ft], at the end of inflation ($t=t_{R_i}$), the inflaton $\varphi$ decays into Standard Model (SM) particles, and reheats Universe to the highest temperature, $T_{R_f}$, at $t=t_{R_f}$. After that, Universe evolves into a radiation-dominated expansion. The duration $t_{R_i}\le t\le t_{R_f}$ refers as the reheating. $\phi$ and $\chi$ are produced from $t=t_{R_i}$. We assume that $\phi$ is light and tightly coupled to the hot plasma filling the background, so $\phi$ can attain and track thermal equilibrium (denoted by the superscript $^{eq}$) before $t= t_{R_f}$. On the other hand, due to the small cross-section, DM attains and tracks thermal equilibrium after the end of reheating ($t_\chi^{eq}> t_{R_f}$), and freeze out at $t=t_{fo}$.
![The evolution of $\phi$ (solid curve) and $\chi$ (dash-dotted curve) abundances in new WIMP paradigm. The dotted and dashed curves, respectively, denotes the expected abundances of $\phi$ and $\chi$ at thermal equilibrium. The region in the black dotted box is the scope of existing WIMP paradigm in which DM production is not considered. []{data-label="fig:capab.pdf"}](capab.pdf){width="48.00000%"}
In new WIMP paradigm, during $y_{R_f}\le y\le 1$, DM abundance takes (see Ref.[@Li:2014era] for a similar computation), $$\label{eq:ychib}
Y_\chi=\frac{1-e^{-2\pi^2[\kappa(y-y_{R_f})+Y_{R_f}]}}{\pi^2\left(1+e^{-2\pi^2[\kappa(y-y_{R_f})+Y_{R_f}]}\right)},$$ where $y\equiv m_\chi T^{-1}$, $Y_\chi\equiv n_\chi T^{-3}$ with $n_\chi$ being DM number density, and $\kappa\equiv m_\chi^3\sigma_0(4\pi^4H_m)^{-1}$ is a dimensionless constant parameter defined with the reduced Hubble parameter $H_m\equiv Hy^2$ and the reduced thermally averaged cross-section $\sigma_0=4\widetilde{\langle\sigma v\rangle}y^{-2}$. $Y_{R_f}$ is DM abundance at $t=t_{R_f}$ [^2], and takes ([*c.f.*]{} Eq.(5) in [@Li:2019cjp]) , $$\label{eq:yrfexi}
Y_{R_f}=T_{R_f}^{-3}\int_{t_{R_i}}^{t_{R_f}}\widetilde{\langle\sigma v\rangle}\left(n_\phi/n_\phi^{eq}\right)^2\left(n_\chi^{eq}\right)^2 dt~.$$
According to Eq.(\[eq:ychib\]), at $y=y^{eq}=\left(2\pi^2\kappa\right)^{-1}$, [*i.e.*]{} $t=t_\chi^{eq}$ in FIG.\[fig:capab.pdf\], DM abundance attains and tracks thermal equilibrium, $Y_\chi^{eq}=\pi^{-2}$. At $y=1$, DM freezes out with a relic abundance, $Y_f=1.7\times 10^{-29} (\sigma_0 m_\chi)^{-1} \text{eV}^{-1}$. By applying the density fraction contributed by DM, $\Omega_\chi=1.18\times 10^{-2} \text{eV}^{-1}\times m_\chi Y_f= 0.26$, the well-known prediction of WIMP paradigm, $\sigma_0=0.3\times 10^{-39}\text{cm}^2$ ([*c.f.*]{} Eq.(3.60) in [@Dodelson:2003ft]), can be re-obtained. However, $m_\chi$ can not be determined with relic abundance.
[*Amplification of $\Phi$ during DM production. –* ]{} In order to constrain $m_\chi$, we compute $\Delta\Phi$ for new WIMP paradigm. To facilitate our analysis, we divide DM evolution into three phases, 1) DM Production ($y_{R_f}\le y\le y_\chi^{eq}$), 2) Equilibrium ($y_{R_f}\le y\le y_\chi^{eq}$) and 3) Freezing-out ($y> y_\chi^{eq}$). We notice that $\Delta\Phi_i$ are generated and accumulated only during DM Production. In Equilibrium, each $\Delta\Phi_i$ generated in the pair production of DM particles is cancelled exactly by its counterpart generated in the pair annihilation, so there is no accumulation of $\Delta\Phi_i$. And in Freezing-out, no pair production or annihilation of DM particles takes place, so no $\Delta\Phi_i$ is generated.
By using the method provided by Ref.[@Li:2019cjp], we obtain the evolution of super-horizon scalar modes of metric perturbation, $\Phi(y)=\Phi_\varphi+\Delta\Phi$, during DM production [^3], $$\label{eq:phevi}
\Phi(y)=\Phi_\varphi \mathcal{G}\left[-y\kappa Y_{R_f}^{-1}\right],$$ where $\mathcal{G}(x)\equiv {_2F_1}\left(\frac{3-\sqrt{17}}{4},\frac{3+\sqrt{17}}{4};\frac{7}{2}; x \right)$ is the Gauss hypergeometric function used in Ref.[@Li:2019cjp], the perturbed FLRW metric in conformal Newtonian gauge, $ g_{\mu\nu}=\{-1-2\Psi(\vec{x},t), a^2(t) \delta_{ij} \left[1+2\Phi(\vec{x},t)\right]\}$, is adopted to describe the curved spacetime, $\Phi(y)$ denotes the super-horizon Fourier modes of $\Phi(\vec{x},t)$, and $\Phi_\varphi$ is the primordial value of $\Phi(y)$ at the end of reheating ($y=y_{R_f}$).
After DM abundance attains thermal equilibrium, $y\ge y_\chi^{eq}$, the amplitude of $\Phi(y)$ is frozen, $$\label{eq:phevf}
\Phi(y)=\Phi(y_\chi^{eq})=\Phi_\varphi \mathcal{G}\left[-\left(2\pi^2Y_{R_f}\right)^{-1} \right].$$
In FIG.\[fig:Phi.pdf\], we plot the evolution of $\Phi(y)$ by using Eq.(\[eq:phevi\]) and Eq.(\[eq:phevf\]). $\Phi(y)$ is amplified significantly during DM production. The amplification in each case is scale-independent, which ensures that the scale-invariance of primordial curvature perturbation can be preserved by such amplification.
![The evolution of $\Phi(y)$ in new WIMP paradigm.[]{data-label="fig:Phi.pdf"}](Phi.pdf){width="48.00000%"}
As shown in FIG.\[fig:Phi.pdf\], there are two salient features for new WIMP paradigm,
1. For different $Y_{R_f}$, the magnitude of amplification is different;
2. For different $\kappa$ (different $m_\chi$), the amplifications are frozen at different time.
As we will show later, the former leads to the new strategy to constrain $m_\chi$ while the latter is related to structure formation at small scale.
[*The Suppression of $r$. –*]{} As tensor modes of metric perturbation are not affected by the pair productions of DM particles, the amplification of $\Phi(y)$, thus, results in a suppression of $r$, $$r\equiv \frac{\mathcal{P}_h}{\mathcal{P}_\zeta}=4\epsilon\times\left(\frac{\Phi_\varphi}{\Phi(y)}\right)^2~,$$ where $\zeta$ is curvature perturbation in comoving gauge, and it takes $\zeta=\frac{3}{2}\Phi$ in radiation-dominated era [@Dodelson:2003ft]. The power spectra of primordial metric perturbations at $y=y_{R_f}$, $\mathcal{P}_{\zeta}(y_{R_f})=\frac{9}{4}\mathcal{P}_{\Phi_\varphi}=\frac{9}{4}\frac{8\pi G}{9 k^3}\frac{H^2_\star}{\epsilon}$ and $\mathcal{P}_h=\frac{8\pi G}{k^3}H^2_\star$, are employed by following the simplest realization of slow-roll inflation [@Dodelson:2003ft], and the slow-roll parameter takes $\epsilon\simeq 0.01$ [@Komatsu:2010fb; @Ade:2015xua].
At $y=y^{eq}$, the amplitude of $r$ is frozen accordingly, $$\label{eq:rzepi}
r=4\epsilon \left\{\mathcal{G}\left[-\left(2\pi^2\right)^{-1} Y_{R_f}^{-1}\right]\right\}^{-2}.$$ which is smaller than the conventional expected value, $4\epsilon$, and predicts a smaller amplitude of PGW to be detected.
In FIG.\[fig:r.pdf\], the suppression of $r$ is illustrated for typical cases. For the cases with the same value of $Y_{R_f}$ (a combination of $m_\chi$ and the reheating parameters), the final amplitudes of $r$ are same. With further elucidation of the reheating process, $m_\chi$ can be essentially constrained in future PGW searches.
![The suppression of $r$ in new WIMP paradigm.[]{data-label="fig:r.pdf"}](r.pdf){width="48.00000%"}
[*A New Relation amongst DM, PGW and Reheating. –*]{} To obtain an explicit relation between $m_\chi$ and $r$, we implement a typical realization of reheating process [@Bassett:2005xm; @Allahverdi:2010xz; @Amin:2014eta] to compute $Y_{R_f}$. By simplifying Eq.(\[eq:yrfexi\]) with the physical conditions in $t_{R_i}\le t\le t_{R_f}$, we have $$\label{eq:ychiexp}
Y_{R_f}=\frac{\sigma_0 m_\chi^2 }{4\pi^4 T_{R_f}^3}\int_{t_{m_\chi}}^{t_{R_f}}[T(t)]^4dt~,$$ where $T(t)$ describes how Universe is reheated, $t_{m_\chi}$ labels the moment $T=m_\chi$ during reheating, and the integral in $t_{R_i}\le t\le t_{m_\chi}$ is negligible [@Li:2019cjp].
By assuming that the inflaton $\varphi$ decays into SM particles effectively with a constant dissipative rate $\Gamma_0$, $\Gamma_0\gg 3(1+w_\varphi)H$ [^4], the equation governing $\rho_\varphi$ takes $$\label{eq:rvpeq}
\frac{d\rho_\varphi}{dt}=-\Gamma_0\rho_\varphi.$$ Substituting Eq.(\[eq:rvpeq\]) into the continuity condition of energy density, $\sum_i\left[\frac{d\rho_i}{dt}+3(1+w_i)\rho_i\right]=0$, we obtain $$\label{eq:rgeq}
\frac{d\rho_\gamma}{dt}+4H\rho_\gamma=\Gamma_0\rho_\varphi,$$ where $\sum_i$ denotes the summation of all relativistic components $\gamma$ and inflaton $\varphi$, and $w$ denotes the Equation of State. On the other side, we have $$\label{eq:rgt}
\rho_\gamma=g_\ast [T(t)]^4,$$ where $g_\ast$ is the relativistic degrees of freedom at $t\ge t_{m_\chi}$. By solving Eq.(\[eq:rvpeq\]), Eq.(\[eq:rgeq\]) and Eq.(\[eq:rgt\]), and substituting their solutions into Eq.(\[eq:ychiexp\]) [^5] , we obtain $$\label{eq:yrffi}
Y_{R_f}=-\frac{\sigma_0 m_\chi^2 T_{R_f}}{4\pi^4\Gamma_0}\ln\left(\frac{4\sqrt{2g_\ast}T_{R_f}^2}{M_p\Gamma_0}\right),$$ where $M_p$ is the reduced Planck mass.
By substituting Eq.(\[eq:yrffi\]) into Eq.(\[eq:rzepi\]), we obtain a new relation amongst DM, PGW, and reheating, $$\label{eq:rzep}
r=4\epsilon \left\{\mathcal{G}\left[\left(\frac{\sigma_0 m_\chi^2 T_{R_f}}{2\pi^2\Gamma_0}\ln\left[\frac{4\sqrt{2g_\ast}T_{R_f}^2}{M_p\Gamma_0}\right]\right)^{-1}\right]\right\}^{-2}.$$
In FIG.\[fig:rmchi.pdf\], this relation is illustrated with examples of $T_{R_f}=10^{12} \text{GeV}$ and $\Gamma_0=\{10^9,10^{12},10^{15}\} \text{GeV}$ with $g_\ast=90$. It shows that, in a given reheating process, $m_\chi$ can be constrained by $r$. For instance, given $T_{R_f}=10^{12}\text{GeV}$ and $\Gamma_0=10^9\text{GeV}$, if PGW is detected at $r=0.6\times 10^{-3}$ in future, $m_\chi$ is predicted to be $10^3\text{GeV}$.
![The relation between $m_\chi$ and $r$ in WIMP paradigm. []{data-label="fig:rmchi.pdf"}](rmchi.pdf){width="48.00000%"}
As shown in FIG.\[fig:rmchi.pdf\], the more pair productions of DM particles take place after reheating (with smaller $m_\chi$ and/or larger $\Gamma_0$), the stronger the suppression of $r$ is. In other words, if no suppression is observed in the future searches of PGW, it implies 1) large $m_\chi$ and/or 2) small $\Gamma_0$ (see the curve plotted with $\Gamma_0=10^9\text{GeV}$ for $m_\chi> 10^4 \text{GeV}$). Although the prediction of this work is still beyond the detectability of current experiments (for example, see the shaded region plotted according to Ref.[@Ade:2015xua]), optimized by future searches of PGW and further elucidation of cosmic reheating, it will be tested in a sizable parameter region, and then constrain $m_\chi$ concretely.
[*Conclusions. –*]{} This work elaborates a new strategy to constrain $m_\chi$ in WIMP paradigm with the search of PGW. By modeling the process of DM production embedded in WIMP paradigm, this work investigates the suppression of $r$ caused by the post-reheating pair production of DM particles, and obtains an explicit relation between $m_\chi$ and $r$, which also contains the reheating parameters, $\Gamma_0$ and $T_{R_f}$. It predicts a smaller amplitude of PGW to be detected, especially, for small $m_\chi$ and/or large $\Gamma_0$. Once $r$ is determined in a future search of PGW, $m_\chi$ and the reheating parameters are then constrained by using this relation. As established by investigating DM density perturbation, this new strategy of determining $m_\chi$ within WIMP paradigm is complementary to other proposals primarily focusing on DM abundance [@Patrignani:2016xqp].
To sum up, we discuss two issues worthy of further study. 1) [*Implications on Small Scale Structure.*]{} As shown in FIG.\[fig:Phi.pdf\], the amplification of $\Phi$ completes at $y=y^{eq}$. For long wavelength modes discussed in this work, the amplifications of them are the same in each case. However, for short wavelength modes, they get less amplification as they re-enter horizon before the end of DM production. It thus leads to a relative suppression of linear matter perturbation at small scale. By using horizon-crossing condition, the critical scale of suppression, $k_\star^{-1}$, can be determined by $y^{eq}$, and $y^{eq}$ only depends on $m_\chi$, thus $k_\star^{-1}$ is solely determined by $m_\chi$. Once $k_\star^{-1}$ is precisely measured in future astrophysical observations [@Weinberg:2013aya], it then imposes another new constraint on $m_\chi$, which is complementary to what we have obtained based on the magnitude of suppression of $r$ in this work. Recently, this issue has been studied for non-thermal DM in Ref.[@Li:2019yyx], and it should be also investigated for WIMP paradigm. 2) [*Generalization beyond WIMP.*]{} The strategy elaborated in this work is not a monopoly of WIMP or standard cosmological model. It is very promising to apply this strategy to various DM candidates [@Tanabashi:2018oca] in alternative cosmological models [@Nojiri:2017ncd] in further study.
[*Acknowledgments. –*]{} I thank Yeuk-kwan Edna Cheung and Xiaheng Xie for useful discussion, careful reading of this manuscript and important suggestion on improving the clarity of presentation. This work has been supported by the National Natural Science Foundation of China (11603018, 11963005, 11775110, 11433004, 11690030), Yunnan Provincial Grants (No.2016FD006), and the Leading/Top Talents of Yunnan Province (2015HA022, 2015HA030).
[10]{}
G. Bertone and D. Hooper, Rev. Mod. Phys. [**90**]{}, no. 4, 045002 (2018) \[arXiv:1605.04909 \[astro-ph.CO\]\]. C. Li, arXiv:1902.05738 \[astro-ph.CO\]. C. Li, JCAP [**1609**]{}, 038 (2016) \[arXiv:1512.06794 \[astro-ph.CO\]\].
G. Bertone, D. Hooper and J. Silk, Phys. Rept. [**405**]{}, 279 (2005) \[hep-ph/0404175\].
B. A. Bassett, S. Tsujikawa and D. Wands, Rev. Mod. Phys. [**78**]{}, 537 (2006) \[astro-ph/0507632\]. R. Allahverdi, R. Brandenberger, F. Y. Cyr-Racine and A. Mazumdar, Ann. Rev. Nucl. Part. Sci. [**60**]{}, 27 (2010) \[arXiv:1001.2600 \[hep-th\]\]. M. A. Amin, M. P. Hertzberg, D. I. Kaiser and J. Karouby, Int. J. Mod. Phys. D [**24**]{}, 1530003 (2014) \[arXiv:1410.3808 \[hep-ph\]\]. K. N. Abazajian [*et al.*]{} \[CMB-S4 Collaboration\], arXiv:1610.02743 \[astro-ph.CO\].
S. Dodelson, Amsterdam, Netherlands: Academic Pr. (2003) 440 p
C. Li, R. H. Brandenberger and Y. K. E. Cheung, Phys. Rev. D [**90**]{}, no. 12, 123535 (2014) \[arXiv:1403.5625 \[gr-qc\]\].
E. Komatsu [*et al.*]{} \[WMAP Collaboration\], Astrophys. J. Suppl. [**192**]{}, 18 (2011) \[arXiv:1001.4538 \[astro-ph.CO\]\]. P. A. R. Ade [*et al.*]{} \[Planck Collaboration\], Astron. Astrophys. [**594**]{}, A13 (2016) \[arXiv:1502.01589 \[astro-ph.CO\]\].
C. Patrignani [*et al.*]{} \[Particle Data Group\], Chin. Phys. C [**40**]{}, no. 10, 100001 (2016).
D. H. Weinberg, J. S. Bullock, F. Governato, R. Kuzio de Naray and A. H. G. Peter, Proc. Nat. Acad. Sci. [**112**]{}, 12249 (2015) \[arXiv:1306.0913 \[astro-ph.CO\]\].
C. Li, arXiv:1903.00701 \[astro-ph.CO\].
M. Tanabashi [*et al.*]{} \[Particle Data Group\], Phys. Rev. D [**98**]{}, no. 3, 030001 (2018). S. Nojiri, S. D. Odintsov and V. K. Oikonomou, Phys. Rept. [**692**]{}, 1 (2017) \[arXiv:1705.11098 \[gr-qc\]\].
[^1]: As a contrast, in Ref.[@Li:2015egy], there is no such fluctuation since the interaction between scalar modes of metric perturbation and DM density perturbation has been neglected for simplicity.
[^2]: Only after $t=t_{R_f}$, $y$ can be used to describe the fully thermalized Universe. To determine $Y_{R_f}$, the process of reheating, $T=T(t)$ during $t_{R_i}\le t\le t_{R_f}$, should be specified.
[^3]: During the radiation-dominated phase, by applying the perturbed Einstein equation to the scalar modes of metric perturbation, the linear DM density perturbation can be determined accordingly. We therefore only present $\Phi(y)$ throughout this work.
[^4]: In so-called “old” reheating theory, by assuming inflaton $\varphi$ decaying as a single-body, a rough estimation can relate $T_{R_f}$ and $\Gamma_0$, $T_{R_f}\simeq\sqrt{\Gamma_0 M_p}$. Thus a single parameter, $T_{R_f}$ or $\Gamma_0$, is enough in such theory (see Ref.[@Bassett:2005xm] for a review). However, in a general framework of cosmic reheating, the concept of single-body decay can not be always applied, so this work takes $T_{R_f}$ and $\Gamma_0$ as two independent parameters.
[^5]: By solving Eq.(\[eq:rvpeq\]), we obtain $\rho_\varphi=\rho_0 e^{-\Gamma_0 t}$, where $\rho_0$ is the value of $\rho_\varphi$ at $t=t_{R_i}=0$. By substituting it into Eq.(\[eq:rgeq\]), we have $$\label{eq:rhovex}
\frac{d\rho_\gamma}{dt}+4H\rho_\gamma=\Gamma_0\rho_0 e^{-\Gamma_0 t}~.$$ For $t\le t_{R_f}$, Eq.(\[eq:rhovex\]) can be simplified as $d\rho_\gamma/dt=\Gamma_0\rho_0 e^{-\Gamma_0 t}$ with $\Gamma_0\gg H$, which solution is $\rho_\gamma=\rho_0 \left(1-e^{-\Gamma_0 t}\right)$. By substituting it into Eq.(\[eq:rgt\]) and Eq.(\[eq:ychiexp\]), we obtain $$\label{eq:yrfint}
Y_{R_f}=\frac{\sigma_0 m_\chi^2 T_{R_f}}{4\pi^4}t_{R_f},$$ where $\Gamma_0 t_{R_f}\gg 1$ and $t_{R_f}\gg t_{m_\chi}$ are used. By applying $d\rho_\gamma/dt |_{t=t_{R_f}}=0$ to Eq.(\[eq:rhovex\]), we have $4H(t_{R_f})\rho_\gamma(t_{R_f})=\Gamma_0\rho_0e^{-\Gamma_0 t_{R_f}}$, and yields $t_{R_f}=-\Gamma_0^{-1}\ln\left[\frac{4\sqrt{2g_\ast}T_{R_f}^2}{M_p\Gamma_0}\right]$, where $H(t_{R_f})=\sqrt{2g_\ast}T_{R_f}^2/M_p$ are used. Then by substituting it into Eq.(\[eq:yrfint\]), we obtain Eq.(\[eq:yrffi\]).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a 2+1 dimensional quantum gauge theory model with correlated fermions that is exactly solvable by bosonization. This model gives an effective description of partially gapped fermions on a square lattice that have density-density interactions and are coupled to photons. We show that the photons in this model are massive due to gauge-invariant normal-ordering, similarly as in the Schwinger model. Moreover, the exact excitation spectrum of the model has two gapped and one gapless mode. We also compute the magnetic field induced by an external current and show that there is a Meissner effect. We find that the transverse photons have significant effects on the low-energy properties of the model even if the fermion-photon coupling is small.'
author:
- Jonas de Woul
- Edwin Langmann
date: 'July 5, 2011'
---
The possible violation of Landau’s Fermi liquid theory in models of strongly interacting fermions has been an actively researched problem for many years. Interest in this topic quickly grew with the discovery of the cuprate high-temperature superconductors in 1986 [@BednorzMueller], and the realization that these materials display many properties not described by Fermi liquid theory [@Bonn]. Early on, it was suggested that models of Hubbard-type capture the strongly correlated physics of cuprates [@AndersonRVB; @Emery; @VSA; @ZhangRice]. While it has proven very difficult to do reliable computations for two dimensional (2D) such models, an excellent theoretical understanding has been obtained for the corresponding one-dimensional (1D) cases. This understanding is largely based on models that are exactly solvable. A famous example is the [*Luttinger model*]{} [@Luttinger; @MattisLieb], which provides an effective description of interacting fermions on a 1D lattice, and which has become a prototype of non-Fermi-liquid behavior [@Haldane; @Tsvelik].
It has been known for quite some time that fermions coupled to dynamical photons can have non-Fermi-liquid behaviour [@Holstein; @Reizer1]. One argument against this mechanism being relevant for real materials is the smallness of the fine-structure constant ($\alpha\approx \frac{1}{137}$), which governs the strength of interactions between matter and transverse photons. However, various scenarios have been proposed and explored in which effective photon-like gauge fields arise in the low-energy limit of models for strongly correlated fermions [@NL; @FranzTesanovic:2001; @Leeetal]; see also [@Tsvelik]. In these instances, the effective coupling constant need not be small. The computations to explore this mechanism in 2D are usually based on approximations that are difficult to justify and, again, things are much better understood in 1D due to the existence of exactly solvable prototype models. For example, the (1+1)D quantum gauge theory obtained by minimally coupling the Luttinger model to dynamical photons is exactly solvable [@GLR]. This model is a generalization of the [*Schwinger model*]{} [@Schw3] and, similarly as in the latter model [@Manton], the photons have a non-zero mass that arises due to gauge-invariant normal ordering.
Despite a concerted effort, much remains to be understood when it comes to strongly correlated fermions in 2D. We believe that there is a need for models that are amenable to exact computations. One pioneer in this direction is Mattis who, already in 1987, presented a 2D interacting fermion model that can be mapped exactly to non-interacting bosons [@Mattis]. Despite its potential, Mattis’ model did not receive much attention (notable exceptions are [@Hlubina; @Luther; @SL]). In recent work, we derived an effective model describing fermions on a square lattice with local interactions and close to half filling [@EL1; @EL2; @dWL1]. Since this model is similar to Mattis’ [@Mattis], we refer to it as the [*Mattis model*]{}. In [@dWL2] we obtained a complete solution of the Mattis model, including all Green’s functions. We also proved that the fermion two-point functions have algebraic decay with exponents that depend on the interaction strength, which is a hallmark of Luttinger liquids [@Haldane]. The Mattis model is not relativistically invariant but, as already noted in [@Mattis], this is natural for effective models of square-lattice fermions.
In this Letter, we present a (2+1)D quantum gauge theory obtained by coupling the Mattis model to a dynamical $U(1)$ gauge field. We refer to the latter as [*photon field*]{}, even though other physical interpretations are possible. Our main result is that this model is exactly solvable by bosonization. We also show that this model has some remarkable properties, including a non-zero photon mass due to gauge-invariant normal ordering, and a Meissner effect. We find that there is a renormalization of the [*bare charge*]{}, which determines the strength of fermion-photon interactions. Furthermore, the gauge theory contains a gapless mode that is strongly affected by the presence of the dynamical photons even if the fermion-photon coupling is weak. The idea of mapping fermions to free bosons has been used before to study 2D interacting fermions coupled to gauge fields [@KHR; @KHM], but no exactly solvable fermion model was clearly identified before.
([*Notation*]{}: $\mu,\nu=0,+,-$ are space-time indices, $s,s'=+,-$ space indices, $r,r'=\pm$ chirality indices; the space-time metric signature is $(-,+,+)$; $x=(ct,{{\bf x}})$ are space-time coordinates with 2D positions ${{\bf x}}=(x^+,x^-)$; $c$ is the velocity of light; $\partial_s={\partial}/{\partial x^s}$ are spatial derivatives; $\psi^{(\dag)}_{r,s}$ are standard fermion field operators; $A_\mu$ is the gauge potential, $E_s$ are the electric field components, and $B=\partial_+A_- - \partial_- A_+$ is the magnetic field. Common argument ${{\bf x}}$ of field operators are suppressed whenever possible.)
The relation of the Mattis model to square-lattice fermions is described using Figure \[FermiSurface\], which shows the Brillouin zone corresponding to a square lattice (dashed large square) divided into regions of non-equal sizes. Close to half filling, mean field theory predicts that the system is partially gapped, and there is an underlying Fermi surface in the so-called [*nodal*]{} regions (the four tilted rectangles) [@dWL2]. We model this underlying surface by straight arcs, which either corresponds to a truncated Fermi surface or the portion of a closed Fermi pocket having dominant momentum occupation. The Mattis model describes the fermion degrees of freedom in the vicinity of these arcs. It is written in terms of four fermion field operators $\psi_{r,s}$, $r,s=\pm$, in one-to-one correspondence with the nodal regions. The quantum field theory limit making this model amenable to bosonization amounts to removing the momentum cutoff orthogonal to the arcs (indicated by the arrow in the nodal $(+,+)$-region, for example), which is possible after normal ordering [@EL2]. The arc-picture underlying our derivation of the Mattis model is supported by renormalization group studies of weakly coupled 2D Hubbard-like systems [@FRS]. It is also a signature feature of the pseudogap phase as observed in angle-resolved photoemission experiments on hole-doped cuprates [@ARPES].
The Hamiltonian of the Mattis model is $$\label{Mattis Hamiltonian}
\begin{split}
H_M = \sum_{r,s=\pm} \int {\mathrm{d}}^2 x\, \Bigl( rv_F :\! \psi^\dag_{r,s}(-{{\rm i}}\partial_s)\psi_{r,s}{^{\phantom\dag}}\!:
+ \sum_{r',s'=\pm} g_{r,s,r',s'} :\!\psi^\dag_{r,s}\psi_{r,s}{^{\phantom\dag}}\!: :\!\psi^\dag_{r',s'}\psi_{r',s'}{^{\phantom\dag}}\!: \Bigr)
\end{split}$$ with $\bigl\{ \psi_{r,s}^{\phantom\dag}({{\bf x}}),\psi_{r',s'}^\dag({{\bf y}})\bigr\}=\delta_{r,r'}\delta_{s,s'}\delta^2({{\bf x}}-{{\bf y}})$, etc., and colons denoting standard fermion normal ordering with respect to a Dirac vacuum [@dWL2]. These definitions are formal since an important UV regularization has been suppressed: while the $x_s$-component of ${{\bf x}}$ in $\psi^{(\dag)}_{r,s}({{\bf x}})$ is continuous, the $x_{-s}$-component is discretized to integer multiples of a UV cutoff $\tilde a$, and thus the integrals and Dirac deltas have to be interpreted as partial Riemann sums and Kronecker deltas [@dWL2]. The coupling constants scale with the UV cutoff as $g_{r,s,r',s'} = {\tilde{a}}\pi v_F\bigl(\gamma_1\delta_{s,s'}\delta_{r,-r'}+ \gamma_2\delta_{s,-s'}/2\bigr)$, with the Fermi velocity $v_F>0$ and dimension-less constants $\gamma_{1,2}$ such that $|\gamma_1|<1$ and $|\gamma_2|<|1+\gamma_1|$. The above scaling of the coupling constants is not only obtained by deriving the model from lattice fermions [@EL2], but it also ensures that $H_M$ has a non-trivial limit as ${\tilde{a}}\to 0$. The restrictions on $\gamma_{1,2}$ are to ensure stability of the Dirac vacuum [@dWL2].
The Hamiltonian of our quantum gauge theory model is obtained by coupling the Mattis Hamiltonian in to a dynamical electromagnetic field, $$\label{Gauged Mattis Hamiltonian}
\begin{split}
H = \sum_{r,s=\pm} \int {\mathrm{d}}^2 x\, \Bigl( rv_F {\stackrel {\scriptscriptstyle \circ}{\scriptscriptstyle \circ}\! }\psi^\dag_{r,s}(-{{\rm i}}\partial_s + {e_{0}}A_s)\psi_{r,s}{^{\phantom\dag}}{\! \stackrel {\scriptscriptstyle \circ}{\scriptscriptstyle \circ}}+ \sum_{r',s'=\pm} g_{r,s,r',s'} {\stackrel {\scriptscriptstyle \circ}{\scriptscriptstyle \circ}\! }\psi^\dag_{r,s}\psi_{r,s}{^{\phantom\dag}}{\! \stackrel {\scriptscriptstyle \circ}{\scriptscriptstyle \circ}}{\stackrel {\scriptscriptstyle \circ}{\scriptscriptstyle \circ}\! }\psi^\dag_{r',s'}\psi_{r',s'}{^{\phantom\dag}}{\! \stackrel {\scriptscriptstyle \circ}{\scriptscriptstyle \circ}}\Bigr)
\\
+ \frac12\int {\mathrm{d}}^2x {\stackrel {\scriptscriptstyle \times}{\scriptscriptstyle \times} \!}\Bigl(E_+^2 + E_-^2 + c^2 B^2\Bigr){\! \stackrel {\scriptscriptstyle \times}{\scriptscriptstyle \times}}\end{split}$$ with $[A_s({{\bf x}}),E_{s'}({{\bf y}})]={{\rm i}}\delta_{s,s'}\delta^2({{\bf x}}-{{\bf y}})$, etc., $e_0$ the bare charge, ${\stackrel {\scriptscriptstyle \times}{\scriptscriptstyle \times} \!}\cdots{\! \stackrel {\scriptscriptstyle \times}{\scriptscriptstyle \times}}$ boson normal ordering, and ${\stackrel {\scriptscriptstyle \circ}{\scriptscriptstyle \circ}\! }\cdots{\! \stackrel {\scriptscriptstyle \circ}{\scriptscriptstyle \circ}}$ a gauge-invariant generalization of fermion normal ordering (see below). The Gauss law operators generating gauge transformations $A_s\to A_s + \partial_s\chi$, etc., are $$\label{Gauss law generator}
G[\chi]=\int {\mathrm{d}}^2x\, \chi \sum_{s=\pm}\Bigl(-\partial_sE_s + {e_{0}}\sum_{r=\pm}{\stackrel {\scriptscriptstyle \circ}{\scriptscriptstyle \circ}\! }\psi^\dag_{r,s}\psi_{r,s}{^{\phantom\dag}}{\! \stackrel {\scriptscriptstyle \circ}{\scriptscriptstyle \circ}}\Bigr),$$ and the physical states are those annihilated by $G[\chi]$, for arbitrary real-valued functions $\chi({{\bf x}})$. Note that, except for the normal-ordering procedures, the Hamiltonian in is obtained from the Mattis Hamiltonian by standard minimal coupling: introduce the action corresponding to the Mattis Hamiltonian in ; couple to Abelian gauge fields using the substitution $-{{\rm i}}\partial_\mu \to -{{\rm i}}\partial_\mu + {e_{0}}A_\mu$; add the Maxwell term $c^2F_{\mu\nu}F^{\mu\nu}/4$; perform the usual Dirac procedure for systems with constraints [@Sundermeyer] to obtain the formal Hamiltonian (i.e. without normal ordering) and the formal Gauss law constraint. We note in passing that the gauge field operators are well-defined without UV regularization (details will be spelled out elsewhere).
As already indicated, fermion normal ordering plays a key role for our model. Indeed, formulating a sensible quantization of the classical theory is non-trivial, even in the absence of gauge fields: In order for the Hamiltonian to be bounded from below (have a ground state), we need to normal-order all fermion bilinears with respect to the Dirac vacuum in which all negative energy states are occupied. An important consequence of normal-ordering is that the fermion densities $J_{r,s} \equiv \, :\!\psi^\dag_{r,s}\psi^{\phantom\dag}_{r,s}\!:$ obey the anomalous commutator relations [@Schw1] $$\label{JJ}
[J_{r,s}({{\bf x}}),J_{r',s'}({{\bf y}})]=r\delta_{r,r'}\delta_{s,s'}({2\pi{{\rm i}}\tilde a})^{-1} \partial_s\delta^2({{\bf x}}-{{\bf y}}),$$ and the non-interacting part of can be expressed in terms of these densities using the operator identity $$\label{Kronig identity}
\int {\mathrm{d}}^2x\, :\!\psi^\dag_{r,s}r \left(- {{\rm i}}\partial_s\right)\psi^{\phantom\dag}_{r,s}\!: \, = \pi{\tilde{a}}\int {\mathrm{d}}^2 x {\stackrel {\scriptscriptstyle \times}{\scriptscriptstyle \times} \!}J_{r,s}^2 {\! \stackrel {\scriptscriptstyle \times}{\scriptscriptstyle \times}}$$ (see Propositon 2.1 in [@dWL2] for proofs of these statements). The commutation relations in imply that $$\label{bosons}
\partial_s\Phi_s = \sqrt{\pi\tilde a}\bigl( J_{+,s}+ J_{-,s} \bigr), \qquad
\Pi_s = \sqrt{\pi\tilde a}\bigl(-J_{+,s} + J_{-,s}\bigr)
$$ define boson operators obeying the usual canonical commutator relations $[\Phi_s({{\bf x}}),\Pi_{s'}({{\bf y}})]={{\rm i}}\delta_{s,s'}\delta^2({{\bf x}}-{{\bf y}})$, etc. It follows, using , that the Mattis Hamiltonian in can be written in terms of free bosons, and this is the key step towards the exact solution of the Mattis model [@dWL2].
The gauge fields are quantized as usual by a partial gauge fixing $A_0=0$, postulating the canonical commutation relations (see below ), and imposing the Gauss law operator constraint on the Hilbert space [@Jackiw]. However, the corresponding quantum Hamiltonian is [*not*]{} obtained as the straightforward quantization of the minimally coupled classical Hamiltonian: due to the anomalous commutators in , the Gauss law operators $G[\chi]$ would in this case no longer commute with the Hamiltonian, and the theory would thus not be gauge invariant. The remedy of this problem is to introduce a manifestly gauge-invariant normal-ordering prescription. We use the point-splitting method pioneered by Schwinger [@Schw1]: start with the gauge-invariant expression $$\label{BL}
\psi^\dag_{r,s}({{\bf x}}-\epsilon{{\bf e}}_s/2) {\,{\rm e}}^{ {{\rm i}}{e_{0}}\int_{-\epsilon/2}^{\epsilon/2} A_s({{\bf x}}+\xi{{\bf e}}_s)d\xi} \psi^{\phantom\dag}_{r,s}({{\bf x}}+\epsilon{{\bf e}}_s/2),$$ which includes a line integral of the gauge field. We define our gauge-invariant fermion normal ordering of bilinears, ${\stackrel {\scriptscriptstyle \circ}{\scriptscriptstyle \circ}\! }\psi^\dag_{r,s}\psi^{\phantom\dag}_{r,s}{\! \stackrel {\scriptscriptstyle \circ}{\scriptscriptstyle \circ}}$, as the limit $\epsilon\to 0$ of after first subtracting off its singular part $r/(2\pi{{\rm i}}\tilde{a}\epsilon)$ (see Equation(4.9) in [@dWL2]). The result is $$\label{JrsA}
{\stackrel {\scriptscriptstyle \circ}{\scriptscriptstyle \circ}\! }\psi^\dag_{r,s}\psi^{\phantom\dag}_{r,s}{\! \stackrel {\scriptscriptstyle \circ}{\scriptscriptstyle \circ}}\; = J_{r,s} + r {e_{0}}A_{s}/({2\pi\tilde{a}}),$$ and similarly $$\label{gauge-invariant fermion normal-ordering with derivative}
\begin{split}
\int d^2x\, {\stackrel {\scriptscriptstyle \circ}{\scriptscriptstyle \circ}\! }\psi_{r,s}^\dag r\left( { - {{\rm i}}\partial _s } + {e_{0}}A_s\right)\psi_{r,s}{^{\phantom\dag}}{\! \stackrel {\scriptscriptstyle \circ}{\scriptscriptstyle \circ}}\; = \int d^2 x\, \Bigl( :\! \psi_{r,s}^\dag r ( { - {{\rm i}}\partial _s })\psi_{r,s}{^{\phantom\dag}}\!: +r{e_{0}}A_s J_{r,s} + \frac{{e_{0}}^2}{4\pi{\tilde{a}}} A_s^2 \Bigr)
\end{split}$$ (computational details will be provided elsewhere). An important feature of is the “bare” photon mass term in the second line; as noted, this is a direct consequence of gauge-invariant normal-ordering. One can verify that and are gauge-invariant expressions.
The Hamiltonian and the Gauss law operators of the gauged model are now bosonized using the above results: $$\begin{split}
\label{HGM}
H = \frac12\int {\mathrm{d}}^2x {\stackrel {\scriptscriptstyle \times}{\scriptscriptstyle \times} \!}\Bigl(v_F\sum_{s=\pm}\Bigl[(1-\gamma_1)(\Pi_s-{e_{R}}A_s)^2
+(1+\gamma_1)(\partial_s\Phi_s)^2
\\
+\gamma_2(\partial_s\Phi_s)(\partial_{-s}\Phi_{-s})\Bigr] + E_+^2 + E_-^2 + c^2 B^2\Bigr){\! \stackrel {\scriptscriptstyle \times}{\scriptscriptstyle \times}}\end{split}$$ and $G[\chi] = \int {\mathrm{d}}^2x\, \chi \sum_{s=\pm} \partial_s(-E_s+{e_{R}}\Phi_s)$, with the [*renormalized*]{} charge ${e_{R}}= {{e_{0}}}/({\sqrt{\pi \tilde a}})$. Note that the scaling of the model parameters and boson fields are such that our gauge theory model remains well-defined in the UV limit ${\tilde{a}}\to 0^+$. The charge renormalization ${e_{0}}\to {e_{R}}$ shows that photons can have stronger influence on physical properties than what superficial arguments might suggest. Under a gauge transformation, $A_s\to A_s+\partial_s\chi$, $E_s\to E_s$, $\Phi_s\to \Phi_s$ and $\Pi_s\to\Pi_s+{e_{R}}\partial_s\chi$, such that $\Pi_s-{e_{R}}A_s$, and thus $H$ in , are gauge invariant.
The Hamiltonian in is quadratic in boson operators and can therefore be diagonalized by a Bogoliubov transformation. To this end, we perform a Fourier transformation, $E_s({{\bf x}})\to \hat E_s({{\bf p}})$, and define longitudinal- and tranverse fields by $|{{\bf p}}|\hat E_L({{\bf p}}) = {{\rm i}}p_+\hat E_+({{\bf p}})+{{\rm i}}p_-\hat E_-({{\bf p}})$ and $|{{\bf p}}|\hat E_T({{\bf p}}) = {{\rm i}}p_+\hat E_-({{\bf p}})-{{\rm i}}p_-\hat E_+({{\bf p}})$ (similarly for $\hat A_s$, $\hat \Pi_s$, and $\hat \Phi_s$). The Gauss law constraint then implies that $|{{\bf p}}|(-\hat E_L+{e_{R}}\hat\Phi_L)$ is zero on the physical space. Fixing the Coulomb gauge, $\hat A_L=0$, and solving the Gauss law, $\hat E_L={e_{R}}\hat\Phi_L$, we obtain the gauge-fixed Hamiltonian $H_{g.f.}$ describing transverse photons ($\hat A_T$) coupled to longitudinal- and transverse [*plasmons*]{} ($\hat\Phi_L$ and $\hat\Phi_T$, respectively). By straightforward methods, this Hamiltonian can be written in the diagonal form $H_{g.f.}=E_0+\sum_{j=1,2,3}\sum_{{{\bf p}}} \omega_j({{\bf p}})b^\dag_j({{\bf p}})b_j({{\bf p}})$ with standard boson operators, i.e. $[b_j({{\bf p}}),b^\dag_{j'}({{\bf p}}')]=\delta_{j,j'}\delta_{{{\bf p}},{{\bf p}}'}$, etc. (see e.g. Appendix C in [@dWL2] for further details), and the groundstate energy $E_0$. The exact dispersion relations $\omega_j({{\bf p}})$ are computed from the eigenvalues of a certain $3\times 3$ matrix. We obtain the following characteristic polynomial of this matrix whose zeros $\lambda=\lambda_j$ are equal to $\omega_j({{\bf p}})^2$ $$\label{cp}
\begin{split}
\lambda(\lambda-\Theta^2-c^2|{{\bf p}}|^2)(\lambda-\Theta^2 -\tilde v_F^2 |{{\bf p}}|^2) +|{{\bf p}}|^4S^2 v_-^2 ( v_+^2(\lambda-c^2|{{\bf p}}|^2)-c^2\Theta^2 )
\end{split}$$ with $v_\pm^2 = v_F^2(1-\gamma_1)(1-\gamma_1\pm\gamma_2)/2$, $\tilde v_F^2 = v_F^2(1-\gamma_1^2)$, $\Theta=\sqrt{v_F(1-\gamma_1)}|{e_{R}}|$, and $S=|\sin(2\varphi)|=|2p_+p_-|/|{{\bf p}}|^2$ ($|{{\bf p}}|^2\equiv p_+^2+p_-^2$). We thus obtain two gapped modes with the same gap proportional to the renormalized charge: $\omega_1(\boldsymbol{0})=\omega_2(\boldsymbol{0})=\Theta$, whereas the third mode is gapless, $\omega_3(\boldsymbol{0})=0$. Moreover, for $c\gg v_F$ and $v_F^2|{{\bf p}}|^2S\ll 1$, $\omega_1({{\bf p}})\approx \sqrt{\Theta^2+c^2|{{\bf p}}|^2}$, $\omega_2({{\bf p}})\approx \sqrt{\Theta^2+\tilde v_F^2|{{\bf p}}|^2}$, and $$\label{omega3}
\omega_3({{\bf p}}) \approx |{{\bf p}}|^2S\sqrt{\frac{c^2v_-^2(\Theta^2+ v_+^2|{{\bf p}}|^2)}{(\Theta^2+c^2|{{\bf p}}|^2)(\Theta^2 + \tilde v_F^2 |{{\bf p}}|^2)}}.$$ This shows that $\omega_1$ gives the energy spectrum of the dressed transverse photon, whereas $\omega_2$ and $\omega_3$ give those of the dressed transverse- and longitudinal plasmons, respectively. Remarkably, the behavior of the last mode is [*qualitatively*]{} different for ${e_{R}}=0$ and ${e_{R}}\neq 0$: in the former case, $\omega_3({{\bf p}}) \approx |{{\bf p}}|S\sqrt{v_+v_-}/\tilde v_F$, which is equal to the lowest-energy mode of the Mattis model [@dWL2], whereas in the latter case, $\omega_3({{\bf p}}) \approx |{{\bf p}}|^2S c v_-/\Theta$ for $p<\Theta/c$. Thus the photons can affect the low-temperature thermodynamical properties of the system, no matter how small $|{e_{R}}|\neq 0$. For example, we found that the temperature ($T$) dependence of the specific heat at low $T$ is linear for ${e_{R}}=0$: $C_v\propto T $, but there are logarithmic corrections for ${e_{R}}\neq 0$: $C_v \propto T\ln(T_0/T)$ with $T_0 = (\pi/{\tilde{a}})^2c\sqrt{v_-}/(1.423\Theta)$.
We also studied the magnetic field response to an external current $J^\mu$ with $J^0=0$, i.e., we computed the linear response function $\hat K(\omega,{{\bf p}})$ in the relation $\langle \hat B(\omega,{{\bf p}})\rangle = \hat K(\omega,{{\bf p}})|{{\bf p}}|\hat J_T(\omega,{{\bf p}})+O(\hat J_T^2)$ ($|{{\bf p}}|\hat J_T$ is the Fourier transform of $\partial_+J_--\partial_-J_+$). The exact result is $$\label{hat K}
\hat K(\omega,{{\bf p}})= \frac{\omega_+^2(\omega_+^2-\Theta^2-\tilde v_F^2|{{\bf p}}|^2) + v_-^2 S^2|{{\bf p}}|^2(\Theta^2+v_+^2|{{\bf p}}^2|)}{(-\omega_+^2+\omega_1({{\bf p}})^2)(-\omega_+^2+\omega_2({{\bf p}})^2)(-\omega_+^2+\omega_3({{\bf p}})^2)}$$ with $\omega_+^2\equiv (\omega + {{\rm i}}0^+)^2$ and $\omega_j({{\bf p}})$ the dispersion relations given above. Expanding $\hat K(\omega,{{\bf p}})$ in partial fractions and inserting the long-distance approximations of $\omega_j({{\bf p}})$ given above, we obtain for $c\gg v_F$, $$\langle \hat B(\omega,{{\bf p}})\rangle \approx \frac{1}{-\omega_+^2+c^2|{{\bf p}}|^2+\Theta^2} |{{\bf p}}|\hat J_T (\omega,{{\bf p}}),$$ ignoring $O(\hat J_T^2)$-terms. This proves that there is a Meissner effect with a London penetration depth $\lambda_L=c/\Theta$. The question if our model has other features of a superconductor, e.g. a fermion condensate, is under investigation.
The exact solution of the model presented in this paper provides a reliable starting point to explore the physics of 2D correlated fermion systems. If the gauge field in this model is to be interpreted as a physical electromagnetic field, one should extend it to three dimensions. Furthermore, if the model is to describe correlated fermions in real materials, one should introduce spin degrees of freedom, and investigate if the gapped fermions affect the ones in the nodal regions. Exactly solvable extensions of our model addressing these remarks will be presented elsewhere.
Acknowledgments {#acknowledgments .unnumbered}
---------------
Useful discussions with F. H. L. Essler, L. B. Ioffe, V. M. Krasnov, J. Lidmar, S. Sachdev, G. W. Semenoff, M. Wallin, L. C. R. Wijewardhana, and K. Zarembo are acknowledged. This work was supported by the Göran Gustafsson Foundation and the Swedish Research Council.
[99]{}
J. G. Bednorz and K. A. Müller, Z. Physik B **64**, 189 (1986)
See e.g.: D. A. Bonn, Nature Physics **2**, 159 (2006)
P. W. Anderson, Science **235** , 1196 (1987)
V. J. Emery, Phys. Rev. Lett. **58**, 2794 (1987)
C. M. Varma, S. Schmitt-Rink, and E. Abrahams, Solid State Comm. **62**, 681 (1987)
F. C. Zhang and T. M. Rice, Phys. Rev. B **37**, 3759 (1988)
J. M. Luttinger, J. Math. Phys. **4**, 1154 (1963)
D. C. Mattis and E. H. Lieb, J. Math. Phys. **6**, 304 (1965)
F. D. M. Haldane, J. Phys. C **14**, 2585 (1981)
See e.g.: A. M. Tsvelik: Quantum Field Theory in Condensed Matter Physics, Cambridge Univ. Press (2003)
T. Holstein, R. E. Norton and P. Pincus, Phys. Rev. B **8**, 2649 (1973)
M. Yu. Reizer, Phys. Rev. B **40**, 11571 (1989)
N. Nagaosa and P. A. Lee, Phys. Rev. Lett. **64**, 2450 (1990)
M. Franz and Z. Tešanović, Phys. Rev. Lett. **87**, 257003 (2001)
For review see: P. A. Lee, N. Nagaosa, X.-G. Wen, Rev. Mod. Phys. **78**, 17 (2006)
H. Grosse, E. Langmann and E. Raschhofer, Annals of Phys. (N.Y.) **253**, 310 (1997)
J. Schwinger, Phys. Rev. **128**, 2425 (1962)
N. S. Manton, Ann. Phys. (N.Y.) **159**, 220 (1985)
D. C. Mattis, Phys. Rev. B **36**, 745 (1987)
R. Hlubina, Phys. Rev. B **50**, 8252 (1994)
A. Luther, Phys. Rev. B **50**, 11446 (1994)
O.F. Sylju[å]{}sen and A. Luther, Phys. Rev. B **72**, 165105 (2005)
E. Langmann, Lett. Math. Phys. **92**, 109 (2010)
E. Langmann, J. Stat. Phys. **141**, 17 (2010)
J. de Woul and E. Langmann, J. Stat. Phys. **139**, 1033 (2010)
J. de Woul and E. Langmann, arXiv:1011.1401v2\[math-phys\]
D. V. Khveshchenko, R. Hlubina and T. M. Rice, Phys. Rev. B **48**, 10766 (1993)
H.-J. Kwon, A. Houghton, and J. B. Marston: Phys. Rev. Lett. **73**, 284 (1994)
N. Furukawa, T. M. Rice, and M. Salmhofer, Phys. Rev. Lett. **81**, 3195 (1998)
For review see: A. Damescelli, Z. Hussain and Z.-X. Shen, Rev. Mod. Phys. **75**, 473 (2003)
See e.g.: K. Sundermeyer: Constrained Dynamics, Lecture Notes in Physics 169, Springer, Berlin (1982)
J. Schwinger, Phys. Rev. Lett. **3**, 296 (1959)
See e.g.: R. Jackiw, Rev. Mod. Phys. **52**, 661 (1980)
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
[Yusuke Ide[^1]]{}\
[Department of Information Systems Creation, Faculty of Engineering, Kanagawa University]{}\
[Kanagawa, Yokohama 221-8686, Japan]{}\
[e-mail: [email protected]]{}\
[Norio Konno]{}\
[Department of Applied Mathematics, Faculty of Engineering, Yokohama National University]{}\
[Hodogaya, Yokohama 240-8501, Japan]{}\
[e-mail: [email protected]]{}\
[Junji Shikata]{}\
[Graduate School of Environment and Information Sciences, Yokohama National University]{}\
[Hodogaya, Yokohama 240-8501, Japan]{}\
[e-mail: [email protected]]{}\
---
[**Abstract**]{}. The Tsallis and Rényi entropies are important quantities in the information theory, statistics and related fields because the Tsallis entropy is an one parameter generalization of the Shannon entropy and the Rényi entropy includes several useful entropy measures such as the Shannon entropy, Min-entropy and so on, as special choices of its parameter. On the other hand, the discrete-time quantum walk plays important roles in various applications, for example, quantum speed-up algorithm and universal computation. In this paper, we show limiting behaviors of the Tsallis and Rényi entropies for discrete-time quantum walks on the line which are starting from the origin and defined by arbitrary coin and initial state. The results show that the Tsallis entropy behaves in polynomial order of time with the parameter dependent exponent while the Rényi entropy tends to infinity in logarithmic order of time independent of the choice of the parameter. Moreover, we show the difference between the Rényi entropy and the logarithmic function characterizes by the Rényi entropy of the limit distribution of the quantum walk. In addition, we show an example of asymptotic behavior of the conditional Rényi entropies of the quantum walk.
Introduction
============
The Tsallis entropy [@Tsallis1988] and the Rényi entropy [@Renyi1961] are important quantities in the information theory, statistics and related fields. The Tsallis entropy is viewed as an one parameter generalization of the Shannon entropy. Also the Rényi entropy includes several useful entropy measures such as the Shannon entropy, Min-entropy and so on, as special cases. In the information theory, information measures such as entropies play a significant role because they measure the quantity of information [@IwamotoShikata2013]. In addition, the (relative) Shannon entropy gives a meaning of the rate function of the Large Deviation Principle (LDP) for random walks.
On the other hand, the discrete-time quantum walks (DTQWs) have been attractive research topics in the last decade as quantum counterparts of the random walks [@Kempe2003; @Kendon2007; @VAndraca2012; @Konno2008b; @AharonovEtAl2001; @AmbainisEtAl2001; @ManouchehriWang2013]. As the random walk plays important roles in various fields, DTQW also plays such roles in various applications, for example, quantum speed-up algorithm [@Ambainis2003; @ChildsEtAl2003; @Ambainis2004; @ShenviEtAl2003] and universal quantum computation [@Childs2009; @LovettEtAl2010]. Recently, the LDP for the one-dimensional DTQW have been shown in [@SunadaTate2012]. In the present stage, the entropic meaning of the rate function for the LDP for DTQW is unknown. Therefore it is important to make clear the roles of entropies in DTQWs.
In this paper, we investigate asymptotic behaviors of the Tsallis and Rényi entropies of DTQWs on the line as a fundamental study. The result for the Rényi entropy is a generalization of the result for the Shannon entropy given by [@IdeEtAl2011]. Therefore it is consistent for the numerical results of [@BrackenEtAl2004; @ChandrashekarEtAl2008] for the Shannon entropy.
The rest of this paper is organized as follows. In Sect. 2, we give the definition of the DTQW and state our results for the Tsallis and Rényi entropies (Theorem [\[thm:renyi\]]{}) and the conditional Rényi entropies (Corollary [\[cor:condrenyi\]]{}). The proof of the Theorem [\[thm:renyi\]]{} is presented in Sect. 3. In this paper, we use two for the base of logarithm functions but the choice of the base is not essential.
Definition and results
======================
The discrete-time quantum walk is a quantum counterpart of the classical random walk which has additional degree of freedom called chirality. For DTQW on the line ${\mathbb{Z}}$ where ${\mathbb{Z}}$ is the set of integers, the chirality takes two values left and right, and it means the direction of the motion of the walker. Now we define the following two dimensional vectors: $$\begin{aligned}
{|L\rangle} =
\left[
\begin{array}{cc}
1 \\
0
\end{array}
\right],
\qquad
{|R\rangle} =
\left[
\begin{array}{cc}
0 \\
1
\end{array}
\right],\end{aligned}$$ where $L$ and $R$ refer to the left and right chirality state, respectively. At each time step, the walker moves one step to the left if it has the left chirality, and if it has the right chirality, it moves one step to the right.
Let $\hbox{U}(2)$ denote the set of $2 \times 2$ unitary matrices. The time evolution of the DTQW on ${\mathbb{Z}}$ is defined by $$\begin{aligned}
U =
\left[
\begin{array}{cc}
a & b \\
c & d
\end{array}
\right] \in \hbox{U}(2),\end{aligned}$$ with $a, b, c, d \in {\mathbb{C}}$ where ${\mathbb{C}}$ is the set of complex numbers. The unitarity of $U$ gives $$\begin{aligned}
|a|^2 + |b|^2 =|c|^2 + |d|^2 =1, \> a \overline{c} + b \overline{d}=0, \> c= - \triangle \overline{b}, \> d= \triangle \overline{a},
\label{konno-eqn:seisitu}\end{aligned}$$ where $\overline{z}$ is the complex conjugate of $z \in {\mathbb{C}}$ and $\triangle = \det U = a d - b c$ with $|\triangle|=1.$ In order to define the dynamics of the model, we divide $U$ into two matrices: $$\begin{aligned}
P =
\left[
\begin{array}{cc}
a & b \\
0 & 0
\end{array}
\right],
\quad
Q =
\left[
\begin{array}{cc}
0 & 0 \\
c & d
\end{array}
\right],\end{aligned}$$ with $U=P+Q$. The matrix $P$ (resp. $Q$) represents the weight of the walker’s movement to the left (resp. right) at each time step. Let $\Xi_{n}(l, m)$ denote the sum of all paths starting from the origin in the trajectory consisting of $l$ steps left and $m$ steps right. In fact, for time $n = l+m$ and position $x=-l + m$, we have $$\begin{aligned}
\Xi_n (l,m) = \sum_{l_j, m_j} P^{l_{1}} Q^{m_{1}} P^{l_{2}} Q^{m_{2}} \cdots P^{l_{n-1}} Q^{m_{n-1}} P^{l_{n}} Q^{m_{n}},\end{aligned}$$ where the summation is taken over all integers $l_j, m_j \ge 0$ satisfying $l_1+ \cdots +l_n=l, \> m_1+ \cdots + m_n = m, \> l_j+ m_j=1$. We should note that the definition gives $$\begin{aligned}
\Xi_{n+1}(l, m) = P \> \Xi_{n}(l-1, m) + Q \> \Xi_{n}(l, m-1).\end{aligned}$$ For example, in the case of $l=3, \> m=1$, we have $$\begin{aligned}
\Xi_4 (3,1) &= QP^3 + PQP^2 + P^2QP + P^3 Q.
$$
The set of initial qubit states at the origin for the DTQW is given by $$\begin{aligned}
\label{eq:initq}
\Phi = \left\{ \varphi =
\alpha {|L\rangle}+\beta {|R\rangle} \in \mathbb C^2 :
|\alpha|^2 + |\beta|^2 =1
\right\}.\end{aligned}$$ The probability that a quantum walker is in position $x$ at time $n$ starting from the origin with $\varphi \in \Phi$ is defined by $$\begin{aligned}
P (X_{n}^{\varphi } =x) = || \Xi_{n}(l, m) \varphi ||^2,\end{aligned}$$ where $n=l+m$ and $x=-l+m$.
The Tsallis entropy $T_{\hat{\alpha }}(X)$ and the Rényi entropy $R_{\hat{\alpha }}(X)$ of order $\hat{\alpha }\in [0,\infty )\setminus \{1\}$ for a random variable $X$ taking values in a finite set $\mathcal{X}$ are given by $$\begin{aligned}
T_{\hat{\alpha }}(X) &= \frac{1}{1-\hat{\alpha }} \left( \sum_{x\in \mathcal{X}} \> P_{X}(x)^{\hat{\alpha }}-1 \right),\\
R_{\hat{\alpha }}(X) &= \frac{1}{1-\hat{\alpha }} \log_2 \left\{ \sum_{x\in \mathcal{X}} \> P_{X}(x)^{\hat{\alpha }} \right\},\end{aligned}$$ where $P_{X}$ is a probability measure on $\mathcal{X}$. Let $P_n^{\varphi } (x) = P (X_{n}^{\varphi } =x)$. Here we define the Tsallis entropy of order $\hat{\alpha }\in [0,\infty )\setminus \{1\}$ for the DTQW at time $n$ as follows: $$\begin{aligned}
\label{eq:defTsallis}
T_{\hat{\alpha }}^{\varphi }(n) = T_{\hat{\alpha }}(X_{n}^{\varphi }) = \frac{1}{1-\hat{\alpha }} \left( \sum_{x=-n}^n \> P_n^{\varphi } (x)^{\hat{\alpha }}-1 \right).\end{aligned}$$ Also we define the Rényi entropy of order $\hat{\alpha }\in [0,\infty )\setminus \{1\}$ for the DTQW at time $n$ as follows: $$\begin{aligned}
\label{eq:defRenyi}
R_{\hat{\alpha }}^{\varphi }(n) = R_{\hat{\alpha }}(X_{n}^{\varphi }) = \frac{1}{1-\hat{\alpha }} \log_2 \left\{ \sum_{x=-n}^n \> P_n^{\varphi } (x)^{\hat{\alpha }} \right\}.\end{aligned}$$ Note that the limit $\lim _{\hat{\alpha }\to 1}R_{\hat{\alpha }}^{\varphi }(n)$ is the Shannon entropy $S_n^{\varphi } = - \sum_{x=-n}^n \> P_n^{\varphi } (x) \log_2 P_n^{\varphi } (x)$. Also $R_{\infty }^{\varphi }:=\lim _{\hat{\alpha }\to \infty}R_{\hat{\alpha }}^{\varphi }(n)$ is equal to $-\log \max _{-n\leq x \leq n}P_n^{\varphi } (x)$ the min-entropy. Remark that the Tsallis entropy and the Rényi entropy defined by Eqs. (\[eq:defTsallis\]) and (\[eq:defRenyi\]) have the following one-to-one correspondence: $$\begin{aligned}
T_{\hat{\alpha }}^{\varphi }(n) = \frac{1}{1-\hat{\alpha }} \left( 2^{(1-\hat{\alpha })R_{\hat{\alpha }}^{\varphi }(n) }-1 \right).\end{aligned}$$ In this paper, we show the following long-time behavior of the Tsallis entropy and the Rényi entropy:
\[thm:renyi\] If the DTQW is determined by $U$ with $abcd \not= 0$, then we have $$\begin{aligned}
\lim_{n\to \infty }\left\{\frac{T_{\hat{\alpha }}^{\varphi }(n) + 1/(1-\hat{\alpha })}{(n/2)^{1-\hat{\alpha }}}-\frac{1}{1-\hat{\alpha }}\right\}
=
\frac{1}{1-\hat{\alpha }}\left( \int_{-|a|}^{|a|} f(x)^{\hat{\alpha }} dx -1 \right),
$$ and $$\begin{aligned}
\lim_{n \to \infty} \log_2 (n/2) \left( \frac{R_{\hat{\alpha }}^{\varphi }(n)}{\log_2 (n/2)} -1 \right)
&= \frac{1}{1-\hat{\alpha }} \log_2 \left\{ \int_{-|a|}^{|a|} f(x)^{\hat{\alpha }} dx \right\}
=:R_{\hat{\alpha }}^{\varphi }(\infty ),
\label{manj2}\end{aligned}$$ for each $\hat{\alpha } \in [0,\infty )\setminus \{1\}$. Here $$\begin{aligned}
f(x) = \frac{|b|}{\pi (1-x^2) \sqrt{|a|^2 - x^2}} \left\{ 1 - \left( |\alpha|^2 - |\beta|^2 + \frac{ a \alpha \overline{b \beta} + \overline{a \alpha} b \beta }{|a|^2} \right) x \right\}.\end{aligned}$$
Theorem \[thm:renyi\] shows that the asymptotic behavior of the Tsallis entropy $T_{\hat{\alpha }}^{\varphi }(n)$ is governed by $(n/2)^{1-\hat{\alpha }}$ while the Rényi entropy $R_{\hat{\alpha }}^{\varphi }(n)$ tends to infinity in order $\log _2(n/2)$ independent of choice of the parameter $\hat{\alpha }$. Moreover, the difference between the Rényi entropy $R_{\hat{\alpha }}^{\varphi }(n)$ and $\log _2(n/2)$ is measured by the Rényi entropy with related to $f(x)$ the limit density function of $X_{n}^{\varphi }/n$ which is obtained by [@Konno2002; @Konno2005]. This situation is consistent with the Shannon entropy case [@IdeEtAl2011]. It is observed that asymptotic behaviors of the Tsallis and Rényi entropies are strongly dependent on that of $X_{n}^{\varphi }/n$ in any choice of the parameter $\hat{\alpha }$.
It is natural to consider the conditional entropy for DTQW for the next step. In this paper, we consider the conditional Rényi entropies. As it is already mentioned by [@IwamotoShikata2013], form the axiomatic point of view, there are five types of definitions of the conditional Rényi entropies of order $\hat{\alpha }\in [0,\infty )\setminus \{1\}$ for the DTQW at time $n$ as follows: $$\begin{aligned}
R_{\hat{\alpha }}^{C}(X_{n}^{\varphi }|Y) &=
\sum_{y\in \mathcal{Y}}P_{Y}(y)R_{\hat{\alpha }}(X_{n}^{\varphi }|Y=y),
\\
R_{\hat{\alpha }}^{JA}(X_{n}^{\varphi }|Y) &=
R_{\hat{\alpha }}(X_{n}^{\varphi },Y) - R_{\hat{\alpha }}(Y),
\\
R_{\hat{\alpha }}^{RW}(X_{n}^{\varphi }|Y) &=
\frac{1}{1-\hat{\alpha }}\max_{y\in \mathcal{Y}} \left[ \log_2 \left\{ \sum_{x=-n}^n \> P_{X_{n}^{\varphi }|Y} (x|y)^{\hat{\alpha }} \right\}\right],
\\
R_{\hat{\alpha }}^{A}(X_{n}^{\varphi }|Y) &=
\frac{\hat{\alpha }}{1-\hat{\alpha }} \log_2 \left[ \sum_{y\in \mathcal{Y}}P_{Y}(y)\left\{ \sum_{x=-n}^n \> P_{X_{n}^{\varphi }|Y} (x|y)^{\hat{\alpha }} \right\}^{1/\hat{\alpha }}\right],
\\
R_{\hat{\alpha }}^{H}(X_{n}^{\varphi }|Y) &=
\frac{1}{1-\hat{\alpha }} \log_2 \left[ \sum_{y\in \mathcal{Y}}P_{Y}(y) \left\{ \sum_{x=-n}^n \> P_{X_{n}^{\varphi }|Y} (x|y)^{\hat{\alpha }} \right\}\right],\end{aligned}$$ where $Y$ is a random variable taking values in a finite set $\mathcal{Y}$.
Now we examine the case that the initial qubit state $\varphi $ of DTQW is randomly chosen by a random variable $Y$ with some probability measure $P_{Y}$ on $\mathcal{Y}\subset \Phi $ defined by Eq. (\[eq:initq\]). In this setting, the conditional probability measure $P_{X_{n}^{\varphi }|Y}$ is the same as $P_n^{\varphi }$ on the condition $\{Y=\varphi \}$. In addition, $ \log_2 \left\{ \sum_{x=-n}^n \> P_n^{\varphi } (x)^{\hat{\alpha }} \right\}=(1-\hat{\alpha })R_{\hat{\alpha }}^{\varphi }(n)$ by the definition of the Rényi entropy Eq. (\[eq:defRenyi\]). Then these conditional entropies are calculated by the following forms: $$\begin{aligned}
R_{\hat{\alpha }}^{C}(X_{n}^{\varphi }|Y) &=
\sum_{\varphi \in \mathcal{Y}}P_{Y}(\varphi )R_{\hat{\alpha }}^{\varphi }(n),
\\
R_{\hat{\alpha }}^{JA}(X_{n}^{\varphi }|Y) &=
\frac{1}{1-\hat{\alpha }} \log_2 \left \{ \sum_{\varphi \in \mathcal{Y}}P_{Y}(\varphi )^{\hat{\alpha }}\cdot 2^{(1-\hat{\alpha })R_{\hat{\alpha }}^{\varphi }(n)} \right \}
- R_{\hat{\alpha }}(Y),
\\
R_{\hat{\alpha }}^{RW}(X_{n}^{\varphi }|Y) &=
\begin{cases}
\displaystyle\max_{\varphi \in \mathcal{Y}}\left \{R_{\hat{\alpha }}^{\varphi }(n) \right \}, &\text{if $0\leq \hat{\alpha } <1$,}\\
\displaystyle\min_{\varphi \in \mathcal{Y}}\left \{R_{\hat{\alpha }}^{\varphi }(n) \right \}, &\text{if $1 < \hat{\alpha }$,}
\end{cases}
\\
R_{\hat{\alpha }}^{A}(X_{n}^{\varphi }|Y) &=
\frac{\hat{\alpha }}{1-\hat{\alpha }} \log_2 \left \{ \sum_{\varphi \in \mathcal{Y}}P_{Y}(\varphi )\cdot 2^{\left((1-\hat{\alpha })/\hat{\alpha } \right)R_{\hat{\alpha }}^{\varphi }(n)} \right \},
\\
R_{\hat{\alpha }}^{H}(X_{n}^{\varphi }|Y) &=
\frac{1}{1-\hat{\alpha }} \log_2 \left \{ \sum_{\varphi \in \mathcal{Y}}P_{Y}(\varphi )\cdot 2^{(1-\hat{\alpha })R_{\hat{\alpha }}^{\varphi }(n)} \right \}.\end{aligned}$$ As a consequence, we have the following corollary by using Theorem \[thm:renyi\]:
\[cor:condrenyi\] If the DTQW is determined by $U$ with $abcd \not= 0$, then we have the limits of the conditional Rényi entropies for each $\hat{\alpha } \in [0,\infty )\setminus \{1\}$. $$\begin{aligned}
\lim_{n\to \infty}\log_{2}(n/2)\left(\frac{R_{\hat{\alpha }}^{C}(X_{n}^{\varphi }|Y)}{\log_{2}(n/2)}-1 \right) &=
\sum_{\varphi \in \mathcal{Y}}P_{Y}(\varphi )R_{\hat{\alpha }}^{\varphi }(\infty ),
\\
\lim_{n\to \infty}\log_{2}(n/2)\left(\frac{R_{\hat{\alpha }}^{JA}(X_{n}^{\varphi }|Y)}{\log_{2}(n/2)}-1 \right) &=
\frac{1}{1-\hat{\alpha }} \log_2 \left \{ \sum_{\varphi \in \mathcal{Y}}P_{Y}(\varphi )^{\hat{\alpha }}\cdot 2^{(1-\hat{\alpha })R_{\hat{\alpha }}^{\varphi }(\infty )} \right \}
- R_{\hat{\alpha }}(Y),
\\
\lim_{n\to \infty}\log_{2}(n/2)\left(\frac{R_{\hat{\alpha }}^{RW}(X_{n}^{\varphi }|Y)}{\log_{2}(n/2)}-1 \right) &=
\begin{cases}
\displaystyle\max_{\varphi \in \mathcal{Y}}\left \{R_{\hat{\alpha }}^{\varphi }(\infty ) \right \}, &\text{if $0\leq \hat{\alpha } <1$,}\\
\displaystyle\min_{\varphi \in \mathcal{Y}}\left \{R_{\hat{\alpha }}^{\varphi }(\infty ) \right \}, &\text{if $1 < \hat{\alpha }$,}
\end{cases}
\\
\lim_{n\to \infty}\log_{2}(n/2)\left(\frac{R_{\hat{\alpha }}^{A}(X_{n}^{\varphi }|Y)}{\log_{2}(n/2)}-1 \right) &=
\frac{\hat{\alpha }}{1-\hat{\alpha }} \log_2 \left \{ \sum_{\varphi \in \mathcal{Y}}P_{Y}(\varphi )\cdot 2^{\left((1-\hat{\alpha })/\hat{\alpha } \right)R_{\hat{\alpha }}^{\varphi }(\infty )} \right \},
\\
\lim_{n\to \infty}\log_{2}(n/2)\left(\frac{R_{\hat{\alpha }}^{H}(X_{n}^{\varphi }|Y)}{\log_{2}(n/2)}-1 \right) &=
\frac{1}{1-\hat{\alpha }} \log_2 \left \{ \sum_{\varphi \in \mathcal{Y}}P_{Y}(\varphi )\cdot 2^{(1-\hat{\alpha })R_{\hat{\alpha }}^{\varphi }(\infty )} \right \},\end{aligned}$$ where $R_{\hat{\alpha }}^{\varphi }(\infty )$ is defined by Eq. (\[manj2\]).
Proof of Theorem [\[thm:renyi\]]{}
==================================
In this section we assume $abcd \not=0$. We consider the following four matrices: $$\begin{aligned}
P =
\left[
\begin{array}{cc}
a & b \\
0 & 0
\end{array}
\right],
\quad
Q =
\left[
\begin{array}{cc}
0 & 0 \\
c & d
\end{array}
\right],
\quad
R =
\left[
\begin{array}{cc}
c & d \\
0 & 0
\end{array}
\right],
\quad
S =
\left[
\begin{array}{cc}
0 & 0 \\
a & b
\end{array}
\right]. \end{aligned}$$ Put $x \wedge y = \min \{x,y \}$. For $l \wedge m \ge 1$, we have $$\begin{aligned}
\Xi_n (l,m)
= a^l \overline{a}^m \triangle^m \sum_{\gamma =1}^{l \wedge m} \left(- \frac{|b|^2}{|a|^2} \right)^{\gamma} {l-1 \choose \gamma-1} {m-1 \choose \gamma -1} \left( \frac{l - \gamma}{a \gamma} P + \frac{m - \gamma}{\triangle \overline{a} \gamma} Q - \frac{1}{\triangle \overline{b}} R + \frac{1}{b} S\right), \end{aligned}$$ by the path counting method [@Konno2002; @Konno2005]. Therefore, for $k \in \{1, \ldots ,[n/2]\}$, we have $$\begin{aligned}
P_{n}^{\varphi}(n-2k)
&
= |a|^{2(n-1)}
\sum_{\gamma =1} ^{k} \sum_{\delta =1} ^{k}
\left(-{|b|^2 \over |a|^2} \right)^{\gamma + \delta}
{k-1 \choose \gamma- 1}
{k-1 \choose \delta- 1}
{n-k-1 \choose \gamma- 1}
{n-k-1 \choose \delta- 1}
\\
&
\qquad \qquad
\times
\left( {1 \over \gamma \delta} \right) \>
\Biggl[
\{ k^2 |a|^2 + (n-k)^2|b|^2 - (\gamma + \delta) (n-k)\} |\alpha|^2
\\
&
\qquad \qquad \qquad \qquad \qquad
+
\{ k^2 |b|^2 + (n-k)^2|a|^2 - (\gamma + \delta) k \} |\beta|^2
\\
&
\qquad \qquad \qquad \qquad \qquad
+ {1 \over |b|^2}
\biggl[
\{ (n-k) \gamma - k \delta + n(2k-n) |b|^2 \} a \alpha \overline{b \beta}
\\
&
\qquad \qquad \qquad \qquad \qquad \qquad
+
\{ -k \gamma +(n-k) \delta + n(2k-n) |b|^2 \} \overline{a \alpha} b \beta +
\gamma \delta
\biggr] \Biggr],\end{aligned}$$ $$\begin{aligned}
P_{n}^{\varphi}(-(n-2k))
&
= |a|^{2(n-1)}
\sum_{\gamma =1} ^{k} \sum_{\delta =1} ^{k}
\left(-{|b|^2 \over |a|^2} \right)^{\gamma + \delta}
{k-1 \choose \gamma- 1}
{k-1 \choose \delta- 1}
{n-k-1 \choose \gamma- 1}
{n-k-1 \choose \delta- 1}
\\
&
\qquad \qquad
\times
\left( {1 \over \gamma \delta} \right) \>
\Biggl[
\{ k^2 |b|^2 + (n-k)^2|a|^2 - (\gamma + \delta) k \} |\alpha|^2
\\
&
\qquad \qquad \qquad \qquad \qquad
+
\{ k^2 |a|^2 + (n-k)^2|b|^2 - (\gamma + \delta) (n-k) \} |\beta|^2
\\
&
\qquad \qquad \qquad \qquad \qquad
+ {1 \over |b|^2}
\biggl[
\{ k \gamma - (n-k) \delta - n(2k-n) |b|^2 \} a \alpha \overline{b \beta}
\\
&
\qquad \qquad \qquad \qquad \qquad \qquad
+
\{ -(n-k) \gamma + k \delta - n(2k-n) |b|^2 \} \overline{a \alpha} b \beta +
\gamma \delta
\biggr] \Biggr],
\\
P_{n}^{\varphi}(n)
&
= |a|^{2(n-1)} \{ |b|^2 |\alpha|^2 +|a|^2 |\beta|^2- (a \alpha \overline{b \beta} + \overline{a \alpha} b \beta ) \},
\\
P_{n}^{\varphi}(-n)
&
= |a|^{2(n-1)} \{ |a|^2 |\alpha|^2 +|b|^2 |\beta|^2+ (a \alpha \overline{b \beta} + \overline{a \alpha} b \beta ) \}.\end{aligned}$$ Here $[x]$ denote the integer part of $x \in {\mathbb{R}}$ where ${\mathbb{R}}$ is the set of real numbers.
Let $P^{\nu, \mu} _n (x)$ denote the Jacobi polynomial which is orthogonal on $[-1,1]$ with respect to $(1-x)^{\nu}(1+x)^{\mu}$ with $\nu, \mu > -1$. Then the following relation holds: $$\begin{aligned}
P^{\nu, \mu} _n (x) = \frac{\Gamma (n + \nu + 1)}{\Gamma (n+1) \Gamma (\nu +1)} \> {}_2F_1(- n, n + \nu + \mu +1; \nu +1 ;(1-x)/2),
$$ where ${}_2F_1(a, b; c ;z)$ is the hypergeometric series and $\Gamma (z)$ is the gamma function. In general, as for orthogonal polynomials, see [@Andrews1999]. Then we have $$\begin{aligned}
\sum_{\gamma =1} ^{k}
\left(- \frac{|b|^2}{|a|^2} \right)^{\gamma -1}
{1 \over \gamma}
{k-1 \choose \gamma- 1}
{n-k-1 \choose \gamma- 1}
&=
\frac{|a|^{-2(k-1)}}{k} P^{1,n-2k} _{k-1}(2|a|^2-1),
\label{yukari1}
\\
\sum_{\gamma =1} ^{k}
\left(- \frac{|b|^2}{|a|^2} \right)^{\gamma -1}
{k-1 \choose \gamma- 1}
{n-k-1 \choose \gamma- 1}
&= |a|^{-2(k-1)} P^{0,n-2k} _{k-1}(2|a|^2-1).
\label{yukari2}\end{aligned}$$ By using Eqs. (\[yukari1\]) and (\[yukari2\]), we see that for $k \in \{1, \ldots ,[n/2]\}$,
$$\begin{aligned}
P_{n}^{\varphi}(\pm (n-2k))
&=
|a|^{2n - 4k -2}|b|^4 / 2\\
&\times
\biggl[
\left\{ {2x^2-2x+1 \over x^2} (P^{1})^2 - {2 \over x} P^{1} P^{0}+
{2 \over |b|^2} (P^{0})^2 \right\} \\
& \qquad \pm \left( {1-2x \over x} \right)
\biggl\{ - {1 \over x}
\{
(|a|^2 - |b|^2)
(|\alpha|^2 - |\beta|^2) + 2 (a \alpha \overline{b \beta} + \overline{a \alpha} b \beta ) \} (P^{1})^2 \\
& \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad -2 \left( |\alpha|^2 - |\beta|^2 -
{ a \alpha \overline{b \beta} + \overline{a \alpha} b \beta \over |b|^2 }
\right) P^{0} P^{1} \biggl\}
\biggr].\end{aligned}$$
where $P^{i} = P^{i,n-2l} _{l-1}(2|a|^2-1) \> (i=0,1)$. Let $a(n) \sim b(n)$ means $a(n)/b(n) \to 1$ as $n \to \infty$. As it is mentioned in [@Konno2002; @Konno2005], if $n\to \infty $ with $k/n\in (-(1-|a|)/2, (1+|a|)/2)$, then $$\begin{aligned}
P^{0}&\sim \frac{2|a|^{2k-n}}{\sqrt{\pi n\sqrt{-\Lambda }}}\cos(An+B), \\
P^{1}&\sim \frac{2|a|^{2k-n}}{\sqrt{\pi n\sqrt{-\Lambda }}}\sqrt{\frac{x}{(1-x)(1-|a|^2)}}\cos(An+B+\theta ),\end{aligned}$$ where $\Lambda =(1-|a|^2)\{(2x-1)^2-|a|^2\}$, $A$ and $B$ are some constants which are independent of $n$, and $\theta \in [0,\pi/2]$ is determined by $\cos \theta =\sqrt{(1-|a|^2)/4x(1-x)}$. By these asymptotics and the Riemann-Lebesgue lemma, we obtain $$\begin{aligned}
R_{\hat{\alpha }}^{\varphi}(n)
&= \frac{1}{1-\hat{\alpha } } \log_2 \left\{ \sum_{x=-n}^n \> P_n^{\varphi} (x)^{\hat{\alpha }} \right\}
\\
&\sim \frac{1}{1-\hat{\alpha }} \log_2 \left\{ \int_{(1-|a|)/2}^{(1+|a|)/2} f(1-2x)^{\hat{\alpha }} dx \times \frac{ 1 }{ n^{ \hat{\alpha }-1 } }\right\}
\\
&= \log_2 n + \frac{1}{1-\hat{\alpha }} \log_2 \left\{ \int_{-|a|}^{|a|} f(x)^{\hat{\alpha }} dx \times 2^{ \hat{\alpha }-1 } \right\}
\\
&= \log_2 (n/2) + \frac{1}{1-\hat{\alpha }} \log_2 \left\{ \int_{-|a|}^{|a|} f(x)^{\hat{\alpha }} dx \right\}.\end{aligned}$$ By the same argument, we have $$\begin{aligned}
T_{\hat{\alpha }}^{\varphi}(n) + \frac{1}{1-\hat{\alpha}}
= \frac{1}{1-\hat{\alpha } } \sum_{x=-n}^n \> P_n^{\varphi} (x)^{\hat{\alpha }}
\sim \left(\frac{n}{2}\right)^{ 1-\hat{\alpha }} \times \frac{1}{1-\hat{\alpha }} \int_{-|a|}^{|a|} f(x)^{\hat{\alpha }} dx,\end{aligned}$$ as $n \to \infty$. This completes the proof.
Summary
=======
In this paper, we show a limit theorem for the Tsallis and Rényi entropies of the DTQWs on ${\mathbb{Z}}$ starting from the origin with arbitrary coin and initial state by using a path counting method. The result shows that asymptotic behaviors of both entropies are strongly dependent on that of scaling limit of the walker’s position in any choice of the parameter. It is an interesting future problem that building some novel information theoretic scheme by using DTQWs and these entropies. Also searching for the entropic meaning of the LDP for the DTQW is one of important future problems to be solved.
[**Acknowledgments.**]{} This work was partially supported by the Grant-in-Aid for Scientific Research (C) of Japan Society for the Promotion of Science (Grant No. 24540116).
[000]{}
Aharonov, D., Ambainis, A., Kempe, J., Vazirani, U.: Quantum walks on graphs. In: Proceedings of the 33rd Annual ACM Symposium on Theory of Computing, pp. 50–59 (2001).
Ambainis, A.: Quantum walks and their algorithmic applications. Int. J. Quantum Inf. [**1**]{}, 507–518 (2003).
Ambainis, A.: Quantum walk algorithm for element distinctness. In: Proceedings of the 45th Annual IEEE Symposium on Foundations of Computer Science, pp. 22–31 (2004).
Ambainis, A., Bach, E., Nayak, A., Vishwanath, A., Watrous, J.: One-dimensional quantum walks. In: Proceedings of the 33rd Annual ACM Symposium on Theory of Computing, pp. 37–49 (2001).
Andrews, G. E., Askey, R., Roy, R.: Special Functions. Cambridge University Press (1999).
Bracken, A. J., Ellinas, D., Tsohantjis, I.: Pseudo memory effects, majorization and entropy in quantum random walks, J. Phys. A : Math. Gen. [**37**]{}, L91–L97 (2004).
Chandrashekar, C. M., Srikanth, R., Laflamme, R.: Optimizing the discrete time quantum walk using a $SU(2)$ coin. Phys. Rev. A [**77**]{}, 032326 (2008).
Childs, A. M.: Universal computation by quantum walk. Phys. Rev. Lett. [**102**]{}, 180501 (2009).
Childs, A. M., Cleve, R., Deotto, E., Farhi, E., Gutmann, S., Spielman, D. A.: Exponential algorithmic speedup by quantum walk. In: Proceedings of the 35rd Annual ACM Symposium on Theory of Computing, pp. 59–68 (2003).
Ide, Y., Konno, N., Machida, T.: Entanglement for discrete-time quantum walks on the line. Quant. Inf. Comput. [**11**]{}, 0855–0866 (2011).
Iwamoto, M., Shikata, J.: Information Theoretic Security for Encryption Based on Conditional Rényi Entropies. The 7th International Conference on Information Theoretic Security (ICITS2013), LNCS 8317, pp.103-121, Springer, November 2013 (2013). The full version is entitled “Revisiting Conditional Rényi Entropies and Generalizing Shannon’s Bounds in Information Theoretically Secure Encryption”, and available at http://eprint.iacr.org/2013/440 .
Kempe, J.: Quantum random walks - an introductory overview. Contemporary Physics [**44**]{}, 307–327 (2003).
Kendon, V.: Decoherence in quantum walks - a review. Math. Struct. in Comp. Sci. [**17**]{}, 1169–1220 (2007).
Konno, N.: Quantum random walks in one dimension. Quant. Inform. Process [**1**]{}, 345–354 (2002).
Konno, N.: A new type of limit theorems for the one-dimensional quantum random walk. J. Math. Soc. Jpn. [**57**]{}, 1179–1195 (2005).
Konno, N.: Quantum Walks. In: Quantum Potential Theory, Franz, U., and Schürmann, M., Eds., Lecture Notes in Mathematics: Vol. 1954, pp. 309–452, Springer-Verlag, Heidelberg (2008).
Lovett, N. B., Cooper, S., Everitt, M., Trevers, M., Kendon, V.: Universal quantum computation using the discrete-time quantum walk. Phys. Rev. A [**81**]{}, 042330 (2010).
Manouchehri, K., Wang, J.B.: Physical Implementation of Quantum Walks. Springer-Verlag, Heidelberg (2013).
Rényi, A.: On measures of entropy and information. In: Proceedings of the 4th Berkeley Symposium on Mathematics, Statistics and Probability 1960, pp. 547–561 (1961).
Shenvi, N., Kempe, J., Whaley, K. B.: Quantum random-walk search algorithm. Phys. Rev. A [**67**]{}, 052307 (2003).
Sunada, T., Tate, T.: Asymptotic behavior of quantum walks on the line. J. Funct. Anal. [**262**]{}, 2608–2645 (2012).
Tsallis, C.: Possible generalization of Boltzmann-Gibbs statistics. J. Stat. Phys. [**52**]{}, 479–487 (1988).
Venegas-Andraca, S.E. Quantum walks: A comprehensive review. Quant. Inform. Process [**11**]{}, 1015–1106 (2012).
[^1]: To whom correspondence should be addressed. E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'H. Ogawa,'
- 'K. Abe,'
- 'M. Matsukura'
- 'and H. Mimura'
title: 'Development of Low Radioactive Molecular Sieves for Ultra-Low Background Particle Physics Experiment'
---
Introduction {#sec:intro}
============
Liquid xenon (LXe) has recently attached the attention of the research due to its application as noble gas detector media for ultra-low background particle physics, such as direct dark matter search experiment [@Aprile]. The scintillation light of LXe is as large as that of NaI. The wavelength of scintillation light is about 175 nm. Xenon purification is an important task in LXe detector because the scintillation light is absorbed by impurities like, for example, oxygen and water.
One major device for impurity absorption is molecular sieve (MS), also called Zeolite, which is well- known and commercially distributed. There are several types of MS such as 3A, 4A, 5A and 13X depending on the pore size. The uniform pore size is enough for the selective absorption of impurity. MEG experiments performed xenon purification by applying MS filtering and liquid circulation [@MEG]. MS can be used for purification of not only xenon but also other noble gases, such as liquid argon, which is also the detector media for dark matter search experiment. Besides, MS adsorbs radioactive impurities, for example radon [@CYG], from the rare gas. Consequently, MS is useful in the purification of noble gas for dark matter search experiment.
For the perspective of the low- background experiment, the commercial MS includes a large amount of radioactive isotope, especially beta and gamma rays originating from the decay of radon are the main background candidate for dark matter search experiment. Radon is originated as gas from the MS and then it diffuses into a detector. In order to use the MS for purification in dark matter search experiments, we reduce the radioactive isotope in MS.
This study develops low radioactive MS, which can be applied for dark matter search experiment. This study is a joint research with UNION SHOWA K.K. The radioactivity of the materials for MS and those developed is measured using HPGe and pin-photo type radon detectors. The rest of this paper is organized as follows. Section \[sec:target\] demonstrates the target radioactivity of MS. Section \[sec:develop\] reports the actual development of 4A MS. Section \[sec:meas\] presents the achievement of low activity MS. Finally, Section \[sec:conc\] concludes the paper.
The Target of Radioactivity for MS Development {#sec:target}
==============================================
The typical target of LXe impurity (water, oxygen and others) is required to be $<$1 ppb. In this study, the performance of the developed MS is targeted to have the same level of adsorption capacity as that of commercial MS. Then, the amount of water absorption needed in MS more than 20 g for a 100 g MS.
As radon source, $^{222}$Rn (with a half-life of 3.82 days) of uranium series and $^{220}$Rn (with a half-life of 55.6 s) of thorium series are considered. In the case of dark matter search experiment, the activity of $^{222}$Rn should be less than 1 mBq/ton in a noble gas detector with the background level ${\rm 1 \times 10^{-4} \, kg^{-1} keV^{-1} day^{-1}}$ for $<$ 100 keV, which covers the energy region for dark matter search. Now we assume that amount of MS in purification filter is about 100 g. The radioactivity in MS requires 10 mBq/kg for $^{226}$Ra which is parent isotope of $^{222}$Rn. $^{220}$Rn has a very short half-life when compared with $^{222}$Rn. It is possible to reduce the decay of $^{220}$Rn in the detector space by, for example, installing the buffer space between detector and filter. Thus the target value of the activity can be relaxed. In this study, a target value of 100 mBq/kg or less in MS (with a margin, $^{222}$Rn target value $\times$ 10) was set for the concentration of $^{220}$Rn and $^{232}$Th.
The radioactivity of commercial MS was measured using a HPGe detector. The radioactivities of $^{226}$Ra and $^{232}$Th in commercial MS is about 5.5 and 7.0 Bq/kg, respectability. Enormous radioactivity was observed, and it introduced the large radon emanation from the MS filter. It is necessary to develop a MS that reduces $^{226}$Ra to about 1/500 or less from a commercial MS. Commercial MS is configured by main component to absorb impurity and the binder component which make the shape using natural clay with large radioactivity. MS without binder still includes about 1.2 Bq/kg concentration of $^{226}$Ra. We select both the low radio-isotope (RI) material for MS and substitution of the binder.
Material Selection and Development of MS {#sec:develop}
========================================
The basic material for MS4A is NaOH, Al(OH)$_{3}$ and silica component. The selected NaOH was developed by Wako Junyaku Co. Ltd. and is a special grade reagent. Al(OH)$_{3}$, also called BHP39, was developed by Nihon Keikinzoku Co. Ltd. and it is a further refinement of the imported Al(OH)$_{3}$. Meanwhile, we selected the Snowtechs ST-30 developed by Nissan Kagaku Co. Ltd., and about 30 $\%$ silica component is included. Silica component was also used as a solidifying agent instead of clay component. Table \[table:selRI\] presents the selected materials and their radioactivities for $^{226}$Ra and $^{232}$Th measured using HPGe detector. The measured materials were transferred into a polyethylene (PE) vessel and the grove box was filled with pure air to guarantee a low radon level by avoiding contamination from the radon in the room air. Furthermore, vessels are inserted into the ethylene-vinyl alcohol copolymer (EVOH) bug. We prepared the empty PE vessel to evaluate the background in HPGe detector measurement. The concentration of $^{226}$Ra takes the value of the upper limit of NaOH and Al(OH)$_{3}$. We checked the various kinds of samples for Al(OH)$_{3}$. BHP39 is a lower concentration of $^{226}$Ra. The significant $^{226}$Ra and $^{232}$Th sources still remain in the silica component. Another candidate of silica component, which is reagent Na$_{2}$SiO$_{3}$ has 10 times the concentration of $^{226}$Ra compared with Snowtechs ST-30. This was selected. The water is required for the MS development. The ion exchange pure water was supplied from Organo Corporation. The concentration of uranium is less than 2 ppt measured using Inductively Coupled Plasma Mass Spectrometry (ICP-MS) supplied by Toshiba Nanoanalysis Co.Ltd.
\[table:selRI\]
material $^{226}$Ra\[mBq/kg\] $^{232}$Th\[mBq/kg\] Company/Commercial name
------------------ ---------------------- ---------------------- --------------------------------
NaOH $<$12.2 $<$8.14 WAKO/NaOH for precise analysis
Al(OH)3 $<$9.1 $<$4.26 Nihon Keikinzoku / BHP39
Silica component 19.6$\pm$0.3 93.4$\pm$4.3 Nissan Kagaku/Snowtechs ST-30
The new MS was developed at Nihon University with the supervision of UNION SHOWA K.K.. All the devises used for the development were washed with pure water to keep them clean and avoid contamination by outside impurities. The four steps for MS production were as follows. First we produced the aluminate configured from NaOH, Al(OH)$_{3}$ and pure water. The component of the developed aluminate was checked by specific gravity and titration measurement. Second, the crystal formation process of MS4A by mixing of aluminate, silica component and pure water. Third, the drying process for making MS powder. Forth, the formation of shape by utilizing a solidifying agent with silica component and pure water using electric furnace. In these processes, the temperature and the time of sample stirring, drying, mixture etc., were controlled based on the recipes provided by UNION SHOWA K.K.. The development processes was performed three times and the developed MS is separated for each developed process. The total amount of developed MS was about 450 g.
The ratio of crystallization and water adsorption capacity were measured for the developed MS4A powder by UNION SHOWA K.K., as presented in Table \[table:UNION\]. In sample 2, the crystal formation process was different when compared with samples 1 and 3. Based on this, water adsorption capacity is reduced too but it is still enough. The performance of the developed MS is comparable to that of a commercial MS.
Sample 1 was solidified after the performance check and HPGe measurement in section \[sec:meas\]. We mixed silica component and pure water with the developed MS powder. The mixture sintered in an electric furnace, and the solidified sample was crushed. A sample of 120 g was collected.
\[table:UNION\]
sample name sample 1 sample 2 sample 3 commercial MS
--------------------------------------- ---------- ---------- ---------- ---------------
water adsorption capacity \[g/100g\] 23.2 18.6 23.3 $>$ 20
the ratio of crystallization \[$\%$\] 69.4 53.7 70.0 $>$ 80
The Radioactivity Measurement of the Developed MS {#sec:meas}
=================================================
The radioactivities of the developed MS4A were measured by applying two methods. One is HPGe detector measurement, whereas the other is pin-photo type radon detector measurement. These devices are owned by Institute for Cosmic Ray Research, University of Tokyo and Kavli Institute for the Physics and Mathematics of the Universe, University of Tokyo, respectively.
For the HPGe measurement, the sample treatment follows the same method as in Section \[sec:develop\]. Since sample 3 was utilized to test the development of the other type MS, only samples 1 and 2 were measured. Additionally, sample 1 was measured again after it solidified and was referred to as sample 1b. Table \[table:devMS4A\] presents the results of the RI measurement. The results of samples 1 and 2 agree with the sum of the results of RI measurement of the material (Table \[table:selRI\]). Moreover, sample 1b has a larger activity. The increase in the $^{226}$Ra and $^{232}$Th in sample 1b can be explained by the content in silica component applied for solidification.
The radon emanation was measured for the MS filter after HPGe measurement. We placed 120 g of sample 1b in the filter housing, which comprises ConFlat (CF) nipples, metal particle filters and valves. The solidified MS was baked at about 200 $^{\circ}$C for 1 h in the filter housing to release water inside the MS. The left picture on Figure \[fig:ema\] shows the setup of $^{222}$Rn emanation measurement. Radon detector and filter housing are connected by Stainless Used Steel (SUS) pipes, valves, dew point meter, flow meter and air-tight sealing compressor.
Radon detector is configured by applying pin-photo dynode with electronic field to capture the daughters of $^{222}$Rn ( $^{218}$Po and $^{214}$Po) [@RnDet]. We detect alpha particle from the decay of radon daughters. Radon concentration was estimated using energy region of alpha from $^{214}$Po decay in the spectrum and collection efficiency that has dew point dependence in the setup. The radon detector, filter housing and all devices were vacuumed to less than 10 Pa. After pure air guaranteed low radon level was filled in the setup by 0.01 MPaG.
After the preparation, measurement of the $^{222}$Rn emanation was started. The air inside the detector and sample vessel was circulated by air circulation pump of about 0.8 L/min. The measurement was continued for about 20 days. The right plot on Figure \[fig:ema\] demonstrates the time profile for $^{222}$Rn emanation measurement. The emanation rate was estimated from time variation of radon rate with a remained background in radon detector by following equation: $$R=A\cdot(1-e^{-t/\tau})+B\cdot e^{-t/\tau},
\label{eq:fitting}$$ where $R$ is the observed radon rate. $A$ denotes the radon emanation rate from the material. Additionally, $B$ denotes the remaining radon rate in the detector at the beginning of the measurement ($\tau = 3.82/ln2 = 5.52$ days is the lifetime of $^{222}$Rn), and $t$ denotes the time since the start of the measurement. The time profile for radon is fitted by applying Equation \[eq:fitting\]. The radon emanation $A$ of sample 1b filter is estimated as 2.1$\pm$0.1 mBq. Background, which is emanated from the radon detector and setup, is independently measured without a sample less than 0.1 mBq. The decrease from the activity of $^{226}$Ra in HPGe measurement is due to the effect of radon diffusion in the MS.
\[table:devMS4A\]
sample name $^{226}$Ra\[mBq/kg\] $^{232}$Th\[mBq/kg\]
------------------------- ---------------------- ----------------------
sample 1 22.6$\pm$7.9 91.1$\pm$8.9
sample 1b (solidifying) 57.0$\pm$14.0 198.4$\pm$16.5
Sample 2 22.8$\pm$9.2 92.4$\pm$10.4
![Left: The setup of $^{222}$Rn emanation measurement. Right: Time profile of $^{222}$Rn emanation measurement for sample 1b. Red dots is measured emanation rate. Green, black dash and red dash lines indicate the fitting result of $R$, first term and second term in Equation \[eq:fitting\], respectively.[]{data-label="fig:ema"}](fig-radon.png){height="45mm"}
Conclusion {#sec:conc}
==========
This study developed a very low radioactive MS that can be applied for dark matter search experiments. The MS manufactured is 4A type. The concentration of $^{226}$Ra and $^{232}$Th in powder type is about 57 and 198 mBq/kg measured with HPGe detector, respectively. The reduction of the radioactivity from the commercial MS is about 99 and 97 percent for Ra and Th, respectively. The estimated radon emanation rate from MS filter is about 2.1 mBq using radon detector measurement. This study demonstrated that low RI MS can be made by applying material-screening techniques.
We need further improvement to lower radioactive MS. $^{226}$Ra and $^{232}$Th still remained in the developed MS. As shown in Table \[table:selRI\], the silica component is main source of RI. For the silica component, we will explore to lower activity one. In addition, the 5A type MS with $\sim$ 5 Å pore size, which is expected for the removal of radon from gas, shown in [@CYG], can be produced by ion-exchange of 4A with Ca compound. Consequently, it is important to obtain or synthesize Ca compound of very low activity.
Acknowledgments {#acknowledgments .unnumbered}
===============
We gratefully acknowledge the cooperation of Organo Corporation and the cooperation of Kamioka Mining and Smelting Company. This work was supported by the Japanese Ministry of Education, Culture, Sports, Science and Technology, the joint research program of the Institute for Cosmic Ray Research (ICRR), Grant-in-Aid for Scientific Research, JSPS KAKENHI Grant Number, 19GS0204, 19H05806, 19K03893 and XMASS collaboration.
[00]{} E. Aprile and T. Doke, *Liquid xenon detectors for particle physics and astrophysics*, *Rev.Mod.Phys.*, [**82**]{} (2010) 2053-2097. S. Mihara et al., *Development of a method for liquid xenon purification using a cryogenic centrifugal pump*, *Cryogenics* [**46**]{} (2006) 688-693. A.C. Ezeribe et al., *Demonstration of radon removal from SF$_{6}$ using molecular sieves*, 2017 *JINST* [**12**]{} P09025. K.Hosokawa et al., *Development of a high-sensitivity 80 L radon detector for purified gases*, *Prog. Theor. Exp. Phys.* (2015) 033H01.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The vanilla Generative Adversarial Networks (GAN) are commonly used to generate realistic images depicting aged and rejuvenated faces. However, the performance of such vanilla GANs in the age-oriented face synthesis task is often compromised by the mode collapse issue, which may result in the generation of faces with minimal variations and a poor synthesis accuracy. In addition, recent age-oriented face synthesis methods use the L1 or L2 constraint to preserve the identity information on synthesized faces, which implicitly limits the identity permanence capabilities when these constraints are associated with a trivial weighting factor. In this paper, we propose a method for the age-oriented face synthesis task that achieves a high synthesis accuracy with strong identity permanence capabilities. Specifically, to achieve a high synthesis accuracy, our method tackles the mode collapse issue with a novel Conditional Discriminator Pool (CDP), which consists of multiple discriminators, each targeting one particular age category. To achieve strong identity permanence capabilities, our method uses a novel Adversarial Triplet loss. This loss, which is based on the Triplet loss, adds a ranking operation to further pull the positive embedding towards the anchor embedding resulting in significantly reduced intra-class variances in the feature space. Through extensive experiments, we show that our proposed method outperforms state-of-the-art methods in terms of synthesis accuracy and identity permanence capabilities, qualitatively and quantitatively.'
author:
- 'Haoyi Wang, Victor Sanchez, Chang-Tsun Li, [^1][^2][^3]'
bibliography:
- 'refs.bib'
title: 'Age-Oriented Face Synthesis with Conditional Discriminator Pool and Adversarial Triplet Loss'
---
[Shell : Bare Demo of IEEEtran.cls for IEEE Journals]{}
age-oriented face synthesis, generative adversarial networks, mode collapse, triplet loss
Introduction
============
face synthesis (AOFS) is a generative task aiming to generate older and younger faces by rendering facial images with natural aging and rejuvenating effects. An efficient AOFS method can be integrated into a wide range of forensic and commercial applications, e.g., tracking persons of interest like suspects or missing children over a long time span, predicting the outcomes of a cosmetic surgery, and generating special visual effects on characters of video games, films and dramas [@fu2010age; @lanitis2002toward]. The synthesis in recent works [@li2019global; @wang2018face; @yang2018learning; @zhang2017age] is usually conducted among age categories (e.g., the 30s, 40s, 50s) rather than specific ages (e.g., 32, 35, 39) since there is no noticeable visual change of a face over a few years.
The vanilla Generative Adversarial Network (GAN) [@goodfellow2014generative] is commonly used as the backbone of several state-of-the-art AOFS methods [@li2019global; @antipov2017face; @genovese2019towards; @pantraki2018face; @zhao2019look]. One of the biggest advantages of the vanilla GAN over other generative methods, like the Variational Autoencoder [@kingma2013auto], is that it can generate sharp and realistic images by playing a minimax game between the generator and the discriminator. However, the vanilla GAN suffers from the mode collapse issue caused by the vanishing gradient due to the involvement of the negative log-likelihood loss [@arjovsky2017towards]. Specifically, once the discriminator converges, the loss does not penalize the generator any further [@che2016mode]. This allows the generator to find a specific mode (i.e., a distribution) that can easily fool the discriminator [@bishop2006pattern]. The mode collapse issue may also occur in the AOFS task, where a mode is represented by an age category. Within this context, the vanilla GAN may generate faces with limited variations as exemplified in Fig. \[fig:demo\], resulting in poor synthesis accuracy.
To boost the state-of-the-art performance in the AOFS task, this work proposes an AOFS method that includes two novel components. Namely, a Conditional Discriminator Pool (CDP) and an Adversarial Triplet loss. The proposed CDP helps to achieve a high synthesis accuracy by alleviating the mode collapse issue. Specifically, it allows learning multiple modes (i.e., age categories) explicitly and independently to generate realistic faces with a wide range of variations. Our CDP comprises multiple feature-level discriminators that learn the transformations from the source age category to the target age category. For each transformation, only the feature-level discriminator associated with the target age category is used. As a result, each feature-level discriminator only needs to learn one age category throughout the entire training process. The proposed Adversarial Triplet loss helps to preserve the identity information in the synthesized faces. This loss, which improves the Triplet loss \[53\], uses an additional ranking operation that can further optimize the distances within a triplet of feature embeddings comprising an *anchor*, a *positive* and a *negative*. Specifically, it helps to bring the *positive* much closer to the *anchor*, while guaranteeing that the distance between the *anchor* and the *negative* is larger than that between the *anchor* and the *positive*. The additional ranking operation forces the triplets to a play zero-sum game \[5\] during training. As a result, our Adversarial Triplet loss yields high-density clusters with dramatically reduced intra-class variances in the feature space.
![A demonstration of face aging. The top row depict images generated by a vanilla GAN suffering from the mode collapse issue. The bottom row depicts images generated by the proposed AOFS method.[]{data-label="fig:demo"}](demo.pdf){width="0.95\linewidth"}
Our contributions can be summarized as follows.
- We study the mode collapse issue in the AOFS task. To the best of our knowledge, our work is the first to tackle the AOFS task from the aspect of mode learning.
- To address the mode collapse issue in the vanilla GAN and attain a high synthesis accuracy, we propose the CDP, which allows our AOFS method to learn multiple modes explicitly and independently.
- To preserve the identity information in the synthesized images, we propose the Adversarial Triplet loss. Smaller intra-class variance can be achieved by forcing triplets to play zero-sum games during training.
- We extensively evaluate the proposed AOFS method on several AOFS benchmark datasets to show that it can precisely transform faces to the target age category while preserving the identity information robustly.
The rest of this paper is organized as follows. In Section 2, we review the related works on GANs, especially those tackling the mode collapse issue. In this Section, we also review the Triplet loss and the state-of-the-art AOFS methods. In Section 3, we present details of the proposed AOFS method including the CDP and the Adversarial Triplet loss. In Section 4, we explain the experimental settings and discuss the performance results on several AOFS benchmark datasets. Finally, we conclude our work in Section 5.
Related Work
============
To lay the foundation of our work, in this section we first discuss the mode collapse issue in the vanilla GAN and some of the solutions that have been proposed to tackle it. Then, we review the Triplet loss and show the main differences between the proposed Adversarial Triplet loss and other variations. Finally, we discuss some state-of-the-art AOFS methods.
Mode collapse in GANs
---------------------
The vanilla GAN, which is introduced by Goodfellow *et al.* [@goodfellow2014generative], is capable of generating sharp and realistic images by playing a minimax game between its generator and its discriminator. When training the vanilla GAN, the generator and the discriminator try to reach a Nash equilibrium [@metz2017unrolled] by minimizing the negative log-likelihood loss and minimizing the JS-divergence [@lin1991divergence]. However, the involvement of the negative log-likelihood loss may cause the discriminator to converge faster than the generator [@heusel2017gans]. Once the discriminator finds its global minima, the loss function stops penalizing the generator [@che2016mode]. This is also known as the vanishing gradient problem [@arjovsky2017towards; @fedus2017many; @jolicoeur2019relativistic] and is the main cause of the mode collapse issue. Since the parameters in the discriminator are not further updated, the generator may then find a specific mode that can easily fool the discriminator. When such an issue occurs, the vanilla GAN can only generate limited varieties of samples. Solving this mode collapse issue has become one of the most trending research topics on GANs.
Since the mode collapse issue is caused by the vanishing gradient problem due to the involvement of the negative log-likelihood loss, one strategy to alleviate it is to use an alternative loss function that minimizes a different divergence. Nowozin *et al.* [@nowozin2016f] first show that the optimization of GANs is a general process that can be done by minimizing any $f$-divergence [@csiszar2004information; @liese2006divergences], which is a family of divergences aiming to minimize the distance between two distributions. Some commonly used members of the $f$-divergence family are the JS-divergence, the Kullback-Leibler divergence (KL-divergence) [@kullback1951information], the squared Hellinger divergence, and the Pearson $\chi^2$ divergence [@pearson1900x]. The authors show that GANs trained with other divergences, like the KL-divergence or the squared Hellinger divergence, can generate images with more variations compared to those generated by the vanilla GAN. Although the work in [@nowozin2016f] does not tackle the mode collapse issue directly, it shows the possibility of using other loss functions to optimize GANs.
Arjovsky *et al.* [@arjovsky2017wasserstein] propose the Wasserstein GAN (WGAN) and use the Wasserstein or Earth-Mover (EM) distance to calculate the distance between distributions of the real and synthesized data. Intuitively, the EM distance computes the cost of transforming one distribution to another, which is more sensitive to the difference between two distributions [@arjovsky2017wasserstein]. Therefore, even if the discriminator is well-trained, it can still keep rejecting the data synthesized by the generator. The Least Square GAN (LSGAN) [@mao2017least], on the other hand, replaces the negative log-likelihood loss by the L1 loss. Minimizing the L1 loss is equivlent to minimizing the Pearson $\chi^2$ divergence, which can produce overdispersed approximations and thus makes the LSGAN less mode-seeking [@dieng2017variational; @mao2018effectiveness].
Although the methods discussed before may alleviate the mode collapse issue, their discriminators still have to learn from all the modes. Therefore, recently proposed methods now focus on modifying the GAN structure. For example, Nguyen *et al.* [@nguyen2017dual] propose the Dual Discriminator Generative Adversarial Nets (D2GAN) where each discriminator favors data from a different distribution. By using this strategy, their method can compute the KL and reverse KL divergence simultaneously, which in turn increases the variety of samples. Based on this idea, Zhang *et al.* [@zhang2018convergence] propose a D2GAN variation with two customized discriminators. Specifically, one discriminator consists of residual blocks to form a deep network aiming to increase the variety of generated samples. The other discriminator uses the scaled exponential linear unit (SELU) function [@klambauer2017self] as the non-linear activation function. Adopting the SELU function guarantees that this discriminator produces a non-zero value even if the distributions of the synthesized and real data are similar. The authors further propose the D2PGGAN [@zhang2019d2pggan] to stabilize the training by leveraging the idea of progressively increasing the complexity of the generator [@karras2018progressive]. Durugkar *et al.* [@durugkar2016generative] propose a GAN with multiple discriminators. Their method may alleviate the mode collapse issue to some extent since the generator has to fool a set of discriminators, which in turn makes the generated samples diverse. It is important to note that by introducing additional discriminators in parallel, the aforementioned methods are also more computationally complex than their plain counterparts (e.g., the vanilla GAN). On the contrary, by selecting a particular discriminator from a discriminator pool, our CDP only uses one discriminator for each transformation, which does not increase the computational complexity.
{width="98.00000%"}
Triplet Loss
------------
The Triplet loss is proposed in [@schroff2015facenet] aiming to learn feature embeddings for images by optimizing the geometric relationship, in the feature space, within a triplet consisting of an *anchor*, a *positive* and a *negative*. Within this context, the *anchor* and *positive* represent feature embeddings of the same class and the *negative* represents a feature embedding of a different class. The goal is to minimize the distance between the *anchor* and the *positive* and simultaneously push the *negative* away from the *anchor*. Since then, a number of variations to this loss have been proposed. For instance, Chen *et al.* [@chen2017beyond] uses an additional *negative* embedding alongside the original triplet to form a quadruplet. Huang *et al.* [@huang2016learning] implement three ranking operations in total by using an *anchor*, a *negative* and three *positives*. Ye *et al.* [@ye2018visible], on the other hand, adopt additional images from other modalities. It is worth noting that all these variants leverage additional samples either within the same or from another modality. Therefore, these losses can no longer help to optimize the geometric relationship within a triplet. This is explained in detail in Section III.D.
We find that the original Triplet loss produces clusters with large intra-class variances that can be further optimized. To produce high-density clusters, we add another ranking operation and propose the Adversarial Triplet loss to pull the *positive* closer to the *anchor*. It is worth noting that compared to the aforementioned Triplet loss variants, our Adversarial Triplet loss still focuses on optimizing distances within triplets without leveraging additional samples.
Age-Oriented Face Synthesis
---------------------------
The first AOFS methods can be traced back to [@mark1983perception; @mark1981perception; @todd1980perception], in which craniofacial growth in young faces is studied. In the early stage, geometry-based methods were a popular choice among researchers, and one of the most representative works is the Active Shape Model (ASM) [@cootes1995active]. The authors model the shape of faces by adjusting the positions of a number of points. Each point marks one part of the face, such as the position of the eyes and the boundary of the face. Synthetic facial images of different shapes and ages can then be obtained by adjusting the position of these points. Another approach to rendering aging or rejuvenating effects is to directly synthesize or remove wrinkles on a given facial image [@bando2002simple; @liu2004image; @mukaida2004extraction; @wu1999simulating; @wu1995dynamic]. Later, Ramanathan and Chellappa [@ramanathan2006modeling] propose an aging-focused method called the craniofacial growth model for synthesizing elderly faces by leveraging facial landmark movements. Another worth-noting early AOFS method is [@tang2018personalized], where the authors use dictionary learning to learn a personalized aging process, and associate an aging dictionary to each subject to represent their aging characteristics.
With the increasing popularity of deep learning, several attempts have been made to tackle the AOFS problem using various network architectures. Both Wang *et al.* [@wang2018face] and Zhang *et al.* [@zhang2017age] use conditional adversarial learning [@mirza2014conditional] to synthesize aged faces. Wang *et al.* further employ an age category classifier to boost the synthesis accuracy and an L2 constraint on the identity-specific features to preserve the identity information. Yang *et al.* [@yang2018learning] propose a GAN framework by implementing a customized discriminator with a pyramid architecture, which leads to more realistic results than a conventional discriminator as images can be discriminated based on features at multiple scales. They further adopt a pre-trained identity classifier to preserve the identity in the synthesized images. AOFS methods based on the Wavelet transform are proposed recently in [@li2019global; @liu2019attribute], where this transform is used to enhance the texture information in the frequency domain so that richer aging and rejuvenating effects can be synthesized. He *et al.* [@he2019s2gan] implement a GAN model with a customized generator, where a number of decoders are implemented, each one learning an age category. All the decoders are associated with a weight factor to control their relative importance in each transformation. Since all the decoders are trained in parallel, the computational complexity of the method is proportional to the number of age categories to be learned.
Our work is different from the aforementioned deep-learning methods as it tackles the AOFS problem from a different angle, i.e., mode learning. Our method can achieve a high synthesis accuracy by learning multiple modes explicitly and independently. Additionally, compared to the L1 loss, the L2 loss, and the simple classifiers used in the those methods, our AOFS method uses the proposed Adversarial Triplet loss to keep the identity information unaltered in the synthesized facial images.
Proposed AOFS Method
====================
In this section, we explain in detail our proposed method by first formulating the problem and explaining the pre-trained Multi-Task Feature Extractor (MTFE) used to extract age-specific and identity-specific features. We then present the proposed CDP and the Adversarial Triplet loss. Finally, we explain the overall loss used to train our method.
Problem Formulation
-------------------
Since the transformation is conducted among age categories rather than specific ages, following the prior work in [@li2019global; @yang2018learning; @liu2019attribute], we divide the data into four categories according to the following age ranges: $30^-$, $31 - 40$, $41 - 50$, and $51^+$. Each category is denoted by $C_i$, where $i\in[1,4]$.
To render aging and rejuvenating effects, the proposed AOFS method takes two faces, $x\in{C_X}$ and $y\in{C_Y}$, and the age label of $\mathbf{\mathit{y}}$, $l_{age}^y$, as the inputs, where ${X}\neq{Y}$. Specifically, $x$ is the face that is to be aged or rejuvenated and $y$ carries the desired age information. Our method aims to generate an aged or rejuvenated $x$, denoted by $\tilde{x}$, which is expected to belong to the same age category as $y$. Moreover, to ensure that the identity information is effectively preserved in $\tilde{x}$, our method also uses other images in the same batch, $\left \{x'\right \}$, to compute the Adversarial Triplet loss. It is worth noting that both $x'$ and $y$ do not share the same identity information of $x$.
In summary, the proposed method achieves three goals simultaneously: 1) To generate realistic aged and rejuvenated faces; 2) to force the synthesized faces to be within the target age category; and 3) to preserve the identity information in the synthesized image. The architecture of our proposed AOFS method is illustrated in Fig. \[fig:architecture\].
Multi-Task Feature Extractor
----------------------------
The CDP and the Adversarial Triplet loss of the proposed AOFS method use age-specific and identity-specific features from input images and synthesized images. To extract and disentangle these features, we use the decomposition method proposed in [@wang2018orthogonal]. Specifically, we use a ResNet-50 [@he2016deep] as the backbone. The architecture of this feature extractor is depicted in Fig. \[fig:fx\]. This model decomposes all the features extracted from a facial image into two components based on a spherical coordinate system, which is formulated as: $$\label{decomposition}
{f}_{sphere} := \left \{r; \textbf{\textit{theta}} \right \},$$ where the ${f}_{sphere}$ is the set of features after the decomposition in which the angular component ***theta*** = $\left \{\theta_1, \theta_2, ..., \theta_k \right \} $ indicate the identity-specific features for $k$ identities, and the radial component $r$ encodes the age-specific features.
We replace the regression loss used to learn age-specific features in [@wang2018orthogonal] with an age regression model [@rothe2018deep; @wang2018fusion] to supervise the age-specific learning process, which has been shown to achieve better performance for the age estimation task. We observe that feature extractors trained in this multi-tasking manner can achieve higher accuracy on both the age category classification and identity classification tasks than single-task networks. Additionally, we use our proposed Adversarial Triplet loss to learn identity-specific features.
Conditional Discriminator Pool
------------------------------
In the vanilla GAN with a single image-level discriminator, the loss function for face synthesis is usually formulated as: $$\label{image_adversarial}
\begin{aligned}
\mathcal{L}_{adv} = &\mathbb{E}_{y}[logD(y)] \\
&+ \mathbb{E}_{x}[log(1-D(G(x))],
\end{aligned}$$ where $G$ is the generator trying to minimize the loss, and $D$ is the discriminator trying to maximize the loss. As mentioned before, GANs based on this loss function suffer from the mode collapse issue. To force the network to learn each mode independently and thus alleviate this issue, one can add more discriminators directly. However, such an strategy may lead to a high computational complexity and redundancy during training, as not all the discriminators are expected to back-propagate the loss during each transformation. Therefore, we propose a mechanism to select the corresponding discriminator for each transformation based on the input label that represents the target age information. Let us recall that our proposed AOFS method treats each age category as a mode, which results in four modes in total. We use the input label, $l_{age}^y$, to select the corresponding discriminator that learns the target age category. Our proposed method implements this mechanism on discriminators at the feature level, which are used to synthesize aging and rejuvenating effects. Therefore, we assemble four feature-level discriminators with an identical architecture to form our CDP. Each feature-level discriminator targets one mode. Our method additionally uses an image-level discriminator to remove artificial effects from the synthesized faces. As illustrated in Fig. \[fig:architecture\], in each transformation, our method leverages the selected feature-level discriminator alongside the image-level discriminator.
![Architecture of our MTFE. After the decomposition, we resize each set of task-specific features to be used by the corresponding feature-level discriminator of the CDP or the Adversarial Triplet loss.[]{data-label="fig:fx"}](extractor.pdf){width="1\linewidth"}
It is important to note that an alternative way to select the feature-level discriminator is by employing an additional classifier. However, within the context of AOFS, the accuracy of classifying age categories may be very low, from 25% to 60% depending on the specific age category in different AOFS benchmark datasets [@wang2018face; @liu2019attribute]. Employing such a low-accuracy classifier may result in a selecting a discriminator that learns an incorrect mode. Instead, we directly use $l_{age}^y$ to select discriminators, which guarantees that, in each transformation, the discriminator associated with the target mode is used. We then formulate the feature-level adversarial loss as follows: $$\label{feature_adversarial}
\begin{aligned}
\mathcal{L}_{adv_{feature}} = \mathbb{E}_{f_{age}^y}[log(FD_{C_i}(f_{age}^y)|l_{age}^y)] \\
+ \mathbb{E}_{f_{age}^{\mathit{\tilde{x}}}}[log(1-(FD_{C_i}(f_{age}^{G(x|l_{age}^y)})|l_{age}^y))],
\end{aligned}$$ where $FD_{C_i}$ is the selected feature-level discriminator trying to maximize the loss; $f_{age}^y$ denotes the age-specific features extracted from the target image, $y$; and $f_{age}^{G(x|l_{age}^y)}$ denotes the age-specific features extracted from the synthesized image, $\tilde{x}$, where $G(x|l_{age}^y)$ is the generator that produces $\tilde{x}$ conditioned on $l_{age}^y$. Finally, $l_{age}^y$ is a one-hot encoded vector indicating the label for the target age category, $C_i$.
{width="98.00000%"}
Adversarial Triplet Loss
------------------------
The Triplet loss [@schroff2015facenet] with three feature embeddings is formulated as: $$\label{triplet}
\begin{aligned}
\mathcal{L}_{Triplet} (a,p,n) =& \sum_{a,p,n} [m + Dist_{a,p} - Dist_{a,n}]_+,
\end{aligned}$$ where $Dist_{j,k}$ indicates the Euclidean distance between embeddings $j$ and $k$ in the feature space and $a, p, n$ are the indices of the *anchor*, the *positive* and the *negative*, respectively. This loss forces $Dist_{a,n}$ to be larger than $Dist_{a,p}$ by at least a margin $m$. However, once this criterion is satisfied, $Dist_{a,p}$ cannot be further minimized, which may lead to large intra-class variances. To overcome this problem, we add another ranking operation to Eq. (\[triplet\]), which forces $Dist_{a,n}$ to be larger than the distance between $n$ and $p$, $Dist_{n,p}$. This additional operation helps to further bring $p$ closer to $a$ by forcing different triplets with the same $a$ and $p$ but different $n$ to play a zero-sum game: $$\label{adversarial-triplet}
\begin{aligned}
\mathcal{L}_{AT} (a,p,n) = & \sum_{a,p,n} [m + Dist_{a,p} - Dist_{a,n}]_+ \\
&+ [Dist_{n,p} - Dist_{a,n}].
\end{aligned}$$
Let us assume there are several triplets with the same $a$ and $p$, but different $n$, where each distinct $n$ is denoted by $n_i$. Under this assumption, the Triplet loss in Eq. (\[triplet\]) can be minimized as long as $Dist_{a,n_{i}} > Dist_{a,p} + m$, which may result in clusters with large intra-class variances. To reduce such variances, $Dist_{a,n_{i}}$ should be larger than $Dist_{n_{i},p}$. Let us take the triplets $a-p-n_1$ and $a-p-n_3$ in Fig. \[fig:adversarial-triplet\] as an example, where $n_1$, $n_2$, $n_3$, and $n_4$ are all from different classes. In this example, both $n_1$ and $n_3$ should maintain their relative position with respect to the $a-p$ cluster in order to also be far from other neighboring clusters. In other words, $n_1$ and $n_3$ should not move towards either $n_2$ or $n_4$. In this case, $\mathcal{L}_{AT} (a,p,n_1)$ tries to pull $p$ towards $n_1$ and minimize $Dist_{n_{1},p}$, while $\mathcal{L}_{AT} (a,p,n_3)$ tries to pull $p$ towards $n_3$ and minimize $Dist_{n_{3},p}$. Therefore, $\mathcal{L}_{AT} (a,p,n_1)$ and $\mathcal{L}_{AT} (a,p,n_3)$ play a zero-sum game as minimizing one loss increases the other. This is also true for $\mathcal{L}_{AT} (a,p,n_2)$ and $\mathcal{L}_{AT} (a,p,n_4)$. In order to minimize all losses in this example, i.e., to have a total loss equal to zero, $p$ should be in the same position as $a$ so that $Dist_{a,n_{i}} = Dist_{n_{i},p}$. In practice, however, our Adversarial Triplet loss pulls $p$ to a position very close to $a$ so that $Dist_{a,n_{i}} \approx Dist_{n_{i},p}$.
![Feature distribution of the MNIST dataset for classification on (a),(c) the training set and (b),(d) the test set when the Triplet loss and the Adversarial Triplet loss are used.[]{data-label="fig:triplet_mnist"}](triplet_mnist.pdf){width="1\linewidth"}
Fig. \[fig:triplet\_mnist\] demonstrates the performance of the Adversarial Triplet loss on a real dataset. In this example, the feature distribution of the MNIST dataset for classification is presented. To this end, we employ an Alexnet [@krizhevsky2012imagenet] as the deep network, but replace all the fully-connected layers, except the output layer, by a single linear layer with two neurons for visualization purposes. From the figure, we can observe that the features learned by the Adversarial Triplet loss dramatically reduce the intra-class variances compared to the features learned by the Triplet loss. The classification accuracy attained by each loss is tabulated in Table \[tab:triplet\_mnist\].
One of the most critical issues in the Triplet loss is that as the number of triplets grows, many triplets can easily satisfy the constraint in Eq. (\[triplet\]), which in turn may lead to poor convergence [@schroff2015facenet]. To overcome this issue in the Adversarial Triplet loss, we adopt a hard negative mining strategy [@hermans2017defense]. Specifically, we use an online hard sample mining method in which each batch consists of samples from $T$ classes, and each class has $S$ samples within one batch, for a batch size of $B=TS$. In this method, each sample in a batch acts as the *anchor* for one triplet, thus, there are a total of $B$ triplets within one batch. For each *anchor*, a hardest *positive* sample with the largest distance and a hardest *negative* sample with the smallest distance are selected to form a triplet. This method does not require pre-defining the triplets and can generate hard triplets in an online manner. After incorporating this hard sample mining strategy, our Adversarial Triplet loss in Eq. (\[adversarial-triplet\]) is as follows: $$\label{adversarial-triplet_modified}
\begin{aligned}
\mathcal{L}_{AT} (a,p,n) = &\sum_{\substack{t=1}}^T\sum_{\substack{s=1}}^S[m + \max_{\substack{p}}Dist_{a,p} - \min_{\substack{n}}Dist_{a,n}]_+ \\
&+ [Dist_{n,p} - \min_{\substack{n}}Dist_{a,n}],
\end{aligned}$$ where $t$ is the class index and $s$ is the image index for each class in one batch.
-- -- --
-- -- --
: Classification accuracy (%) n on the MNIST dataset.[]{data-label="tab:triplet_mnist"}
Since we are trying to optimize the identity-specific features on the synthesized faces when training our AOFS method, we use the identity-specific features, $f_{id}^{x}$, from the source image as the *anchor* and the identity-specific features, $f_{id}^{\tilde{x}}$, from the synthesized image as the *positive*. In addition, we use all other images in the same batch that do not share the same identity with the source image as the *negatives*. The Adversarial Triplet loss of our AOFS method with the hard sample mining strategy is then formulated as: $$\label{adversarial-triplet_AOFS}
\begin{aligned}
\mathcal{L}_{AT} &(f_{id}^{x},f_{id}^{\tilde{x}},\left \{ f_{id}^{x'},f_{id}^{y} \right \}) = \sum_{\substack{t=1}}^T\sum_{\substack{s=1}}^S \\
&[m + Dist_{f_{id}^{x},f_{id}^{\tilde{x}}} - \min_{\substack{\left \{f_{id}^{x'},f_{id}^{y}\right \}}}Dist_{f_{id}^{x},\left \{f_{id}^{x'},f_{id}^{y}\right \}}]_+ \\
&+ [Dist_{\left \{f_{id}^{x'},f_{id}^{y}\right \},f_{id}^{\tilde{x}}} - \min_{\substack{\left \{f_{id}^{x'},f_{id}^{y}\right \}}}Dist_{f_{id}^{x},\left \{f_{id}^{x'},f_{id}^{y}\right \}}],
\end{aligned}$$ where $\left \{f_{id}^{x'}\right \}$ are the identity-specific features of images within the same age category as the source image but carrying different identity information, and $f_{id}^{y}$ are the identity-specific features of images within the target age category. It is worth noting that the above equation do not have the $max$ operation as in Eq. (\[adversarial-triplet\_modified\]) since the *positive* in this case, $f_{id}^{\tilde{x}}$, is synthesized thus cannot be selected.
Overall Loss
------------
The image-level adversarial loss in our AOFS method is formulated as: $$\label{feature_adversarial}
\begin{aligned}
\mathcal{L}_{adv_{image}} = &\mathbb{E}_{y}[logD(y)] \\
&+ \mathbb{E}_{x}[log(1-D(G(x|l_{age}^y))].
\end{aligned}$$
The overall loss function, $\mathcal{L}_{overall}$, to train our method is a weighted summation of several losses, with $\mathcal{L}_{adv_{image}}$ removing ghost artifacts, $\mathcal{L}_{adv_{feature}}$ synthesizing ageing and rejuvenating effects and attaining a high synthesis accuracy, and $\mathcal{L}_{AT}$ preserving the identity information:
$$\label{overall}
\begin{aligned}
\mathcal{L}_{overall} =& \mathcal{L}_{adv_{image}} + \lambda_{adv_{feature}}\mathcal{L}_{adv_{feature}} \\
&+ \lambda_{AT}\mathcal{L}_{AT},
\end{aligned}$$
where $\lambda_{adv_{feature}}$ and $\lambda_{AT}$ control the relative importance among learning objectives.
{width="18.4cm"}
{width="18.4cm"}
Experiments
===========
In this section, we first briefly describe the two AOFS benchmark datasets used in our experiments followed by the implementation details of our method. Then, we compare our method with state-of-the-art methods and conduct ablation studies, both qualitatively and quantitatively, to show that our method can achieve a high synthesis accuracy while preserving the identity information on the synthesized facial images.
AOFS benchmark datasets
-----------------------
We use the MORPH II dataset [@ricanek2006morph] and the Cross-Age Celebrity Dataset (CACD) [@chen2014cross] to train the MTFE and evaluate our method. The MORPH II dataset contains about 55,000 facial images of individuals with ages ranging from 16 to 77. The CACD contains more than 160,000 facial images of individuals with ages ranging from 16 to 62. Most of the images in the MORPH II dataset are mugshots, while images in the CACD contain Pose, Illumination, and Expression (PIE) variations. Each image in both datasets is associated with an age label and an identity label.
All images are cropped to 128$\times$128 pixels and aligned based on the location of the eyes. Since not all images can be aligned by using this technique, in the end, 55,062 images from the MORPH II dataset and 159,226 images from the CACD are used in our experiments. For each dataset, we use 80% of the images for training and the remaining 20% for testing. The number of training images for each age category in the MORPH dataset is 19,949, 12,496, 8,982, and 2,622, for the categories $\left \{30^-, 31-40, 41-50, 51^+ \right \}$, respectively. For the CACD, the number of training images of each age category is 39,416, 33,742, 30959, and 23,262, respectively. There is no identity overlap between the training and test sets.
We conduct a five-fold cross validation for all our experiments. For the MORPH II dataset, each fold has about 2,550 subjects with 3,989, 2,499, 1,796, and 524 images within each age category, respectively. For the CACD, each fold contains about 400 subjects with 7,883, 6,748, 6,191 and 4,652 images within each age category, respectively.
[p[0.06]{}|p[0.13]{}|p[0.11]{}|p[0.09]{}]{}\
& & &\
& & &\
& & &\
\
& & &\
& & &\
& & &\
\
& & &\
& & &\
& & &\
[p[0.06]{}|p[0.13]{}|p[0.11]{}|p[0.09]{}]{}\
& & &\
& & &\
& & &\
& & &\
& & &\
& & &\
\
& & &\
& & &\
& & &\
& & &\
& & &\
& & &\
AOFS quality criteria
---------------------
There are two criteria commonly used to measure the quality of synthesized images [@yang2018learning; @liu2019attribute] in the AOFS task. Under the first criterion, synthesized images are fed into an age category classifier to evaluate whether the depicted face has been transformed to the target age category. The second criterion measures the identity permanence and relies on face verification models to validate whether the synthesized image and the source image depict the same person.
![Visual comparison of a baseline model, six state-of-the-art works, and our proposed method on two benchmarks. The top two rows show the results on the MORPH II dataset and the bottom two rows show the results on the CACD. The input image is within the youngest group and the results are expected to be within the eldest group.[]{data-label="fig:baseline"}](visual_comparison.pdf){width="1.\linewidth"}
To evaluate our method and demonstrate its robustness, we use another two large-scale benchmark datasets to train two separate validation networks, one for each criterion. In particular, we use the AgeDB dataset [@moschoglou2017agedb], which is widely used for age estimation, to train the network that evaluates the synthesis accuracy and a face recognition benchmark dataset, the VGGFace2 dataset [@cao2018vggface2], to train the network that evaluates the identity permanence capabilities. In addition, we use the commonly used ResNet-50 as the backbone for both evaluation networks.
Network architecture
--------------------
We employ the architecture from [@zhu2017unpaired] for our generator. The generator has six residual blocks and each convolutional and deconvolutional layer is followed by an instance normalization and a ReLU function. For the image-level discriminator, we implement a patch discriminator [@isola2017image] with five convolutional layers, each followed by an instance normalization and a LeakyReLU function. Each feature-level discriminator has the same architecture as that of the image-level discriminator but consists of fully-connected layers.
The details of the architectures of the generator and discriminators in our AOFS method are tabulated in Tables \[tab:arch\_generator\] and \[tab:arch\_discriminator\], respectively. In both tables, for each convolutional and deconvolutional layer, $k$ indicates the kernel size, $s$ indicates the stride, and $p$ indicates the padding size. In Table \[tab:arch\_discriminator\], the second column for the feature-level discriminators tabulates the dimensions of the corresponding layer.
-- ------ ------ ------ -- ------ ------ ------
**** **** **** **** **** ****
-- ------ ------ ------ -- ------ ------ ------
\[tab:aging\]
Data augmentation
-----------------
When training the MTFE and validation networks, we use a combination of rotation, flip, and crop operations to augment the data. Specifically, we first randomly rotate each image by a angle between +10 deg. and -10 deg., and then randomly flip the rotated image with a probability of 0.5. Finally, we pad the image on all sides with 10 pixels and crop the padded image at a random location to the original image size (i.e. $128 \times 128$ pixels). When training the proposed AOFS method, in order to increase the size of the training set without introducing additional variance to the dataset, we only use the flip operation.
Hyper-parameter setting
-----------------------
When training the MTFE, we set the batch size to 128 and the initial learning rate to $0.002$ for both datasets. We train it for 500 epochs while decreasing the learning rate by 0.1 every 150 epochs. When training the AOFS method, we set the batch size to 8 and the initial learning rate to $0.0002$. The learning rate decreases linearly after the first 25 epochs. We empirically set $\lambda_{adv_{feature}}$ to $1$ and $\lambda_{AT}$ to $0.001$. The margin hyper-parameter, $m$ in Eq. (\[adversarial-triplet\_AOFS\]), is set to 0.3. We use the PyTorch framework [@paszke2017automatic] for the implementation and run each experiment for 50 epochs. All experiments are run on a single NVIDIA GTX2080Ti GPU.
Synthesis accuracy
------------------
We first qualitatively evaluate the synthesized facial images based on their visual quality. We then present quantitative results based on age category classification accuracy, image quality and the degree of mode collapse. We perform these evaluations for our AOFS method and several state-of-the-art methods.
-- ------ ------ ------ -- ------ ------ ------
****
**** **** **** **** ****
-- ------ ------ ------ -- ------ ------ ------
\[tab:rejuvenating\]
-- ------ ------
**** ****
-- ------ ------
: ResNet Score and Fréchet ResNet Distance on the MORPH II dataset.[]{data-label="tab:score_morph"}
-- ------ ------
**** ****
-- ------ ------
: ResNet Score and Fréchet ResNet Distance on the CACD.[]{data-label="tab:score_cacd"}
-- ------ ------
**** ****
-- ------ ------
: Degree of mode collapse as measured by the KL divergence.[]{data-label="tab:kl"}
### Visual Quality
Fig. \[fig:progress\_aging\] and \[fig:progress\_rejuvenating\] show some sample images synthesized by our AOFS method. Fig. \[fig:progress\_aging\] shows aging results for 6 subjects from the MORPH II dataset and 6 from the CACD using a source image from the youngest category ($30^-$). We can see that our method turns hair gray or white, introduces forehead wrinkles and nasolabial folds, and makes the skin to appear rough. Fig. \[fig:progress\_rejuvenating\] shows rejuvenating results for 6 subjects from each dataset using a source image from the oldest category ($51^+$). We can see that for these cases, our method removes wrinkles and gray/white hair.
We also evaluate six state-of-the-art methods, namely the method by Antipov *et al.* [@antipov2017face], the Identity-Preserving Conditional Generative Adversarial Networks (IPCGAN) [@wang2018face], the S$^2$GAN [@he2019s2gan], and the methods by Liu *et al.* [@liu2019attribute], Li *et al.* [@li2019global], and Yang *et al.* [@yang2018learning]. To have a fair comparison, we replace the feature extractors in these methods with our pre-trained MTFE and use the same number of residual blocks in their generator expect for the method in [@antipov2017face], as there is no residual block originally involved in this particular method.
Since the synthesis accuracy of our AOFS method depends on the CDP, we also evaluate a baseline model without the CDP (hereinafter called *w/o CDP*) as part of an ablation study. The *w/o CDP* model replaces the CDP with a simple feature-level discriminator, which makes this model similar to a vanilla GAN but with two discriminators, one at the feature level and the other at the image level.
Fig. \[fig:baseline\] depicts the visual results of these evaluations. Note that it is visually evident that the results generated by the *w/o CDP* model do not contain much aging and rejuvenating effects as this model suffers from the mode collapse issue. On the contrary, our proposed method can synthesize the aging and rejuvenating effects realistically. Among all state-of-the-art methods, Yang *et al.* [@yang2018learning] is able to synthesize the most realistic effects due to the use of a multi-level feature discriminator.
-- -- ------ ------ ------ -- ------ ------ ------
**** **** **** **** **** ****
**** **** **** **** **** ****
-- -- ------ ------ ------ -- ------ ------ ------
\[tab:identity\]
### Age category classification accuracy
Table \[tab:aging\] and \[tab:rejuvenating\] tabulate the age category classification accuracies of various methods on the synthesized images when images from the $30^-$ and $51^+$ categories are used as source images, respectively. In these tables, the *Natural Faces* row tabulates the accuracy attained when using the original facial images. Since [@antipov2017face] uses a relatively shallow generator compared to other works, its performance is hence below others by a significant margin. IPCGAN uses the age labels as conditions in the GAN learning process and incorporates an age category classification loss. However, due to the fact that the classification error is high (the classifier is noisy), the gradient for the age information is not accurate. As a result, although its performance is higher than that of [@antipov2017face], it is still lower than the one attained on the original facial images by a large margin. The recently proposed S$^2$GAN attains a higher accuracy by implementing a customized generator where each age category is associated with a decoder. The methods of Liu *et al.* [@liu2019attribute] and Li *et al.* [@li2019global] achieve similar accuracy since both use the Wavelet transform. Among all the other evaluated methods, the one proposed by Yang *et al.* [@yang2018learning] achieves the best performance by using a multi-level feature discriminator. By adding a feature-level discriminator to the vanilla GAN, the baseline *w/o CDP* model achieves a comparable performance to that achieved by IPCGAN. Our proposed AOFS method outperforms all evaluated methods for the majority of age categories.
### Image Quality
The synthesis accuracy is also related to the quality of the generated images [@wang2018face]. The quality and diversity of the synthesized images are usually measured in terms of the Inception Score (IS) and the Fréchet Inception Distance (FID). IS measures the image quality and diversity by computing the KL divergence between the real and the generated class distributions. On the other hand, FID uses a multivariate Gaussian distribution to model the data distribution and the mean and the covariance from two distributions to compute their distance. Since we use a ResNet-50 to evaluate the identity permanence capabilities (see Section IV.G), we rename these two metrics as the ResNet Score (RS) and the Fréchet ResNet Distance (FRD). The RS and FRD are tabulated in Table \[tab:score\_morph\] and Table \[tab:score\_cacd\], respectively, for our AOFS method and several state-of-the-art methods. Since our AOFS method can render more realistic aging and rejuvenating effects than other evaluated methods and has stronger identity permanence capabilities, it achieves the best performance for both metrics, especially for the FRD, which is sensitive to the mode collapse issue.
### Degree of Mode Collapse
Since our method tackles the AOFS task from the aspect of mode learning, we also measure the degree of mode collapse by computing the KL divergence between the distribution of the synthesized images and the expected distribution. We compute this divergence for all synthesized images within each fold.
As shown in Table \[tab:kl\], the proposed AOFS method significantly outperforms the baseline model and the method in [@antipov2017face], which use the negative log-likelihood loss from the vanilla GAN. By using different discriminators to learn different modes, our method also achieves a lower divergence value compared to other methods that leverage the least square loss from the LSGAN.
Identity permanence
-------------------
To evaluate the identity permanence on the synthesized images, we design a new baseline, the *Triplet* model. Specifically, in the *Triplet* model, we replace the Adversarial Triplet loss with the original Triplet loss to directly compare these two loss functions. The identity permanence capabilities are measured in terms of the face verification accuracy, i.e.. whether the synthesized image and the original image depict the same person. To this end, we define three input settings based on three different target age categories for each synthesis process. Specifically, the query images are the original facial images from the datasets, while the gallery images are the synthesized images that are expected to be within the target age category, as tabulated in Table \[tab:identity\] with the column headings *S31-40*, *S41-50*, and *S51$^+$* for the aging process and headings *S41-50*, *S31-40*, and *S30$^-$* for the rejuvenating process. For example, *S31-40* refers to the synthesized images expected to be within the $31-40$ category. We use the *cosine similarity* to measure the distance of each pair of query and gallery images.
As tabulated in Table \[tab:identity\], all the state-of-the-art methods achieve a similar accuracy since they all use a similar strategy, namely, minimizing the distance between two identity-specific features using the L1 or L2 loss. Li *et al.* [@li2019global] slightly outperforms other methods as it uses a combination of these two losses. The subtle difference in accuracy among these methods may also be due to the quality of the images, since the identity information may be distorted in images of poor quality. By replacing the L1 or L2 loss with the Triplet loss, the identity permanence capability can be remarkably boosted by about 3 % on both datasets. Our AOFS method, which uses the Adversarial Triplet loss, reduces intra-class variances within each age category in the feature space. Consequently, it achieves the highest accuracy among all evaluated methods.
Conclusion
==========
In this paper, we tackle the Age-Oriented Face Synthesis task from the aspect of the mode learning. Specifically, we present an AOFS method that incorporates a novel Conditional Discriminator Pool to alleviate the mode collapse issue in the vanilla GAN. Our method also incorporates a novel Adversarial Triplet loss to attain strong identity permanence capabilities. By using the proposed CDP, only the target feature-level discriminator that learns the current mode is deployed, which does not increase the computational complexity during training. Our CDP then allows learning multiple modes explicitly and independently. As a result, our proposed AOFS method outperforms several state-of-the-art methods on AOFS benchmark datasets. In the future, we will investigate into improving the aging and rejuvenating effects by including the synthesis and removal of wrinkles and face shape manipulation among different age categories. Improving these aspects of the synthesis process is expected to further boost the synthesis accuracy and have the potential to simulate a more personalized aging and rejuvenating process.
Acknowledgment
==============
This work is supported by the EU Horizon 2020 - Marie Sklodowska-Curie Actions through the project Computer Vision Enabled Multimedia Forensics and People Identification (Project No. 690907, Acronym: IDENTITY).
[^1]: H. Wang and V. Sanchez are with the Department of Computer Science, University of Warwick, Coventry, CV4 7AL, UK (e-mail: [email protected], [email protected].)
[^2]: C-T. Li is with the School of Information Technology, Deakin University, Geelong VIC 3216, Australia (e-mail: [email protected].)
[^3]: © 2020 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We construct Morse-Smale-Witten complex for an effective orientable orbifold. For a global quotient orbifold, we also construct a Morse-Bott complex. We show that certain type of critical points of a Morse function has to be discarded to construct such a complex, and gradient flows should be counted with suitable weights. The homology of these complexes are shown to be isomorphic to the singular homology of the quotient spaces under the self-indexing assumptions.'
author:
- 'Cheol-Hyun Cho and Hansol Hong'
title: 'Orbifold Morse-Smale-Witten complex'
---
[^1]
Introduction
============
Morse theory is one of the most important tools to understand the topology of manifolds. A modern approach to the Morse theory, Morse-Smale-Witten complex has been very popular as its infinite dimensional application, Floer homology theory, has proven to be a very powerful tool in the area of symplectic and differential geometry. Morse-Smale-Witten complex is a free module generated by critical points of a Morse function graded by their indices, and the differential on this complex is given by counting (signed) number of gradient lines between critical points of index difference one.
The notion of a differentiable orbifold was introduced by Satake [@Sa] in the fifties under the name V -manifold, as a natural generalization of the notion of differentiable manifold. An effective orbifold is a space which is locally the quotient space of a smooth manifold by the effective action of a finite group.
In this paper, we develop Morse-Smale-Witten complex for effective orbifolds. Morse functions are given by invariant functions whose local lift is Morse. Morse theory on orbifolds, such as Morse inequalities or informations about local Morse data has been known for a while since the work of Lerman-Tolman [@LT]. But the construction of Morse-Smale-Witten complex has not been available.
We construct such a complex, and show that its homology is isomorphic to the singular homology of the quotient space (with certain assumptions). There are a few interesting differences from the case of manifolds. The first one is that a broken trajectory (assume only one breaking) may be a limit of several different families of smooth trajectories, whereas for manifolds, there exist only one family of smooth trajectories converging to such a broken trajectory. This is because there can be several different lifts of broken trajectories, which are not equivalent via local group action and we analyze them carefully in section 4.
The second one is that one should count with suitable weights to define a chain complex depending on the order of related isotropy groups. Namely, each gradient trajectory, which was counted as one in the case of manifolds, should be counted with weights depending on the stabilizer of the end points and that of trajectory itself.
Most interesting of all is probably that one has to discard certain critical points (namely unorientable critical points) and consider a subcomplex generated by only orientable critical points, to define Morse-Smale-Witten complex. This is related to the observation, already in [@LT] that when the local group action does not preserve the orientation of the unstable directions, then there exists no change of topology when passing through such a critical point.
One drawback is that it is known to be very difficult to study Morse-Smale condition(transversality between any unstable and stable manifolds of critical points) together with the local invariance condition. For example, Morse functions on orbifolds are dense among smooth functions (by [@Wa], [@H]) but they may not be Morse-Smale. Hence throughout the paper, we only consider the case when the given function is indeed Morse-Smale. It is also difficult to carry out an analogue of the usual invariance proof of the homology of the chain complex due to such issues. (We hope to return to this issues in the future research.) Instead, we will prove that the homology is isomorphic to the singular homology of the quotient space.
The work of Dixon, Harvey, Vafa and Witten on string theory on orbifolds[@DHVW], and the discovery of new ring structure on cohomology of inertia orbifolds by Chen and Ruan[@CR], and orbifold Gromov-Witten invariants [@CR2], has prompted many exciting new developments on the study of orbifolds in the last decades. But the Fukaya category theory on orbifolds has not been developed yet. This paper lays a foundation to define orbifold Fukaya category theory. Namely, we expect that the new phenomenons which appeared in this paper on Morse theory should also be present in orbifold Fukaya category theory, and they should be dealt in a similar way as in this paper. We hope to explore this elsewhere jointly with Mainak Poddar. We remark that for toric orbifolds, Lagrangian Floer theory for smooth Lagrangian torus fibers has been developed in [@CP].
We also remark that the equivariant cohomology version of Morse-Bott theory has been defined by Austin and Braam [@AB] for compact (connected) Lie group $G$, but their construction does not immediately generalize for finite $G$ unless the group action preserves the orientation of critical submanifolds and orientations of unstable directions.
We consider local coefficient system on each critical submanifolds from the determinant bundle of normal directions, and take an invariant subcomplex (where invariance involves local coefficients) to define a Morse-Smale-Witten complex in such a case. This gives a correct analogue of their construction of equivariant cohomology Morse-Bott complex in the case of finite group $G$.
Here is the outline of the paper. In section 2, the case of global quotient is developed with careful examination of orientation issues. In section 3, we reformulated the construction of section 2 in a more intrinsic form. In section 4, we define a Morse-Smale-Witten complex for effective orbifolds and show that $\partial ^2 =0$. In section 5, we compare the homology of Morse-Smale-Witten complex with singular homology of the quotient space. In section 6, we consider the case of non-orientable manifolds and also Morse-Bott case.
Morse-Smale-Witten complex for global quotients
===============================================
Let $M$ be a closed oriented connected manifold and suppose the finite group $G$ acts on $M$ effectively in an orientation preserving way. Set ${\boldsymbol{X}}$ to be the global quotient orbifold $M/G$ and denote by $X$ the quotient space and $\pi : M \to X$ the natural projection to the orbit space. By [@Wa], we can take a $G$-invariant Morse function $f : M \to {\mathbb{R}}$. Suppose further that $f$ is Morse-Smale with respect to a $G$-invariant metric. In this section, we construct a Morse-Smale-Witten complex for $ \bar{f}:{\boldsymbol{X}}\to {\mathbb{R}}$, where $f = \bar{f} \circ \pi$. A more general Morse-Bott case will be discussed later, but we explain this case here for the introduction and to present a clearer picture on orientation issues.
We begin by recalling that Morse homology of orbifolds has been studied by Lerman and Tolman [@LT], where they analyzed the local Morse data near a critical point in orbifold setting and proved Morse inequalities for orbifolds. The equivariant cohomology complex of $M$ with respect to $G$-action in terms of Morse theory has been defined by Austin and Braam in [@AB] using Cartan model together with Morse-Bott complex, when $G$ is compact connected Lie group. (See remark \[ABrelation\] for more discussions on this).
Although it is not stated in [@AB], they implicitly assume that $G$ is connected compact Lie group, because first, they are using Cartan model, and second they do not discuss orientation issues of the group action, which becomes trivial for connected Lie group actions.
In fact, as we will see, the orientation issues are very important even to set up Morse-Witten-Smale complex for finite group actions. Even though the group action is assumed to preserve the given orientation of manifold, it may not preserve the orientation of unstable directions.
The critical points of $\bar{f}$ will be divided into two types, orientable and non-orientable critical points (see Definition \[def:type\] below). We will discard the non-orientable critical points and show that invariant chains in the complex made of orientable critical points define Morse homology of the orbifold ${\boldsymbol{X}}$.
We recall that a smooth function on orbifold by definition has smooth invariant lift on each uniformizing chart.
\[orbM\] A smooth function $\bar{f}:{\boldsymbol{X}}\to {\mathbb{R}}$ is called Morse if every point $\bar{x}$ in the orbifold has a uniformizing chart $({\widetilde{U}}_{ \bar{x} } , G_{\bar{x}}, \pi_{\bar{x}})$ such that $\bar{f} \circ \pi_{\bar{x}}$ is Morse on ${\widetilde{U}}_{\bar{x}}$.
From now on, we consider global quotient orbifold ${\boldsymbol{X}}= M/G$ as above. Let $f$ be a Morse function on $M$ which is in addition $G$-invariant, then $f$ induces a Morse function $\bar{f}$ on ${\boldsymbol{X}}$. Denote the set of critical points of $f$ and $\bar{f}$ by ${\textit{crit}}(f)$ and ${\textit{crit}}(\bar{f})$, respectively. i.e. $\bar{p} \in {\textit{crit}}(\bar{f})$ if there exists $p \in {\textit{crit}}(f)$ such that $\pi (p) = \bar{p}$. As usual, we define $Cf_\ast (M)$ as a complex of ${\mathbb{R}}$-vector spaces freely generated by each $p \in {\textit{crit}}(f)$. We write $W^+ (p)$ and $W^- (p)$ to denote stable and unstable manifolds at $p$, respectively(see for example [@Ni]). Also, denote the set of all critical points of $f$ with index $i$ by ${\textit{crit}}_i (f)$, for simplicity.
We construct a Morse-Smale-Witten complex associated to $\bar{f}$ as a certain subcomplex of $C f _\ast (X)$ as follows: First, orient $W^{-}(p)$ for each $p \in Crit(f)$.
\[def:type\] We define the type of a critical point $\bar{p} \in {\textit{crit}}(\bar{f})$ as follows. If $G_{p}$ -action on the unstable manifold $W^-(p)$ at $p \in \pi^{-1} (\bar{p})$ is orientation preserving, then $p$ is called [*orientable*]{} critical point, and non-orientable otherwise. Denote by ${\textit{crit}}^+ (\bar{f})$ (resp. ${\textit{crit}}^- (\bar{f})$ ) the set of all orientable (resp. non-orientable) critical points of $\bar{f}$.
We use the similar notation for critical points of $f$ .
The idea of non-orientable critical points was considered already in [@LT] and in several subsequent works such as [@H]. As observed in [@LT], this is very natural in terms of local Morse data. Indeed, we will see later that attaching cells which arise at non-orientable critical points do not contain any topological information for the quotient space (see Corollary \[disk\]).
If $G_{p}$ is orientation preserving for one of $p \in \pi^{-1} (\bar{p})$, then it is so for other $p' \in \pi^{-1} (\bar{p})$.
Let ${\textit{crit}}_i^\pm ({f}) = {\textit{crit}}_i ({f}) \cap {\textit{crit}}^\pm ({f})$, which induces the decomposition $Cf_i (M) = C f_i^+ (M) \oplus C f_i^- (M)$. $G$-action not only preserves ${\textit{crit}}(f)$ but also preserves a index and the type (orientable or not) of a critical point and hence $G$ naturally acts on $Cf_i^+(M)$. Define $$C \bar{f}^+_i (X) := Cf^+_i (M)^G,$$ $G$-invariant chains in $Cf_\ast (X)$ consisting of orientable critical points of degree $i$. For $\bar{p} \in {\textit{crit}}(\bar{f})$, we formally write $$\label{barp}
[\bar{p}] := \displaystyle\sum_{ p \in \pi^{-1} (\bar{p}) } p.$$ $C \bar{f}^+ (X)(= \displaystyle\oplus_i C \bar{f} _i^+ (X) )$ is freely generated by such $[\bar{p}]$’s for $\bar{p} \in {\textit{crit}}^+(\bar{f})$.
We will define a boundary map $\partial_i : C \bar{f}_i^+ (X) \to C \bar{f}_{i-1}^+ (X) $ to make it a chain complex. For each orientable critical point $\bar{p}$ of $\bar{f}$, take a $G$-invariant orientation on $W^- (p)$, for all $p \in \pi^{-1} (\bar{p})$. For non-orientable $\bar{p}$, and take arbitrary orientation for $W^- (p)$.
As $f$ is a Morse function on the manifold $M$, we have Morse-Smale-Witten differential $\partial_i : Cf_i (M) \to Cf_{i-1} (M)$ associated to $f$ (defined by using the above choice of orientations), defined as follows:
For $p,q \in {\textit{crit}}(f)$, define ${\widetilde{\mathcal{M}}} (p,q)$ to be the set of all negative gradient flow lines from $p$ to $q$ and by taking quotient under time translation, $$\mathcal{M} (p,q) = {\widetilde{\mathcal{M}}} (p,q)/{\mathbb{R}}.$$ Then, we define $$\partial_{\mu(p)} p = \sum_{q, \mu(q) = \mu(p)-1} \# \mathcal{M} (p,q) \,\, q.$$ Here, $\mu(p)$ is the Morse index of $p$ and $\# \mathcal{M} (p,q)$ is the signed number of oriented moduli space which is a finite set. (i.e. $\# \mathcal{M} (p,q)$ is the ‘signed’ number of negative gradient flow lines from $p$ to $q$. For the sign rule, see below.)
Now, we define a differential $\partial$ for $\bar{f}:{\boldsymbol{X}}\to {\mathbb{R}}$ on $[\bar{p}]$ from the formula \[barp\] and using the differential for $f: M \to {\mathbb{R}}$. We claim that this defines a differential for $C \bar{f}^+ (X)$. To show this, we need the following two crucial lemmas:
\[cancel\] If $\bar{p} \in {\textit{crit}}^+ (\bar{f})$, then $$\label{bdy1}
\partial [\bar{p}]= \partial \left( \displaystyle\sum_{ p \in \pi^{-1} (\bar{p})} p \right) \in C\bar{f}^+(X),$$ i.e., $\partial [\bar{p}]$ has nonzero coefficients only at orientable critical points of $f$.
Let $\bar{p}$ be of index $i$. Then, every $p$ in the sum has index $i$. We will show that the coefficient at an arbitrary non-orientable critical point $q \in {\textit{crit}}_{i-1}^- (f)$ is zero. Set $\mathcal{M} (\bar{p}, q)$ be the set of negative gradient flow lines from any $p \in \pi^{-1} (\bar{p})$ to $q$. Then, $\mathcal{M} (\bar{p}, q)$ is a signed set. i.e. there is a prescribed sign for each element $\gamma \in \mathcal{M} (\bar{p}, q)$ according to the standard sign rule in Morse homology theory.
We briefly recall the sign rule for reader’s convenience. For an oriented manifold $M$ and a Morse function $f$, fix an orientation of each unstable manifold, it will orient every stable manifolds so that for each critical point $p$ of $f$, $[W^{+}(p)][W^{-}(p)]=[M]$, where $[\;]$ means the oriented frames of the tangent spaces at $p$. Hence, $W^- (p) \cap W^+(q)$ has an induced orientation (we follow the orientation conventions of [@GP]).
Let $\gamma$ be a negative gradient flow line connecting $p$ and $q$. Then, ${\rm{im}} \, \gamma \subset W^- (p) \cap W^+(q)$. If the negative gradient flow orientation of $\gamma$ matches the induced orientation, then it is counted as $+1$ and otherwise as $-1$.
The following convention also gives the same sign. Fix $s$ so that $f(q) < s < f(p)$. Orient the set $f^{-1}(s)$ so that $[\nabla f][f^{-1}(s)]=[M]$. Consider $S^-(p) := W^- (p) \cap f^{-1}(s) $ which is oriented as a boundary of $D^- (p) := W^-(p) \cap f^{-1}([s,\infty))$. Similarly, consider $S^+(q)$ which is oriented as a boundary of $D^+ (q)$. In fact, they are diffeomorphic to $S^{i-1}$ and $S^{n-i}$, respectively. Because $ S^- (p)$ and $ S^+(q) $ are of complementary dimensions to each other in the slice $f^{-1} (s)$, we can count their signed intersection number. One can check that this sign agrees with the above convention (following sign rules of [@GP]).
Suppose that $q$ is a non-orientable critical point. We split $\mathcal{M} (\bar{p}, q)=\cup_{p} \mathcal{M} (p, q)$ into a disjoint union $\mathcal{M} (\bar{p},q)= \mathcal{M} (\bar{p},q)^+ \sqcup \mathcal{M} (\bar{p},q)^-$ with respect to their signs. Clearly, the sum of these signs will be the coefficient of $q$ in and hence we have to show that $|\mathcal{M} (\bar{p}, q)^+| = |\mathcal{M} (\bar{p}, q)^-|$.
Pick any $g \in G_{q}$ which reverses the orientation of $W^- (q)$. Then $g$ will give a permutation of $\mathcal{M} (\bar{p}, q)$, which also written as $g$, since $g$ preserves $\pi^{-1} (\bar{p})$. We claim that $g$ sends $\mathcal{M} (\bar{p}, q)^+$ to $\mathcal{M} (\bar{p}, q)^-$. To see this, we consider the action of $g$ on $f^{-1} (s)$. Note that the $G$-action preserves each slice of $f$. If $g \cdot p = p'$, $g$ sends $S^- (p)$ to $S^- (p')$ and preserves $S^+ (q)$ (note that $g \in G_{q}$ is an automorphism of $f^{-1} (s)$). Let $x \in S^- (p) \cap S^+(q) $ represent $\gamma \in \mathcal{M} (\bar{p}, q)^+$. By the sign rule given above, $S^- (p)$ and $S^+ (q)$ intersect positively at $x$ in $f^{-1}(s)$, meaning that $$[ S^- (p)]_x[ S^+ (q) ]_x=[f^{-1}(s)]_x.$$ Here the subscript $x$ means they are oriented frames of tangent spaces at $x$. Now, let’s pass through the automorphism of $f^{-1} (s)$ induced by $g$. $G$-invariant orientations on unstable manifolds at $p$’s implies that: $$g \cdot [S^- (p)]_x =[S^- (p') ]_{gx}.$$ and, by the assumption that $g$ reverses the orientation of unstable manifold at $q$, $$g \cdot [S^+ (q)]_x = - [S^+ (q)]_{gx}.$$ Since $g$ preserves the orientation of $M$ and $f$ is $g$-invariant, hence $g$ preserves the orientation of $f^{-1} (s)$, or $g \cdot [f^{-1}(s)]_x = [f^{-1}(s)]_{gx}$. In conclusion, considering the oriented frames at $g\cdot x$, we have $$[ S^- (p')]_{gx}[S^+ (q)]_{gx}
= (g \cdot [S^- (p)]_x)(-g \cdot [S^+ (q)]_x)
= - g \cdot [ f^{-1} (s) ]_x = - [ f^{-1} (s) ]_{gx}.$$ This means $ S^- (p')$ and $ S^+(q)$ intersect negatively at $g \cdot x$. Because the point $g \cdot x$ represent $g \cdot \gamma$, the sign of $g \cdot \gamma$ should be minus. This proves our claim. By the same argument $g^{-1}$ sends $\mathcal{M} (\bar{p},q)^-$ to $\mathcal{M} (\bar{p}, q)^+$. But since $g$ and $g^{-1}$ are inverse to each other, we get a bijection $g$ from $\mathcal{M} (\bar{p}, q )^+$ to $\mathcal{M} (\bar{p},q)^-$. In particular $|\mathcal{M} (\bar{p}, q)^+| = |\mathcal{M} (\bar{p}, q)^-|$.
The following expression (\[bdy1\]), $
\partial \left( \sum_{p \in \pi^{-1} (\bar{p})} p \right)
$ is $G$-invariant if $\bar{p}$ is orientable.
By the previous lemma, only consists of orientable critical points. Consider two orientable points $q$ and $q' := g \cdot q$ appearing non trivially in . We need to show that coefficients of $q$ and $q'$ are equal. However, this is obvious since $g$ and $g^{-1}$ give the sign preserving isomorphisms between $\mathcal{M} (\bar{p}, q) $ and $\mathcal{M} (\bar{p}, q')$. This is because we chose orientations on unstable manifolds at each orientable critical points of $f$ so that they are $G$-invariant.
We have shown that the Morse boundary map $\partial$ preserves $Cf^+_{\ast} (M)^G \subset Cf_\ast (M)$. Thus, we conclude that $ C \bar{f}^+_\ast (X) = Cf^+_{\ast} (M)^G$ is a subcomplex of $ Cf_\ast (M)$. We denote $ C \bar{f}^+_\ast (X)$ by $CM_\ast ({\boldsymbol{X}}, \bar{f})$ for simplicity and use the same notation $\partial$ for the restriction of $\partial : Cf_\ast (M) \to Cf_\ast (M)$ to $CM_\ast ({\boldsymbol{X}},\bar{f})$. Note that $\partial^2 =0$ automatically follows from that of $M$.
In fact, the resulting homology group $HM_\ast ({\boldsymbol{X}}, \bar{f})$ is isomorphic to the singular homology of the orbit space(quotient space). We postpone its proof to section 4 proposition \[comp\] where we prove it in more general case.
$HM_\ast ({\boldsymbol{X}}, \bar{f} ) \cong H_\ast (M/G) = H_\ast (X)$.
\[broken heart\] Consider the famous heart $S^2$ with the Morse function $h$ given by the height function $h$ with two maximum $p, q$ , minimum $s$ and one saddle point $r$. We assume that the heart $S^2$ admits an ${\mathbb{Z}}/2{\mathbb{Z}}$-action given by $180$ degree rotation which interchanges $p$ and $q$ and fixes $r$ and $s$ and assume that this ${\mathbb{Z}}/2{\mathbb{Z}}$-action preserves $h$. Then, the quotient space $S^2/({\mathbb{Z}}/2{\mathbb{Z}})$ is again topologically $S^2$. A naive $G$-invariant Morse chain complex is $$0 \;\to \;<(p+q)>\; \to\; <r> \;\to\; <s>\; \to \;0.$$ where $<(p+q)>$ denotes the one dimensional vector space generated by $p+q$. Here, the differential all vanishes, and the resulting homology does not equal to the singular homology of $S^2$.
{height="1.5in"} \[hannah\]
A correct Morse chain complex for the quotient orbifold is $$0 \;\to\; <(p+q)> \;\to\; 0 \;\to <s>\; \to \;0.$$ Here, critical points $p, q, s$ is orientable whereas the critical point $r$ is non-orientable as the half-rotation reverses the orientation of the unstable manifold at $r$. Hence, we discard $<r>$ and do not use it as a generator in the above complex. In this way, we obtain a Morse-homology of $S^2/({\mathbb{Z}}/2{\mathbb{Z}})$ isomorphic to the singular homology of $S^2$.
In fact, this is $G$-invariant part of the Morse chain complex of $M$ when the critical points of $M$ are decorated with orientation sheaf of unstable manifolds. Such a point of view will be explained in more detail in the last section in the Bott-Morse setting.
But this approach of taking invariant subcomplex obviously does not work for general orbifolds which are not global quotients.
Intrinsic formula
=================
In order to extend the result to general orbifolds, we will reformulate the formula of $\partial$ for the global quotient orbifold ${\boldsymbol{X}}$ to be more intrinsic form. To do this, consider two orientable critical points $\bar{p}$ and $\bar{q}$ of $\bar{f}$ of indices $k$ and $k-1$, respectively and suppose there exists a negative gradient flow line $\bar{\gamma}$ from $\bar{p}$ to $\bar{q}$ in $X$. We want to find the contribution of $\bar{\gamma}$ to the coefficient at $[\bar{q}]$ in $\partial [\bar{p}]$. Recall that $[\bar{p}] = \sum_{ p \in \pi^{-1} (\bar{p}) } p$. Let $\gamma$ be one of its liftings of $\bar{\gamma}$.
\[flow\] Given a negative gradient flow line $\gamma$ in $M$, the isotropy groups $G_{x}$ are all isomorphic for any point $x \in {\rm{im}} \, \gamma$.
Indeed, they are all the same because the diffeomorphism $\Phi_t$ of $M$ induced by the negative gradient vector field of $f$ is $G$-equivariant. More precisely, let $y$ be another point in $\gamma$. Since both $x$ and $y$ are not critical points of $f$, there exists $t$ such that $\Phi_t (x) = y$. Then, for any $g \in G$, $G$-equivariance of $\Phi_t$ implies $\Phi_t (g \cdot x ) = g \cdot \Phi_t (x)=g \cdot y$ and hence $G_x = G_y$.
By the above lemma, we may denote $G_x$ by $G_\gamma$ for $x \in {\rm{im}} \, \gamma$. It is natural to define $G_{\bar{\gamma}}$ as the conjugacy class of $G_\gamma$. Then, $|G_{\bar{\gamma}}|$ is well-defined.
Note that the number of lifts of $\bar{\gamma}$ in $M$ is $ |G| / |G_{\bar{\gamma}}| $. Hence, there exist $\left( \sum_{ \bar{\gamma}: \bar{p} \rightarrow \bar{q}} |G| / |G_{\bar{\gamma}}| \right) $-negative flow lines connecting critical points of $\bar{p}$ and $\bar{q}$. We want the coefficient of $[\bar{q}]= \sum_{ q \in \pi^{-1} (\bar{q}) } q$ instead of one of single $q$, and therefore we divide $\sum_{\bar{ \gamma }: \bar{p} \rightarrow \bar{q}} |G| / |G_{\bar{\gamma}}| $ by the number of $q$’s $= |\pi^{-1} (\bar{q})|$. Note that all coefficients of $q$’s in the sum are equal because of $G$-action. Therefore:
$$\begin{aligned}
\partial [\bar{p}] &=& \sum_{\bar{q} \in {\textit{crit}}_{i-1}^+ (\bar{f})} \sum_{\bar{\gamma} : \bar{p} \rightarrow \bar{q}} \epsilon(\bar{\gamma}) \dfrac{1}{|\pi^{-1} (\bar{q})|} \dfrac{|G|}{|G_{\bar{\gamma}}|} \cdot[\bar{q}] \\
&=& \sum_{\bar{q} \in {\textit{crit}}_{i-1}^+ (\bar{f})} \sum_{\bar{\gamma} : \bar{p} \rightarrow \bar{q}} \epsilon(\bar{\gamma}) \dfrac{|G_{\bar{q}}|}{|G|} \dfrac{|G|}{|G_{\bar{\gamma}}|} \cdot [\bar{q}] \\
&=& \sum_{\bar{q} \in {\textit{crit}}_{i-1}^+ (\bar{f})} \sum_{\bar{\gamma} : \bar{p} \rightarrow \bar{q}} \epsilon(\bar{\gamma}) \dfrac{|G_{\bar{q}}|}{|G_{\bar{\gamma}}|} \cdot[\bar{q}].\end{aligned}$$
Here, by $G_{\bar{q}}$, we denote the conjugacy class of $G_q$, $q \in \pi^{-1} (\bar{q})$. The sum is taken over all orientable critical points $\bar{q}$ of index $k-1$. Also $ \epsilon(\bar{\gamma}) = \pm 1$ assigned to $\bar{\gamma}$ is from the sign convention explained before. From now on, we use $\bar{p}$ itself instead of $[\bar{p}]$ for simplicity.
We denote $$n(\bar{p},\bar{q}) :=\sum_{\bar{\gamma} : \bar{p} \rightarrow \bar{q}} \epsilon(\bar{\gamma}) \frac{|G_{\bar{q} } |}{|G_{\bar{\gamma}}|} ,
\;\; \nu_{\bar{q}} (\bar{\gamma}):= \epsilon(\bar{\gamma}) \frac{|G_{\bar{q}}|}{|G_{\bar{\gamma}}|}.$$ On a minimal chart around $\bar{q}$, the preimage of $\bar{\gamma}$ is $|\nu_{\bar{q}} (\bar{\gamma})|$ copies of gradient flow lines which can be obtained by $G_{\bar{q}}$-action to a single lifting $\gamma$. (namely, the chart $({\widetilde{U}}_{\bar{q}}, G_{\bar{q}}, \pi_{\bar{q}})$ in which ${\widetilde{U}}_{\bar{q}}$ is an open subset of an Euclidean space equipped with linear $G_{\bar{q}}$-action and we assume that there is unique lifting $q$ of $\bar{q}$ which is the origin).
So, $\nu_{\bar{q}} (\bar{\gamma})$ can be regarded as a multiplicity of $\bar{\gamma}$ at $\bar{q}$ and $n(\bar{p},\bar{q})$ can be seen as the number of negative gradient flow lines from $\bar{p}$ to $\bar{q}$ counted with multiplicity or weight. We also denote $\nu_{\bar{p}} (\bar{\gamma}) = \epsilon(\bar{\gamma}) \frac{|G_{\bar{p}}|}{|G_{\bar{\gamma}}|} $, the number of liftings of $\bar{\gamma}$ in an uniformizing chart around $\bar{p}$ counting with signs.
Hence we obtain: $$\begin{aligned}
\label{fin def}
\partial \bar{p} &=& \sum_{\bar{q}} \left( \sum_{\bar{\gamma} : \bar{p} \rightarrow \bar{q}} \epsilon(\bar{\gamma}) \dfrac{|G_{\bar{q}}|}{|G_{\bar{\gamma}}|} \right) \bar{q} \\
&=& \sum_{\bar{q}} \sum_{\bar{\gamma} : \bar{p} \rightarrow \bar{q}} \nu_{\bar{q}} (\bar{\gamma}) \,\, \bar{q} = \sum_{\bar{q}} n(\bar{p},\bar{q}) \,\, \bar{q}. \nonumber\end{aligned}$$
We emphasize that $\nu_{\bar{q}} (\bar{\gamma})$ and $n (\bar{p},\bar{q})$ make sense for an arbitrary orbifold ${\boldsymbol{X}}$ with a given Morse-Smale function. Namely, coefficients of are intrinsic, only considering the critical points of $\bar{f}$, gradient flow lines in the orbit space, and the local groups at each critical points. Also note that if the group action is trivial we get the usual formula of the Morse boundary operator. In the next section, we define a Morse-Smale-Witten complex of a general orbifold using the above formula.
We would like to introduce an alternative formula of the Morse boundary operator which is also intrinsic in the above sense. We simply use $$\left<\bar{p} \right> := \frac{|G_{\bar{p}}|}{|G|} \sum_{ p \in \pi^{-1}} (\bar{p})$$ instead of $[\bar{p}]$. Note that $\left< \bar{p} \right>$ can be seen as the average of $p$ with respect to $G$-action since $\frac{|G|}{|G_{\bar{p}}|}$ is the cardinality of the orbit containing $p$. With this slight change of generators, the boundary operator is computed as follows. $$\partial \left<\bar{p} \right> = \sum_{\bar{q} \in {\textit{crit}}_{i-1}^+ (\bar{f})} \sum_{\bar{\gamma} : \bar{p} \rightarrow \bar{q}} \epsilon(\bar{\gamma}) \dfrac{|G_{\bar{p}}|}{|G_{\bar{\gamma}}|} \cdot \left< \bar{q} \right>,$$ which has a nice shape as much as the old one. To avoid the confusion, we will use $\underline{\partial}$ for general orbifolds to denote the operator coming arise from the above choice of generators. i.e. $$\underline{\partial} \bar{p} := \sum_{\bar{q} \in {\textit{crit}}_{i-1}^+ (\bar{f})} \sum_{\bar{\gamma} : \bar{p} \rightarrow \bar{q}} \epsilon(\bar{\gamma}) \dfrac{|G_{\bar{p}}|}{|G_{\bar{\gamma}}|} \cdot \bar{q}.$$ Note that two homology groups are obviously isomorphic by the scaling $$\psi : \bar{p} \mapsto |G_{\bar{p}}| \cdot \bar{p},$$ which is a chain map with respect to $(\partial, \underline{\partial})$ since we are using ${\mathbb{R}}$-coefficients for both sides.
Morse-Smale-Witten complex of General Orbifolds
===============================================
From now on, let ${\boldsymbol{X}}$ be a compact oriented connected $n$-dimensional effective orbifold, which may not be a global quotient orbifold. It is known that we can still choose a Morse function on ${\boldsymbol{X}}$ (see [@H]) in the sense of definition \[orbM\]. Unfortunately, it is not always possible to choose it to be a Morse-Smale function. We assume that we already have a Morse-Smale function $\bar{f}$ on ${\boldsymbol{X}}$. We write $X$ the underlying orbit space of ${\boldsymbol{X}}$ and the same $\bar{f}$ for the underlying continuous map on the orbit space $X \to {\mathbb{R}}$.
We define the orientability of critical points as in Definition \[def:type\]. The index of critical points can be defined, and these two notions are well-defined independent of choice of local uniformizing charts.
Again, we denote by ${\textit{crit}}_k^+ (\bar{f})$ by the set of all orientable critical points of $\bar{f}$ of index $k$. Regardless of $\bar{p} \in {\textit{crit}}(\bar{f})$ being orientable or not, we fix an orientation of $W^- (p)$ for future reason. It can be happen that $g \in G_{\bar{p}}$ reverses the given orientation on $W^-(p)$. Let $CM_k ({\boldsymbol{X}}, \bar{f})$ be the ${\mathbb{R}}$-vector space generated by ${\textit{crit}}_k^+ (\bar{f})$.
By using the notation of , we define
$$\partial \bar{p} = \sum_{\bar{q} \in {\textit{crit}}_{k-1}^+ (\bar{f})} n(\bar{p},\bar{q}) \, \bar{q} = \sum_{\bar{q} \in {\textit{crit}}_{k-1}^+ (\bar{f})} \, \sum_{\bar{\gamma} \in \mathcal{M} (\bar{p},\bar{q})} \nu_{\bar{q}} (\bar{\gamma}) \, \bar{q},$$ for $\bar{p} \in {\textit{crit}}_k^+ (\bar{f})$.
The main theorem of this section is
\[thm:main\] $( CM_\ast ({\boldsymbol{X}}, \bar{f}), \partial)$ defines a chain complex. i.e $\partial^2=0$
The proof of this theorem will occupy the rest of the section. Before we proceed for its proof, we first explain the main difference from the case of manifolds.
![limits of gradient flows for (a) manifolds and (b) orbifolds[]{data-label="flow1"}](flow){height="2in"}
Recall that the standard Morse homology argument to show $\partial^2=0$ uses the compactifications of moduli spaces of negative gradient flow lines between critical points whose index difference is $2$. The same analysis still works on the uniformizing covers, but there is a crucial difference. As illustrated in the figure a given broken trajectory (representing $\partial ^2$) on $X$ can become limits of several smooth flow trajectories (different even after the local group action). This can be easily seen as in the following example.
Let $\bar{\gamma}$ be the negative gradient flow lines from $\bar{p}$ to $\bar{q}$ and $\bar{\delta}$ be the one from $\bar{q}$ to $\bar{r}$. Assume $G_{\bar{\gamma}} = G_{\bar{\delta}} = 1$, for simplicity. Then, on an uniformizing neighborhood $({\widetilde{U}}_{\bar{q}}, G_{\bar{q}}, \pi_{\bar{q}})$, there are $|G_{\bar{q}}|$-flow lines which lift $\bar{\gamma}$ and also $|G_{\bar{q}}|$-flow lines which lift $\bar{\delta}$.
Choose $\gamma$, $\delta$ to be one of the flow trajectories covering $\bar{\gamma}$ and $\bar{\delta}$ in the cover ${\widetilde{U}}_{\bar{q}}$. For each $g_1,g_2 \in G_{\bar{q}}$, the choice of lifts $g_1 \cdot \gamma$ together with $g_2 \cdot \delta$ gives another broken trajectory in the cover.
In this way, we find $|G_{\bar{q}}|^2$-broken trajectories of the lifting $f$ of $\bar{f}$ in $\tilde{U}_{\bar{q}}$ and hence, we have $|G_{\bar{q}}|^2$-families of smooth gradient flow lines converging to each of $|G_{\bar{q}}|^2$-broken trajectories in ${\widetilde{U}}_{\bar{q}}$. Since $G_{\bar{q}}$-action on the set of broken trajectories in ${\widetilde{U}}_{\bar{q}}$ is free and so is on the set of (local) gluings in ${\widetilde{U}}_{\bar{q}}$, we get $ |G_{\bar{q}}|(=|G_{\bar{q}}|^2 / |G_{\bar{q}}| )$-families after quotient by $G_{\bar{q}}$-action. We see that there are $|G_{\bar{q}}|$ different families of smooth trajectories converging to a single broken trajectory $(\bar{\gamma}, \bar{\delta})$ near $\bar{q}$ in this case.
Recall that the convergence of 1-parameter family of negative gradient flow lines to a broken trajectory is given as uniform convergence in locally compact subsets. Also, the standard gluing lemma which guarantees that there exists an 1-parameter family of flow lines which converges to the given broken trajectory is also carried out locally. So, these facts are still true for general orbifolds. We are going to use these facts freely.
For effective orbifolds, gradient vector fields can be studied in the following way, which we learned from E. Lerman. For any orbifold ${\boldsymbol{X}}$, its Frame bundle $Fr({\boldsymbol{X}})$ is a smooth manifold with a smooth, effective, and almost free $O(n)$-action. Then, ${\boldsymbol{X}}$ is naturally isomorphic to the quotient orbifold $Fr({\boldsymbol{X}})/O(n)$ ( Theorem 1.23 of [@ALR]). Or consider any manifold $M$ and compact Lie group $G$ with $X = M/G$ with smooth, effective, almost free $G$-action. Then, we can lift our Morse function $\bar{f} :X \to {\mathbb{R}}$ to a function $\tilde{f} : M \to {\mathbb{R}}$ simply by setting $\tilde{f}:= \bar{f} \circ \pi$. Note that the critical submanifolds of $\tilde{f}$ typically look like $\pi^{-1} (\bar{p}) \cong G / G_{\tilde{p}}$ for $\tilde{p} \in \pi^{-1} (\bar{p})$, $\bar{p} \in {\textit{crit}}(\bar{f})$. The hessian of $\tilde{f}$ along the normal direction to the critical submanifold $\pi^{-1} (\bar{p})$ precisely equals to that of $\bar{f}$ at $\bar{p}$. Therefore, we can conclude that the lift $\tilde{f}:M \to {\mathbb{R}}$ is Morse-Bott. Thus, the properties of gradient trajectories in $M$ are already well-known (see[@AB]) including analytic properties of gradient trajectories of $\tilde{f}$ such as convergence, gluing and so on. This implies the desired properties for gradient trajectories of $\bar{f}$.
Another way to define Morse-Smale-Witten complex for an orbifold ${\boldsymbol{X}}$ would be to use $G$-equivariant Morse-Bott complex for $\tilde{f}:M \to {\mathbb{R}}$. As discussed in remark \[ABrelation\], the construction in [@AB] provides such a complex if the assumptions of [@AB] are met, such as critical submanifolds are orientable, $G$-action preserves orientations of unstable and stable manifolds, and the submersiveness of evaluation maps from the trajectories. Even when all these conditions are met, the complex in [@AB] uses Cartan model of $BG$, and it is not clear what is the relation of the differential there and geometric counting of gradient flow-lines. For ${\mathbb{R}}$ or ${\mathbb{Q}}$-coefficients, the $G$-equivariant cohomology of $M$ in this setting is isomorphic to the singular homology of $M/G$ (Proposition 2.12 [@ALR]). Hence it would be interesting to find a relation between the construction of [@AB] and the construction in this paper in the above setting.
Consider the moduli space of gradient flow trajectories between critical points of index difference two, which is of dimension one. The above example illustrates that the compactified moduli space in this case topologically, near each broken trajectory, is given by a join of several copies of interval $[0,1)$ at $0$’s equipped with a $G_{\bar{q}}$ action. Also note that before compactification, the moduli space should be understood as an orbifold as the trajectories lies in the uniformizing covers with group actions on them. We also remark that the orbifold structures of each limiting trajectories to a given broken trajectory may not be isomorphic to each other in non-abelian cases.
To prove the main theorem, we first prove a couple of lemmas on stabilizers of the gradient flow trajectories.
First, we set the notations as follows: Consider $\bar{p} \in {\textit{crit}}_{k}^+ (\bar{f}), \bar{q},\bar{q}' \in {\textit{crit}}_{k-1}(\bar{f}), \bar{r} \in {\textit{crit}}_{k-2}^+(\bar{f})$. Note that $\bar{q}$ and $\bar{q}'$ are not assumed to be orientable. Let $\bar{\gamma}$ (resp. $\bar{\gamma}'$) be negative gradient flow lines from $\bar{p}$ to $\bar{q}$ (resp. $\bar{q}'$) and let $\bar{\delta}$(resp. $\bar{\delta}'$) be flow lines from $\bar{q}$ (resp. $\bar{q}'$) to $\bar{r}$. Suppose that two broken trajectories $(\bar{\gamma}, \bar{\delta})$ and $(\bar{\gamma}' , \bar{\delta}')$ are connected by 1-parameter family of negative gradient flow lines from $\bar{p}$ to $\bar{r}$. Take the set of flows lines in the above 1-parameter family and call it ${\mathcal{P}}$.
Even if ${\mathcal{P}}$ flows between two orientable critical point (and hence will be oriented as we shall see below), breaking points (either $\bar{q}$ or $\bar{q}'$) of limiting broken trajectories are not necessarily orientable. This is the reason why we didn’t impose any condition on the orientability of $\bar{q}$ and $\bar{q}'$. Indeed, example \[broken heart\] already shows this phenomenon. We shall see that, however, each of limiting broken trajectories has a well defined sign as a boundary of ${\mathcal{P}}$.
\[inteq\] ${\mathcal{P}}$ is an one dimensional oriented orbifold whose stabilizers $G_{\bar{\gamma}}$ are all isomorphic for each $\bar{\gamma} \in \mathcal{P}$. (This is an ineffective orbifold for nontrivial $G_{\bar{\gamma}}$.)
We first explain how to obtain a natural orbifold structure on ${\mathcal{P}}$. Fix $\bar{\gamma} \in {\mathcal{P}}$ and $t \in {\mathbb{R}}$. Let $({\widetilde{U}},G)$ be a uniformizing chart of some neighborhood $U \ni \bar{\gamma}(t)$ with the quotient map $\pi :{\widetilde{U}} \to U$ and $f$ be the lifting of $\bar{f}$ on ${\widetilde{U}}$. Consider the level set of $f = \bar{f}\circ \pi$, given as $f^{-1} (\bar{f} (\bar{\gamma}(t))$. Consider the intersection of the level set $f^{-1}(\bar{f} (\gamma(t))$ with the corresponding gradient flow trajectories (from $\mathcal{P}$) in ${\widetilde{U}}$ and denote the intersection by $\mathcal{T}$. From the standard Morse theory, $\mathcal{T}$ is an open one dimensional manifold. Then, as we consider the moduli space (before compactification, after ${\mathbb{R}}$-quotient), ${\mathcal{P}}$ is given by $\mathcal{T}/G$ in the neighborhood $U$. Hence, we can give $\mathcal{P}$ an orbifold structure locally as a suborbifold of ${\boldsymbol{X}}$. This does not depend on the choice of $t \in {\mathbb{R}}$ by lemma \[flow\]. And the resulting orbifold is oriented as we consider flows between orientable critical points as explained in section 2.
Now we show that stabilizers are isomorphic to each other. But this is clear, since any connected one dimensional orientable orbifold satisfies such a property since a finite group action on an interval say $(-1,1)$ is either identity or $x \mapsto -x$ up to diffeomorphism. But the latter cannot be orientation preserving. Hence local groups act trivially and hence the stabilizers are isomorphic to each other.
Next, we consider compactification ${\overline{{\mathcal{P}}}}$ of each component ${\mathcal{P}}$ by adding limit broken trajectories $(\bar{\gamma}, \bar{\delta})$, $(\bar{\gamma}', \bar{\delta}')$ to ${\mathcal{P}}$. To consider the orbifold structure of ${\overline{{\mathcal{P}}}}$, we compare the stabilizers of the limiting trajectories and that of its limit.
Consider the uniformizing chart $(\tilde{U}_{\bar{q}}, G_{\bar{q}}, \pi_{\bar{q}})$ around $\bar{q}$ with $U_{\bar{q}} = \pi_{\bar{q}} (\tilde{U}_{\bar{q}} )$. Let $\Gamma$ be the set of all liftings of $\bar{\gamma} \cap U_{\bar{q}}$ and $\Delta$ be that of $\bar{\delta} \cap U_{\bar{q}}$. Then $G_{\bar{q}}$ naturally acts on $\Gamma \times \Delta$ by the diagonal action. Recall that there is the unique gluing for a given broken trajectory in the uniformizing cover, and hence by diagonal action, the quotient set $\Gamma \times \Delta/G_{\bar{q}}$ can be seen as the set of all possible smooth trajectories converging to $(\bar{\gamma}, \bar{\delta})$ in ${\boldsymbol{X}}$. Here, we can observe that there may be several different gluings for the single broken trajectory $(\bar{\gamma}, \bar{\delta})$. Hence,
\[CP\] ${\mathcal{P}}$ determines an element of $\Gamma \times \Delta/G_{\bar{q}}$, say $ [ \gamma, \delta ] \in \Gamma \times \Delta/G_{\bar{q}}$ and this correspondence is one to one locally around $\bar{q}$.
Now, consider $G_{\gamma}$, $G_\delta$ the isotropy groups of $\gamma$ and $\delta$ respectively. Their intersection $G_{\gamma} \cap G_{\delta} \subset G_{\bar{q}}$ is regarded as the isotropy group at the boundary point $(\bar{\gamma}, \bar{\delta})$ of ${\mathcal{P}}$. We denote its conjugacy class as $G_{[\gamma, \delta ]}$. In general, the limit of isotropy groups are always a subgroup of the isotropy group at the limit point. For the moduli space of gradient flow trajectories, we also have the converse, which is crucial in proving $\partial^2=0$.
\[bdy\] $G_{[\gamma, \delta ]} \cong G_{\bar{x}}$ for any $\bar{x} \in \mathcal{P}$.
We prove that $G_{[\gamma, \delta] } \cong G_{\bar{x}}$, for sufficiently close $\bar{x}$ to the boundary point $(\bar{\gamma}, \bar{\delta})$. This will be enough by lemma \[inteq\]. Take a uniformizing neighborhood around $\bar{q}$, $({\widetilde{U}}_{\bar{q}}, G_{\bar{q}}, \pi_{\bar{q} })$ and consider the lifting of one parameter family ${\widetilde{\mathcal{P}}}$ converging to one of liftings $(\gamma, \delta)$ of $(\bar{\gamma}, \bar{\delta})$. Taking two different slice of $f$ one of which meets ${\gamma}$ and ${\delta}$ respectively, the usual continuity argument shows that $G_{x} \subset G_{\gamma}$ and $G_{x} \subset G_{\delta}$ for $x$ in the slice of $f$ project down to $\bar{x}$. Therefore $G_x \subset G_{[\gamma, \delta]}$. Conversely, assume there exists $g \in G_{\bar{q}}$ which fixes $(\gamma, \delta)$ but does not fix ${\widetilde{\mathcal{P}}}$. Then, $g \cdot {\widetilde{\mathcal{P}}}$ would be a different family from ${\widetilde{\mathcal{P}}}$ converging to the same limit $(\gamma, \delta )$. This is impossible from the standard Morse theory on the uniformizing chart.
Note that $G_{\gamma}$’s are conjugate to each other for liftings $\gamma$ of $\bar{\gamma}$ but the intersection $G_{\gamma} \cap G_{\delta}$ depends on each choice of lift and its cardinality may depend on the choice of the lift.
Therefore, the set $\overline{\mathcal{P}}$ maybe considered as ineffective orbifolds, where the same isotopy group acts on every points trivially. Also, it carries a natural orientation. As in the standard Morse theory, this can be used to prove that the natural orientation at the boundary broken trajectories of $\overline{\mathcal{P}}$ are opposite to each other. To be more precise, we introduce a sign rule for “boundary" of $\overline{{\mathcal{P}}}$. We will in fact show that $\Gamma \times \Delta$ inherits a sign rule from $\Gamma \times \Delta$. Namely, we have:
For $(\gamma, \delta) \in \Gamma \times \Delta$, define $\epsilon(\gamma, \delta)$ as the product of $\epsilon(\gamma)$ and $\epsilon(\delta)$. Whether $\bar{q}$ is orientable or not, $G_{\bar{q}}$-action on $\Gamma \times \Delta$ preserves $\epsilon (\gamma, \delta)$. i.e. $$\epsilon (g \cdot \gamma, g \cdot \delta) = \epsilon(\gamma, \delta)$$ for all $g \in G_{\bar{q}}$.
Suppose $g \in G_{\bar{q}}$ reverses the orientation of $W^- (q)$. (Otherwise, there’s nothing to prove.) Since $\bar{p}$ and $\bar{q}$ are both orientable, exactly the same argument in lemma \[cancel\] shows that $\epsilon (g \cdot \gamma) = - \epsilon(\gamma)$ and $\epsilon (g \cdot \delta) = - \epsilon(\delta)$. This proves the lemma.
From the lemma, the following sign rule makes sense.
For $[ \gamma, \delta] \in \Gamma \times \Delta / G_{\bar{q}}$, we assign a sign to it as $$\epsilon [\gamma, \delta] := \epsilon(\gamma, \delta) = \epsilon(\gamma) \cdot \epsilon(\delta).$$
If $\bar{q}$ is orientable, we can give $\bar{\gamma}$ and $\bar{\delta}$ well-defined signs. Clearly, $\epsilon[\gamma, \delta] = \epsilon(\bar{\gamma}) \cdot \epsilon(\bar{\delta})$ for all $[\gamma, \delta] \in \Gamma \times \Delta / G_{\bar{q}}$ since $G_{\bar{q}}$-action preserves all signs in concern. So, the orientation of broken trajectories $(\bar{\gamma}, \bar{\delta})$ as a boundary of any component of ${\mathcal{P}}$ converging to it is given by $\epsilon(\bar{\gamma}) \cdot \epsilon(\bar{\delta})$.
Consequently, the orientation issue of $\overline{{\mathcal{P}}}$ can be rephrased as follows:
If there is an one parameter family $\mathcal{P}$ which corresponds to $[\gamma, \delta]$ and $[\gamma', \delta']$ in the sense of lemma \[CP\], $\epsilon[\gamma,\delta]$ and $\epsilon[\gamma',\delta']$ should be opposite.
As we said, the proof is not different from the classical one at all. (see [@AB] for example). This sign cancellation is the base of proving $\partial^2 =0$ in smooth case. But, to count gradient flow trajectories and to describe cancellation phenomenon in orbifold case correctly, we should take a weighted sum to take into account the orbifold structure.
For the compactification $\overline{\mathcal{P}}$ as above, the following expression will be called the weighted boundary of $\overline{\mathcal{P}}$: $$\partial {\overline{\mathcal{P}}} = \dfrac{ \epsilon[\gamma, \delta] }{|G_{[\gamma, \delta]}|} (\bar{\gamma}, \bar{\delta}) + \dfrac{\epsilon[\gamma', \delta']}{|G_{[\gamma', \delta']}|} (\bar{\gamma}', \bar{\delta}').$$ We call the numbers $\frac{ \epsilon[\gamma, \delta] }{|G_{[\gamma, \delta]}|} $, $\frac{\epsilon[\gamma', \delta']}{|G_{[\gamma', \delta']}|}$ the weights and write them as $\omega_{\mathcal{P}}(\bar{\gamma}, \bar{\delta})$, $\omega_{\mathcal{P}}(\bar{\gamma}', \bar{\delta}')$, respectively.
Now, the standard arguments of proving $\partial^2=0$ in the smooth case together with the above choice of weights gives the following equation. (lemma \[bdy\]) $$\label{bdysum1}
\sum_{(\bar{\zeta},\bar{\eta}) \in \partial {\overline{\mathcal{P}}}} \omega_{\mathcal{P}}( \bar{\zeta},\bar{\eta} ) = 0$$
Denote by ${\overline{{\mathcal{M}}}}(\bar{p},\bar{r})$ the compactified moduli space of negative gradient flow lines from $\bar{p}$ to $\bar{r}$. Geometrically, as explained in the beginning of the section, this is given by several copies of compact intervals (equipped with trivial actions of corresponding isotropy groups) which are joined at boundary points if they define families of flow lines whose limits at that boundary coincide. Also note that the limiting flows to a fixed broken trajectory might have non-isomorphic stabilizers by lemma \[bdy\]. So we cannot really think of ${\overline{{\mathcal{M}}}}(\bar{p},\bar{r})$ as orbifold with boundary. This is somewhat different from the smooth case where compactified moduli spaces are manifolds with corners. Denote $\partial {\overline{\mathcal{M}}} (\bar{p},\bar{r}): = {\overline{\mathcal{M}}} (\bar{p},\bar{r}) - {\mathcal{M}}(\bar{p},\bar{r})$.
If $ (\bar{\zeta},\bar{\eta}) \in \partial {\overline{\mathcal{M}}}(\bar{p},\bar{r})$, we define $$\label{weight}
\omega ( \bar{\zeta},\bar{\eta}) := \sum_{\mathcal{P} \, \textrm{with}\, (\bar{\zeta},\bar{\eta}) \in \partial {\overline{\mathcal{P}}}} \omega_{\mathcal{P}} (\bar{\zeta},\bar{\eta}),$$ where the sum is taken over all 1-parameter family $\mathcal{P}$ one of whose boundary is $(\bar{\zeta},\bar{\eta})$. Finally, we denote the sum of all weight associated to the gluings converging to one of broken trajectories through $\bar{p}$, $\bar{q}$ and $\bar{r}$ as $$\label{brk}
\omega(\bar{p},\bar{q},\bar{r}) := \sum_{(\bar{\zeta},\bar{\eta}) \in \mathcal{M}(\bar{p},\bar{q}) \times \mathcal{M}(\bar{q},\bar{r} )} \omega(\bar{\zeta},\bar{\eta}).$$ For $ \partial {\overline{\mathcal{M}}}(\bar{p},\bar{r})$ (which is the set of all broken trajectories from $\bar{p}$ to $\bar{r}$), note that we have $\mathcal{M}(\bar{p},\bar{q}) \times \mathcal{M}(\bar{q},\bar{r}) \subset \partial {\overline{\mathcal{M}}} (\bar{p},\bar{r})$ for any $\bar{q}$.
Since all intervals contained ${\overline{\mathcal{M}}} (\bar{p},\bar{r})$ are oriented so that they are compatible with their boundary orientations, all terms in the sum of have the same signs.
From , we get the following equality. $$\label{bdysum2}
\sum_{\bar{q} \in {\textit{crit}}_{k-1} (\bar{f})} \omega(\bar{p},\bar{q},\bar{r}) =\sum_{(\bar{\zeta},\bar{\eta}) \in \partial {\overline{\mathcal{M}}} (\bar{p},\bar{r})} \omega(\bar{\zeta},\bar{\eta})= \sum_{\mathcal{P}} \sum_{(\bar{\zeta},\bar{\eta}) \in \partial {\overline{\mathcal{P}}} } \omega_{\mathcal{P}} (\bar{\zeta},\bar{\eta}) = 0$$
The following figure explains how we sum up weighted contributions near orientable critical points. The lines in the figure represent (oriented)1-dimensional moduli spaces, and these converge to broken trajectories, which are drawn as $\circ$’s. In the figure $(b)$, $\lambda_1 + \lambda_2+ \lambda_3 +\lambda_4 + \lambda_5$ contributes to $\omega (\bar{p}, \bar{q}, \bar{r})$, and summation on a neighboring dotted circle contributes to $\omega (\bar{p}, \bar{q}', \bar{r})$
![Shape of 1-dimensional moduli space near an orientable critical points (a) topologically and (b) considering orbifold structures[]{data-label="1moduli"}](moduli){height="3in"}
Near an unorientable critical points, additional cancellation phenomenon as in lemma \[cancel\] occurs, and this will be explained below in .
Now, we are ready to prove our main theorem \[thm:main\]
Observe that $$\nu_{\bar{r}} (\bar{\delta}) = \epsilon (\bar{\delta}) \dfrac{ |G_{\bar{r}}|}{|G_{\bar{\delta}}|} = \dfrac{ |G_{\bar{r}}|} {|G_{\bar{q}}|} \cdot \nu_{\bar{q}} (\bar{\delta}).$$ Therefore, $$\begin{aligned}
\partial^2 \bar{p} &=& \partial \left(\sum_{\bar{q} \in {\textit{crit}}_{k-1}^+ (\bar{f}) } \sum_{\bar{\gamma} \in \mathcal{M} (\bar{p},\bar{q}) }\nu_{\bar{q}} (\bar{\gamma}) \right) \\
&=& \sum_{\bar{r} \in {\textit{crit}}_{k-2}^+ (\bar{f})} \left( \sum_{\bar{q} } \sum_{(\bar{\gamma}, \bar{\delta}) \in \partial {\overline{\mathcal{M}}} (\bar{p},\bar{r}) } \nu_{\bar{q}} (\bar{\gamma}) \nu_{\bar{r}} (\bar{\delta}) \right) \bar{r} \\
&=& \sum_{\bar{r} \in {\textit{crit}}_{k-2}^+ (\bar{f})} |G_{\bar{r}}| \left( \sum_{\bar{q} } \sum_{(\bar{\gamma}, \bar{\delta}) \in \partial {\overline{\mathcal{M}}} (\bar{p},\bar{r}) } \dfrac{ \nu_{\bar{q}} (\bar{\gamma}) \nu_{\bar{q}} (\bar{\delta}) }{|G_{\bar{q} } |} \right) \bar{r} \\
&=& \sum_{\bar{r} \in {\textit{crit}}_{k-2}^+ (\bar{f})} |G_{\bar{r}}| \left( \sum_{\bar{q} \in {\textit{crit}}_{k-1}^+ (f)} \omega (\bar{p}, \bar{q}, \bar{r}) \right) \bar{r}\end{aligned}$$ where the last sum is taken over all broken trajectories $(\bar{\gamma}, \bar{\delta})$ through $\bar{p}$, $\bar{q}$ and $\bar{r}$. The last equality follows from the lemma below, which directly implies the theorem.
If $\bar{q}$ is orientable, then $$\sum_{ (\bar{\gamma}, \bar{\delta}) } \dfrac{ \nu_{\bar{q}} (\bar{\gamma}) \nu_{\bar{q}} (\bar{\delta}) }{|G_{\bar{q}}|}
= \omega(\bar{p},\bar{q},\bar{r}),$$ and if $\bar{q}$ is unorientable, then $\omega(\bar{p},\bar{q},\bar{r}) =0.$ Therefore, $$\label{eq:can}
\sum_{\bar{q} \in {\textit{crit}}_{k-1}^+ (\bar{f})} \omega(\bar{p},\bar{q},\bar{r}) = \sum_{\bar{q} \in {\textit{crit}}_{k-1} (\bar{f})} \omega(\bar{p},\bar{q},\bar{r}) =0.$$
The first identity is nothing but a direct application of the weighted version of the orbit counting lemma \[count\]. As before, in the uniformizing chart $({\widetilde{U}}_{\bar{q}}, G_{\bar{q}}, \pi_{\bar{q}})$ around $\bar{q}$ let $\Gamma$ be the set of all liftings of $\bar{\gamma} \cap \pi ( {\widetilde{U}}_{\bar{q}})$ and $\Delta$ be that of $\bar{\delta} \cap \pi ( {\widetilde{U}}_{\bar{q}} )$. The quotient space is the set of all gluings of $(\bar{\gamma}, \bar{\delta})$ in $X$. From \[brk\], $\omega(\bar{p},\bar{q},\bar{r})$ is equivalent to the “weighted" number of elements of the space $\bigcup \, \Gamma \times \Delta / G_{\bar{q}}$ for all broken trajectories $(\bar{\gamma}, \bar{\delta})$ through $\bar{q}$, where the weight $[\gamma, \delta] \in \Gamma \times \Delta / G_{\bar{q}}$ is given by $\frac{\epsilon [\gamma,\delta] }{ |G_{[ \gamma, \delta ]}|}$. Thus, the weighted number of elements in $\Gamma \times \Delta / G_{\bar{q}}$ is
$$\dfrac{1}{|G_{\bar{q}}|} \sum_{ (\gamma, \delta ) \in \Gamma \times \Delta } \dfrac{\epsilon [\gamma, \delta]}{|G_{[ {\gamma}, {\delta}]}|} \cdot | \{ g \in G_{\bar{q}} | g \cdot \gamma = \gamma, g \cdot \delta = \delta \}|$$
by lemma \[count\] and since $| \{ g \in G_{\bar{q}} | g \cdot \gamma = \gamma, g \cdot \delta = \delta \} | = |G_{\gamma} \cap G_{\delta}| = |G_{[ {\gamma}, {\delta}]}|$, it equals to $$\frac{1}{|G_{\bar{q}}|} \sum_{ (\gamma, \delta ) \in \Gamma \times \Delta } \epsilon [\gamma, \delta]
=\frac{1}{|G_{\bar{q}}|} \sum_{ (\gamma, \delta ) \in \Gamma \times \Delta } \epsilon(\gamma) \cdot \epsilon(\delta) .$$ If $\bar{q}$ is orientable, $\epsilon(\bar{\gamma}) \cdot \epsilon (\bar{\delta})$ is constant for all $(\gamma, \delta) \in \Gamma \times \Delta$ so that
$$\begin{aligned}
\dfrac{1}{|G_{\bar{q}}|} \sum_{ (\gamma, \delta) \in \Gamma \times \Delta } \epsilon [\gamma, \delta]
&=&
\dfrac{\epsilon (\bar{\gamma})\epsilon (\bar{\delta})}{|G_{\bar{q}}|} \sum_{ (\gamma, \delta ) \in \Gamma \times \Delta} 1
\\
&=& \dfrac{\epsilon (\bar{\gamma})\epsilon (\bar{\delta})}{|G_{\bar{q}}|} \cdot |\Gamma \times \Delta| \\
&=& \dfrac{\epsilon (\bar{\gamma})\epsilon (\bar{\delta})}{|G_{\bar{q}}|} \cdot \dfrac{ |G_{\bar{q}}| }{|G_{\bar{\gamma}}|} \cdot \dfrac{ |G_{\bar{q}}| }{|G_{\bar{\delta}}|} \\
&=& \dfrac{ \nu_{\bar{q}} (\bar{\gamma}) \nu_{\bar{q}} (\bar{\delta}) }{|G_{\bar{q} }|} .\end{aligned}$$
Therefore, $\omega(\bar{p},\bar{q},\bar{r}) = \sum_{ (\bar{\gamma}, \bar{\delta}) } \dfrac{ \nu_{\bar{q}} (\bar{\gamma}) \nu_{\bar{q}} (\bar{\delta}) }{|G_{\bar{q}}|}$, if $\bar{q}$ is orientable.
On the other hand, suppose $\bar{q}$ is unorientable. Pick any $g \in G_{\bar{q}}$ which reverses the orientation of $W^-(p)$. Then, $g$ gives a permutation $\Gamma \times \Delta$ by $g \cdot (\gamma, \delta) := (g \cdot \gamma, \delta).$ Note that $$\epsilon (g \cdot \gamma) \cdot \epsilon( \delta) = -\epsilon (\gamma) \cdot \epsilon( \delta).$$ By the same argument in the case of global quotients (lemma \[cancel\]), the number of elements in $\Gamma \times \Delta$ which have positive signs should agree with the number of elements with negative signs. Thus, $ \sum_{ (\gamma, \delta ) \in \Gamma \times \Delta } \epsilon(\gamma) \cdot \epsilon(\delta)=0 $ and $\omega(\bar{p},\bar{q},\bar{r}) =0$ when $\bar{q}$ is not orientable.
\[count\] Let $X$ be a finite set on which a finite group $G$ acts and suppose $X/G$ is a weighted set so that each element $[x] \in X/G$ has the weight $\lambda_{[x]}$. Then, $$\sum_{[x] \in X/G} \lambda_{[x]} = \dfrac{1}{|G|} \sum_{x \in X} \lambda_{[x]} \cdot |G_x|.$$
Just follow the standard proof of the Burnside’s lemma.
The proof of $\underline{\partial}^2=0$ is similar. Indeed, this is automatic since we have a $(\partial, \underline{\partial})$-chain map $\psi : \bar{p} \mapsto |G_{\bar{p}}| \cdot \bar{p}$ which is an (${\mathbb{R}}$-vector space) isomorphism.
Comparison with the Homology of the Orbit Space
===============================================
In this section, we show that the homology of the Morse-Smale-Witten complex of general orbifolds $(CM_*({\boldsymbol{X}}, \bar{f}), \partial)$ equals the singular homology of the orbit space. We assume in this section that $\bar{f}$ is self-indexing, meaning that $\bar{f} (\bar{p}_i) = \lambda_i$, $\lambda_i$ the Morse index of $\bar{p}_i$.
For the general case without self-indexing assumption, one may use the filtration $$X_k := \displaystyle\bigcup_{{\rm{ind}} (\bar{p}) \leq k} \underline{W}^- (\bar{p}),$$ $$\phi=X_{-1} \subset X_0 \subset X_1 \subset \cdots \subset X_{n-1} \subset X_n = X$$ instead of the one described below and proceed as in [@S], where $\underline{W}^- (\bar{p})$ is given in . Note that $X_k$ is compact.
We will apply the topological method of [@Ni] which uses the cell structure of ${\boldsymbol{X}}$ induced by Morse data of $f$. This kind of cell structure was already revealed by several authors, for example [@LT] and [@H].
[[(Theorem 7.6 of [@H])]{}]{} Let $\bar{p} \in {\textit{crit}}_k(\bar{f})$ and $\bar{f} (\bar{p})=c$. Suppose $\bar{p}$ is the only critical point in $\bar{f}^{-1} [c- \epsilon , c+ \epsilon]$ for small $\epsilon \ll \frac{1}{2}$. Then, $\bar{f}^{-1} (-\infty, c+\epsilon]$ is homotopic to $\bar{f}^{-1} (-\infty, c-\epsilon]$ along $D^k / G_{\bar{p}}$ attached with $\partial D^k / G_{\bar{p}}$, Here, $D^k$ is a small invariant disc in the unstable manifold in a uniformizing chart around $\bar{p}$ and hence endowed with $G_{\bar{p}}$-action.
See theorem 7.6 of [@H] and compare it with 3.2 of [@M].
We need an elementary fact of equivariant topology to compute homological information of attaching cells.
[[(Theorem 2.4 of [@BR])]{}]{} \[br\] Let $K$ be a (regular) $G$-simplicial complex with $G$ finite and $L$ be a subcomplex. Then, $$H_\ast (K,L ; {\mathbb{R}})^G \cong H_\ast (K/G, L/G ; {\mathbb{R}}),$$ where the left hand side means the subset of $H_\ast (K,L ; {\mathbb{R}})$ fixed by $G$.
\[disk\] Let $D^n$ be the $n$-dimensional disc and the finite group ${\Gamma}$ act on $(D^n, \partial D^n)$. Then, the homology group $H_\ast (D^n / {\Gamma}, \partial D^n / {\Gamma};{\mathbb{R}})$ are given as follows. $$H_i (D^n / {\Gamma},\partial D^n/ {\Gamma};{\mathbb{R}})=
\left\{
\begin{array}{ll}
{\mathbb{R}}& \mbox{ if}\,\,\, i=n \mbox{ and} \,\,\,{\Gamma}\mbox{ preserves
the orientation of} \,\,\,D^n\\
0 & \, \mbox{otherwise}
\end{array}\right.$$
It suffices to note that there exists a ${\Gamma}$-invariant triangulation of $(D^n, \partial D^n)$ by [@IL] and that we can achieve the regularity condition in the theorem by subdivisions.
Recall that in smooth case, the coefficient of $q$ in $\partial p$ is defined from the (relative) intersection number between the unstable of $p$ and the stable manifold of $q$(see for example [@Ni]). In what follows, we will use instead the integration of Thom forms as they are more suitable in the orbifold setting. Recall from [@CR] (or [@ALR]) that Thom form of suborbifold $\bold{N}$ of ${\boldsymbol{X}}$ is defined locally as an invariant Thom forms of the preimage of $\bold{N}$ in each uniformizing chart. Let $N$ denote the underlying space of $\bold{N}$.
\[Thom\] On an uniformizing chart, integration of Thom form (which defined by that of the preimage ${\widetilde{N}}$ of $N$) along a normal fiber of ${\widetilde{N}}$ at $p \in {\widetilde{N}}$ is $1$ where $ \pi (p) = \bar{p}$, according to the usual definition of Thom forms on Euclidean spaces. Hence, the orbifold integration of the Thom form along a normal fiber of $\bold{N}$ in ${\boldsymbol{X}}$ at $\bar{p}$ is $1 / |G_{\bar{p}}|$. See [@ALR] for more details about Thom forms and Poincare duals of suborbifolds.
We will also need the Stokes theorem for orbifolds, which goes back to [@Sa]. We recall it here for readers convenience. A $C^\infty$ singular simplex $\bar{s}$ of dimension $k$ in ${\boldsymbol{X}}$ is defined by a smooth map $\bar{s}$ from a $k$-dimensional simplex $\Delta_k $ to $X$. Suppose the image of $\bar{s}$ lies in a single uniformizing chart $({\widetilde{U}}, G, \pi)$ so that it admits a lifting $s : \Delta_k \to {\widetilde{U}}$ with $ \pi \circ s = \bar{s}$. Consider a $k$-form $\bar{\omega}$ on $\pi({\widetilde{U}})$, which is given by an invariant $k$-form $\omega$ on ${\widetilde{U}}$. We define $$\int_{\bar{s}} \bar{\omega} = \int_{\Delta_k} s^\ast \omega.$$ For general $\bar{s}$, use a partition of unity to define $\int_{\bar{s}} \bar{\omega} $. One can prove Stokes formula: $$\int_{\bar{s}} d \bar{\omega} = \int_{\partial \bar{s}} \bar{\omega}.$$
\[comp\] The homology of $(CM_*({\boldsymbol{X}}, \bar{f}), \partial)$, which is constructed in section 4, equals the singular homology of the underlying space $X$.
We begin with the filtration of singular homology of $X$. Let $X_k = \bar{f}^{-1} (-\infty, k+1- \epsilon)$, $0<\epsilon \ll 1$, and $Y_k = \bar{f}^{-1} [k-\epsilon, k+1-\epsilon]$. Then, $$\label{filt}
C_\ast ( X_0 ; {\mathbb{R}}) \subset C_\ast ( X_1 ;{\mathbb{R}}) \subset \cdots \subset C_\ast ( X_n ; {\mathbb{R}})= C_\ast ( X ; {\mathbb{R}})$$ gives a filtration on the singular chain complex $C_\ast ( X;{\mathbb{R}})$. For a critical point $\bar{p}$ with $\bar{f}(\bar{p})=k$, let $\underline{W}^\pm (\bar{p})$ the stable and unstable manifolds at $\bar{p}$, respectively. i.e. $$\label{undW}
\underline{W}^\pm (\bar{p}) = \{ \bar{x} \in X : \lim_{t \to \pm \infty} \bar{\Phi}_t (\bar{x}) = \bar{p} \}.$$ Set $\underline{D}^\pm (\bar{p}) = \underline{W}^\pm (\bar{p}) \cap Y_k$. Then, topologically (since $\epsilon$ is small enough) $$\underline{D}^\pm (\bar{p}) \cong D^\pm (p) / G_{\bar{p}},$$ where $D^\pm (p)$ are small invariant neighborhoods of $p \in \pi^{-1} (\bar{p})$ in stable and unstable manifolds of $p$ with respect to the lift $f$ of $\bar{f}$. By $\partial \underline{D}^\pm (\bar{p})$, we mean the image of $\{ \partial D^\pm (p) \} / G_{\bar{p}}$, equivalently $$\partial \underline{D}^+ (\bar{p}) = \underline{D}^+ (\bar{p}) \cap \{ \bar{f} = k+1-\epsilon \}.$$ $$\partial \underline{D}^- (\bar{p}) = \underline{D}^- (\bar{p}) \cap \{ \bar{f} = k-\epsilon \}.$$ By the excision, we have (see [@H]) $$H_\ast (X_k, X_{k-1} ;{\mathbb{R}})=
\left\{
\begin{array}{ll}
\displaystyle\bigoplus_{\bar{p} \in {\textit{crit}}_k (\bar{f})} H_k (\underline{D}^- (\bar{p}) , \partial \underline{D}^- (\bar{p}) ; {\mathbb{R}}) \,\, & \ast=k \\
\qquad \qquad \qquad \,\,\,\,\, 0 & \, \mbox{otherwise}
\end{array}\right.$$ From \[disk\], $H_k (\underline{D}^- (\bar{p}) , \partial \underline{D}^- (\bar{p}) ; {\mathbb{R}}) \cong H_k (D^- (p), \partial D^- (p) )^{G_{\bar{p}} }$ is isomorphic to ${\mathbb{R}}$ if $\bar{p}$ is orientable and vanishes otherwise.
The $E^1$-terms of are the following chain complex. $$\cdots \to H_{k+1} (X_{k+1}, X_{k} ;{\mathbb{R}}) \to H_k (X_k, X_{k-1} ;{\mathbb{R}}) \to H_{k-1} (X_{k-1}, X_{k-2} ;{\mathbb{R}}) \to \cdots ,$$ where the boundary map is given by the composition $$H_{k} (X_{k}, X_{k-1} ;{\mathbb{R}}) \to H_{k-1} (X_{k-1} ;{\mathbb{R}}) \to H_{k-1} (X_{k-1}, X_{k-2} ;{\mathbb{R}})$$ Let $st (\underline{W}^-(\bar{p}) )$ be the conjugacy class represented by the subgroup $$\{ g \in G_{\bar{p}} : g \cdot x = x , \forall x \in D^- (p) \}$$ of $G_{\bar{q}}$. We now choose a generator of $H_{k} (X_{k}, X_{k-1} ;{\mathbb{R}}) $ to be a multiple of $$(\underline{D}^- (\bar{p}) , \partial \underline{D}^- (\bar{p}) ) \subset (X_{k}, X_{k-1} )$$ for $\bar{p} \in {\textit{crit}}_k^+ (\bar{f})$. More precisely, we denote “$\frac{1}{| st (\underline{W}^-(\bar{p}) )|}$-times this generator" by $\left| \bar{p} \right>$ to get rid of the influence of the ineffective action of $G_{\bar{p}}$ on $D^- (p)$. i.e. $$\left| \bar{p} \right> := \frac{1}{| st (\underline{W}^-(\bar{p}) )| } \cdot \left[ (\underline{D}^- (\bar{p}) , \partial \underline{D}^- (\bar{p}) ) \right] \in H_{k} (X_{k}, X_{k-1} ;{\mathbb{R}}).$$
Similarly, we set $\left| \bar{q} \right> = \frac{1}{| st (\underline{W}^-(\bar{q}) )| } \cdot \left[ (\underline{D}^- (\bar{q}) , \partial \underline{D}^- (\bar{q}) ) \right]$ for $\bar{q} \in {\textit{crit}}_{k-1}^+ (\bar{f})$, generating $H_{k} (X_{k-1}, X_{k-2} ;{\mathbb{R}}) $. Then, there exist real numbers $a_{\bar{q}}$ for each $\bar{q}$ such that $$\label{ket}
\partial \left| \bar{p} \right> = \sum_{ \bar{q} \in {\textit{crit}}_{k-1}^+(\bar{f})} a_{\bar{q}} \, \left| \bar{q} \right>,$$ or, equivalently, there exist $a_{\bar{q}}$ for each $\bar{q}$ satisfying $$\label{ket'}
\frac{1}{| st (\underline{W}^-(\bar{p}) )| } \cdot \left[ \partial \underline{D}^- (\bar{p}) ) \right]
= \sum_{ \bar{q} \in {\textit{crit}}_{k-1}^+(\bar{f})} \frac{a_{\bar{q}}}{| st (\underline{W}^-(\bar{q}) )| } \cdot \left[ (\underline{D}^- (\bar{q}) , \partial \underline{D}^- (\bar{q}) ) \right].$$ We remark that the following proof is rather complicated than that of [@Ni] for the case of manifolds, as we are working with quotient spaces in the above filtration.
Now, it is enough to show that $$a_{\bar{q}} = n (\bar{p}, \bar{q}) = \sum_{\bar{\gamma} \in \mathcal{M} (\bar{p},\bar{q}) } \epsilon(\bar{\gamma}) \dfrac{|G_{\bar{q}}|}{|G_{\bar{\gamma}}|} .$$
For this, we will consider chains $\partial \left| \bar{p} \right>$, and $\left| \bar{q} \right>$ in the subspace $Y_{k-1}$ of $X$ and use Thom form of $\underline{D}^+(\bar{q})$ to identify the constant $a_{\bar{q}} = n (\bar{p},\bar{q})$. More precisely, consider the following two sets, $$\mathring{Y}_{k-1} := Y_{k-1} \setminus \{ \bar{f}=k-1-\epsilon \}$$ and $$\mathring{\underline{D}}^+(\bar{q}):=\underline{D}^+ (\bar{q}) \setminus \partial \underline{D}^+ (\bar{q}) \subset \mathring{Y}_{k-1}.$$ Since $\bar{f}$ is invariant, $\mathring{Y}_{k-1}$ carries a natural orbifold structure inherited from ${\boldsymbol{X}}$ and $\mathring{\underline{D}}^+(\bar{q})$ can be considered as a suborbifold of $\mathring{Y}_{k-1}$. So one can take $\eta_{\bar{q}}$, a Thom form of $\mathring{\underline{D}}^+(\bar{q})$ on $\mathring{Y}_{k-1}$. On the other hand, setting $\underline{D}'^- (\bar{p}) := W^-(\bar{p}) \cap \bar{f}^{-1} [k-\epsilon', \infty)$ and $\partial \underline{D}'^- (\bar{p}) := \underline{D}^- (\bar{p}) \cap \bar{f}^{-1} (k-\epsilon')$ for $ \epsilon < \epsilon'$, we may identify $$\partial \left| \bar{p} \right> = \frac{1}{|st (\underline{W}^- (\bar{p}) )|} \partial \underline{D}'^- (\bar{p}),$$ by flowing $\partial \left| \bar{p} \right>$ along negative gradient flows from $\bar{f}^{-1}(\epsilon)$ to $\bar{f}^{-1}(\epsilon')$. This is to consider possible intersections of $\partial \underline{D}'^- (\bar{p})$ and $\mathring{\underline{D}}^+(\bar{q})$.
For simplicity, we denote $$b_{\bar{p}}:=|st (\underline{W}^- (\bar{p}))|,\; b_{\bar{q}}:=|st (\underline{W}^- (\bar{q}))|.$$
Now, by the definition of $a_{\bar{q}}$ in the identity , we have a formal sum of simplicial complexes $K$ which maps (say, via $\tau$) to $Y_{k-1}$, whose boundary $\tau:\partial K \to Y_{k-1}$ is given by $$\frac{1}{b_{\bar{p}}} \partial \underline{D}'^- (\bar{p}) \cup \bigcup_{\bar{q}} \frac{a_{\bar{q}}}{b_{\bar{q}}} \underline{D}^- (\bar{q}),$$ (with the opposite orientation on the first component) union some subset of $X_{k-2}$ as its image.
Here, we consider $\tau:K \to Y_{k-1}$, $\partial \underline{D}'^- (\bar{p})$ and $\underline{D}^- (\bar{q})$ as singular chains on $ \mathring{Y}_{k-1}$. The rational coefficients can appear as we work with ${\mathbb{R}}$-coefficients in singular homology. By subdividing simplices repeatedly if necessary, we may assume that the map $\tau$ when restricted to each simplex in $K$ has a lift in some uniformizing charts of ${\boldsymbol{X}}$.
Then, Stokes theorem of [@Sa] tells us that $$\int_{\partial K} \tau^\ast \eta_{\bar{q}} = \int_K d( \tau^\ast \eta_{\bar{q}}) = \int_K \tau^\ast (d \eta_{\bar{q}}) =0$$
We will compute the integral on the left hand side to get $a_{\bar{q}}$. Since the support of $\eta_{\bar{q}}$ can be shrunken so that it lies in an arbitrary small open neighborhood of $\mathring{\underline{D}}^+(\bar{q})$, (and since for each $\bar{q}$, $\underline{D}^+(\bar{q})$’s are disjoint,) $$\label{IJ}
\displaystyle\int_{\partial K} \tau^\ast \eta_{\bar{q}} = - \int_I \tau^\ast \eta_{\bar{q}} \,\, + \,\, \int_ J \tau^\ast \eta_{\bar{q}}=0,$$ where $I$ and $J$ are sub-complexes of $\partial K$ mapping to $\left|\bar{p} \right> =\frac{1}{b_{\bar{p}}} \partial \underline{D}'^- ( {\bar{p}} )$ and $a_{\bar{q}} \left|\bar{q} \right> =\frac{a_{\bar{q}}}{b_{\bar{q}}} \underline{D}'^- ({\bar{q}})$, respectively. The numbers $1 / b_{\bar{p}}$ and $ a_{\bar{q}} / {b_{\bar{q}}}$ here are considered to be coefficients of singular chains.
To compute the integral over $I$, we may assume $\epsilon$ is so small that there exists an uniformizing chart $({\widetilde{U}}_{\bar{p}}, G_{\bar{p}}, \pi_{\bar{p}})$ around $\bar{p}$ with $\partial \underline{D}'^- ( {\bar{p}} ) \subset \pi_{\bar{p}} ({\widetilde{U}}_{\bar{p}})$. Let $\partial {D}'^- ( p )$ be the preimage of $\partial \underline{D}'^- ( {\bar{p}} )$ and ${\widetilde{\eta_{\bar{q}} }}$ represent $\eta_{\bar{q}}$ on ${\widetilde{U}}_{\bar{p}}$. Then, the preimage of $\mathring{\underline{D}}^+(\bar{q})$, say $\mathring{D}^+(q)$, on $U$ will meet $\partial {D}'^- ( p )$, $\left( \sum_{\bar{\gamma} : \bar{p} \to \bar{q} }\frac{ \epsilon(\bar{\gamma}) |G_{\bar{p}}| }{ |G_{\bar{\gamma}} |} \right)$-times.
Obviously, $\int_I \tau^\ast \eta_{\bar{q}}$ equals to $1 / b_{\bar{p}}$-times the integral of $\eta_{\bar{q}}$ over $\partial \underline{D}'^- (\bar{p})$, where we mean by $\int_{\partial \underline{D}'^- (\bar{p})} \eta_{\bar{q}}$ the integration of ${\widetilde{\eta_{\bar{q}} }}$ over $\partial {D}'^- ( p )$ divided by $|G_{\bar{p}}| / |b_{\bar{p}}|$. Here, the division is due to the degree of the projection map $\pi_{\bar{p}} | : \partial D'^- (p) \to \partial \underline{D}' (\bar{p})$ since $G_{\bar{p}} / st( \underline{W}^- ({\bar{p}}))$ acts on $\partial D'^-(p)$ effectively. By the definition of $\eta_{\bar{q}}$, we get $$\begin{aligned}
\int_I \tau^\ast \eta_{\bar{q}} &=& \frac{1}{b_{\bar{p}} } \cdot \frac{ b_{\bar{p}} }{|G_{\bar{p}}|} \int_{\partial {D}'^- ( p ) } {\widetilde{\eta_{\bar{q}} }}\\
&=& \frac{1}{|G_{\bar{p}}|} \cdot \left( \sum_{\bar{\gamma} : \bar{p} \to \bar{q} }\frac{ \epsilon(\bar{\gamma}) |G_{\bar{p}}| }{ |G_{\bar{\gamma}} |} \right) \cdot \int_{F} {\widetilde{\eta_{\bar{q}} }} \\
&=& \sum_{\bar{\gamma} : \bar{p} \to \bar{q} }\frac{ \epsilon(\bar{\gamma}) }{ |G_{\bar{\gamma}} |},\end{aligned}$$ where $F$ denotes a general fiber of the normal bundle of $\mathring{D}^+(q)$ in ${\widetilde{U}}_{\bar{p}}$. Here, we use the transversality at intersection points of $\partial \underline{D}'^-(\bar{p})$ and $\underline{D}^+ (\bar{q})$.
On the other hand, using an uniformizing chart ${\widetilde{V}}_{\bar{q}}$ around $\bar{q}$ where the preimage $D^- (q)$ of $\underline{D}^- (\bar{q})$ is used to calculate $\int_J \tau^\ast \eta_{\bar{q}}$, $$\begin{aligned}
\int_ J \tau^\ast \eta_{\bar{q}} &=& \frac{a_{\bar{q}}} {b_{\bar{q}}} \cdot \frac{ b_{\bar{q}} }{|G_{\bar{q}}|} \int_{D^- (q)} {\widetilde{\eta_{\bar{q}} }} \\
&=& \frac{ a_{\bar{q}} }{|G_{\bar{q}}|} \int_F {\widetilde{\eta_{\bar{q}} }} \\
&=& \frac{ a_{\bar{q}} }{|G_{\bar{q}}|} .\end{aligned}$$ (We abbreviate ${\widetilde{\eta_{\bar{q}} }}$ to denote the representative of $\eta_{\bar{q}}$ on ${\widetilde{V}}_{\bar{q}}$.)
By comparing both integrals and , we conclude that
$$a_{\bar{q}} = \sum_{\bar{\gamma} \in \mathcal{M} (\bar{p},\bar{q}) } \epsilon(\bar{\gamma}) \dfrac{|G_{\bar{q}}|}{|G_{\bar{\gamma}}|}.$$
Note that there are several points in the proof where we use the fact that $CM_\ast({\boldsymbol{X}}, \bar{f})$ is defined on the field coefficient, although it would be still a chain complex using ${\mathbb{Z}}$-coefficients.
Therefore, under the existence of a Morse-Smale function $\bar{f}$ on ${\boldsymbol{X}}$, we can prove the Poincare duality of the singular homology of the orbit space $X$ by considering $-\bar{f}$. However, the inner product which gives the Poincare pairing between $HM_\ast ({\boldsymbol{X}}, \bar{f})$ and $HM_\ast ({\boldsymbol{X}}, -\bar{f})$ induced by slight different pairing $$<\, \, \, , \, \, > : CM_\ast ({\boldsymbol{X}}, \bar{f}) \otimes CM_\ast ({\boldsymbol{X}}, -\bar{f}) \to {\mathbb{R}},$$ where $< \bar{p}, \bar{p}> = 1 / |G_{\bar{p}}|$ and $<\bar{p},\bar{q}> = 0$ if $\bar{p} \neq \bar{q}$. Let $\partial_+$ and $\partial_-$ be boundary operators of $CM_\ast ({\boldsymbol{X}}, \bar{f})$ and $CM_\ast ({\boldsymbol{X}}, -\bar{f}) $, respectively. Then, $$\begin{aligned}
< \partial_+ \bar{p} , \bar{q} >&=& n(\bar{p},\bar{q}) <\bar{q},\bar{q}> \\
&=& \dfrac{1}{|G_{\bar{q}}|} \sum_{\bar{\gamma} \in \mathcal{M} (\bar{p},\bar{q}) } \epsilon(\bar{\gamma}) \dfrac{|G_{\bar{q}}|}{|G_{\bar{\gamma}}|} \\
&=& \sum_{\bar{\gamma} \in \mathcal{M} (\bar{p},\bar{q}) } \dfrac{ \epsilon(\bar{\gamma})}{|G_{\bar{\gamma}}|},\end{aligned}$$ and similarly, $$<\bar{p}, \partial_- \bar{q}> = n(\bar{q},\bar{p}) <\bar{p},\bar{p}> = \sum_{\bar{\gamma} \in \mathcal{M} (\bar{p},\bar{q}) } \dfrac{\epsilon(\bar{\gamma}) }{|G_{\bar{\gamma}}|}$$ so that $$< \partial_+ \bar{p} , \bar{q} >= <\bar{p}, \partial_-\bar{q}>$$ and $<\,\,\, , \,\,>$ induces a pairing on homologies.
To get a similar pairing between $(CM_\ast ({\boldsymbol{X}}, \bar{f} ), \underline{\partial})$ and $ (CM_\ast ({\boldsymbol{X}}, -\bar{f}) , \underline{\partial}) $, one should modify $<\,\,\, , \,\,>$ such that $< \bar{p}, \bar{p}> = |G_{\bar{p}}|$.
Morse-Bott case for global quotient orbifolds
=============================================
We generalize the construction of the section 2 of Morse functions on global quotients to the Morse-Bott case. Let a finite group $G$ act effectively on the compact smooth connected oriented manifold $M$ and assume that there is a $G$-invariant Morse-Bott function $f : M \to {\mathbb{R}}$.
$f$ is called Morse-Bott if ${\textit{crit}}(f)$ is a finite union of compact connected submanifold of $M$ and the Hessian of $f$ is nondegenerate at each critical submanifold in normal direction. In addition, we require the Morse-Smale condition that each unstable manifold and stable manifold (for each critical submanifold) intersect transversally.
\[ABrelation\] The construction in this section is related to the construction of equivariant Morse-Smale-Witten complex by Austin and Braam in [@AB].
1. The construction of Austin and Braam is for equivariant cohomology for manifold $M$ with $G$-action. We believe that authors [@AB] implicitly assume that $G$ is connected Lie group, as they have used the Cartan model, and have not discussed orientation issues.
2. For the case of finite $G$, the associated Lie-algebra is trivial, and the construction of [@AB] would provide a complex, built on invariant differential forms of critical submanifolds. But in general, $G$-action may not preserve the orientations of unstable directions (of critical submanifolds) or even orientations of critical submanifolds, and in such cases, one cannot obtain the singular cohomology of the quotient space (as discussed so far). Hence, the construction in this subsection may be considered as an extension of [@AB] for the case of finite $G$ with orientation issues considered.
3. We also consider slightly generalized setting without assuming that the end point map from the moduli space of the connecting trajectories (i.e. flow lines at $\pm \infty$) is submersive.
There are several approaches for Morse-Bott homology such as [@F], [@L] etc. Banyaga and Hurtubise assumed different transversality condition (called the Morse-Bott-Smale transversality condition) which enables them to go through issues about fiber product without perturbation. See [@BH].
If $S$ is a critical submanifold, the unstable manifold and the stable manifold over $S$ is defined as follows. $$W^- (S) := \{ x \in M : \lim_{t \to -\infty} \Phi_t (x) \in S\}$$ $$W^+ (S) := \{ x \in M : \lim_{t \to \infty} \Phi_t (x) \in S\}$$
By now, there exist several versions of Morse-Bott theory, and we will follow the one using currents as in [@Hu] (which is based on the construction by Fukaya[@F]). We do not use the de Rham version as in the setting of [@AB] because non-trivial assumptions about relative orientability of critical submanifolds and unstable manifolds over them in order to define fiberwise integration of differential forms. The currents were introduced to deal with the following problems: the end point map $e_+ : \mathcal{M} (S_1, S_2) \to S_2$ is not a submersion in general and hence, we cannot integrate differential forms along the fiber of this map. Furthermore, the moduli space $\mathcal{M} (S_1, S_2):= W^-(S_1) \cap W^+ (S_2) / {\mathbb{R}}$ might not be (invariantly) orientable, even when $S_1$, $S_2$ and $M$ are all orientable.
We will introduce the Morse-Bott chain complex of $M$(following that of Hutchings [@Hu]) and then restrict to $G$-invariant subcomplex.
Let $\sigma$ be a simplex in $S_1$ and define the following fiber product $$\overline{ \mathcal{M} (\sigma, S_2)} :=\sigma \times_{S_1} \overline{ \mathcal{M} (S_1, S_2)},$$ the set of flow lines from $\sigma$ to $S_2$. Moreover, we have the following diagram of fiber product. $$\label{fiberproduct}
\begin{diagram}
\node{ \overline{ \mathcal{M} (\sigma, S_2)}}\arrow{s,r}{}\arrow{e,t}{i} \node{\overline{\mathcal{M} (S_1, S_2) }} \arrow{s,r}{e_-} \arrow{e,t}{e_+} \node{S_2}\\
\node{\Delta^k}\arrow{e,t}{\sigma} \node{S_1} \end{diagram}$$ We require that $\sigma$ is transversal to $e_-$ because we want $\overline{ \mathcal{M} (\sigma, S_2)}$ to be a manifold with boundary (or with corners). Indeed we will think of $$\overline{\mathcal{M} (\sigma, S_2)} \xrightarrow{ i \circ e_+} S_2$$ as a current on $S_2$, given the genericity assumption on $\sigma$ which will be explained later. We will write this current as $ (i \circ e_+) \left[ \overline{\mathcal{M} (\sigma, S')} \right]$.
This is always possible by generic perturbation (when $G$-action is not involved). Note that if $e_+$ is a submersion, the desired transversality follows automatically.
For $\gamma \in \mathcal{M}(\sigma, S_2)$, orientations of $\sigma$, $W^- (p_1)$ and $W^- (p_2)$ determine a local orientation of $\mathcal{M} (\sigma, S_2)$ where $\gamma$ runs from $p_1$ to $p_2$. (see ).
We introduce the fiberwise orientation sheaf of the pair $(W^- (S),S)$ for a critical submanifold $S$ of $M$. It is a locally constant sheaf $\mathcal{O}$ whose stalk is defined by $$\mathcal{O}_p := H_{\mu (S)-1} (W^-(p) \setminus p) \simeq {\mathbb{R}},$$ where $\mu (S)$ is the number of negative eigenvalues of the Hessian of $f$ at a point in $S$ ,or simply, the Morse-Bott index of $S$. Let $C_\ast^{sing} (S, \mathcal{O})$ be the space of singular chains with coefficients in $\mathcal{O}$. i.e. $$C_\ast^{sing} (S, \mathcal{O}) := {\mathbb{R}}[ (\sigma, o) : \sigma \,\, \mbox{a simplex in}\,\, S, \, o \in \Gamma( Im (\sigma),\mathcal{O} ) ] / \sim ,$$ where $(\sigma, -o) \sim -(\sigma, o)$.
\[trans\] To define the fiber products, we consider a subspace of currents spanned by pairs $(\sigma, o)$ where $\sigma$ satisfies conditions, below.
1. $\sigma$ is smooth,
2. each face of $\sigma$ is transverse to $e_+$ of all moduli spaces of flow lines between critical submanifolds and all iterated fiber products thereof.
$\sigma$ satisfying the above conditions is generic. We will call such a simplex as a generic simplex.
Let $C_\ast (S, \mathcal{O})$ denote the resulting chain complex equipped with the differential defined in a standard way. We now define a chain complex as follows. The $k$-th chain group is $$C_k^{Bott} :=\displaystyle\bigoplus_S C_{k- \mu (S)} (S, \mathcal{O}).$$ If $\sigma \in C_\ast (S, \mathcal{O})$ is a generic simplex and for $S' \neq S$, we have a well-defined current $$(i \circ e_+) \left[ \overline{\mathcal{M} (\sigma, S')} \right] \in C_\ast (S', \mathcal{O}) .$$ For simplicity, from now on, we write $e_+$, instead of $i \circ e_+$ and also sometimes write $\sigma$ instead of $(\sigma, o)$ to denote elements of $C_\ast (S, \mathcal{O})$.
Define $$\label{D}
D \sigma := \partial{\sigma} + \displaystyle\sum_{S' \neq S} e_+ \left[ \overline{\mathcal{M} (\sigma, S')} \right].$$ One can easily check that $D : C_k^{Bott} \to C_{k-1}^{Bott}$ and the standard argument using the convergence of gradient flows shows $D^2 =0$.
Now, we consider finite group $G$ action on them. Since $G$ acts on singular chains of $S$ and on local sections of $\mathcal{O}$, we have an well-defined action of $G$ on $C_k^{Bott} $. Note that $G$ obviously preserves transversality condition of singular chains in Definition \[trans\], as the moduli spaces of gradient flows also admit $G$-action on them ($\overline{\mathcal{M}} (S_1, S_2) \xrightarrow{e_-} S_1$ is $G$-equivariant). We check whether the $G$-action is compatible with the differential $D$ or not.
The boundary operator $D$ is $G$-equivariant so that $$C_k^{Bott, G} := \{ \alpha \in C_k^{Bott} : g \cdot \alpha = \alpha \,\,\, \forall g \in G \}$$ forms a sub-complex of $C_k^{Bott}$. Similarly, we define invariant subcomplex $C_\ast^G (S, \mathcal{O})$ of $C_\ast (S, \mathcal{O})$.
As mentioned earlier, $\sigma$ in should be thought of as $(\sigma, o)$, where $o$ determines continuously varying orientations on fibers of $W^- (S) \to S$ over $\sigma$.
It suffices to show that $D ( g \cdot (\sigma, o)) = g \cdot D (\sigma, o)$ for $g \in G$ and $(\sigma, o) \in C_k^{Bott}$, which implies that the boundary operator $D$ preserves $C_\ast^{Bott, G}$. It is easy to see that the only non-trivial part of the above is the sign. We explain the related signs more explicitly.
$o$ will induce a fiberwise orientation of $e_+ \left[ \overline{ \mathcal{M} (\sigma, S') } \right]$, which is denoted as $D (o)$ (see ). Then we need to show $$\left(e_+ \left[ \overline{ \mathcal{M} ( (g \cdot \sigma, g \cdot o ) , S') }\right] , D(g \cdot o ) \right) = \left( g \cdot e_+ \left[ \overline{ \mathcal{M} ((\sigma,o), S') } \right] , g \cdot D (o) \right)$$ As sets without concerning orientations, the above identity holds because $G$-action is compatible with gradient flows of $f$. To see the equivalence of signs, we recall the sign rule in [@Hu](see [@F] also). The sign or orientation of $e_+ \left[ \overline{ \mathcal{M} (\sigma, S') } \right]$ is determined simply by the following equality of oriented vector spaces: $$T_{p_1} \sigma \oplus T_{p_1} W^- (p_1) \cong T_{\gamma} \mathcal{M} (\sigma, S') \oplus T\gamma \oplus T_{p_2} W^- (p_2),$$ where $p_1 \in \sigma$ and the flow line $\gamma$ runs from $p_1$ to $p_2 \in S'$. The sign of $(\sigma, o)$ is exactly described as the orientation of the left hand side. Since $T \gamma$ has a canonical orientaion (in negative gradient direction), the orientation on $T_{\gamma} \mathcal{M} (\sigma, S') \oplus T_{p_2} W^- (p_2)$ is determined by $(\sigma,o)$, which corresponds to the orientation of $e_+ \left[ \overline{ \mathcal{M} (\sigma, S') } \right]$ itself together with the fiberwise orientation of unstable manifold over $e_+ \left[ \overline{ \mathcal{M} (\sigma, S') } \right]$ at $p_2$. Now apply $g \in G$ to this equation . Then, we obtain: $$\label{si}
T_{g \cdot p_1} g \cdot \sigma \oplus T_{g \cdot p_1} W^- (g \cdot p_1) \simeq g \cdot \left( T_{\gamma} \mathcal{M} (\sigma, S') \oplus T_{p_2} W^- (p_2) \right) \oplus T ( g \cdot \gamma).$$ The left hand side determines $D(g \cdot o)$ whereas the orientation of the right hand side corresponds to $g \cdot D(o)$.
Suppose $f$ is not only Morse-Bott, but Morse. Then, for a non-orientable critical points $p \in {\textit{crit}}^- (f)$, consider the element with orientation, $(p, o) \in C_k^{Bott}$. Non-orientability of $p$ implies that some $g \in G_p$ (in fact half of $G_p$) reverses the orientation $o$, and hence, the average of $(p,o)$ via $G_p$ in fact vanishes. Thus $(p,o)$ is not an element of $C_k^{Bott, G}$. Hence, $C_k^{Bott, G}$ agrees with our earlier invariant Morse complex when $f$ is Morse. i.e. we have $$\begin{aligned}
\sum_{g \in G} g \cdot (p, o) &=&\sum_{g \in G^+} g \cdot (p, o) + \sum_{h \in G^-} h \cdot (p, o) \\
&=& \sum_{G^+} (p, o) - \sum_{G^-} (p, o) =0,\end{aligned}$$ where $G^\pm$ denote the set of elements preserving or reversing orientation, respectively.
Now, we show that the homology of $(C_*^{Bott, G},\partial)$ is isomorphic to the singular homology of $M/G$ with the assumption that $f$ is furthermore self-indexing. We first recall the compactification of unstable manifolds of $f$ in [@AB]. For a critical submanifold $S_i$ of index $i$ ($\mu (S_i)= i$) and a sequence $i_0<i_1<\cdots<i_m = i$, define $$Y_{i_0, i_1, \dots, i_m} = \mathcal{M} ( S_{i_m}, S_{i_{m-1}} ) \times_{S_{i_{m-1}}} \mathcal{M} ( S_{i_{m-1}}, S_{i_{m-2}} ) \times_{S_{i_{m-2}}}\cdots \times_{S_{i_1}} W^- (S_{i_0} ),$$ where the index of points in $S_{i_k}$ is $i_k$. Naively, $Y_{i_0, i_1, \dots, i_m}$ is a product of $W^- (S_{i_0} )$ with the moduli space of broken trajectories passing through $S_{i_{m-1}}, \cdots, S_{i_1}$ to $S_{i_0}$.
\[cpt\] [[(lemma 3.3 of [@AB])]{}]{} The unstable manifold $W^- (S_i)$ can be compactified so that $$\partial \overline{ W^- (S_i) } = \displaystyle\bigcup_{i_0<i_1<\cdots<i_m=i} Y_{i_0, i_1, \dots, i_m}.$$ Here, the codimension of the stratum $ Y_{i_0, i_1, \dots, i_m}$ in $ \overline{ W^- (S_i) }$ is exactly $m$. The compactified space has the structure of a manifold with corners.
Moreover, there is a natural immersion $\iota_{S_i}$ of the $\overline{ W^- (S_i) }$ into $M$, which is not injective nor proper in general. For example, the immersion of $\partial \overline{ W^- (S_i) }$ restricted on the boundary component $ Y_{i_0, i_1, \dots, i_m}$ is just the projection from $ Y_{i_0, i_1, \dots, i_m}$ to the last factor $ W^- (S_{i_0} )$. Also we have $$\partial \overline{ W^- (S) } = \bigcup_{S'} \overline{\mathcal{M} (S,S')} \times_{S'} W^- (S').$$ For $\tau \in C_k (S_i) \in C_{k+i}^{Bott}$, let $${W^-(\tau)} := \tau \times_{S_i} {W^- (S_i)}.$$ From our generecity assumption, this define a current on $M$. Note that $$\begin{aligned}
\dim {W^-(\tau) } &=& \dim \tau + \dim {W^- (S_i)} - \dim S_i \\
&=& k + i + \dim S_i - \dim S_i \\
&=& k + i\end{aligned}$$ Since the element of $C_k (S_i)$ also contains the information of the local orientation of fiber of $W^- (S_i) \to S_i$ around the image of simplex, say $o$, $(\tau, o)$ defines an orientation on ${W^-(\tau) } $. Therefore, we get a map $ \Phi : C_k^{Bott} = \oplus_i C_{k-i} (S_i) \to C_k (M)$ sending $\tau \mapsto {W^- (\tau)}$. Here, we regard $C_k (M)$ as a singular chain complex of $M$ generated by currents.
As a current on $M$, $W^- (S_i)$ is equivalent to its closure in $M$ or embedding of its compactification since they only differ from measure zero subsets of $M$.
We can show that $\Phi$ becomes isomorphism of homology groups when restricted to $C_\ast^{Bott, G}$. First, we prove
$\Phi : C_k^{Bott} \to C_k (M)$ is a $G$-equivariant chain map.
Consider $S \subset {\textit{crit}}(f)$ and $\sigma $, one of generators of $C_k (S, \mathcal{O})$. Then, we need to show: $$\Phi ( D \sigma) = \partial \Phi (\sigma),$$ where $\partial$ is the usual boundary operator of the singular chain complex of $M$. We will only check the identify as a set, and the check on orientation can be done without difficulty. So we will simply write $\sigma$ instead of $(\sigma, o)$. The left hand side can be written as $$\label{PhiD}
\Phi ( D \sigma) = \partial \sigma \times_S {W^- (S)} + \sum_{S' \neq S} e_+ \left[ \overline{\mathcal{M} (\sigma, S')}\right] \times_{S'} {W^- (S')}$$ On the other hand, $$\begin{aligned}
\partial \Phi (\sigma) &=& \partial \left( \sigma \times_S \overline{W^- (S)} \right) \\
&=& \partial \sigma \times_S \overline{W^- (S)} + \sigma \times_S \partial \overline{W^- (S)} \\
&=& \partial \sigma \times_S \overline{W^- (S)} + \sum_{S' \neq S} \sigma \times_S \left( \overline{\mathcal{M} (S, S')} \times_{S'} W^- (S') \right)\end{aligned}$$ Here, only codimension $1$ boundary components of $\overline{W^-(S)}$ appear since we are considering them as currents. The map $\overline{\mathcal{M} (S, S')} \times_{S'} W^- (S') \to S$ is induced by $e_-: \overline{\mathcal{M} (S, S')} \to S$. Now, the identification easily follows by considering the diagram .
Since $\Phi$ is $G$-equivariance, we get a map $\Phi : C_\ast^{Bott, G} \to C_\ast^G (M)$ by restricting original $\Phi$, where $C_\ast^G (M)$ denotes $G$-invariant singular chains in $M$. From now on, assume further that the Morse-Bott function $f$ satisfies the self-indexing condition. i.e $f(S_i) = i$. Then, we get a filtration of complex $C_\ast^G (M)$ as follows. Let $M_k = f^{-1} ( -\infty, k + \frac{1}{2})$. Each $M_k$ is $G$-invariant and gives a filtration on $C_\ast^G (M)$, $$0 \subset C_\ast^G (M_0) \subset C_\ast^G (M_1) \subset
\cdots \subset C_\ast^G (M_n) = C_\ast^G (M), \,\mbox{i.e.} \left( C_p^G (M) \right)^k =C_p^G (M_k)$$ where $n=\dim M$. For $C_\ast^{Bott, G}$, we take the following filtration: $$\left( C_p^{Bott, G} \right)^k = \bigoplus_{i \leq k} C_{p-i}^G (S_i, \mathcal{O}_i),$$ where $\mathcal{O}_j$ means the fiberwise orientation sheaf on $(W^-(S_j), S_j)$.
We prove that $\Phi$ induces an isomorphism between cohomology groups. As the homology of $C_\ast^G (M)$ is the homology of $M/G$ (see Theorem \[br\]), we have
\[isom\] $\Phi_* : H_*^{Bott, G} (M) \xrightarrow{\sim} H_* ( M/G )$ is an isomorphism.
We are going to show that $\Phi$ induces an isomorphism between the homology groups of associated graded complexes of $C_\ast^{Bott, G}$ and $C_\ast^G (M)$. Since our complexes are bounded, spectral sequences of $C_\ast^{Bott, G}$ and $C_\ast^G (M)$ (with respect to the filtrations given above) will converge to associated graded groups of homologies. Thus if we can prove $\Phi$ induces an isomorphism between $E^1$-terms of spectral sequences which are homology groups of associated graded complexes of $C_\ast^{Bott, G}$ and $C_\ast^G (M)$, then we will get the desired result. Let $$G {C'}_p^k = C_p^G (M_k) / C_p^G (M_{k-1})$$ and $$G C_p^k = C_{p-k}^G (S_k, \mathcal{O}_k).$$ They are precisely the $E^1$-terms of our complexes. The homology of the first complex is simply the $G$-invariant relative homology group of the pair $(M_k, M_{k-1})$. i.e. $$H_p (G {C'}_\ast^k ) = H_p^G (M_k, M_{k-1}).$$ For $ S_k = {\textit{crit}}(f) \cap (M_k \setminus M_{k-1} ) )$, define $\mathcal{F}_{k} = W^- (S_k) \cap \{x : k-\frac{1}{2} \leq f(x) \leq k+ \frac{1}{2} \}$. $\mathcal{F}_k$’s are $G$-invariant, obviously and hence there is natural $G$-action on $\mathcal{F}_k$. Combining theorem \[br\] and excision, we get: $$\begin{aligned}
H_p^G (M_k, M_{k-1}) &\simeq& H_p (M_k /G, M_{k-1} /G) \\
&\simeq& H_p (\mathcal{F}_k /G , \partial \mathcal{F}_k /G) \\
&\simeq& H_p^G (\mathcal{F}_k, \partial \mathcal{F}_k) \end{aligned}$$ (The gradient flow defines a deformation retract of the pair $(M_k, M_{k-1})$ onto $(\mathcal{F}_k \cup_{\partial \mathcal{F}_k} M_{k-1}, M_{k-1})$ and hence, a retract between pairs of $G$-quotients.)
Thus, what we are left with is to show the following morphism of chain complexes induces the isomorphism on homology groups of them: $$\Phi' : C^G_{p-k} (S_k, \mathcal{O}_k) \to C_p^G (\mathcal{F}_k, \partial \mathcal{F}_k).$$ Keeping track of our isomorphisms used above, one can find that this map is defined as $$\Phi' ( \sigma ) = \mathcal{F}_k (\sigma) := \sigma \times_{S_k} \mathcal{F}_k.$$ We again emphasize that chain $\sigma$ contains information of the coherent orientation of the fiber of $e_- | : \mathcal{F}_k \to S_k$ and that the genericity condition on $\sigma$ guarantees the well-definedness of this map. So, $\mathcal{F}_k (\sigma)$ carries a natural orientation from this information.
Observe that $\Phi'$ induces the homology Thom isomorphism, or dual to well-known cohomology Thom isomorphism defined by the integration along fiber. Even if neither a bundle nor a base is not orientable, the Thom isomorphism still holds true. However, we have to use a relative orientation sheaf as a coefficient ring (as we introduced $\mathcal{O}_k$). See theorem 7.10 of [@BT] or [@Sk] for more details.
D.M. Austin and P.J. Braam, [*Morse-Bott theory and Equivariant cohomology*]{} The Floer memorial volume, 123‚Äì183. Progr. Math. 133, Birkhäuser, Basel, 1995.
G. Bredon, [*Introduction to Compact Transformation Groups*]{} Academic Press (1972) R. Bott and L.W. Tu, [*Differential Forms in Algebraic Topology*]{} Grad. Text. Math. 82, Springer C.-H. Cho and M. Poddar [*Holomorphic orbidiscs and Lagrangian Floer cohomology of toric orbifolds*]{} in preparation. W. Chen and Y. Ruan [*A new cohomology theory of orbifold*]{}, Comm. Math. Phys. 248 (2004), no. 1, 1–31.
W. Chen and Y. Ruan [*Orbifold Gromov–Witten theory,* ]{} Orbifolds in mathematics and physics (Madison, WI, 2001), pp. 25–85, Contemp. Math. 310, Amer. Math. Soc., Providence, RI, 2002.
L. Dixon, V. Harvey, C. Vafa and E. Witten [*Srings on orbifolds,*]{} Nucl. Phys. B 261 (1985), no. 4, 678–686. K. Fukaya [*Floer homology of connected sum of homology 3-spheres*]{} Topology, 35, no.1, 89-136, 1996 V. Guillemin and A. Pollack, [*Differential Topology*]{}, Prentice-Hall 1974. R. Hepworth, [*Morse Inequalities for Orbifold Cohomology*]{} Algebr. Geom. Topol., **9** (2009) 1105-1175 M. Hutchings, [*Lecture notes on Morse homology*]{} Available http://math.berkeley.edu/ hutching/ S. Illman [*Smooth Equivariant Triangulations of G-Manifolds for G a Finite Group*]{} Math. Ann. 233, 199–220 (1978) E. Lerman and S. Tolman, [*Hamiltonian Torus Actions on Symplectic Orbifolds and Toric Varieties*]{} Trans. Amer. Math. Soc. 349 (1997) 4201-4230 J. Milnor, [*Morse Theory*]{} Annals of Math. Studies 51, Princeton University Press, 1973
L. Nicolasescu, [*An Invitation to Morse Theory*]{} Universitext. Springer, New York, 2007.
A. Adem, J. Leida and Y. Ruan, [*Orbifolds and Stringy Topology*]{} Cambridge Tracts in Mathematics, 171. Cambridge University Press, Cambridge, 2007.
F. Sch[ä]{}tz [*The Morse-Smale complex*]{} Master thesis at university of Vienna, 2005.
I. Satake, [*On a generalization of the Notion of Manifold*]{} Proc. Nat. Acad. Sci. USA 42 (1956), 359-363. E.G. Sklyarenko, [*The Thom isomorphism for nonorientable bundles (in Russian)*]{} The Journal of Fundamental and Applied Mathematics, 9(4), 2003, 55-103; translated in Journal of Mathematical Sciences, 136(5), 2006, 4166–4200.
A.G. Wasserman, [*Equivariant Differential Topology*]{} Topology **8** (1969), 127-150.
U. Frauenfelder, [*The Arnold-Givental conjecture and moment Floer homology*]{} Int. Math. Res. Notices (2004), no. 42, 2179–2269.
J. Latschev, [*Gradient flows of Morse–Bott functions*]{} Math. Ann. 318 (2000), 731– 759.
A. Banyaga and D. Hurtubise, [*Morse-Bott homology*]{} Trans. Amer. Math. Soc. 362 (2010), no. 8, 3997–4043.
[^1]: This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MEST)(No. 2011-0001181)
| {
"pile_set_name": "ArXiv"
} |
---
title: ' Agent-Based Proof Design via Lemma Flow Diagram'
---
We discuss an agent-based approach to proof design and implementation, which we call [*Lemma Flow Diagram*]{}. This approach is based on the multicut rule with $shared$ cuts. This approach is modular and easy to use, read and automate. Some examples are provided.
proof design, flow proof, lemma flow diagram.
Introduction {#sec:intro}
============
Proofs typically have a great level of complexity and therefore require a careful [*proof design*]{} to reduce their complexity. Despite its importance, proof design has received little attention. This is in contrast to the situation where proof procedures – natural deduction, resolution and sequent systems, etc – are well-studied.
In this paper, we introduce a new [*proof design*]{} tool called [*Lemma Flow Diagram*]{} ([LFD]{}). The most central aspect of proof design is a set of well-chosen lemmas. A set of well-chosen lemmas is important in two aspects:
- It improves the readability of the proof.
- It improves the search for the proof in automated reasoning.
In fact, it is a set of lemmas that makes proofs reative, diverse and interesting. In contrast, proofs without lemmas are rather mechanical, unimaginative and quite awkward.
As we shall see, our design tool is natural and very easy to use, read and automate. This is analogous to the situation where people have introduced software design tools – UML, etc – to design complex software. [LFD]{} is modeled after the interactive computing model of Computability Logic [@Jap0]–[@JapCL12] and its distributed variant with agent parameters [@kwon19d].
Our starting point for building lemma-based proofs is the multicut(MC) rule. The MC rule expresses a form of modus pones – or lemma-based reasoning – which is the key element in compact proof design.
Suppose we have well chosen lemmas $B_1,\ldots,B_n$. Then the (distributed version of) MC rule is of the form:\
$${{{{\sbox{\infRuleBot}{\ensuremath{\Delta_1,\ldots,\Delta_n, \Gamma\Entails^m C}}
\sbox{\infRuleLabel}{\textsf{}}
\sbox{\infRule}{{\ensuremath{\begin{array}[b]{@{}c@{}}\Delta_1\Entails^{m_1} B_1 \ldots \quad \Delta_n\Entails^{m_n} B_n
\quad B_1,\ldots,B_n,\Gamma\Entails^{m} C\\\hline\usebox{\infRuleBot}\end{array}}}}
\settowidth{\infRuleWidth}{\usebox{\infRule}}
\settoheight{\infRuleCenter}{\usebox{\infRuleBot}}
\settoheight{\infRuleLblHeight}{\usebox{\infRuleLabel}}
\addtolength{\infRuleCenter}{-0.5\infRuleLabel}
\addtolength{\infRuleCenter}{-0.4ex}
\makebox{\usebox{\infRule} \raisebox{\infRuleCenter}{\usebox{\infRuleLabel}}}}}}}$$ Here, $\Gamma, \Delta$ are a set of formulas, $B,C$ are formulas, and each $m,m_1,\ldots,m_n$ is an agent/machine. That is, each $m_i$ tries to prove $B_i$ with respect to $\Delta_i$.
The premise of the above rule is easy to implement by distributedly processing each sequent in the premise but there is some unnaturalness in the above rule. That is, each $B_i$ can be “shared” between two agents but there is no indication of this fact.
A second, alternative way is to represent the MC rule with “sharing”. This rule, which we call [*the MC’ rule*]{}, is shown below:
$${{{{\sbox{\infRuleBot}{\ensuremath{\Delta_1,\ldots,\Delta_n, \Gamma\Entails^m C}}
\sbox{\infRuleLabel}{\textsf{}}
\sbox{\infRule}{{\ensuremath{\begin{array}[b]{@{}c@{}}\Delta_1\Entails^{m_1} \ldots \quad \Delta_n\Entails^{m_n}
\quad B_1^{m_1},\ldots,B_n^{m_n},\Gamma\Entails^m C\\\hline\usebox{\infRuleBot}\end{array}}}}
\settowidth{\infRuleWidth}{\usebox{\infRule}}
\settoheight{\infRuleCenter}{\usebox{\infRuleBot}}
\settoheight{\infRuleLblHeight}{\usebox{\infRuleLabel}}
\addtolength{\infRuleCenter}{-0.5\infRuleLabel}
\addtolength{\infRuleCenter}{-0.4ex}
\makebox{\usebox{\infRule} \raisebox{\infRuleCenter}{\usebox{\infRuleLabel}}}}}}}$$ Here, $\Delta_i$ acts as the knowledgebase of $m_i$ and $B_i^{m_i}$ acts as a query $B_i$ to $m_i$.
This MC’ rule leads to the diagram in Figure 1.
![[LFD]{} for the multicut](1f.jpg){width="50.00000%"}
In the diagram, note that each lemma $B_i$ is shared between $m_i$ and $m_{i+1}$. We call this diagram [LFD]{} with the following features:
- A regular arrow to box B with circle F means that some unknown identity (e.g. Nature) provides a regular service $F$ to B. Regular services include $sunny$ or $prime(2)$.
- A regular arrow from box A to box B with circle F in the middle means that A provides a querying service F to B.
- Knowledgebase(KB) of an agent is marked with $\bullet$s. In contrast, queries to the agent is marked without $\bullet$s.
- We assume that the top-level operator associated with an agent is ${\hspace{2pt}\mbox{\small $\vee$}\hspace{2pt}}$. That is, an agent with multiple KB services $KB_1,\ldots,KB_m$ and a query services $Q$ should be understood as an agent with a query service ${\mbox{\small $\neg$}}KB_1 {\hspace{2pt}\mbox{\small $\vee$}\hspace{2pt}}\ldots {\hspace{2pt}\mbox{\small $\vee$}\hspace{2pt}}{\mbox{\small $\neg$}}KB_m {\hspace{2pt}\mbox{\small $\vee$}\hspace{2pt}}Q$.
[LFD]{} can be seen as a modular alternative to the “flow proof” which is popular in mathematical education.
Examples {#sec:modules}
========
As an example of [LFD]{}, we will look at Peano Arithmetic (PA). PA is formulated with the usual $0,1,+,\times,=$. This is shown below.
$PA$ rules.\
(1)${\mbox{$\forall$}}x{\mbox{$\forall$}}y(x+1=y+1 {\hspace{2pt}\mbox{\small $\rightarrow$}\hspace{2pt}}x=y)$.\
(2)${\mbox{$\forall$}}x(x+1\neq 0)$.\
(3)$0+1=1$.\
(4)${\mbox{$\forall$}}x(x+0=x)$.\
(5)${\mbox{$\forall$}}x{\mbox{$\forall$}}y(x+(y+1)=(x+y)+1)$.\
(6)${\mbox{$\forall$}}x(x\times 0=0)$.\
(7)${\mbox{$\forall$}}x{\mbox{$\forall$}}y(x\times(y+1)=(x+y)+x)$.\
(8)${\mbox{$\forall$}}Q ((Q(0){\hspace{2pt}\mbox{\small $\wedge$}\hspace{2pt}}{\mbox{$\forall$}}x(Q(x){\hspace{2pt}\mbox{\small $\rightarrow$}\hspace{2pt}}Q(x+1))) {\hspace{2pt}\mbox{\small $\rightarrow$}\hspace{2pt}}{\mbox{$\forall$}}x Q(x))$.\
Rule (8) is called the [*Induction*]{} rule.
![[LFD]{} for the example](2f.jpg){width="50.00000%"}
Now we want to prove the following:
$${\mbox{$\forall$}}x (x+1=1+x)$$
The proof requires mathematical induction with $Q(x)$ being $(x+1=1+x)$. The base case $Q(0)$ which is $(0+1=1+0)$ follows from axiom 3 and 4. The induction step follows from induction hypothesis and axiom 5.
We display the [LFD]{} of the above example in Figure 2. This diagram is straightforward to understand.
Distributed first-order logic {#s2tb}
=============================
To represent the ‘querying’ formulas in the MC’ rule, we need a new representation language, because they are not allowed in classical logic. For this reason, we introduce which provides an elegant solution in this regard.
is based on the game semantics. For this reason, we review the the game-semantical meanings of first-order logic[@Japfi].
First of all, classical propositions are viewed as special, [*elementary*]{} sorts of games that have no moves and are automatically won by the machine if true, and lost if false. For example, the proposition $prime(5)$ is true and is automatically won by the machine.
There are two special atoms ${\top}$ and ${\bot}$. ${\top}$ is always true, and ${\bot}$ is always false.
Negation ${\mbox{\small $\neg$}}$ is a role-switch operation: ${\mbox{\small $\neg$}}A$ is obtained from $A$ by turning ${\top}$’s moves and wins into ${\bot}$’s moves and wins, and vice versa. For example, if [*Chess*]{} means the game of chess from the view of the black player, then ${\mbox{\small $\neg$}}$[*Chess*]{} is the same game from the of view of the white player.
The parallel operations deal with concurrent computations. Playing $A{\hspace{2pt}\mbox{\small $\wedge$}\hspace{2pt}}B$ or $A{\hspace{2pt}\mbox{\small $\vee$}\hspace{2pt}}B$ means playing, in parallel, the two games $A$ and $B$. In $A{\hspace{2pt}\mbox{\small $\wedge$}\hspace{2pt}}B$, ${\top}$ is considered the winner if it wins in both of the components, while in $A{\hspace{2pt}\mbox{\small $\vee$}\hspace{2pt}}B$ it is sufficient to win in one of the components.
The reduction operation $A{\hspace{2pt}\mbox{\small $\rightarrow$}\hspace{2pt}}B$ is understood as ${\mbox{\small $\neg$}}A {\hspace{2pt}\mbox{\small $\vee$}\hspace{2pt}}B$.
The blind quantifiers ${\mbox{$\forall$}}x A$ and ${\mbox{$\exists$}}x A$ model information hiding with ${\mbox{$\exists$}}x$ meaning ${\top}$’s choices, and ${\mbox{$\forall$}}x$ meaning choices by ${\bot}$. In both cases, the value of $x$ is unknown or invisible to the opponent.
Following the ideas in [@kwon19d], we introduce ${\mbox{FOL$^\Omega$}}$, a slight extension to FOL with agent parameters. Let $F$ be a formula in first-order logic. We introduce a new [*env-annotated*]{} formula $F^\omega$ which reads as ‘play $F$ against an agent $\omega$. For $F^\omega$, we say $\omega$ is the [*matching*]{} environment of $F$.
In introducing agent parameters to a formula $F$, some formulas turn out to be difficult to process. ‘Environment-switching’ formulas are such examples. In them, the machine initially provides a service $F$ against agent $w$ and then switches to provide another service to another agent $u$ inside $F$. This kind of formulas are difficult to process. This leads to the following definition:
The class of ${\mbox{FOL$^\Omega$}}$-formulas is defined as the smallest set of expressions such that (a) For any ${\mbox{FOL}}$-formula $F$ and any agent $\omega$, $F^\omega $ is in it and, (b) if $H$ and $J$ are in it, then so are ${\mbox{\small $\neg$}}H, H{\hspace{2pt}\mbox{\small $\wedge$}\hspace{2pt}}J, H{\hspace{2pt}\mbox{\small $\vee$}\hspace{2pt}}J, H{\hspace{2pt}\mbox{\small $\rightarrow$}\hspace{2pt}}J$.
Given a ${\mbox{FOL$^\Omega$}}$-formula $H$, the diagram of $H$ can be easily obtained.
Given a ${\mbox{FOL$^\Omega$}}$-formula $H$, the diagram of $H$ – denoted by $diagram(H)$ – is obtained by the rules in Figure 3. There, $F$ is a ${\mbox{FOL}}$-formula and $H,J$ are ${\mbox{FOL$^\Omega$}}$-formulas. Further, $A$ is the current machine, $W$ is an environment.
{width="50.00000%"}
An Example: Encoding the Peano Axiom {#sec:modules}
====================================
The diagram in Section 2 corresponds to the following . The (KB of) agent $m_1$ is shown below.
$agent\ m_1$.\
${\mbox{$\forall$}}x{\mbox{$\forall$}}y(x+1=y+1 {\hspace{2pt}\mbox{\small $\rightarrow$}\hspace{2pt}}x=y)$.\
${\mbox{$\forall$}}x(x+1\neq 0)$.\
$0+1=1$.\
${\mbox{$\forall$}}x(x+0=x)$.\
${\mbox{$\forall$}}x{\mbox{$\forall$}}y(x+(y+1)=(x+y)+1)$.\
${\mbox{$\forall$}}x(x\times 0=0)$.\
${\mbox{$\forall$}}x{\mbox{$\forall$}}y(x\times(y+1)=(x+y)+x)$.\
Our language permits ‘querying knowledge’ of the form $Q^\omega$ in KB. This requires the current machine to invoke the query $Q$ to the agent $\omega$. Now let us consider the $m$ agent which handles mathematical induction. It contains the induction rule and two querying knowledges $Q(0)^{m_1}$ and $({\mbox{$\forall$}}x(Q(x){\hspace{2pt}\mbox{\small $\rightarrow$}\hspace{2pt}}Q(x+1)))^{m_1}$. Note that the induction rule is technically not first-order but it does not affect our main argument.
$agent\ m$.\
$Q(0)^{m_1}$.\
$({\mbox{$\forall$}}x(Q(x){\hspace{2pt}\mbox{\small $\rightarrow$}\hspace{2pt}}Q(x+1)))^{m_1}$.\
${\mbox{$\forall$}}Q ((Q(0){\hspace{2pt}\mbox{\small $\wedge$}\hspace{2pt}}{\mbox{$\forall$}}x(Q(x){\hspace{2pt}\mbox{\small $\rightarrow$}\hspace{2pt}}Q(x+1))) {\hspace{2pt}\mbox{\small $\rightarrow$}\hspace{2pt}}{\mbox{$\forall$}}x Q(x))$.
Last, the theorem to prove corresponds to the query ?- ${\mbox{$\forall$}}x (x+1=1+x)$ with respect to the $m$ agent.
It can be easily observed that the above code can be automated. This leads to the notion of [*Lemma Flow Theorem Proving*]{}. Now, solving the above goal has the following strategy: first solve both the query $Q(0)^{m_1}$ and $({\mbox{$\forall$}}x(Q(x){\hspace{2pt}\mbox{\small $\rightarrow$}\hspace{2pt}}Q(x+1)))^{m_1}$ with respect to $m_1$ and then solve ${\mbox{$\forall$}}x (x+1=1+x)$ with respect to $m$. Note that all three attempts must succeed for ${\mbox{$\forall$}}x (x+1=1+x)$ to be true.
Conclusion {#sec:conc}
==========
In this paper, we have considered a proof design tool called [LFD]{} and a distributed first-order logic . Our diagram allows modular, distributed proof design. It is therefore useful for reading/writing/automating proofs. has a simple syntax and semantics and is well-suited to representing distributed KB [@Loke].
Acknowledgements
================
This work was supported by DongA University Research Fund.
[00]{}
Japaridze G. Introduction to computability logic. [*Annals of Pure and Applied Logic*]{}, 2003, 123(1/3): 1-99.
Japaridze G. Computability logic: a formal theory of interaction. In [*Interactive Computation: The New Paradigm*]{}, Goldin D, Smolka S A, Wegner P (eds.), Springer, 2006, pp. 183-223.
Japaridze G. In the beginning was game semantics. In [*Games: Unifying Logic, Language and Philosophy*]{}, Majer O, Pietarinen A -V, Tulenheimo T (eds.), Springer, 2009, pp. 249-350.
Japaridze G. Towards applied theories based on computability logic. [*Journal of Symbolic Logic*]{}, 2010, 75(2): 565-601.
Japaridze G. On the system CL12 of computability logic. http://arxiv.org/abs/1203.0103, June 2013.
Kwon K. Towards distributed logic programming based on computability logic. http://arxiv.org/abs/1909.07036, 2019.
S.W. Loke and A. Davison: LogicWeb: Enhancing the Web with Logic Programming. [*Journal of Logic Programming*]{}, 1998, 36(3): 195-240.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A considerable amount of machine learning algorithms take instance-feature matrices as their inputs. As such, they cannot directly analyze time series data due to its temporal nature, usually unequal lengths, and complex properties. This is a great pity since many of these algorithms are effective, robust, efficient, and easy to use. In this paper, we bridge this gap by proposing an efficient representation learning framework that is able to convert a set of time series with equal or unequal lengths to a matrix format. In particular, we guarantee that the pairwise similarities between time series are well preserved after the transformation. The learned feature representation is particularly suitable to the class of learning problems that are sensitive to data similarities. Given a set of $n$ time series, we first construct an $n\times n$ partially observed similarity matrix by randomly sampling $O(n \log n)$ pairs of time series and computing their pairwise similarities. We then propose an extremely efficient algorithm that solves a highly non-convex and NP-hard problem to learn new features based on the partially observed similarity matrix. We use the learned features to conduct experiments on both data classification and clustering tasks. Our extensive experimental results demonstrate that the proposed framework is both effective and efficient.'
author:
- |
Qi Lei$^{\natural}$ Jinfeng Yi$^{\dagger}$ Roman Vaculin$^{\dagger}$ Lingfei Wu$^{\dagger}$ Inderjit S. Dhillon$^{\natural}$\
$^{\natural}$The University of Texas at Austin, Austin, TX, USA\
$^{\dagger}$IBM Thomas J. Watson Research Center, Yorktown Heights, NY, USA\
bibliography:
- 'ref.bib'
title: Similarity Preserving Representation Learning for Time Series Analysis
---
Introduction
============
Modeling time series data is an important but challenging task. It is considered by [@yang200610] as one of the $10$ most challenging problems in data mining. Although time series analysis has attracted increasing attention in recent years, the models that analyze time series data are still much fewer than the models developed for static data. The latter category of models, which usually take instance-feature matrices as their inputs, cannot directly analyze time series data due to its temporal nature, usually unequal lengths, and complex properties [@langkvist2014review]. This is a great pity since many static models are effective, robust, efficient, and easy to use. Introducing them to time series analysis can greatly enhance the development of this domain.
In this work, we bridge this gap by proposing an efficient unsupervised representation learning framework that is able to convert a set of time series data with equal or unequal lengths to a matrix format. In particular, the pairwise similarities between the raw time series data are well preserved after the transformation. Therefore, the learned feature representation is particularly suitable to the similarity-based models in a variety of learning problems such as data clustering, classification, and learning to rank. Notably, the proposed framework is flexible to any time series distance or similarity measures such as Mikowski distance, cross-correlation, Kullback-Leibler divergence, dynamic time warping (DTW) similarity, and short time series (STS) distance. In this work, we use DTW similarity by default since it is known as the best measure for time series problems in a wide variety of domains [@rakthanmanon2013addressing].
Given a total of $n$ time series, our first step is to generate an $n\times n$ similarity matrix $\A$ with $\A_{ij}$ equaling to the DTW similarity between the time series $i$ and $j$. However, computing all the pairwise similarities requires to call the DTW algorithm $O(n^2)$ times, which can be very time-consuming when $n$ is large. As a concrete example, generating a full similarity matrix when $n=150,000$ takes more than $28$ hours on an Intel Xeon $2.40$ GHz processor with $256$ GB of main memory. In order to significantly reduce the running time, we follow the setting of matrix completion [@sun2015guaranteed] by assuming that the similarity matrix $\A$ is of low-rank. This is a very natural assumption since DTW algorithm captures the co-movements of time series, which has shown to be driven by only a small number of latent factors [@stock2005implications; @basulow]. According to the theory of matrix completion, only $O(n\log n)$ randomly sampled entries are needed to *perfectly* recover an $n \times n$ low-rank matrix. This allows us to only sample $O(n\log n)$ pairs of time series to generate a partially observed similarity matrix $\tilde \A$. In this way, the time spent on generating similarity matrix is significantly reduced by a factor of $O(n/\log n)$, a number that scales almost linearly with respect to $n$. When $n = 150,000$, it only takes about 3 minutes to construct a partially observed similarity matrix with $[20 n \log n]$ observed entries,[^1] more than $500$ times faster than generating a full similarity matrix.
Given the generated partially observed similarity matrix $\tilde \A$, our second step learns a new feature representation for $n$ time series such that their pairwise DTW similarities can be well approximated by the inner products of new features. To this end, we solve a symmetric matrix factorization problem to factorize $\tilde \A$, i.e., learning a feature representation $\X\in {\mathbb{R}}^{n\times d}$ such that $P_{\Omega}{(\tilde \A)}\approx P_{\Omega}{(\X\X^\top)}$, where $P_{\Omega}$ is a matrix projection operator defined on the observed entry set $\Omega$. Despite its relatively simple formulation, this optimization problem is hard to solve since it is highly non-convex and NP-hard. To address this challenge, we propose a very efficient exact cyclic coordinate descent algorithm. By wisely updating variables and taking the advantage of sparse observed entries in $\tilde \A$, the proposed algorithm incurs a very low computational cost, and thus can learn new feature representations in an extremely efficient way. To evaluate the performance of the learned feature representation, we use more than $10$ real-world data sets to conduct experiments on data classification and clustering tasks. Our results show that classical static models that are fed with our learned features outperform the state-of-the-art time series classification and clustering algorithms in both accuracy and computational efficiency. In summary, our main contributions of this work are two-fold:
1. We bridge the gap between time series data and a great amount of static models by learning a new feature representation of time series. The learned feature representation preserves the pairwise similarities of the raw time series data, and is general enough to be applied to a variety of learning problems.
2. We propose the first, to the best of our knowledge, optimization parameter free algorithm to solve symmetric matrix factorization problem on a partially observed matrix. The proposed algorithm is highly efficient and converges at a very fast rate.
Related Work
============
In this section, we review the existing work on learning feature representations for time series data. Among them, a family of methods use a set of derived features to represent time series. For instance, [@nanopoulos2001feature] proposed to use the mean, standard deviation, kurtosis, and skewness of time series to represent control chart patterns. The authors in [@wang2006characteristic] introduced a set of features such as trend, seasonality, serial correlation, chaos, nonlinearity, and self-similarity to partition different types of time series. [@deng2013time] used some easy to compute features such as mean, standard deviation and slope temporal importance curves to guide time series classification. In order to automate the selection of features for time series classification, the authors in [@fulcher2014highly] proposed a greedy forward method that can automatically select features from thousands of choices. Besides, several techniques have been proposed to represent time series by a certain types of transformation, such as discrete Fourier transformation [@faloutsos1994fast], discrete wavelet transformation [@chan1999efficient], piecewise aggregate approximation [@keogh2001dimensionality], and symbolic aggregate approximation [@lin2007experiencing]. In addition, deep learning models such as Elman recurrent neural network [@elman1990finding] and long short-term memory [@schmidhuber2015deep] are capable of modeling complex structures of time series data and learn a layer of feature representations. Due to their outstanding performance on a number of applications, they have become increasingly popular in recent years.
Despite the remarkable progress, the feature representations learned by these algorithms are usually problem-specific, and are not general enough for applications in multiple domains. Besides, the learned features cannot preserve similarities of the raw time series data, thus they are not suitable to the problems that are sensitive to the data similarity. These limitations inspire us to propose a problem-independent and similarity preserving representation learning framework for time series data.
Similarity Preserving Representation Learning for Time Series Analysis
======================================================================
In this section, we first present a general approach of our similarity preserving time series representation learning framework. We then propose an extremely efficient algorithm that is significantly faster than a naive implementation.
Problem Definition and General Framework
----------------------------------------
Given a set of $n$ time series $\mathcal{T}=\{T_1, \cdots, T_n\}$ with equal or unequal lengths, our goal is to convert them to a matrix $\X\in {\mathbb{R}}^{n\times d}$ such that the time series similarities are well preserved after the transformation. Specifically, we aim to learn a mapping function $f: T\rightarrow {\mathbb{R}}^d$ that satisfies $$\begin{aligned}
\mathrm{S}(T_i, T_j)\approx \langle f(T_i),f(T_j)\rangle\ \ \forall{i,j}\in [n],\label{eq:1}\end{aligned}$$ where $\langle \cdot , \cdot\rangle$ stands for the inner product, one of the most commonly used similarity measure in analyzing static data. $\mathrm{S}(\cdot, \cdot)$ denotes the pairwise time series similarity that can be computed by a number of functions. In this work, we use dynamic time warping (DTW) algorithm to measure the similarity by default. By warping sequences non-linearly in the time dimension, the DTW algorithm can calculate an optimal match between two given temporal sequences with equal or unequal lengths. Due to its superior performance, DTW has been successfully applied to a variety of applications, including computer animation [@muller2007dtw], surveillance [@sempena2011human], gesture recognition [@celebi2013gesture], signature matching [@efrat2007curve], protein sequence alignment [@vial2009combination], and speech recognition [@muda2010voice]. Normally, DTW algorithm outputs a pairwise distance between two temporal sequences, thus we need to convert it to a similarity score. Since the inner product space can be induced from the normed space using $\langle x,y\rangle =(\|x\|^2+\|y\|^2-\|x-y\|^2)/2$ [@adams2004knot], we generate the DTW similarity by $$\begin{aligned}
\mathrm{S}(T_i,T_j)\!=\!\frac{\DTW(T_i,\emph{0})^2\!\!+\!\DTW(T_j,\emph{0})^2\!\!-\!\DTW(T_i,T_j)^2}{2}\nonumber,\end{aligned}$$ where $\emph{0}$ denotes the length one time series with entry $0$. We note that the similarity computed via the above equation is a more robust choice than some other similarity measures such as the reciprocal of distance. This is because when two time series are almost identical, their DTW distance is close to $0$ and thus its reciprocal tends to infinity.
In order to learn the matrix $\X$, an intuitive idea is to factorize the similarity matrix $\A\in {\mathbb{R}}^{n\times n}$ where $\A_{ij}=\mathrm{S}(T_i,T_j)$. In more detail, this idea consists of two steps, i.e., a similarity matrix construction step and a symmetric matrix factorization step. In the first step, it constructs a similarity matrix $\A$ as follows $$\A_{ij}=\left\{
\begin{array}{cc}
[{b_i^2+b_j^2-\DTW^2(T_i,T_j)}]/2, & \mathrm{if\ } i<j\\
(b_i^2\ +\ b_j^2)/2, & \mathrm{if\ } i=j\\
\A_{ji} & \ \mathrm{if\ } i>j,\\
\end{array}\right.\label{eq:2}$$ where $b_i=\DTW(T_i,\emph{0}),\ i=1,\cdots\!,n$. In the second step, it learns an optimal data-feature matrix $\X$ by solving the following optimization problem $$\min_{\X= (\x_1, \cdots, \x_n)}\ \ \sum_{i=1}^n\sum_{j=i+1}^n \|\A_{ij}-\langle\x_i,\x_j\rangle\|^2,\label{eq:3}$$ which can be further expressed in a matrix form $$\min_{\X\in {\mathbb{R}}^{n\times d}}\ \ \ \|\A-\X\X^\top\|_F^2.\label{eq:4}$$ The problems and can be solved by performing eigen-decomposition on $\A$, i.e., $$\X=\Q_{1:n,1:d}\times \sqrt{\bm\Lambda_{1:d,1:d}}\ ,$$ where $\A=\Q\bm\Lambda\Q^\top$ and the notation $\Q_{1:k,1:r}$ represents the sub-matrix of $\Q$ that includes its first $k$ rows and the first $r$ columns.
Although the inner products of the learned data points $\x_1, \cdots, \x_n$ can well approximate the DTW similarities of the raw time series, the idea described above is not practical since both two steps are painfully slow when $n$ is large. In order to generate a $n\times n$ similarity matrix, we need to call the DTW algorithm $O(n^2)$ times.[^2] In addition, a naive implementation of eigen-decomposition takes $O(n^3)$ time. Although we can reduce this cost by only computing the $d$ largest eigenvalues and the corresponding eigenvectors, it is still computational intensive with a large $n$. As a concrete example, when $n=150,000$ and the length of time series is $30$, it takes more than $28$ hours to generate the similarity matrix and more than $3$ days to compute its $30$ largest eigenvalues on an Intel Xeon $2.40$ GHz processor with $256$ GB of main memory.
A Parameter-free Scalable Algorithm
-----------------------------------
In this subsection, we propose an extremely efficient approach that significantly reduces the computational costs of both steps while learns the new feature representation with a high precision. In addition to efficiency, another nice property is that the proposed approach is *optimization parameter free*, that is, the user does not need to decide any optimization parameter such as step length or learning rate.
To significantly improve the efficiency of the first step, we make the key observation that the similarity matrix $\A$ should be of low-rank. This is due to the fact that the DTW algorithm measures the level of co-movement between time series, which has shown to be dictated by only a small number of latent factors [@stock2005implications; @basulow]. Indeed, we can verify the low-rankness in another way. Since the matrix $\A$ is a special case of Wigner random matrix, the gaps between its consecutive eigenvalues should not be small [@marvcenko1967distribution]. This implies the low-rank property since most of its energy is concentrated in its top eigenvalues [@erdHos2009local].
Based on the theory of matrix completion [@sun2015guaranteed], only $O(n \log n)$ randomly sampled entries are needed to perfectly recover an $n\times n$ low-rank matrix. Thus, we don’t need to compute all the pairwise DTW similarities. Instead, we randomly sample only $O(n \log n)$ pairs of time series, and then compute the DTW similarities only within the selected pairs. In other words, we generate a partially observed similarity matrix $\tilde \A$ with $O(n \log n)$ observed entries as $$\begin{aligned}
\tilde \A_{ij}=\left\{
\begin{array}{ll}
\mathrm{S}(T_i, T_j) & \ \text{if\ }\ \Omega_{ij}=1 \\
\text{unobserved} & \ \text{if }\ \Omega_{ij}=0,
\end{array}
\right.\end{aligned}$$ where $\Omega\in\{0,1\}^{n\times n}$ is a binary matrix indicating the indices of sampled pairs. In this way, the running time of the first step is significantly reduced by a factor of $O(n/\log n)$. Since this factor scales almost linearly with $n$, we can greatly save the running time when $n$ is large. For instance, when $n=150,000$, it only takes $194$ seconds to construct a partially observed similarity matrix with \[$20 n\log n$\] observed entries, more than $500$ times faster than generating a full similarity matrix. Given the partially observed similarity matrix $\tilde \A$, our second step aims to learn a new feature representation matrix $\X$. Instead of first completing the full similarity matrix $\A$ and then factorize it, we propose an efficient symmetric factorization algorithm that is able to directly factorize the partially observed similarity matrix $\tilde \A$, i.e., find a $\X\in {\mathbb{R}}^{n\times d}$ that minimizes the following optimization problem $$\begin{aligned}
\min_{\X\in {\mathbb{R}}^{n\times d}}\ \ \ \|P_{\Omega}\ (\tilde\A\ -\X\X^T)\|_F^2,
\label{eqn:fact}\end{aligned}$$ where $P_{\Omega}: {\mathbb{R}}^{n\times n} \to {\mathbb{R}}^{n\times n}$ is a projection operator defined as $$\begin{aligned}
[P_{\Omega}\ (\B)]_{ij}=\left\{
\begin{array}{ll}
\B_{ij} & \ \text{if\ }\ \Omega_{ij}=1\\
0 & \ \text{if }\ \Omega_{ij}=0.
\end{array}
\right.\end{aligned}$$
The objective function does not have a regularization term since it already bounds the Frobenius norm of $\X$. Despite its relatively simple formulation, solving problem is non-trivial since its objective function is highly non-convex and the problem is NP-hard. To address this issue, we propose a very efficient optimization algorithm that solves problem based on exact cyclic coordinate descent (CD). Although [@ge2016matrix] shows that the symmetric matrix factorization problem can be solved via (stochastic) gradient descent algorithm as well, our proposed coordinate descent algorithm has the following two advantages: (i) for each iteration, our CD algorithm directly updates each coordinate to the optimum. Thus, we do not need to decide any optimization parameter; and (ii) by directly updating coordinates to the optimums using the most up-to-date information, our CD algorithm is highly efficient and converges at a very fast rate.
At each iteration of the exact cyclic CD method, all variables but one are fixed, and that variable is updated to its optimal value. One of the main strengths of our algorithm is its capacity to update variables in an extremely efficient way. Besides, the proposed algorithm takes the advantage of sparse observed entries in $\tilde \A$ to further reduce the computational cost. To be precise, our algorithm consists of two loops that iterate over all the entries of $\X$ to update their values. The outer loop of the algorithm traverses through each column of $\X$ by assuming all the other columns known and fixed. At the $i$-th iteration, it optimizes the $i$-th column $\X_{1:n,i}$ by minimizing the following subproblem $$\begin{aligned}
\|\textbf{R}-P_{\Omega}(\X_{1:n,i}\X_{1:n,i}^T)\|_F^2,\end{aligned}$$ where $\textbf{R}$ is the residual matrix defined as $\textbf{R} = P_{\Omega}(\tilde \A-\sum_{j\neq i}\X_{1:n,j}\X_{1:n,j}^T)$.
In the inner loop, the proposed algorithm iterates over each coordinate of the selected column and updates its value. Specifically, when updating the $j$-th entry $\X_{ji}$, we solve the following optimization problem $$\begin{aligned}
\begin{split}
&\ \ \ \ \min_{\X_{ji}}\ \|\textbf{R}-P_{\Omega}(\X_{1:n,i}\X_{1:n,i}^T)\|_F^2\\
\!\!\!\Longleftrightarrow&\ \ \ \ \min_{\X_{ji}}\ \|\textbf{R}\|_F^2\!-\!2\langle \textbf{R}, P_{\Omega}(\X_{1:n,i}\X_{1:n,i}^T)\rangle \\
& \ \ \ \ \ \ \ \ \ \ \ \ +\|P_{\Omega}(\X_{1:n,i}\X_{1:n,i}^T)\|_F^2\\
\!\!\!\Longleftrightarrow&\ \ \ \ \min_{\X_{ji}}\ \X_{ji}^4+2(\!\!\!\!\sum_{k\in \Omega_j,\ k\neq j}\!\!\!\! \X_{ki}^2-\textbf{R}_{jj})\X_{ji}^2\\
\!\!\!& \ \ \ \ \ \ \ \ \ \ \ \ -4(\!\!\!\!\sum_{k\in \Omega_j,\ k\neq j}\!\!\!\! \X_{ki}\textbf{R}_{jk})\X_{ji}+C\\
\!\!\!\Longleftrightarrow&\ \ \ \ \min_{\X_{ji}}\ \psi(\X_{ji})=\X_{ji}^4+2p\X_{ji}^2+4q\X_{ji}+C,\label{eq:9}
\end{split}\end{aligned}$$ where $\Omega_i,\ i=1,\!\cdots\!, n$ contains the indices of the observed entries in the $i$-th row of matrix $\tilde \A$. $C$ is a constant that is independent of $\X_{ji}$. Since the $\psi(\X_{ji})$ is a fourth-degree polynomial function, $\X_{ji}$, as will be shown later, can be updated in a very efficiently way. Algorithm \[alg:1\] describes the detailed steps of the proposed exact cyclic CD algorithm.
- $\tilde \A\in {\mathbb{R}}^{n\times n}$: partially observed similarity matrix
- $\Omega_i,\ i=1,\!\cdots\!, n$: indices of the observed entries in the $i$-th row of matrix $\tilde \A$
- $I$: number of iterations
- $d$: dimension of features
<!-- -->
- $\X^{(0)}\leftarrow \mathbf{0_{n\times d}}$
- $\textbf{R} \leftarrow P_{\Omega}(\tilde \A-\X^{(0)}\X^{(0)\top})=P_{\Omega}(\tilde\A)$
$\X^{(t)}\leftarrow \X^{(t-1)}$ $\textbf{R}\leftarrow \textbf{R}+P_{\Omega}(\X^{(t)}_{1:n,i}\X_{1:n,i}^{(t)\top})$\[step:add\] $p\leftarrow\sum_{k\in \Omega_j} \X_{ki}^{(t)2}-\X_{ji}^{(t)2}-\textbf{R}_{jj}$\
$q\leftarrow-\sum_{k\in \Omega_j}\X_{ki}^{(t)}\textbf{R}_{jk}+\X_{ji}^{(t)}\textbf{R}_{jj}$ $\X_{ji}^{(t)}\leftarrow
{\arg\!\min}\{\X_{ji}^4+2p\X_{ji}^2+4q\X_{ji}\}$\[step:root\] $\textbf{R}\leftarrow \textbf{R}-P_{\Omega}(\X^{(t)}_{1:n,i}\X_{1:n,i}^{(t)\top})$\[step:subtract\]
$\X^{(I)}$
\[alg:1\]
The proposed algorithm incurs a very low computational cost in each iteration. Lines $7$-$11$ of the algorithm can be computed in $O(n\log n)$ operations. This is because the costs of computing $p$ and $q$ are only proportional to the cardinality of $\Omega_j$. Besides, the derivative $\nabla\psi(\X_{ji})$ is a third-degree polynomial, thus its roots can be computed in closed form. By using Cardano’s method [@cardano1993ars], the optimal solution of $\X_{ji}$ can be calculated in a constant time given the computed $p$ and $q$. Likewise, lines $6$ and $12$ of the algorithm also take $O(n\log n)$ time since matrix $\textbf{R}$ can be updated by only considering the observed entries. To sum up, the proposed algorithm has a per-iteration cost of $O(d n\log n)$, significantly faster than the recent matrix factorization algorithms that take at least $O(dn^2)$ time in each iteration [@DBLP:journals/corr/VandaeleGLZD15; @yu2012scalable].
![[]{data-label="fig:1"}](POC_last.png){width="0.98\columnwidth"}
In addition to a low per-iteration cost, we also expect that the proposed algorithm yields a fast convergence. This is because our algorithm always uses the newest information to update variables and each variable is updated to the optimum in a single step. This hypothesis is verified by a convergence test conducted on the UCR Non-Invasive Fetal ECG Thorax1 testbed [@UCRArchive]. This testbed contains a total of $3,765$ time series with a length of $750$. In this test, we generate a full similarity matrix $\A$ by computing the DTW similarities between all the time series pairs, and then randomly sample $[20 n \log n]$ of its entries to generate a partially observed matrix $\tilde \A$. We call the proposed algorithm to factorize matrix $\tilde \A$ by setting $d=30$. To measure the performance of the proposed method, we compute two error rates, i.e., the observed error $\|P_{\Omega} (\tilde\A\ -\X\X^T)\|_F/\|P_{\Omega} (\tilde\A)\|_F$ and the underlying true error $\|\A\ -\X\X^T\|_F/\|\A\|_F$, at each iteration. Figure \[fig:1\] shows how they converge as a function of time. This figure clearly demonstrates that the proposed exact cyclic CD algorithm converges very fast – it only takes $1$ second and $8$ iterations to converge. Besides, the construction accuracy is also very encouraging. The observed error and the underlying true error rates are close to each other and both of them are only about $0.1\%$. This result not only indicates that the inner products of the learned features well approximate the pairwise DTW similarities of the raw time series, but also verifies that we can learn accurate enough features by only computing a small portion of pairwise similarities. In addition, this test validates the low-rank assumption. It shows that a $3,765 \times 3,765$ DTW similarity matrix can be accurately approximated by a rank $30$ matrix.
Experiments
===========
In this section, we evaluate the proposed framework, i.e., Similarity PreservIng RepresentAtion Learning (SPIRAL for short), on both classification and clustering tasks. For both tasks, we learn new feature representations by setting the number of features $d=30$, the number of iterations $I=20$, and the sample size $|\Omega|=[20n\log n]$. We then fed the learned features into some static models and compare them with the state-of-the-art time series classification and clustering algorithms. In our experiments, we set a same DTW window size $\min(40,[\mathrm{average\ time\ series\ length}/10])$ for all the DTW-related algorithms evaluated here. Although we can improve the DTW performance by tuning a problem-dependent warping window size [@mueen2016extracting], we skip this step in our experiments to ensure that the superior performance is achieved by the proposed representation learning framework instead of tuning the DTW algorithm. All the results were obtained on a Linux server with an Intel Xeon $2.40$ GHz CPU and $256$ GB of main memory.
Classification Task
-------------------
Six real-world time series data sets are used in our classification analysis. As a research institute, we have partnered with one of the world’s largest online brokers on a challenge problem of predicting its clients’ propensity of trading options. We are provided with a historical records data of $64,523$ sampled clients with the lengths of the time series range from $1$ month to $67$ months. The data contains $76$ dynamic attributes with none of them related to option trading. Given this data, our task is to predict whether the clients will trade options in the next $3$ months. Besides, we also conduct experiments on $5$ benchmark data sets, i.e., FordA, ECG5000, PhalangesOutlinesCorrect (POC), ItalyPowerDemand (IPD), and HandOutlines (HO), from the UCR Time Series Repository [@UCRArchive]. The data sets used in our experiments have widely varying sizes and encompass multiple domains such as finance, healthcare, and manufacture. Table \[tab:1\] summarizes the statistics of these data sets.
Given the feature representations learned by the proposed SPIRAL framework, we fed them into the algorithms of *XGBoost* [@chen2016xgboost] and *$\ell_2$-regularized logistic regression* that is implemented in LibLinear [@fan2008liblinear]. These approaches, denoted as *SPIRAL-XGB* and *SPIRAL-LR*, respectively, are compared with the state-of-the-art time series classification algorithms *NN-DTW* [@wang2013experimental] and *Long Short-Term Memory (LSTM)* [@schmidhuber2015deep]. For the XGBoost algorithm, we use the default parameters specified in its source code. The parameter of logistic regression is determined by a 5-fold cross validation. We use AUC (area under the ROC Curve) to measure the classification performance. All the experiments in this study are repeated five times, and the AUCs averaged over the five trials are reported in Table \[tab:class\]. From this table, we first observe that XGBoost and logistic regression algorithms that are fed with our learned features outperform the two baseline algorithms on all the data sets. In particular, the method SPIRAL-XGB yields the best performance on five out of six data sets, and the method SPIRAL-LR performs the best on the other data set. More encouragingly, SPIRAL-LR can achieve better classification performance than NN-DTW and LSTM on almost all the data sets. This verifies that the feature representation learned by the proposed method is powerful – even classical models such as logistic regression can achieve satisfactory performance on it.
--------------- ------------- ------------- -------------
\# training \# testing length of
time series time series time series
Online Broker 51,618 12,905 1 – 67
ECG5000 1,320 3,601 500
FordA 1,320 3,601 500
POC 1,800 858 80
IPD 67 1,029 24
HO 370 1,000 2,709
--------------- ------------- ------------- -------------
: *Statistics of classification data sets*[]{data-label="tab:dataset"}
\[tab:1\]
--------------- ---------------------------------------------- ---- ---- ----
Online Broker **[0.84]{} & 0.83 & 0.76 & 0.79\
ECG5000 & [**0.91**]{} & 0.85 & 0.76 & 0.76\
FordA & [**0.73**]{} & 0.66 & 0.56 & 0.59\
POC & **[0.77]{} & 0.66 & 0.68 & 0.66\
IPD & 0.94 & [**0.97**]{} & 0.95 & 0.91\
HO & **[0.87]{} & 0.81 & 0.77 & 0.61\
******
--------------- ---------------------------------------------- ---- ---- ----
: *Average AUC Scores of the classification algorithms*[]{data-label="tab:class"}
On the other hand, although LSTM has been successfully applied to a number of sequence prediction tasks, it fails to deliver strong performance in our empirical study. We conjecture that this is because the data sets are not large enough for LSTM to model complex structures. Besides, NN-DTW is usually a very strong baseline that is considered as hard-to-beat in the literature [@xi2006fast; @wang2013experimental; @mueen2016extracting]. However, it also yields an overall worse performance than SPIRAL-LR and SPIRAL-XGB. One possible reason is that NN-DTW uses 1-nearest neighbor classifier, thus is sensitive to noises and outliers. This observation shows another advantage of the proposed method: by learning a feature representation instead of directly developing a time series model, our method is more flexible and can exploit the strengths of different learning algorithms.
In addition to the superior performance, SPIRAL-LR and SPIRAL-XGB are very efficient as well. They have a significantly lower running time than that of the baseline algorithms. For example, it takes NN-DTW more than $2$ days to classify the FordA data set while the running time for SPIRAL-LR and SPIRAL-XGB is only about $5$ minutes.
Clustering Task
---------------
Similar to time series classification, time series clustering is also an important task that has found numerous applications. To further test our learned features on the clustering task, we conduct experiments on another 7 UCR time series data sets. Since data clustering is an unsupervised learning task, we merge their training and testing sets together. Statistics of these data sets are summarized in Table \[tab:2\].
We fed the features learned by the proposed framework into the $k$Means algorithm as our clustering method, and compare it to the state-of-the-art time series clustering algorithm *$k$-Shape* [@paparrizos2015k], which has been shown to outperform many state-of-the-art partitional, hierarchical, and spectral clustering approaches. Besides, we also compare our method with clustering algorithms *$k$Means-DTW* and *CLDS* [@li2011time] since our ideas are similar in some respects. $k$Means-DTW is a popular time series clustering algorithm that uses DTW algorithm to measure pairwise distances between data points. Although it looks similar to the idea of our SPIRAL-$k$Means that also utilizes the DTW and $k$Means algorithms, it is less desirable than SPIRAL-$k$Means mainly because: (i) $k$Means-DTW suffers from a very high computational cost since it needs to compute the pairwise DTW distances between all the time series and all the cluster centers at each iteration; and (ii) the DTW distance does not satisfy the triangle inequality, thus can make the cluster centers computed by averaging multiple time series drift out of the cluster [@niennattrakul2007inaccuracies]. By designing an efficient algorithm that only needs to call the DTW function $O(n \log n)$ times and by embedding time series data to the Euclidean space while preserving their original similarities, the proposed method SPIRAL successfully addresses both these issues. Similar to the idea of representation learning presented in this paper, CLDS also learns a new feature representation of the time series data, and then partition the learned representation via an EM-like algorithm.
In this study, we use the normalized mutual information (NMI for short) to measure the coherence between the inferred clustering and the ground truth categorization. NMI scales from 0 to 1, and a higher NMI value implies a better partition. Each experiment is repeated five times, and the performance averaged over the five trials is reported in Table \[tab:cluster\]. Compared to all the baseline algorithms, our clustering method SPIRAL-$k$Means yields the best performance on all the seven data sets, indicating that it delivers the state-of-the-art performance. In addition, SPIRAL-$k$Means has a significantly lower running time than all the baseline clustering algorithms evaluated here. For instance, clustering the ED data set takes $k$-Shape, CLDS, and $k$Means-DTW $20$, $57$, and $114$ minutes, respectively. As a comparison, our clustering algorithm SPIRAL-$k$Means only spends one minute and a half to partition the ED data set. We finally note that $k$Means is just one choice of clustering algorithms that can take our learned features as the input. By replacing it with some more advanced clustering algorithms, it is expected to achieve an even better clustering performance.
-------------- ------------- ------------- -----------
number of length of number of
time series time series clusters
Swedish Leaf 1,125 128 15
Cricket\_X 780 300 12
uWGLX 4,478 315 8
50words 905 270 50
SLC 9,236 1,024 3
SC 600 60 6
ED 16,637 96 7
-------------- ------------- ------------- -----------
: *Statistics of clustering data sets*[]{data-label="tab:2"}
-------------- ------------------------------------------------- -- -- --
Swedish Leaf **[0.61]{} & 0.56 & 0.60 & 0.50\
Cricket\_X & [**0.32**]{} & 0.31 & 0.26 & 0.22\
uWGLX & [**0.47**]{} & 0.38 & 0.28 & 0.32\
50words & [**0.68**]{} & 0.50 & 0.56 & 0.53\
SLC & **0.60** & 0.56 & 0.41 & N/A\
SC & [**0.80**]{} & 0.78 & 0.51 & 0.70\
ED & **[0.35]{} & 0.23 & 0.2 & 0.24\
****
-------------- ------------------------------------------------- -- -- --
: *Average NMI Scores of all the clustering algorithms. N/A indicates that the clustering task cannot be completed due to memory limitations..*[]{data-label="tab:cluster"}
Conclusions
===========
In this paper, we propose a similarity preserving representation learning framework for time series analysis. Given a set of $n$ time series, the key idea is to first generate a partially observed similarity matrix with $O(n \log n)$ observed DTW similarities, and then factorize this matrix to learn a new feature representation. To this end, we propose the first optimization parameter free algorithm for factorizing partially-observed symmetric matrix. The proposed algorithm updates variables via exact cyclic coordinate descent, and incurs a very low computational cost in each iteration. The feature representation learned by the proposed framework preserves the DTW similarities of the raw time series data, and is general enough to be applied to a variety of learning problems. Our empirical studies on both classification and clustering tasks verify the effectiveness and efficiency of the proposed method.
[^1]: Since this similarity matrix is symmetric, we only need to call the DTW algorithm about $[10 n \log n]$ times to generate this matrix.
[^2]: Indeed, we need to call the DTW algorithm at least $n(n+1)/2$ times to generate full similarity matrix $\A$. This number is achieved when we pre-compute all the $b_i,\ i= 1, \cdots\!, n$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper presents the use of a novel modelling technique based around intermittent transport due to filament motion, to interpret experimental profile and fluctuation data in the scrape-off layer (SOL) of JET during the onset and evolution of a density profile shoulder. A baseline case is established, prior to shoulder formation, and the stochastic model is shown to be capable of simultaneously matching the time averaged profile measurement as well as the PDF shape and autocorrelation function from the ion-saturation current time series at the outer wall. Aspects of the stochastic model are then varied with the aim of producing a profile shoulder with statistical measurements consistent with experiment. This is achieved through a strong localised reduction in the density sink acting on the filaments within the model. The required reduction of the density sink occurs over a highly localised region with the timescale of the density sink increased by a factor of 25. This alone is found to be insufficient to model the expansion and flattening of the shoulder region as the density increases, which requires additional changes within the stochastic model. An example is found which includes both a reduction in the density sink and filament acceleration and provides a consistent match to the experimental data as the shoulder expands, though the uniqueness of this solution can not be guaranteed. Within the context of the stochastic model, this implies that the localised reduction in the density sink can trigger shoulder formation, but additional physics is required to explain the subsequent evolution of the profile.'
author:
- |
N. R. Walkden$^{1}$, A. Wynn$^{1,2}$, F. Militello$^{1}$, B. Lipschultz$^{2}$, G. Matthews$^{1}$, C.Guillemaut$^{1}$, J. Harrison$^{1}$, D. Moulton$^{1}$ and JET Contributors${*}$\
\
\
\
\
Email: `[email protected]`
bibliography:
- '../Bibliography.bib'
title: 'Interpretation of scrape-off layer profile evolution and first-wall ion flux statistics on JET using a stochastic framework based on filamentary motion'
---
Introduction
============
The scrape-off layer (SOL) of a tokamak defines the interface between the hot plasma core and cold material surfaces which must be protected. Understanding the properties of the SOL has proven historically difficult due to the complexities of cross-field transport. The transport is known to be highly non-diffusive [@GarciaJNM2007; @NaulinJNM2007] which makes capturing it in typical 2D transport codes such as SOLPS [@SOLPS], EDGE2D [@EDGE2D] or UEDGE [@UEDGE] a challenge. In these cases profiles in the upstream SOL are conventionally specified as input with transport coefficients adapted to match. Whilst this approach is of great practical use a first-principles understanding of the SOL cannot be established on this basis, for which a detailed knowledge and decent parameterisation of the transport processes occurring in the upstream SOL are needed. Measurements on many tokamaks worldwide have revealed that SOL transport is intermittent [@AntarPoP2003; @BoedoPoP2003; @GarciaNF2007; @DudsonPPCF2005; @XuNF2009], with this intermittency resulting from the radial propagation of coherent structures commonly termed blobs or filaments [@D'IppolitoReview]. The propagation of a filament is a complex non-linear problem [@KrashenninikovPLA2001] for which numerical simulation is often required [@YuPoP2003; @GarciaPPCF2006; @YuPoP2006]. Whilst the study of filamentary motion is still an ongoing area of active research, many advancements have been made in recent years. These include the role of 3D effects in simple slab geometries [@AngusPRL2012; @AngusPoP2012; @EasyPoP2014] and more complex magnetic geometries [@WalkdenPPCF2013; @WalkdenNF2015]; the role of finite ion temperature (and associated FLR effects) [@MadsenPoP2011; @WeisenbergerPoP2014; @OlsenPPCF2016; @NielsenPPCF2017] and dynamic electron temperature [@WalkdenPPCF2016]; the role of electromagnetic effects [@LeePoP2015] and magnetic shear [@StepanenkoPoP2017]. Indeed the present state of simulations of filamentary motion now permit direct comparison with experiment, as conducted on TORPEX [@RivaPPCF2016] and MAST [@MilitelloPPCF2016-2]. To properly address the process of SOL formation or evolution, however, many such simulations would be required, or fully saturated turbulence simulations are required [@GarciaPoP2005; @MilitelloPPCF2013; @RicciPPCF2012]. These are extremely computationally expensive and are therefore difficult to run interpretively over parameter spaces required by experiment. It is therefore desirable to have intermediate models that parameterise some of the complexity of the fully non-linear simulations, but are simplified in nature such that they can be effectively ‘fit’ to experimental data.\
Garcia [@GarciaPRL2012] introduced a statistical characterisation of time-series from single point measurements made in the SOL. This ‘shot-noise’ model treats filaments in the measurement as an uncorrelated train of pulses with an exponential distribution of amplitudes which are ejected in time following a Poisson process. The predictions of the model agree well with measurements on several devices [@GarciaPoP2013; @GarciaNF2015; @GarciaNME2017], in particular managing to capture the non-Gaussian features of the probability distribution function (PDF) [@GravesPPCF2005]. The structure of the PDF is strongly invariant to many changes in plasma conditions [@GravesPPCF2005; @WalkdenNF2017], though as shown on JET, this invariance is due to a temporal balance between the duration of bursts and the time between bursts hitting the probe [@WalkdenNF2017]. This stochastic model is limited however in that it deals with measurements made at a single spatial point. Recently Militello and Omotani [@MilitelloNF2016; @MilitelloPPCF2016] have developed a similar model, based on a statistical description of filamentary motion, to relate spatial profiles in the SOL to the dynamics of filaments. For brevity this model will be referenced herein as the Militello-Omotani (MO) model. The MO model should be viewed as a framework since it is possible to specify a variety of different parameterisations for attributes of the filaments (see the next section for a fuller description) which can then be used to model profiles in the SOL. In such a manner, several possible causes of the experimentally observed non-exponential profiles in the SOL were introduced. Non-exponential profiles, profile flattening or shoulder formation (all used synonymously) describe the tendency of the density profile in the SOL to exhibit distinct regions where the gradient deviates from an exponential profile, often leading to the distinction between a ‘near’ and ‘far’ SOL. This behaviour is remarkably universal across many machines [@LipschultzPPCF2002; @RudakovNF2005; @GarciaNF2007; @CarraleroNF2014; @MilitelloNF2015; @WynnEPS; @WynnPaper] and is induced by increasing fuelling rate or reducing the plasma current . Recently these features of the density profile have been linked to changes in the dynamics of filaments on ASDEX-Upgrade [@CarraleroNF2014], which the MO model is ideally suited to investigate.\
This paper employs a numerical implementation of the MO model to investigate the process of shoulder formation in the SOL density profile of JET. Using data from the Lithium Beam Emission (LiBES) diagnostic, profiles can be directly compared between the model and experiment. Furthermore time-series data collected at the outer wall by a static Langmuir probe (LP) will be used to compare statistical features of the model with experiment. This paper can therefore be viewed as presenting a methodology for the interpretation of features of the SOL using the MO model with particular emphasis placed on interpreting the formation of the density shoulder.
Experimental reference {#Sec:Exp}
======================
Data taken from JET pulse 89350 will be used as a reference for comparison with the stochastic modelling conducted within this paper. This is an Ohmic L-mode horizontal target plasma with a 2MA plasma current and 2T toroidal field. A fuelling ramp is conducted to systematically increase the separatrix density. The measurements used are density profiles from the LiBES, which is a subset of that presented by Wynn et al [@WynnEPS; @WynnPaper], and ion-saturation current time-series from a wall-mounted LP which is a subset of the that presented by Walkden et al [@WalkdenNF2017]. In particular the profile shape, PDF shape and autocorrelation function have been measured from three different periods in the density ramp, shown in figure \[Fig:Exp\_measurements\].
![LiBES density profiles during JET pulse 89350 (left) at three different points during the fuelling ramp. Also shown are the PDF shape (centre) and autocorrelation function (right) from a wall-mounted LP. The PDFs and autocorrelation functions are color coded corresponding to the density profile measured at the same time.[]{data-label="Fig:Exp_measurements"}](JET_exp_data.png){width="\textwidth"}
The experimental measurements show a change in the profile structure with the formation of a shoulder in the outer SOL as the separatrix density increases, consistent with previous measurements on JET [@CarraleroJNM2015] alongside many other machines [@LipschultzPPCF2002; @RudakovNF2005; @GarciaNF2007; @CarraleroNF2014; @MilitelloNF2015]. During this flattening of the profile, the PDF shape remains invariant despite a contraction of the autocorrelation function. This can be achieved because the detection rate of filaments on the probe adapts to compensate the shorter duration time [@WalkdenNF2017]. It is these three features, the profile shape, PDF shape and autocorrelation function, that will be used for comparison within this paper. For more information on the experimental analysis of the density profile structure see ref [@WynnEPS], and for analysis of the fluctuation characteristics see ref [@WalkdenNF2017].
Stochastic modelling framework {#Sec:Stoch}
==============================
In this paper the MO model [@MilitelloNF2016; @MilitelloPPCF2016] is adopted. A one-dimensional (radial) density signal is constructed by the superposition of a series of statistically distributed pulses with a given spatial waveform propagating radially away from the separatrix with an evolving amplitude. These pulses will be referred to here as filaments. Following Garcia [@GarciaPRL2012] the number of filaments ejected in a given interval of time, $\Delta t$, is assumed to behave according to the Poisson distribution such that the the signal at any point along the radial dimension can be described as a ‘shot noise’ process. Physically this means that the ejection of one filament is unaffected by any other ejection event and filaments remain uncorrelated. The ejection rate of the process is given by $f_{f} = N_{f}/\Delta t$ where $N_{f}$ is the total number of filaments ejected with a uniform probability in the time interval given by $\Delta t$. The spatial profile of the i’th filament as defined in ref [@MilitelloPPCF2016], is given by $$\eta_{i}\left(x,t\right) = \eta_{0,i}F_{i}\left(x,t\right)\psi\left(x - \int_{0}^{t}v_{i}\left(x,t'\right)dt',\delta_{i}\right)$$ where $\eta_{i}\left(x,t\right)$ is the filament radial profile, $\eta_{0,i}$ is the amplitude of the filament density at its ejection time , $F_{i}\left(x,t\right)$ is a function which describes the spatial and temporal evolution of the filament amplitude due to sources/sinks. Here this is characterised by an exponential time-scale such that $$F_{i}\left(x,t\right) = \exp\left(\frac{t_{i} - t}{\tau\left(x,t\right)}\right)$$ where $t_{i}$ is the ejection time of filament $i$ within the ensemble and $\tau(x,t)$ is the time-scale of the source/sink (negative for a source, positive for a sink) which can in turn depend on spatial position or time. For brevity, $\tau(x,t)$ will be referred to going forwards as the sink timescale. Through this function features such as drainage due to parallel flows or re-fuelling due to ionisation can be introduced.\
$\psi\left(x,\delta\right)$ is the waveform of the filament and depends on the radial velocity of the filament, $v_{i}\left(x,t\right)$ and the width of the filament, $\delta_{i}$. For all simulations conducted here the filament waveform is taken as a truncated exponential such that $\psi\left(x,\delta\right) = \exp\left(x/\delta\right)H\left(-x\right)$ where $H\left(x\right)$ is the Heaviside function. This shape is motivated by the observation of front formation in filaments as they propagate radially [@D'IppolitoReview]. The filament amplitude, $\eta_{0}$, width, $\delta$ and velocity $v\left(x,t\right)$ must be specified and in general here will be drawn from a set of statistical distributions. The spatial and temporal evolution of the velocity and the sink timescale $\tau\left(x,t\right)$ must also be specified. With the required functions and statistical distributions specified a signal can be generated by a summation of the filaments within the ensemble $$n\left(x,t\right) = \sum_{i=1}^{N}\eta_{i}\left(x,t-t_{i}\right)$$ Here signals will be produced numerically, and measurements of these signals can be made. With the aim of basing this study on a comparison with the experimental observations outlined in section \[Sec:Exp\], two aspects of the synthetically produced signals will be measured. Time-averaged profiles will be constructed by averaging each individual radial point over the time frame of the simulation, whilst single point time series will be analysed by taking the signal at each point in time at a single point in radius, here taken at $R - R_{sep} = 5cm$ corresponding approximately to the outer-wall gap for the experiments described in section \[Sec:Exp\]. Figure \[Fig:sigs\_schema\] provides an example of the synthetic signal produced as a function of radius and time, and the respective time-averaged radial profile and single point time-series at the wall radius.
![Example of a synthetically produced signal as output from the stochastic modelling described in section \[Sec:Stoch\]. To the right is an example of the time-series measurement taken at $R-R_{sep} = 5$cm and to the bottom is the time-averaged radial profile measurement.[]{data-label="Fig:sigs_schema"}](PAPER_schematic_signal.png){width="8cm"}
\
It is well known theoretically that the velocity of a filament has a complex dependancy on its width [@YuPoP2003; @YuPoP2006; @MyraPoP2006; @D'IppolitoReview] and its amplitude [@GarciaPoP2006; @OmotaniPPCF2015] among other parameters. In the ‘inertial regime’ of filament dynamics the filaments radial $\textbf{E}\times\textbf{B}$ velocity is regulated by the ion-polarisation drift and the characteristic filament velocity scales like $v \propto \delta_{\perp}^{1/2}$ where $v$ is the characteristic filament velocity and $\delta_{\perp}$ is the filament width. In the ‘sheath limited regime’ sheath currents regulate the radial velocity which obtains the scaling $v \propto \delta_{\perp}^{-2}$. When resistivity is introduced, the sheath limited scaling is modified to $v \propto (1 + \Lambda)\delta_{\perp}^{-2}$ where $\Lambda$ is the normalised plasma resistivity integrated along the magnetic field line (to be defined later) [@MyraPoP2006; @EasyPoP2016]. Following the example by Militello and Omotani [@MilitelloPPCF2016] a Pade’s approximation is used to derive an analytic form for the filament velocity as a function of the filament width $$\label{Eqn:vel}
v = v_{0}\tilde{v}\left(x,t\right)\eta_{0}^{1/2}\frac{\left(\delta/\delta^{*}\right)^{1/2}}{1 + \left(1 + \Lambda\right)^{-1}\left(\delta/\delta^{*}\right)^{5/2}}$$ where $v_{0}$ is a scaling parameter for the velocity and $\tilde{v}(x,t)$ contains any spatial and temporal dependance of the filament velocity. Here also the dependancy of the velocity on the filament density amplitude, $\eta_{0}$ has been included. Formally this square-root dependancy holds in the inertial regime, whilst the sheath regime has a more complex dependancy of $v \propto \eta/\left(1 + \beta\eta\right)$ [@OmotaniPPCF2015] where $\beta \approx 0.31$ was found to provide a good match to numerical simulation in ref [@OmotaniPPCF2015]. Since this more complex dependance does not deviate too significantly from the square-root scaling over a range of a factor $\sim 5$ [@OmotaniPPCF2015] it has not been adopted here, however it is possible to include in the Pade’s approximation. Finally the parameter $\delta^{*}$ (defined later) determines the transition from the inertial to sheath limited regimes [@YuPoP2003]. Figure \[Fig:Vels\_example\] shows the scaling of the normalised velocity, $v/(\tilde{v}v_{0})$ as a function of the width $\delta$ for various values of the control parameters $\delta^{*}$ and $\Lambda$.
![Scaling curves of the normalised filament velocity, $v/v_{0}$, as a function of the filament width $\delta$ across variation in the normalised resistivity $\Lambda$ with $\delta^{*} = 1$cm (left) and critical filament width $\delta^{*}$ with $\Lambda = 1$ (right).[]{data-label="Fig:Vels_example"}](PAPER_vel_profiles.png){width="16cm"}
\
Finally a choice of statistical distributions must be made for the filament width, $\delta$ and amplitude $\eta_{0}$ at the filament ejection time. For the amplitude distribution function, measurements made with Langmuir probes [@GarciaNF2015; @WalkdenNF2017] and gas-puff imaging [@GarciaJNM2013; @GarciaPoP2013] across different machines and conditions show an exponential behaviour. For this reason an exponential amplitude distribution function is adopted here such that $$\label{Eqn:amp_dist}
P_{\eta_{0}}\left(\eta_{0}\right) = \frac{1}{\eta_{0}^{*}}\exp\left(-\frac{\eta_{0}}{\eta_{0}^{*}}\right)$$ For the width distribution function, a log-normal distribution has been chosen. There is less physical basis for this choice due to the inherent difficulties in measuring filament shape experimentally, though distributions of filament widths measured on MAST [@AyedPPCf2009; @KirkPPCF2016] and NSTX [@ZwebenPPCF2016] show a tendency towards a positively skewed asymmetric distribution. Furthermore the choice of a log-normal distribution may be considered appropriate for a positive definite variable such as the width, $\delta$, which may be expected to have a probability that decays to $0$ as the width approaches $0$. As such the distribution of filament widths is taken to be $$\label{Eqn:width_dist}
P_{\delta}\left(\delta\right) = \frac{1}{\sigma_{\delta}\delta\sqrt{2\pi}}\exp\left(-\frac{\left(\ln\left(\delta\right) - \ln\left(\delta_{0}\right)\right)^{2}}{2\sigma_{\delta}^{2}} \right)$$ where $\delta_{0}$ specifies the most probable value of $\delta$ and $\sigma_{\delta}$ is used to specify the width of the distribution.\
Table \[Tbl:Params\] tabulates the inputs to the stochastic model employed here that must be set before a simulation can be run.
Input Description
----------------------------- ----------------------------------------------------------------------------------------
$f_{f}$ Filament ejection rate
$\Delta t$ Simulation time
$ N_{s} $ Number of samples in $\Delta t$
$\delta^{*}$ Inertial - Sheath limited transition width in filament velocity relation (\[Eqn:vel\])
$\Lambda$ Normalised resistivity parameter in velocity relation (\[Eqn:vel\])
$v_{0}$ Velocity scale parameter in relation (\[Eqn:vel\])
$\delta_{0}$ Peak width in the width distribution (\[Eqn:width\_dist\])
$\sigma_{\delta}$ Spread parameter in the width distribution (\[Eqn:width\_dist\])
$\eta_{0}^{*}$ e-folding amplitude of the amplitude distribution (\[Eqn:amp\_dist\])
$\tilde{v}\left(x,t\right)$ Spatio-temporal evolution function for the filament velocity
$\tau\left(x,t\right)$ Spatio-temporal density sink timescale
: Table of inputs to the stochastic model used in this paper based off of the Militello-Omotani model.[]{data-label="Tbl:Params"}
\
There are a significant number of inputs to the model, some of which can be estimated from experimental conditions, some of which can be constrained by the experimental measurements presented in section \[Sec:Exp\], and some of which will be based on assumptions. In the case of $\Delta t$ and $N_{s}$, these are user defined settings which here are set to $N_{s} = 600,000$ and $\Delta t = 1.8s$ which were found to provide statistically converged results. Whilst the profile shape is relatively insensitive to statistical noise, the PDF shape and autocorrelation function are highly sensitive and require long time-series for a suitably accurate result. In appendix \[App:Conv\] this statistical convergence is demonstrated more clearly.\
In the next section a baseline simulation will be established to match the experimental measurements early in the density ramps (blue curves in figure \[Fig:Exp\_measurements\]). Sensitivity of this baseline case to assumptions about the inputs will be assessed . From the baseline case, possible mechanisms by which shoulder formation in the SOL profile can occur will be investigated.
Establishing a baseline case
============================
In this section a baseline simulation will be established to match the characteristics of the experimental measurements at the beginning of the density ramp. This corresponds to the blue data points in figure \[Fig:Exp\_measurements\] where the density profile shows minimal deviation from an exponential.\
Several of the inputs in table \[Tbl:Params\] can be estimated from conditions typical to JET Ohmic L-mode operation. Firstly the sink function is assumed to be dominated by convective losses due to parallel drainage. The sink timescale can therefore be approximated as $\tau = L_{||}/c_{s}$ [@Stangeby] where $c_{s} = \sqrt{(T_{e} + T_{i})/m_{i}}$ is the Bohm sound speed. Typical temperatures at the divertor target for JET pulses similar to that used here are in the approximate range $5eV < T_{e} < 30eV$ [@WynnEPS; @WynnPaper] across the radial profile. The radial increase of $\tau$ resulting from the radial decrease in $T_{e}$ is somewhat compensated by a drop in $L_{||}$ across the SOL. For the JET shot studied here $L_{||}$ drops from 37m in the near SOL at the vertical position of the OLP to 20m in the outer SOL. $\tau$ can therefore be estimated as being within the range $0.48ms < \tau < 0.66ms $ so here it is taken as $\tau = 0.57ms$ across the radial profile. The effect of possible variation in $\tau$ will be covered later in this section.\
Next the parameters $\Lambda$ and $\delta^{*}$ are estimated. These are defined as [@YuPoP2003; @MyraPoP2006] $$\label{Eqn:dstar}
\delta^{*} = 2\rho_{s}\left(\frac{L_{||}^{2}}{\rho_{s}R}\right)^{1/5}$$ and $$\label{Eqn:Lambda}
\Lambda = \frac{L_{||}}{c_{s}\tau_{e}}\frac{\Omega_{e}}{\Omega_{i}}$$ where $\Omega_{e}$ and $\Omega_{i}$ are the electron and ion gyro-frequencies respectively, $\tau_{e}$ is the electron collision time, $\rho_{s} = c_{s}/\Omega_{i}$ is the Bohm gyro-radius , $R$ is the major radius and $L_{||}$ and $c_{s}$ are defined above. Taking an estimate for the separatrix temperature of $T_{i} = T_{e} = 60$eV give $\delta^{*} = 2.1$cm. The estimation of $\Lambda$ is harder to make since it has such a strong dependancy on the conditions at the divertor target and can easily vary by orders of magnitude. Recent studies on ASDEX-Upgrade [@CarraleroNF2014], JET [@WynnEPS] and TCV [@VianelloPrePrint] have shown that $\Lambda \approx 0.1$ is typical for the early stages of a density ramp, prior to the formation of the profile shoulder. For this reason $\Lambda = 0.1$ is taken here initially.\
Finally, $\eta_{0}^{*} = 1.0$ and $\sigma_{\delta} = 0.3$ are used initially, and both the filament velocity $\tilde{v}$ and the sink timescale $\tau$ are constant in space and time. With these estimations and assumptions in mind an iterative procedure has been followed to optimize the remaining parameters of the model as follows:
- First a choice of $\delta_{0}$ is made
- $v_{0}$ is varied to match the profile gradient for the given choice of $\delta_{0}$
- $f_{f}$ is varied to match the PDF shape for the given $\delta_{0}$ and $v_{0}$
- The autocorrelation function is compared and a new choice of $\delta_{0}$ is made
This process is iterated until a good match for the profile gradient, PDF shape and autocorrelation function is achieved. At present this is done manually due to the inherent statistical noise associated with the stochastic modelling making it challenging to automate the process. A good match between the experimental data at the beginning of the density ramp and the simulated measurements is found for $\delta_{0} = \delta^{*} = 2.1$cm, $v_{0} = 50$m/s and $f_{f} = 42$kHz. This is shown in figure \[Fig:baseline\].
![SOL density profile normalised to the separatrix density (left), PDF shape (centre) and autocorrelation function (right) compared between the experimental cases (coloured) and the simulation (black). The experimental data is a repeat of that shown in figure \[Fig:Exp\_measurements\]. In the central plot, the range of the PDF values taken across the full experimental density scan is given, rather than individual measurements, since the PDF shape collapses strongly.[]{data-label="Fig:baseline"}](JET_baseline.png){width="\textwidth"}
\
Figure \[Fig:baseline\] shows that the stochastic model is able to simultaneously reproduce a consistent profile shape, PDF structure and autocorrelation function compared to that of the experiment. Both the filament ejection frequency and the velocity scale factor are found to have values that are in reasonable agreement with recent measurements of filament properties on JET with the ITER-like wall [@SilvaNF2014]. Furthermore the value of $\delta_{0}$ here (which should be interpreted as a filament full-width) matches that found by Xu *et al* [@XuNF2009] within a reasonable margin, though those measurements were made in inner-wall limited, carbon-wall conditions so are not directly comparable to the measurements considered here. It is worth noting that the value of $v_{0} = 50$m/s obtained here, and effective $\textbf{E}\times\textbf{B}$ velocities observed by Silva *et al* [@SilvaNF2014] are somewhat lower than those often found in other devices which typically are on the order of $0.5 - 1$ km/s [@D'IppolitoReview]. This may be partly the result of a larger major radius in JET, providing reduced curvature drive compared to smaller sized machines. In terms of the modelling conducted here, $v_{0}$ is determined by the gradient of the Density profile, given a fixed sink timescale. For higher values of $v_{0}$ to be achieved, a stronger sink of the filament density is required which would necessitate larger values of $v_{0}$ to maintain the same profile gradient, however this also impacts the structure of the autocorrelation function. Both a reduction in $\tau$ and an increase in $v_{0}$ lead to a contraction of the autocorrelation function. This contraction can be compensated for by increasing $\delta_{0}$ thus making filaments larger, however this process then becomes naturally limited by the spatial scale of the filaments assuming that it is unreasonable to expect filaments to have spatial scales comparable with the outer-wall gap of the machine (here 5cm which is significantly larger than any direct observation found in literature). In figure \[Fig:Width\_scan\] the value of $\delta_{0}$ required to fit the experimental data is shown as a function of $v_{0}$ holding the product $v_{0}\tau = 0.0285$ fixed to leave the profile gradient unchanged. Also shown is the range of the autocorrelation function produced, showing that it remains approximately fixed and the system is approximately invariant.
![left: Filament width, $\delta_{0}$ required to keep the autocorrelation function approximately fixed as a function of filament velocity, keeping $v_{0}\tau = 0.0285$. Right: Autocorrelation functions for each simulated case showing the approximate similarity in all cases.[]{data-label="Fig:Width_scan"}](v0_d0_autocorr_invariance.png){width="80.00000%"}
\
As can be seen, with the product $v_{0}\tau$ fixed to maintain the shape of the profile, the filament widths increase non-linearly with the velocity. Both the increase in velocity and the decrease in drainage time (to keep $v_{0}\tau$ constant) have the effect of contracting the autocorrelation function. The only actuator left to counter this effect is the filament width $\delta_{0}$. This means that, if a specific autocorrelation function is sought (as is the case here) then the required increase to $\delta_{0}$ gets larger as filaments move faster, as evidenced in figure \[Fig:Width\_scan\]. Practically, this limits the velocity and sink timescale that can be used as input. Taking here a limit of $\delta_{0} < 5$cm gives $v_{0} < 80$m/s and $\tau > 0.356$ms.\
Having established a baseline simulation the next stage of the analysis conducted here is to test whether the stochastic model is capable of reproducing the experimental measurements at the onset of the density profile shoulder, and during its subsequent evolution. Militello and Omotani give three principle means by which non-exponential profiles can be introduced in the stochastic model [@MilitelloPPCF2016]. These are:
1. [Through changes to the structure of statistical distributions]{}
2. [Through changes to the spatial and/or temporal behaviour of the source/sink function]{}
3. [Through changes to the spatial and/or temporal behaviour of the filament velocity]{}
In the next few sections these three different methods will be investigated in turn.
Profile modification within the stochastic model
================================================
In this section the three methods by which a shoulder can be introduced into the SOL density profile will be investigated in turn. It is noted that in principle the different mechanisms may all occur simultaneously in experimental conditions, however this creates a vast parameter space to search manually and is unfeasible here. Instead each mechanism will be investigated in isolation to establish its merit with regards to matching the experimental data.
Variation of statistical distributions
--------------------------------------
Through the velocity relation (\[Eqn:vel\]) both the amplitude and width distributions can impact the distribution of filament velocities, which can in turn impact the mean profile shape. With the amplitude distribution shape fixed as an exponential, and the scaling exponent of the filament velocity on the filament amplitude in equation (\[Eqn:vel\]) fixed at $0.5$, changes to the e-folding amplitude of the distribution, $\eta_{0}^{*}$, cannot change the profile shape in the manner required to match experiment. To see this note that $v_{0}$ can simply be scaled by $1/\sqrt{\eta_{0}^{*}}$ which then leaves the velocity distribution un-changed. As a result increasing the e-folding amplitude (i.e larger amplitude filaments) is entirely equivalent to scaling $v_{0}$. This may decrease the profile gradient by increasing the velocity of filaments, but it cannot change the shape of the profile in the manner required during the density shoulder formation since any new profiles can always be re-mapped onto the baseline case by a simple scaling of the profile gradient, through $v_{0}$.\
The width distribution and relationship between the filament velocity and the filament width may be more impactful to the density profile because of the different scalings of the filament velocity in the sheath and inertial regimes. The spread of the width distribution, $\sigma_{\delta}$, as well as the parameters $\delta_{0}$, $\delta^{*}$ and $\Lambda$ may all affect the profile structure via equation (\[Eqn:vel\]). To test some of these effects, a set of simulations has been run varying both $\Lambda$ and $\delta_{0}$. $\Lambda$ takes the values $\Lambda = 0.1, 1, 10$ which are observed to be approximately the range taken in experiment when shoulder formation and expansion in the density profile is observed [@CarraleroNF2014; @WynnEPS; @WynnPaper]. $\delta_{0}$ values are taken with respect to $\delta^{*}$ as $\delta_{0} = \delta^{*}/2$, $\delta_{0} = \delta^{*}$ (which corresponds to the baseline case at $\Lambda = 0.1$) and $\delta_{0} = 2\delta^{*}$. Figure \[Fig:Lambda\_delta\_scan\] shows the profiles and autocorrelation functions calculated for the nine simulations in the scan described here.
![Upper: Normalised profile shape with $\sigma_{\delta} = 0.3$ fixed but varying both $\Lambda$ and $\delta_{0}$ over the ranges given in the text. Lower: Autocorrelation functions measured in the same scenarios as the profiles in the upper row. []{data-label="Fig:Lambda_delta_scan"}](PAPER_Lambda_delta_scan.png){width="80.00000%"}
\
For the case where $\delta_{0} < \delta^{*}$ variation in $\Lambda$ has minimal impact. This arrises because, as showin in figure \[Fig:Vels\_example\], below $\delta \sim \delta^{*}$ changing $\Lambda$ has little impact on the filament velocity. Thus only the filaments in the ensemble with widths $ \delta > \sim\delta^{*}$ are significantly impacted by a change in $\Lambda$. It is therefore clear that the width distribution with $\delta_{0} = \delta^{*}/2$ will be the least affected by varying $\Lambda$. Experimentally, $\Lambda$ is often found to be a correlated parameter with the onset of profile shoulder formation [@CarraleroNF2014; @WynnEPS; @VianelloPrePrint] though the causality of the relation remains un-determined. Figure \[Fig:Lambda\_delta\_scan\] suggest that if $\Lambda$ is a determining factor in the onset of broadening, then a filament width distribution with a significant fraction of the widths occupying the region $\delta > \delta^{*}$ is required. This should be verified experimentally where $\Lambda$ is thought to be an important parameter. Furthermore $\Lambda$ has very little impact on the autocorrelation function, with significant change only being introduced by a variation in $\delta_{0}$.\
The third parameter that influences the width distribution is $\sigma_{\delta}$ which determines the spread in the width distribution. In figure \[Fig:Lambda\_sigma\_scan\] the profiles and autocorrelation functions are shown for the three values of $\Lambda$ studied above, this time fixing $\delta_{0} = \delta^{*}$ and varying $\sigma_{\delta} = 0.1, 0.3$ and $0.5$. At $\sigma_{\delta} = 0.5$ the spread of the distribution is beyond what might reasonably be expected from experiment, with a small bit not insignificant proportion of filament widths $\delta > 8$cm. This provides a good test of the sensitivity of the modelling to $\sigma_{\delta}$, which is a relatively unconstrained parameter given the challenge of measuring the filament width distribution experimentally.
![Upper: Normalised profile shape with $\delta_{0} = \delta^{*}$ fixed but varying both $\Lambda$ and $\sigma_{\delta}$ over the ranges given in the text. Lower: Autocorrelation functions measured in the same scenarios as the profiles in the upper row. []{data-label="Fig:Lambda_sigma_scan"}](PAPER_Lambda_sigma_scan.png){width="80.00000%"}
\
$\sigma_{\delta}$ has very little impact on the autocorrelation function regardless of the value taken by $\Lambda$. In the profile shape, there is a general trend towards flatter profiles as $\sigma_{\delta}$ decreases, with the importance of this effect decreasing as $\Lambda$ increases. This occurs because, with $\sigma_{\delta}$ taking smaller values, a higher proportion of filaments in the ensemble have $\delta \sim \delta^{*}$ and obtain a larger radial velocity. The effect is not dramatic however and suggests that the results presented here are not sensitive to the value of $\sigma_{\delta}$ within the range $0.1<\sigma_{\delta}<0.5$, which supports the used of $\sigma_{\delta} = 0.3$ in establishing the baseline case.\
In both figures \[Fig:Lambda\_delta\_scan\] and \[Fig:Lambda\_sigma\_scan\] there is a general trend towards flatter profiles and (slightly) contracted autocorrelation functions as $\Lambda$ increases. A notable feature of the profile shape though is a lack of a shoulder region. The profile is slightly sub-exponential and does flatten as $\Lambda$ increases, but does not exhibit the experimental behaviour for any of the values of $\delta_{0}$ and $\sigma_{\delta}$ studied here. Furthermore only a variation in $\delta_{0}$ was able to induce experimentally relevant changes to the autocorrelation function by virtue of reducing the spatial width of the filaments. This suggests that within the stochastic model used here, varying $\Lambda$ or indeed other aspects of the width and amplitude distributions is not sufficient to trigger the formation of the shoulder in the SOL profile. It is worth noting though that varying $\Lambda$ does provide a threshold for changes to the profile and SOL statistics by virtue of the $1 + \Lambda$ factor in the denominator in equation (\[Eqn:vel\]). The investigation here does not rule out the role of $\Lambda$ or indeed any other of the parameters, but it does suggest that they are not sufficient in isolation and must be accompanied by a change to either the filament velocity or filament drainage. It should also be noted that the parameters tested in this section were held constant across the radial dimension. Allowing variation in these parameters would be a good avenue of future work. This should be treated sensitively however, since it is not presently clear from simulations whether filament motion depends on local conditions spatially, or the conditions present at the formation of the potential dipole that drives the motion. Such a study should be conducted before spatial variation of the parameters in equation (\[Eqn:vel\]) is introduced. Such a study is beyond the scope of this paper.
Filament velocity and drainage
------------------------------
It was highlighted in the establishment of the baseline simulation that in the stochastic model used here, as in many studies of scrape-off layer profiles [@Stangeby], the profile shape is set by the competition between parallel and perpendicular transport, which here translates into the variation of $\tilde{v}\tau$ in space and time. The required change in $\tilde{v}\tau$ can be estimated since the local gradient scale length of the profile is approximately $\frac{1}{n}\frac{d n}{dR} = 1/\lambda_{ne} \propto \tilde{v}\tau$ so by evaluating the change in the profile gradient from the baseline case to the broadened profile in the experimental data, the required change to $\tilde{v}\tau$ can be estimated. This is shown in figure \[Fig:vtau\_exp\].
![Left: Normalised density profiles showing the onset of profile broadening following the color coding introduced in figure \[Fig:Exp\_measurements\]. Centre: Gradient scale length,$\lambda$, for the density profiles shown to the left. Right: Relative increase observed in $\lambda_{ne}$ compared to the baseline case, $\lambda_{ne,bl}$.[]{data-label="Fig:vtau_exp"}](JET_ne_gradients.png){width="\textwidth"}
In the broadened region of the profile an increase in $\lambda \propto \tilde{v}\tau$ by a factor for 2 to 4 is required, whilst in the far SOL a reduction by a half is observed compared to the baseline case. This indicates that either a localised increase in the filament velocity or a localised increase in the density sink timescale is required to capture the shoulder region of the profile.\
To parameterise the radial dependance of $\tilde{v}$ and $\tau$ used as input to the stochastic model an asymmetric Gaussian function has been employed such that $$g(R) = \begin{cases}
1 + \Delta_{g}\exp\left(-\left(R - R_{0}\right)^{2}/2\sigma_{g,-}^{2}\right), & \text{for } R \leq R_{0} \\
g_{0} + \left(\Delta_{g} + 1 - g_{0}\right)\exp\left(-\left(R - R_{0}\right)^{2}/2\sigma_{g,+}^{2}\right), & \text{for } R > R_{0}
\end{cases}$$ where $g$ represents either the sink timescale or the velocity and $\sigma_{g,-} \neq \sigma_{g,+}$. Once again a manual optimisation of the velocity and drainage function parameters has been used, such that the results may be considered reasonable but not fully optimised. Starting from the baseline case the velocity function and the sink timescale have been varied respectively in separate simulations to try and match the profile shape at the onset of the shoulder formation (green data points in figure \[Fig:Exp\_measurements\]). Figure \[Fig:veloc\_drain\_var\_1\] shows the comparison between the profile structure, as well as the PDF and autocorrelation functions for the cases where the velocity and drainage time have been adapted respectively in separate simulations.
![Comparison of the profile shape (upper right), PDF (lower left) and autocorrelation function (lower right) for a modified velocity profile (corresponding to the black curves in all figures) and a modified drainage time (red curves in all figures), which are both shown in the upper right (note the different axes). The profile, PDF range and autocorrelation functions from the baseline case, and the experimental case at the onset of the density shoulder formation are shown for comparison, with colours adhering to the color coding from figure \[Fig:Exp\_measurements\]. []{data-label="Fig:veloc_drain_var_1"}](Vel_drain_comp.png){width="80.00000%"}
\
Figure \[Fig:veloc\_drain\_var\_1\] shows that within the stochastic model used here it is possible to induce an experimentally consistent shoulder in the density profile through either a localised increase in the filament velocity or in the density sink timescale. The peaks of the $\tilde{v}$ and $\tau$ are offset from one another. This is the result of non-local effects that occur due to the spatial structure of the filaments. When filaments are accelerated they encounter a less severe drainage so transport a higher density into the outer SOL, raising the profile in that region. However filaments in the acceleration zone become more rarified such that on time-averaging, the density in that region can be reduced. The exact opposite effects are true in the case of deceleration, as is present in the outer SOL in figure \[Fig:veloc\_drain\_var\_1\]. The density sink on the other hand has a spatially localised affect on the filaments since their loss rate is determined at each point in the profile by the sink at that point. It is also notable that in order to match the experimental profile by changing the sink timescale, an increase by a factor 25 is required. If the density sink is taken as the result of parallel losses, this implies a near stagnation of any flows that are providing this loss. Alternatively this may imply a refuelling of the filament via, for example, ionisation processes. The stochastic model is not able to determine the physical mechanism at play beyond noting that it impacts the density sink.\
Although both filament acceleration and an increase in the density loss timescale give reasonable profile shapes and leave the PDF shape invariant, the autocorrelation function serves to distinguish the two cases. In the outer part of the density profile, the gradient increases outside of the shoulder. To capture this increase, either a deceleration of the filament velocity or a reduction in the loss timescale is required. These have opposing effects on the autocorrelation function, with the former leading to an expansion and the latter leading to a contraction. Comparing this important difference with the experimental case suggests that the localised acceleration/deceleration of filaments is not consistent with experiment, whilst the localised increase in the sink timescale does produce consistent results. Indeed the value of $\tau$ that produces consistent profile behaviour also produces a consistent reduction in the autocorrelation time without optimising any other parameters (filament width for example). Thus in the context of the stochastic model used in this paper, a radially localised increase in the density sink timescale is the simplest (in the sense that only a single parameter is varied) method by which a profile shoulder can be triggered whilst maintaining experimentally consistent statistics at the outer-wall.\
It is not a simple matter to extend the comparison based on variation of the density sink timescale to the highest density case in the experimental reference (the red data traces in figure \[Fig:Exp\_measurements\]). Figure \[Fig:drain\_var\] shows the result of sequentially increasing the parameter $\Delta_{\tau}$ which determines the level to which the $\tau$ is increased, compared to the high density experimental profile.
![Changes to the simulated profile shape as the degree to which the parallel drainage time is increased with the radially localised profile shown in figure \[Fig:veloc\_drain\_var\_1\]. The experimental data shown (with errorbars) is the highest density case in the experimental reference shown in figure \[Fig:Exp\_measurements\].[]{data-label="Fig:drain_var"}](drainage_variation.png){width="50.00000%"}
As $\tau$ is increased the level to which the profile is affected lessens and the experimental profile cannot be reproduced. Furthermore, and importantly, because the profile gradient in the outer-SOL is moderately decreased in the high density case, the sink timescale cannot be altered in a manner that simultaneously captures the gradient decrease *and* the contraction of the autocorrelation function. This implies that additional physics is required to capture the subsequent evolution of the profile following shoulder formation. It has been possible to generate a stochastic signal that gives reasonable agreement with the experimental data after the shoulder formation by allowing for simultaneous variation of the sink timescale and velocity profiles. This is shown in figure \[Fig:veloc\_drain\_var\_2\] where spatial variation in both $\tilde{v}$ and $\tau$ are included in the same simulation.
![Comparison of the stochastic model with the highest density experimental case allowing for the simultaneous spatial variation of the drainage time and the filament velocity. In the upper left, the profiles of the velocity and density sink (noting the different axes used for each) are shown. In the three comparisons with experiment, black curves indicate the simulated measurements, with the experimental cases following the color coding in figure \[Fig:Exp\_measurements\]. []{data-label="Fig:veloc_drain_var_2"}](JET_compare_2nd_stage.png){width="80.00000%"}
A mild acceleration of the filament velocity alongside the strong increase in the loss timescale couple to provide conditions that match both the profile shape and autocorrelation function. This is a plausible manner in which the profile evolution can be captured within the stochastic model, but an exhaustive search has not been conducted. What is clear from this study is that there is an additional physics requirement to capture the profile evolution, but that the stochastic model does have the capability to do this consistently with experiment.
Discussion
==========
The comparison made with the experimental data on JET, presented here, shows that the stochastic model is capable of reproducing consistent statistical measurements as well as profile measurements at various stages of the profile evolution. Whilst there are certainly limitations to this modelling approach, this level of agreement is encouraging and should motivate wider use of stochastic modelling applied to SOL phenomena. Of particular note is the observation here that a drastic increase in the density loss timescale was found the be the simplest (in terms of requiring no other changes to the model) consistent way of matching the change in profile shape and change in autocorrelation function at the onset of the profile shoulder formation. This comparison was based on probe measurements made at the wall radius where the auto-correlation function contracts as the profile shoulder forms. In the centre of the profile, where the gradient relaxes as the shoulder forms and the loss timescale is drastically increased and the opposite trend is evident. Indeed the variation of the density sink radially leads to a radial variation in the autocorrelation function. This is shown in figure \[Fig:AutoCorr\_Radial\] which shows the autocorrelation function and PDF shape as a function of radius at the onset of the shoulder formation.
![Left: Radial profile taken from the case at the onset of shoulder formation, shown in figure \[Fig:veloc\_drain\_var\_1\], using a modification of the sink function. Centre: Contours of the autocorrelation as a function of the major radius. Right: Contours of the PDF as a function of the major radius.[]{data-label="Fig:AutoCorr_Radial"}](test_autocorr.png){width="120.00000%"}
\
In the centre of the profile, where the shoulder forms, the autocorrelation function is at its widest since the density loss timescale as longest, whilst the PDF shape becomes increasingly skewed further away from the separatrix. Measurements made on ASDEX-Upgrade [@CarraleroNF2014] and TCV [@VianelloPrePrint] have shown such an increase in the autocorrelation time within the shoulder of the profile. A good test of this theory may also be available on MAST, where profile shoulder formation was observed [@MilitelloNF2015] without the secondary increase in the profile gradient in the far SOL. In this case, if the filament drainage is the responsible actuator for the profile shoulder formation then a probe situated anywhere in the far SOL should measure a widening of the autocorrelation function as opposed to the contraction measured on JET.\
The physical mechanism that may lead to such a strong increase in the loss timescale is not something that the stochastic modelling can determine. Militello and Omotani have suggested that strong momentum loss along the magnetic field line due to charge exchange collisions with neutral atoms in the divertor [@MilitelloPPCF2016] may be a possible cause. Alternatively strong ionisation in the divertor may induce a reduction and eventual reversal of parallel flows. Notably this would imply a dependancy on the geometry of the divertor and in particular the poloidal orientation of the divertor plate. Such a dependancy has been noted on JET where plasmas run in the vertical target (VT) configuration showed a lack of profile shoulder formation compared to those in the horizontal target (HT) configuration [@WynnEPS; @WynnPaper]. Another implied dependancy is on the neutral density. Again in JET plasmas with similar divertor conditions driven by Deuterium seeding and Nitrogen seeding respectively, plasmas with N-seeding were significantly more resistant to the shoulder formation than plasmas with D-seeding. There is therefore an evidence base to support the assertion that density sink function may be responsible for shoulder formation. Furthermore strong D$_{\alpha}$ radiation in the divertor of JET has been postulated to be associated with the presence of a shoulder in the SOL density profile [@WynnPaper], however it is not clear from experiment what role is being played. It is also worthwhile noting that the stochastic modelling suggests that an increase in the loss time can trigger shoulder formation, but cannot capture the subsequent profile evolution in isolation. This suggests the evolution of the shoulder and the formation of the shoulder are related, but different events and should possibly be viewed separately from one another. One such mechanism was found to be a radial change in the filament velocity. One possible trigger for such a change could the profile of $\Lambda$ measured at the divertor. In this paper $\Lambda$ was kept constant radially, however in experiment $\Lambda$ is observed to vary from the separatrix outwards and may provide a mechanism for the widening of the shoulder region, as asserted by Carralero *et al*[@CarraleroNF2014]. Such a variation in $\Lambda$ could be included in the stochastic model here, and would be a good avenue of future work.
Conclusions
===========
This paper has presented a first attempt at interpreting experimental scrape-off layer density profiles and ion-flux statistics to the outer wall in JET using a stochastic model based on filament dynamics. The stochastic model is used to match measurements of the profile simultaneously with measurements of the autocorrelation function and PDF shape of the ion-saturation current measured at the outer-wall. It treats filaments stochastically with an amplitude, ejection time and width drawn from statistical distributions and with a velocity determined by a theoretically motivated relationship with the filament width. By establishing a baseline case prior to the onset of density profile shoulder formation, the stochastic model has been manually optimised to match the experimental measurements. The degree of agreement between the model and the measurements is encouraging and shows that with sufficient inputs, the model is capable of simultaneously capturing the time-averaged profile and the spatially localised statistical measurements. A sensitive study was carried out to assess how the model varied as some of the poorly constrained input parameters were varied. In particular, with the product of the velocity and the timescale of the density sink of filaments fixed to maintain a fixed profile gradient, the filament width was varied to keep the autocorrelation function fixed as the velocity was increased. The required filament width grew exponentially and placed reasonable strict limits on the possible velocities that could be used as input without requiring unrealistically large filaments.\
With the baseline case established, inputs to the stochastic model were varied to recreate the formation of a shoulder in the density profile alongside a contraction of the autocorrelation function. Changes to the input statistical distributions and parameters associated with the filement velocity-width relationship were shown to be unable to capture the change in the profile shape on their own. By employing a spatial variation in either the filament velocity or the loss timescale of the density, the modified profile shape was captured. The contraction of the auto-correlation function was then used to eliminate the spatial variation of velocity since this lead to an expansion which opposed the experimental measurement, whilst the loss time variation lead to the correct degree of contraction in the autocorrelation function without further optimisation. Within the stochastic modelling framework this suggests that a strong increase in the density sink timescale is the simplest way to trigger the onset of density profile shoulder formation. The required increase in the loss time is strongly spatially localised in the region of shoulder formation and represents a factor 25 increase over the baseline level.\
It was not possible to match the profile structure later in the density scan, as the shoulder region widened and flattened, with just an increase in the filament density loss time. This implies that, within the stochastic model, shoulder formation can be triggered by the increase density loss timescale, but the subsequent evolution of the profile requires additional physics. Including in addition a radial variation in the filament velocity alongside the localised increase in the sink timescale was shown to be sufficient to match the profile structure, PDF shape and autocorrelation function at the higher density case. The uniqueness of this solution cannot be guaranteed, but this does imply that additional physics, beyond an increase in the density sink, is required for an expansion of the density shoulder in the stochastic model.\
The stochastic model is shown here to be a potentially useful tool allowing for analysis of physics in the SOL on the basis on non-local intermittent transport without the demands of non-linear fluid codes. It can recreate profile shapes and consistent statistical measurements with appropriate inputs. Whilst there is certainly more development of the method required, it is hoped that this motivates its use going forwards.
Acknowledgements
================
This work has been carried out within the framework of the EUROfusion Consortium and has received funding from the Euratom research and training programme 2014-2018 under grant agreement No 633053 and from the RCUK Energy Programme \[grant number EP/I501045\]. To obtain further information on the data and models underlying this paper please contact [email protected]. The views and opinions expressed herein do not necessarily reflect those of the European Commission
References
==========
Appendix: Statistical convergence of stochastic modelling signals {#App:Conv}
=================================================================
In order to assess the statistical convergence of the measurements made in the stochastic model, simulations have been run with the simulation duration increased through two orders of magnitude. In each case, the simulations are repeated 10 times and the range of the measurements of the profile, PDF shape and autocorrelation functions are shown in figures \[Fig:App\_1\], \[Fig:App\_2\] and \[Fig:App\_3\] respectively.
![Measurement ranges from the convergence study carried out with a simulation time of 18ms. The absolute and relative profiles are shown in the upper row, whilst the PDF shape and autocorrelation function is shown in the bottom row. In each case, the range shown corresponds to the range produced in 10 identical simulations due to statistical variation.[]{data-label="Fig:App_1"}](Stat_convergence_1.png){width="80.00000%"}
![Measurement ranges from the convergence study carried out with a simulation time of 180ms. The absolute and relative profiles are shown in the upper row, whilst the PDF shape and autocorrelation function is shown in the bottom row. In each case, the range shown corresponds to the range produced in 10 identical simulations due to statistical variation.[]{data-label="Fig:App_2"}](Stat_convergence_10.png){width="80.00000%"}
![Measurement ranges from the convergence study carried out with a simulation time of 1.8s. The absolute and relative profiles are shown in the upper row, whilst the PDF shape and autocorrelation function is shown in the bottom row. In each case, the range shown corresponds to the range produced in 10 identical simulations due to statistical variation.[]{data-label="Fig:App_3"}](Stat_convergence_100.png){width="80.00000%"}
This study shows that the profile can be considered converged for relatively short simulation times ($<18ms$) however long time-series ($>1.8s$) are required for decent statistical convergence of the PDF and the autocorrelation function.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'To know the total width of the recently discovered Higgs boson particle, it is important to measure the invisible decay width of the Higgs boson. However, the signal for this measurement at the LHC, i.e., a charged lepton pair and missing energy in the final state, cannot be definitely understood as the product of the intermediate produced $Z$ and $H$ bosons due to the possible interaction between dark matter and a $Z$ boson or quarks, which can be described by representative effective operators. First, we consider the relic abundance, the LUX direct detection experiment and the result of searching for Higgs boson invisible decay at the LEP and LHC to find the allowed parameter region for the effective operators. Then we investigate the transverse momentum distribution of the missing energy and propose two observables that can be used to distinguish the different underlying processes. Moreover, with these two observables, we may be able to determine the masses of invisible particles.'
author:
- Jian Wang
bibliography:
- 'Mass\_of\_inv.bib'
title: 'Determining the masses of invisible particles: Application to Higgs boson invisible decay'
---
Introduction
============
It is important to precisely measure the properties of the Higgs boson after its discovery [@Aad:2012tfa; @Chatrchyan:2012ufa]. The total width of the Higgs boson is difficult to measure at a hadron collider due to the unmeasurable partonic center-of-mass energy. As a result, one can only get information on the total width from the global fit and sum over various decay channels of the Higgs boson [@Barger:2012hv] [^1]. The largest decay channel is $H\to b\bar{b}$, which suffers from overwhelming QCD backgrounds. Although it becomes possible to detect $H\to b\bar{b}$ by using a delicate method proposed in Ref. [@Butterworth:2008iy] and the ATLAS and CMS collaborations have searched for the $b\bar{b}$ decay in the $V(W,Z)H$ associated productions, the statistical uncertainties of the results are still too large [@ATLAS-CONF-2013-079; @CMS-PAS-HIG-13-012]. The decay channel $H\to gg$ faces the same problem, but there is no b-tagging technology that can be used in $H\to b\bar{b}$. The decay width of this channel can only be obtained from a global fit of the relevant Higgs couplings.
In addition, there are still possibilities for the Higgs boson decaying to invisible particles [@Shrock:1982kd; @Choudhury:1993hv; @Gunion:1993jf; @Eboli:2000ze; @Belotsky:2002ym; @Godbole:2003it; @Davoudiasl:2004aj; @Patt:2006fw; @Aad:2012tfa; @Chatrchyan:2012ufa]. To measure the invisible decay width of the Higgs boson at colliders, it is necessary to consider the associated production of the Higgs boson with some visible particles. The $Z$ boson is a good choice due to its large coupling to the Higgs boson, large production rate and clear signature at colliders. Because of the limited center-of-mass energy, the experiment at the LEP has only excluded the Higgs boson mass range below 114.4 GeV via the Higgs-strahlung process $e^+e^- \to HZ$ [@Searches:2001ab]. Recently, the ATLAS collaboration at the LHC carried out a similar search in the $HZ$ associated production with $H$ decaying invisibly and the $Z$ boson decaying into a charged lepton pair. The present result shows no deviation from the Standard Model (SM) expectation and constrains the invisible branching fractions to be less than $65\%$ at the $95\%$ confidence level [@ATLAS-CONF-2013-011]. The CMS collaboration has obtained a similar result [@CMS-PAS-HIG-13-018].
We are interested in the question of, if some deviation from the SM expectation is observed in the future as the integrated luminosity of the LHC is increased, if can we determine the invisible decay width of the Higgs boson. In fact, the experimentally observed final state is just a charged lepton pair and missing energy. And the analysis is based on the assumption that a resonance and a $Z$ boson have been associated produced with the cross section predicted by the SM and the resonance totally decaying invisibly. It is reasonable that the charged lepton pair can be attributed to the decay product of the intermediate $Z$ boson if the invariant mass of the charged lepton pair is around the $Z$ boson mass $M_Z$. However, it is not very convincible to interpret the missing energy as a decay product of the intermediate Higgs boson because there are other possible origins of the missing energy. For example, it is likely that some dark matter (DM) can interact with the $Z$ boson or quarks, which has been also studied extensively in the recent years. They will appear at the LHC with the same signature. Thus, it is essential to extract more information about the missing energy.
In this work, we perform such a study toward this direction, focusing on the mass of the missing particles. If one can determine the mass of the missing particle, one can be more confident to judge whether the intermediate particle is the Higgs boson or not. The candidate for the Higgs boson decay products should have a mass lower than one-half of the Higgs boson mass. This study is closely related to searching for new physics. In the $R$-parity violating supersymmetry model, some sparticles would decay into SM particles and a neutralino, which becomes invisible if it is lighter than the other supersymmetry particles and stable enough [@Barbier:2004ez]. In the large extra dimension model, the Kaluza-Klein graviton can escape from the detection at colliders, manifesting itself as missing energy [@Giudice:1998ck]. Cosmology observation has confirmed the existence of DM in our Universe [@Hinshaw:2012aka]. It can be produced at colliders if it is light and has interactions with SM particles. Because it is stable, it is not detectable at colliders [@Cao:2009uw]. Searching for all these kinds of new physics requires analyzing the events with missing energy and recoiling particles, e.g., a photon [@Gao:2009pn; @Wang:2012qi; @Huang:2012hs; @Chatrchyan:2012tea; @Aad:2012fw], a lepton [@Bai:2012xg], a jet [@Karg:2009xk; @Fox:2012ru], a $W/Z$ boson [@Kumar:2010ca; @Bell:2012rg; @Carpenter:2012rg], or a top quark [@Andrea:2011ws; @Wang:2011uxa]. Therefore, it would be helpful to distinguish them if we can know more about the missing energy. In this paper, we focus on the case of missing energy and a $Z$ boson associated production, which is important to determine the invisible decay width of the Higgs boson. The method employed here can be generalized to other cases.
![Leading Feynman diagrams for a $Z$ boson and missing energy associated production. The diagram $(a)$ is the signal process in the analysis of the ATLAS experiment and the diagram $(c)$ is the main irreducible background. The diagrams $(b)$ and $(d)$ are induced by the effective operators in Eq.(\[eqs:oz\]). []{data-label="fig:ZH"}](ZH.eps "fig:"){width="0.95\linewidth"}\
Effective operators
===================
The Higgs boson is special in the SM. Its mass term is not fixed by the gauge invariance and provides an opportunity for this particle to couple with some SM gauge singlets still with renormalizable interactions [@Patt:2006fw]. The simplest case is adding to the SM a new real scalar $S$ with the Lagrangian given by [@Silveira:1985rk; @McDonald:1993ex; @Burgess:2000yq; @Barger:2007im] $$\mathcal{L}_{\rm min}=\mathcal{L}_{\rm SM}+\frac{(\partial_{\mu} S)^2}{2}
-\frac{m_0^2}{2}S^2 - \lambda S^2 |H|^2 - \frac{\lambda_S}{4!}S^4.$$ The new real scalar can account for the observed DM density if a global $Z_2$ symmetry is imposed, under which all the SM particles are singlets while the new scalar can transform nontrivially, i.e.,$S\to -S$. To maintain this $Z_2$ symmetry, the new scalar should have no vacuum expectation value (VEV), $\langle S \rangle=0$. Thus, its mass $m_S$ is given by $m_S^2=m_0^2+\lambda v^2$, where $(0,v)^T/\sqrt{2}$ is the VEV of the Higgs field. When $2m_S$ is smaller than the Higgs boson mass $m_H$, the Higgs boson can decay into an $S$ pair, namely, Higgs boson invisible decay. This would modify the total width of Higgs boson and therefore affect the cross section of Higgs boson production and decay into other final states. Through the mixing with the Higgs boson, the $S$ pair can annihilate to SM particle pairs, for which the cross section is constrained by the DM relic density. With the same interaction as in annihilation, $S$ can scatter with nucleons, which can be measured by DM direct detection experiments. Given that the Higgs boson mass is around 125 GeV, this model receives strong constraint after considering the results of searching for Higgs boson at the LHC, cosmological relic density and the DM direct detection [@Raidal:2011xk; @Mambrini:2011ik; @He:2011de; @Fox:2011pm; @Low:2011kp; @Djouadi:2011aa; @Cheung:2012xb]. Therefore, more complicated models , in which more particles are included, are proposed [@Abada:2011qb; @Pospelov:2011yp; @Abada:2013pca; @Greljo:2013wja]. There are no rules, in principle, that these additional particles should be very light [@Abada:2011qb]. And they are perhaps heavy since a lighter particle is often much easier to find either in the decay product or direct production at colliders. In this case, the role played by these particles can be described by effective operators, $$\begin{aligned}
\label{eqs:oz}
& & \mathcal{O}_Z=\frac{m_Z^2}{4\Lambda_Z^2} Z^{\mu}Z_{\mu} S^2 , \qquad \mathcal{O}_q=\frac{m_q}{2\Lambda_q^2}\bar{q} q S^2.\end{aligned}$$ They can induce the production processes described by the Feynman diagrams $(b)$ and $(d)$ in Fig.\[fig:ZH\], generating the same signature at hadron colliders as the process assumed in the experimental analysis, e.g., the Feynman diagrams $(a)$ and $(c)$. Note that $\mathcal{O}_Z$ and $\mathcal{O}_q$ are not gauge invariant. In fact, one can write down gauge invariant operators with higher dimensions and get the above two operators after symmetry breaking. For example, the first operator $\mathcal{O}_Z$ can be generated from the gauge invariant dimension-6 effective operator: $$\label{eqs:OZdim6}
\mathcal{O}_Z^{(6)}= \frac{\kappa}{\Lambda^2} (D^{\mu}H)^{\dagger}(D_{\mu}H)S^2,$$ where $D_{\mu}$ is the usual covariant derivative, and $H$ denotes the SM Higgs field. After the Higgs field gets a nonvanishing vacuum expected value, $\mathcal{O}_H^{(6)}$ would deduce to $\mathcal{O}_Z$ and the associate similar operator $$\mathcal{O}_W=\frac{M_W^2}{2\Lambda_W^2} W^{+\mu}W^{-}_{\mu} S^2$$ because the Higgs field couples with $W$ and $Z$ bosons simultaneously. The second operator $\mathcal{O}_q$ can be generated from the gauge invariant operator of dimension 6 $$\mathcal{O}_q^{(6)} = \frac{\kappa \lambda_q }{\Lambda^2} \bar{Q}_LH q_R S^2 + H.c.,$$ where $\lambda_q$ is the Higgs-quark-quark Yukawa couplings in the SM.
Here we choose these two effective operators in Eq.(\[eqs:oz\]) because they are the only ones in the leading power order, and representative, one of them inducing an $s$-channel and the other inducing a $t$-channel process; see Feynman diagrams $(b)$ and $(d)$ in Fig.\[fig:ZH\]. The detailed discussion on the possible gauge symmetry breaking mechanism is beyond the scope of this paper. The above effective operators have been discussed in the study of DM [@DelNobile:2011uf; @Goodman:2010ku].
The main difference between Feynman diagrams $(a)$ and $(b)$ is whether the missing particles have a fixed invariant mass. In some phase space points where the invariant mass of the $S$ pair is around the Higgs boson mass, Feynman diagram $(b)$ would be equivalent to Feynman diagram $(a)$ up to some constant coefficients. However, the phase space for missing particles cannot be fixed experimentally and must be integrated in any observable. It is the purpose of this study to find a way to reveal this difference.
Numerical Discussion
====================
To distinguish the different processes shown in Fig.\[fig:ZH\] and extract precise information on Higgs boson invisible decay, we should understand the properties of the cross sections first. At the moment, we neglect the interference among the Feynman diagrams $(a,b,c,d)$. This is reasonable because the total widths of the Higgs and $Z$ bosons are small enough so that we can use narrow width approximation in calculating the Feynman diagrams $(a,c)$ and thus do not need to consider their interference with the Feynman diagrams $(b,d)$. In numerical discussion, we have taken $m_H=125$ GeV, $M_Z=91.18$ GeV. We use the Monte Carlo method to calculate the cross section for $Z$ boson and missing energy associated production at the LEP and LHC, and find agreement with MadGraph5v1.3.3 [@Alwall:2011uj] for processes at the LHC. The CTEQ6L1 parton distribution function (PDF) set [@Pumplin:2002vw] is chosen and the renormalization and factorization scales are set to be 200 GeV, which is about the sum of $m_H$ and $M_Z$.
Before proceeding, we have to find the allowed parameter space of ($\Lambda_{Z/q}, m_S$). In the following, we will discuss the DM relic abundance and direct detection experiment LUX as well as the $Z$ boson and missing energy associated production at the LEP and LHC.
Constraint from relic abundance
-------------------------------
The DM relic abundance is a precision observable in cosmology and imposes constraint on any DM model. It is determined by the annihilation cross section of DM to SM particles, given by [@Kolb:1990vq] $$\Omega_{\rm DM} h^2 \approx \frac{1.07\times 10^9 {\rm GeV}^{-1}x_f}{M_{\rm Pl}g_{*}^{1/2}(a+3b/x_f)},$$ where $\Omega_{\rm DM}$ is the cold DM energy density of the Universe normalized by the critical density, and $h=0.700 \pm 0.022$ is the scaled Hubble parameter. $a$ and $b$ are the coefficients in the partial wave expansion of the DM annihilation cross section, $\sigma_{\rm an}v_{\rm M\o l}=a+bv_{\rm M\o l}^2+O(v_{\rm M\o l}^4)$. $v_{\rm M\o l}$ is called M$\o$ller velocity, defined as [@Gondolo:1990dk] $$v_{\rm M\o l}=\sqrt{|\textbf{v}_1-\textbf{v}_2|^2-|\textbf{v}_1 \times \textbf{v}_2|^2},$$ where $\textbf{v}_1$ and $\textbf{v}_2$ are the velocities of colliding DMs in the cosmic comoving frame. Note that this velocity is different from that in collisions at colliders, since the colliding DMs in the Universe are not necessarily moving along a line. And $v_{\rm M\o l}$ cannot be transformed into the center-of-mass frame of the two colliding DMs because the thermally averaged total annihilation cross section $\langle \sigma_{\rm an}v_{\rm M\o l} \rangle$ has to be evaluated in a common frame for all collisions. In general, it is difficult to calculate $\langle \sigma_{\rm an}v_{\rm M\o l} \rangle$ due to the complex definition of $v_{\rm M\o l}$. Fortunately, it is proved that [@Gondolo:1990dk] $$\langle \sigma v_{\rm M\o l} \rangle = \langle \sigma v_{\rm lab} \rangle^{\rm lab}\neq \langle \sigma v_{\rm cm} \rangle^{\rm cm},$$ in which $v_{\rm lab}=|\textbf{v}_{1,\rm lab}-\textbf{v}_{2,\rm lab}|$ is the relative velocity in the rest frame of one of the incoming particles and $v_{\rm cm}$ is the velocity in the center-of-mass frame of the two colliding DMs. The DM is moving at nonrelativistic velocities when freezing out; thus, $v\ll 1$. $g_*$ is the number of relativistic degrees of freedom available at the freeze-out epoch $x_f$. And $x_f$ is evaluated by [@Kolb:1990vq] $$\begin{aligned}
x_f &=& \ln A - 0.5 \ln \ln A + \ln \left( 1+\frac{6b}{a\ln A} \right)\end{aligned}$$ with $A=0.038ag/g_*^{1/2}M_{\rm Pl}m_S$. $g$ counts the internal degree of freedom and is equal to 1 for a real scalar in our case.
For $SS\to ZZ$ or $q\bar{q}$, the annihilation cross section is $$\begin{aligned}
\sigma_{\rm an}^Z&=&\frac{1}{32\pi s}\left( \frac{M_Z^2}{\Lambda_Z^2} \right)^2\frac{\sqrt{s-4M_Z^2}}{\sqrt{s-4m_S^2}}
\left( \frac{s^2}{4M_Z^4}-\frac{s}{M_Z^2}+3 \right), {\nonumber}\\
\sigma_{\rm an}^q&=&\frac{N_c}{8\pi}\left( \frac{m_q}{\Lambda_q^2} \right)^2\frac{\sqrt{s-4m_q^2}}{\sqrt{s-4m_S^2}} \frac{s-4m_q^2}{s}.\end{aligned}$$ These results are in agreement with those in Refs.[@Beltran:2008xg; @Cao:2009uw; @Yu:2011by]. Substituting $s\approx 4m_S^2 + m_S^2 v^2 + 3 m_S^2 v^4/4$, we obtain $$\begin{aligned}
a^Z &=&\frac{1}{32\pi }\left( \frac{M_Z^2}{\Lambda_Z^2} \right)^2 \frac{\sqrt{m_S^2-M_Z^2} \left(4 m_S^4-4m_S^2 M_Z^2+3
M_Z^4\right)}{2m_S^3 M_Z^4} , {\nonumber}\\
b^Z &=&\frac{1}{32\pi }\left( \frac{M_Z^2}{\Lambda_Z^2} \right)^2\frac{ 3(4 m_S^4-8 m_S^2 M_Z^2+5 M_Z^4)}{16 m_S^3 M_Z^2 \sqrt{m_S^2-M_Z^2}}. {\nonumber}\\\end{aligned}$$ and $$\begin{aligned}
a^q &=&\sum_{q}^{m_q<m_S} \frac{N_c}{4\pi}\left( \frac{m_q}{\Lambda_q^2} \right)^2 \left( 1-\frac{m_q^2}{m_S^2} \right)^{3/2} , {\nonumber}\\
b^q &=&\sum_{q}^{m_q<m_S} \frac{N_c}{4\pi}\left( \frac{m_q}{\Lambda_q^2} \right)^2 \left( 1-\frac{m_q^2}{m_S^2} \right)^{1/2}
\frac{5m_q^2/m_S^2-2}{8}.{\nonumber}\\\end{aligned}$$
The latest WMAP data and distance measurements from baryon acoustic oscillations in the distribution of galaxies and Hubble constant measurements give the constraint [@Hinshaw:2012aka] $$\Omega_{\rm DM} h^2 = 0.1157 \pm 0.0023.$$ Requiring that the DM relic abundance is lower than the observed central value at the $2\sigma$ level, the allowed parameter space of the effective operator is determined. We show the parameter space in Figs. \[fig:relicZ\] and \[fig:relic\] for the operators $\mathcal{O}_Z$ and $\mathcal{O}_q$ , respectively. Because of the heavy mass of the $Z$ boson, we show in Fig. \[fig:relicZ\] only the DM mass larger than 92 GeV. For $m_S=100$ GeV, the upper limit of $\Lambda_Z$ is about 500 GeV. For $m_S < M_Z $, the annihilation cross section of $SS\to ZZ$ is vanishing. Another annihilation channel is required to satisfy the relic abundance constraint, e.g. $SS \to e^+ e^-$. This can be established in a UV-complete model. We are working in the effective operator picture here and will not discuss this further. What we must keep in mind is that the relic density provides an upper limit on $\Lambda_Z$ while the direct detection experiments and processes at colliders impose lower limits. There is some parameter space for the DM signal mimicking the Higgs invisible decay as long as the upper limit is not smaller than the lower limit.
![The allowed space of ($\Lambda_{Z}, ~m_S$) from relic abundance . The region below the solid blue line is allowed. The gray region corresponds to $\Lambda_Z < m_S$. []{data-label="fig:relicZ"}](relicZ.eps "fig:"){width="0.89\linewidth"}\
![The allowed space of ($\Lambda_{q}, ~m_S$) from relic abundance and the direct detection experiment. The region below the solid blue line is allowed by the relic abundance. The region below the dashed red line is excluded by the LUX direct detection experiment [@Akerib:2013tjd]. The gray region corresponds to $\Lambda_q < m_S$. []{data-label="fig:relic"}](relic.eps "fig:"){width="0.89\linewidth"}\
Constraint from direct detection
--------------------------------
DM around the Earth can scatter elastically with atomic nuclei, resulting in recoiling movements of nuclei. These events, if observed, can be explained by DM-nucleon collisions, which can be divided into spin dependent and spin independent according to the DM-quark interactions. Generally, the spin independent elastic scattering cross section has a large value and thus gets a more stringent constraint from direct detection experiments. The most stringent limit on the spin independent elastic scattering cross section comes from the LUX experiment [@Akerib:2013tjd]. Therefore in this analysis, we only consider the data of LUX in order to obtain the allowed parameter space.
The DM-proton spin independent elastic scattering cross section [^2] is given by [@Belanger:2008sj; @Cao:2009uw] $$\sigma_{Sp}^{SI}=\frac{m_p^2}{4\pi(m_S+m_p)^2}[f_{Sp}^{(p)}]^2,$$ where $$f_{Sp}^{(p)}=\sum_{q=u,d,s} f_{T_q}^{(p)}C_{Sq}\frac{m_p}{m_q}
+\frac{2}{27}f_{T_g}^{(p)}\sum_{q=c,b,t}C_{Sq}\frac{m_p}{m_q},$$ with [@Ellis:2000ds] $$\begin{aligned}
&& f_{T_u}^{(p)}\approx 0.020\pm 0.004, \quad
f_{T_d}^{(p)}\approx 0.026\pm 0.005, \quad {\nonumber}\\
&& f_{T_s}^{(p)}\approx 0.118\pm 0.062, \quad
f_{T_g}^{(p)}\approx 1- f_{T_u}^{(p)}- f_{T_d}^{(p)}- f_{T_s}^{(p)}. {\nonumber}\\\end{aligned}$$ In our case, $C_{Sq}=m_q/\Lambda_q^2$. After comparing with the LUX data [@Akerib:2013tjd], we obtain the allowed parameter region in Fig. \[fig:relic\]. It is observed that the combination of the relic abundance and direct detection experiment results in very stringent constraints on $\mathcal{O}_q$. Only a limited parameter space with $m_S<5.5$ GeV and $\Lambda_q<$110 GeV is allowed. We notice that there are no such constraints on the parameters of the operator $\mathcal{O}_Z$ since a nucleon does not contain a $Z$ boson. Therefore, we will discuss only the case of $\mathcal{O}_Z$ in the following part.
Constraints from $Z$ boson and missing energy associated production
-------------------------------------------------------------------
Apart from DM annihilation and elastic scattering with nucleons, DM can be produced at the colliders. Here we are interested in the process of $Z$ boson and missing energy associated production at the LEP and LHC, i.e., $e^+ e^- (pp) \to Z^* \to SSZ$. For the LEP experiments, we take the upper limit on the invisible Higgs boson production cross section at 206.0 GeV, given in Ref. [@Searches:2001ab]. We show the excluded parameter space in Fig. \[fig:totX\]. It is found that the region excluded by the current LHC has covered all that by LEP, indicating the better sensitivity of the LHC to the new physics. As the increasing of the data accumulated at the LHC, the excluded region will enlarge if no signal of new physics is observed. We show in Fig. \[fig:totX\] as well the curves corresponding to the upper limit on the cross section of $HZ$ production at the LHC with a Higgs boson branching fraction of $65\%,10\%, {\rm ~and~} 5\%$, respectively.
On the other hand, if an excess of events is found in the future, it may be induced by $\mathcal{O}_Z$ with the parameter combination lying on the curve that corresponds to a fixed Higgs boson branching fraction. Then it is essential to judge that this excess results from Higgs boson invisible decay or DM associated production with a $Z$ boson.
![The allowed space of ($\Lambda_{Z}, ~m_S$) from $Z$ boson and missing energy associated production. The shaded region is excluded by the limit on the missing energy and $Z$ boson production at the LEP [@Searches:2001ab] and the LHC [@ATLAS-CONF-2013-011]. The three curves from the bottom up correspond to the cross section of $HZ$ production at the LHC with a Higgs boson branching fraction of $65\%,10\%,{\rm ~and ~}5\%$, respectively.[]{data-label="fig:totX"}](kappaz.eps "fig:"){width="0.89\linewidth"}\
Missing transverse momentum distribution
----------------------------------------
Now we show the normalized $p_T^{\rm miss}$ distributions for missing energy and $Z$ boson production at the 8 TeV LHC in Fig.\[fig:pt\_Z\_SSZB\] for the process induced by $\mathcal{O}_Z$. We see that the shape of the process induced by $\mathcal{O}_Z$ with $m_S=$ 30 GeV is very similar to the SM process of $H({\rm invisible~decay})Z$ production. This similarity suggests that one cannot simply interpret the signal of missing energy and a charged pair production as the associated $HZ$ production with Higgs boson invisible decay.
![The $p_T^{\rm miss}$ distribution for missing energy and $Z$ boson production at the 8 TeV LHC induced by $\mathcal{O}_Z$. We also show the SM processes of $H({\rm invisible~decay})Z$ and $Z({\rm invisible~decay})Z$ production. []{data-label="fig:pt_Z_SSZB"}](mpt.eps "fig:"){width="0.99\linewidth"}\
From Fig.\[fig:pt\_Z\_SSZB\], we find that the peak position for the process induced by $\mathcal{O}_Z$ is moving toward the large $p_T^{\rm miss}$ region, and the tail of the distribution drops more slowly as the DM mass increases. To make it explicit, we present the fitted formulas describing the peak positions: $$\begin{aligned}
\label{eqs:pTpeak1}
p_{T,\mathcal{O}_Z}^{\rm peak} & = & 15.83 {~\rm GeV} + 6.67{~\rm GeV} \frac{m_S}{10\rm~GeV}.\end{aligned}$$ We notice that this formula is fitted from the partonic simulation results, not including the parton shower and hadronization effects, high-order QCD corrections, etc. The exact peak positions may be different after taking into account all these effects. However, we still use this formula as long as the difference between the fitted value and the exact one is not very large. For comparisons, we also list the values of the peak positions for the processes in the SM: $$\begin{aligned}
\label{eqs:pTpeak2}
p_{T,HZ}^{\rm peak} & = & 50.0 {~\rm GeV}, \\
p_{T,ZZ}^{\rm peak} & = & 32.5 {~\rm GeV}.
\label{eqs:pTpeak3}\end{aligned}$$
At the low integrated luminosity of the LHC, the observed events may not be enough to provide a full description of the $p_T^{\rm miss}$ distribution. For this reason, we propose two observables that can be used to distinguish the underlying processes at the early stage of the LHC, defined as $$\begin{aligned}
\label{eqs:R1}
R_1 &\equiv& \frac{\sigma(p_T^{\rm miss}<p_T^{\rm peak})}{\sigma(p_T^{\rm miss}>p_T^{\rm peak})}, \\
R_2 &\equiv& \frac{\sigma(p_T^{\rm miss}<p_T^{\rm cut})}{\sigma(p_T^{\rm miss}>p_T^{\rm cut})}.
\label{eqs:R2}\end{aligned}$$ Here $p_T^{\rm peak}$ is defined theoretically as the $p_T$ value around which the distribution is the largest (the experimental value of $p_T^{\rm peak}$ is discussed in the following). It may take different values for the signal and background events. And the default value of $p_T^{\rm cut}$ is chosen to be $150 ~{\rm GeV}$. $R_1$ describes the profile of the peak region while $R_2$ incorporates the information on the tail region. It is more convenient to adopt these two variables rather than the full distributions to understand the underlying processes. Moreover, because of the ratios in $R_1$ and $R_2$, the coefficients of the operators are canceled. As a consequence, $R_1$ and $R_2$ are functions of only the DM mass, providing a handle to the masses of invisible particles. In particular, many effects that may change the leading-order prediction, such as the factorization and renormalization scales, PDF sets, parton shower, and higher order corrections are supposed to be canceled substantially in these observables as well.
We emphasize again that we are interested in how to distinguish the signal processes between Higgs invisible decay and DM associated production after the discovery of the signal. In this case, we can divide the total events observed experimentally to the background (mainly $ZZ$ production) and signals (possible $HZ$ production or $SSZ$ production). And therefore, the two observables, $R_1$ and $R_2$, can be measured separately for the background and signals.
The theoretical prediction for the dependence of $R_{1}$ and $R_{2}$ on the DM mass is shown in Fig.\[fig:R\_SSZB\]. In the plot of $R_1$, the curves are obtained by using Eq. (\[eqs:R1\]) with the corresponding $p_T^{\rm peak}$ given in Eqs.(\[eqs:pTpeak1\]$\sim$\[eqs:pTpeak3\]).
![The dependence of $R_{1}$ and $R_{2}$ on the DM mass. The narrow bands indicate the scale uncertainties by varying the default scale by a factor of 2. The wide bands in the upper plot are obtained by changing the $p_{T}^{\rm peak}$ by $\pm 5$ GeV. The blue and green dashed lines represent the results obtained by using CT10 and MSTW2008 PDF sets, respectively. The $R_2$ value for $Z({\rm invisible~decay})Z$ production has been divided by a factor of 3 in the bottom plot. In the example illustrating the estimation of $\Delta m_S$ in the bottom plot, $\Delta R_2$ is set to be $0.1 R_2$. []{data-label="fig:R_SSZB"}](R1.eps "fig:"){width="0.9\linewidth"}\
![The dependence of $R_{1}$ and $R_{2}$ on the DM mass. The narrow bands indicate the scale uncertainties by varying the default scale by a factor of 2. The wide bands in the upper plot are obtained by changing the $p_{T}^{\rm peak}$ by $\pm 5$ GeV. The blue and green dashed lines represent the results obtained by using CT10 and MSTW2008 PDF sets, respectively. The $R_2$ value for $Z({\rm invisible~decay})Z$ production has been divided by a factor of 3 in the bottom plot. In the example illustrating the estimation of $\Delta m_S$ in the bottom plot, $\Delta R_2$ is set to be $0.1 R_2$. []{data-label="fig:R_SSZB"}](R2.eps "fig:"){width="0.9\linewidth"}\
It can be seen that $R_1$ is insensitive to the DM mass for the processes induced by $\mathcal{O}_Z$ and takes discrepant values between the SM processes and processes induced by $\mathcal{O}_Z$. On the other hand, the experiment would give $p_T$ distributions in bins for both the background and signal, from which we can obtain a rough estimate of $p_T^{\rm peak}$. But because of the limited signal events, the bin is perhaps too wide (for example 10 GeV) to precisely determine $p_T^{\rm peak}$. In this case, we choose $p_T^{\rm peak}$ to be the left (right) edge of the maximum bin if the left (right) neighbor of the maximum bin is larger than the right (left) one. Then we count the left and right bins, obtaining the experimental value for $R_1$. Given that the definitions of $p_T^{\rm peak}$ are different from the theory and experiment sides, it is possible that they are not equal to each other. This uncertainty would dilute the precision in determining the value of DM mass. From Eq.(\[eqs:pTpeak1\]), $\Delta m_S \approx 1.5 \Delta p_{T,Z}^{\rm peak}$. Thus, an uncertainty of 5 GeV in $p_{T,Z}^{\rm peak}$ would result in an uncertainty of about 7.5 GeV in $m_S$. So we will not use $p_{T,Z}^{\rm peak}$ to probe the value of $m_S$. However, we can still use $R_1$ to separate the $HZ$ production with Higgs boson invisible decay and $Z$ boson with DM associated production processes. In Fig.\[fig:R\_SSZB\], we show the uncertainty of $R_1$ by changing the $p_{T}^{\rm peak}$ by $\pm 5$ GeV. It can be seen that the values of $R_1$ for the DM associated production and Higgs invisible decay processes do not overlap, except for a very small region. $R_2$ is very sensitive to the DM mass for the process induced by $\mathcal{O}_Z$, especially in the DM mass range $m_S<50$ GeV. If the experimentally measured value of $R_2$ intersects with the curve corresponding to the process induced by $\mathcal{O}_Z$, we can determine the mass of DM. The uncertainties arising from the variation of scales and PDF sets are also calculated, explicitly shown in Fig.\[fig:R\_SSZB\], which turn out to be small, as expected. After taking into account the scale uncertainties, which are much larger than those from the PDF sets, the accuracy of the determined DM mass is estimated to be $\Delta m_S= 3.4\sim 6.8 $ GeV as $R_2$ changes from 10 to 2 (corresponding to $m_S$ from 15.4 to 65.0 GeV); see Fig.\[fig:error150\]. This estimation is obtained without considering the effects of parton shower, hadronization and detector simulation but under the assumption that the measured $R_2$ is accompanied with an uncertainty of $0.1 R_2$. As the integrated luminosity is increased, the uncertainty of $R_2$ would be reduced. We also show the situation in which the uncertainty of $R_2$ is $0.05 R_2$ in Fig.\[fig:error150\]. Then, the accuracy would be $\Delta m_S= 2.4\sim 4.5 $ GeV as $R_2$ changes from 10 to 2.
![The determined $m_S$ as a function of $R_{2}$ in the case of $p_T^{\rm cut}=150$ GeV. The error of $R_2$ is set to be $0.1 R_2$ and $0.05 R_2$ for the large and small error bars respectively, while the error of $m_S$ is derived as illustrated in Fig.\[fig:R\_SSZB\]. []{data-label="fig:error150"}](error150p.eps "fig:"){width="0.9\linewidth"}\
![The determined $m_S$ as a function of $R_{2}$ in the case of $p_T^{\rm cut}=100$ and 200 GeV. The error of $R_2$ is set to be $0.1 R_2$ while the error of $m_S$ is derived as illustrated in Fig.\[fig:R\_SSZB\]. []{data-label="fig:error100"}](error100.eps "fig:"){width="0.9\linewidth"}\
![The determined $m_S$ as a function of $R_{2}$ in the case of $p_T^{\rm cut}=100$ and 200 GeV. The error of $R_2$ is set to be $0.1 R_2$ while the error of $m_S$ is derived as illustrated in Fig.\[fig:R\_SSZB\]. []{data-label="fig:error100"}](error200.eps "fig:"){width="0.9\linewidth"}\
Then we discuss the impact from the choice of $p_T^{\rm cut}$ in the definition of $R_2$ on the determination of the DM mass. We have chosen $p_T^{\rm cut}=150$ GeV in the above numerical results. But it is possible to choose a different value of $p_T^{\rm cut}$ as long as it is in the tail region (much greater than $p_T^{\rm peak}$). The two cases of $p_T^{\rm cut}=100$ and 200 GeV are also shown in Figs.\[fig:R\_SSZB\] and \[fig:error100\]. We see that for larger $p_T^{\rm cut}$, the scale uncertainty of $R_2$ at fixed $m_S$ is larger. But $R_2$ drops faster as $m_S$ increase at the same time. The net effect results in a similar accuracy in estimating $m_S$. We also notice that too large $p_T^{\rm cut}$ would induce a large statistical uncertainty, and too small $p_T^{\rm cut}$ would reduce the sensitivity to $m_S$ [^3]. As a consequence, $p_T^{\rm cut}=150$ GeV is a good choice.
conclusion
==========
To know the total width of the recently discovered Higgs boson particle, it is important to measure the invisible decay width of the Higgs boson. However, the signal for this measurement at the LHC, i.e., a charged lepton pair and missing energy in the final state, cannot be definitely understood as the product of the intermediate produced $Z$ and $H$ bosons due to the possible interaction between DM and $Z$ boson or quarks, which can be described by representative effective operators. First, we consider the relic abundance, the LUX direct detection experiment and the result of searching for Higgs boson invisible decay at the LEP and LHC in order to find the allowed parameter region space for the effective operators. We discover that the interaction between DM and quarks is stringently constrained. Then we investigate the transverse momentum distribution of the missing energy and propose two observables that can be used to distinguish the different underlying processes. Moreover, with these two observables, we may be able to determine the mass of invisible particles.
In this paper, we only consider the Higgs boson invisibly decaying into scalar DM, inspired by the possible renormalizable extension of the SM. And we also assume the $Z$ boson and quarks also interact with this kind of DM. Given that the DM takes more parts of the energy in the Universe than ordinary matter, it is likely that there are many kinds of DM and the kind coupling with the $Z$ boson and quarks differs from that with the Higgs boson. For example, the $Z$ boson and quarks connect with fermionic DM, while the Higgs boson decays into scalar DM. We will explore these scenarios in the future.
acknowledgement
===============
This work was supported by the Cluster of Excellence [*Precision Physics, Fundamental Interactions and Structure of Matter*]{} (Grant No. PRISMA-EXC 1098).
[^1]: We notice that a constraint is presented on the total width of the Higgs boson, i.e., $\Gamma_{\rm H}< 22 {\rm ~MeV}$ at the $95\%$ confidence level [@Khachatryan:2014iha], by using its relative on-shell and off-shell production and decay rates to a pair of $Z$ bosons [@Caola:2013yja].
[^2]: In practice, the DM-neutron spin independent elastic scattering cross section is almost identical to the DM-proton one [@Cao:2009uw].
[^3]: This sensitivity can be understood as $dR_2/dm_S$.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
\
Institute for Theoretical Physics, University of Wroclaw, 50-204 Wroclaw, Poland\
Bogoliubov Laboratory for Theoretical Physics, Joint Institute for Nuclear Research, RU-141980 Dubna, Russia\
E-mail:
title: Cluster virial expansion for quark and nuclear matter
---
Introduction
============
Let us consider a many-particle system with strong interactions which lead to the formation of bound states (clusters) of different size. One example is nuclear matter, where nucleons are the elementary degrees of freedom and the clusters such as deuterons, tritons, helions, alphas etc. coexist in a nuclear statistical equilibrium (NSE). Another example is quark matter, where mesons and baryons form a spectrum hadronic resonances with the specifics that due to quark confinement no free quark states shall get populated at low densities. The problem we shall consider in this contribution is that the clusters may interact with each other thus forming higher order correlations and their formation itself shall modify the properties of their constituents. These aspects have to be dealt with in a consistent way, obeying conservation laws (sum rules) despite the fact that necessarily approximations have to be applied such as truncations of the hierarchy of many-particle Greens functions describing the correlations.
The problem to define conserving approximations to the many-body problem has been solved in general by the so-called $\Phi-$derivable approach [@KB] which could be applied to the problem of systems with composite particles [@David2; @David3] but in fact has been used in the approximation that all degrees of freedom have no substructure [@David4; @David5; @David6; @David1]. On the other hand, a quantum statistical approach has been developed for the problem of bound state formation (and their dissociation by the Mott effect) in many-particle systems based on a cluster decomposition of the selfenergy [@RMS]. This approach has been developed further to include both, bound and scattering states of the two-particle spectrum in a dense medium in a consistent way given by a generalization [@Zimmermann:1985ji; @SRS] of the well-known Beth-Uhlenbeck approach [@Beth:1936zz; @Beth:1937zz]. A next step in this development was the inclusion of larger clusters [@Typel:2009sy; @Ropke:2014fia] and their mutual interactions to lowest order (second virial coefficient) within a cluster virial expansion [@Ropke:2012qv].
In the present contribution we will show how both approaches can be joined, i.e., how a $\Phi$ functional must be introduced in order to define the cluster virial expansion. As an example serves the formation of deuterons in nuclear matter described on the basis of a nonrelativistic Greens function approach employing a separable potential. Another example is the formation of mesons in quark matter described within the field-theoretic Nambu–Jona-Lasinio model, where the present approach allows to define the contribution of mesons to the quark selfenergy, analoguous to the case of a Dirac fermion coupled to a scalar field [@Blaizot:1991kh], see also [@Kitazawa:2014sga]. This is an effect missing up to now in the Beth-Uhlenbeck description [@Hufner:1994ma; @Zhuang:1994dw; @Blaschke:2013zaa] of the thermodynamics of the quark-meson system, see also [@Yamazaki:2012ux; @Wergieluk:2012gd]. We give an outlook to the case of higher order clusters in nuclear matter as well as to a consistent description of the system of diquarks, mesons and baryons in quark matter.
Selfconsistent approximation scheme and cluster decomposition
==============================================================
Within the $\Phi-$ derivable approach [@KB] the grand canonical thermodynamic potential for a dense fermion system with two-particle correlations is given as $$\begin{aligned}
\label{2PI}
\Omega = -{\rm Tr}\,\{ \ln (-G_1)\} -{\rm Tr} \{\Sigma_1 G_1\}
+{\rm Tr}\,\{ \ln (-G_2)\} +{\rm Tr} \{\Sigma_2 G_2\} +\Phi[G_1,G_2]~,\end{aligned}$$ where the full propagators obey the Dyson-Schwinger equations $$\begin{aligned}
G_1^{-1}(1,z)&=& z - E_1(p_1) - \Sigma_1(1,z);~
G_2^{-1}(12,1'2',z) =z-E_1(p_1)-E_2(p_2)-\Sigma_2(12,1'2',z),\end{aligned}$$ with selfenergies $$\begin{aligned}
\Sigma_1(1,1') &=& \frac{\delta \Phi}{\delta G_1(1,1')}~;~~
\Sigma_2(12,1'2',z)=\frac{\delta \Phi}{\delta G_2(12,1'2',z)}~,\end{aligned}$$ which are defined by the choice for the $\Phi-$ functional, a two-particle irreducible set of diagrams such the one in Fig. \[fig:1\].
![Left panel: Two-particle irreducible $\Phi-$ functional describing two-particle correlations (double line with arrow) of elementary fermions (single arrowed lines); Right panel: The $\Phi-$ functional for the general case of $A-$particle correlations in a many-fermion system.[]{data-label="fig:1"}](1){width="100.00000%"}
![Left panel: Two-particle irreducible $\Phi-$ functional describing two-particle correlations (double line with arrow) of elementary fermions (single arrowed lines); Right panel: The $\Phi-$ functional for the general case of $A-$particle correlations in a many-fermion system.[]{data-label="fig:1"}](2){width="100.00000%"}
The functional for the thermodynamic potential (\[2PI\]) is constructed such that the requirement of its stationarity in thermodynamic equilibrium is equivalent to [@David1] $$\begin{aligned}
\label{n}
n = - \frac{1}{V} \frac{\partial \Omega}{\partial \mu}
= \frac{1 }{V} \sum_{1}
\int_{-\infty}^\infty\frac{d \omega}{ \pi} f_{1}(\omega) A_1(1,\omega)\,,
%\nonumber\\
% S_1(1,\omega) &=& {2 {\rm Im}\,\Sigma_1(1,\omega-i0)
%\over(\omega - E(1)- {\rm Re}\, \Sigma_1(1,\omega))^2 +
%({\rm Im}\,\Sigma_1(1,\omega-i0))^2 }\end{aligned}$$ where $A_1(1,\omega)=2 \Im G_1(1,\omega+i\eta)$ is the fermion spectral function and Eq. (\[n\]) expresses particle number conservation in a system with volume $V$. This approach is straightforwardly generalized to $A-$particle correlations in a many-fermion system $$\begin{aligned}
\Omega &=& \sum_{A} (-1)^A \left[{\rm Tr} \ln \left(-G_A^{-1}\right)
+ {\rm Tr} \left(\Sigma_A~G_A \right) \right]+\sum_{A,B}\Phi[G_A,G_B,G_{A+B}]~,
%\nonumber
\\
G_{A}^{-1}&=&G_{A}^{(0)^{-1}} - \Sigma_A~,~~
\Sigma_A(1\dots A,1^\prime \dots A^\prime,z_A) =
\frac{\delta \Phi}{\delta G_A(1\dots A,1^\prime \dots A^\prime,z_A)}~.
%\nonumber
\label{Phi-A}\end{aligned}$$ The $\Phi-$ functional for this general case is depicted diagrammatically in the right panel of Fig. \[fig:1\] Herewith we have generalized the notion of the $\Phi$ derivable approach to that of a system where the hierarchy of higher order Green functions is built successively from the tower of all Greens functions starting with the fundamental one $G_1$. The open question is how to define the vertex functions joining the Greens functions.
Cluster virial expansion within the $\Phi$ derivable approach
=============================================================
Having introduced the notion of a cluster expansion of the $\Phi-$ functional we want to suggest a definition which eliminates the unknown vertex functions in favour of the $T_{A+B}-$matrix which describes the nonperturbative binary collisions of $A-$ and $B-$ particle correlations in the channel $A+B$, see Fig. \[fig:2\]. In this way we can draw the connection between the cluster virial expansion of Ref. [@Ropke:2012qv] with the $\Phi-$ derivable approach [@KB].
![The $\Phi-$ functional for the cluster virial expansion (left), where the $T_{A+B}-$ matrix for binary collisions in the channel with the partition $A, B$ replaces the higher order Green function $G_{A+B}$ and the corresponding vertex functions according to the scheme given in the right panel.[]{data-label="fig:2"}](Fig3){width="100.00000%"}
![The $\Phi-$ functional for the cluster virial expansion (left), where the $T_{A+B}-$ matrix for binary collisions in the channel with the partition $A, B$ replaces the higher order Green function $G_{A+B}$ and the corresponding vertex functions according to the scheme given in the right panel.[]{data-label="fig:2"}](Fig4new){width="100.00000%"}
Deuterons in nuclear matter
---------------------------
The application of this scheme to the simplest case of two-particle correlations in the deuteron channel in nuclear matter results in the selfenergy [@SRS] $$\label{cluster-SE}
\Sigma(1,z)=\sum_2\int\frac{d\omega}{2\pi} A(2,\omega)\bigg\{
f(\omega)V(12,12) -\int\frac{dE}{\pi} \Im T(12,12;E+i\eta) \frac{f(\omega)+g(E)}{E-z-\omega}\bigg\}~,$$ where $f(\omega)=[\exp(\omega/T)+1]^{-1}$ is the Fermi function and $g(\omega)=[\exp(\omega/T)-1]^{-1}$ the Bose function. The decomposition (\[cluster-SE\]) corresponds to a cluster decomposition of the nucleon density $$n(\mu,T)=n_{\rm free}(\mu,T) + 2n_{\rm corr}(\mu,T)~,$$ where the correlation density contains besides the bound state a scattering state contribution $$\label{ncont}
n_{\rm sc}=\int \frac{dE}{2\pi} g(E) 2 \sin^2\delta(E) \frac{d\delta(E)}{dE}~.$$ The one-particle density of free quasiparticle nucleons is reduced in order to fulfil the baryon number conservation in the presence of deuteron correlations and contains a selfenergy contribution due to the deuteron correlations in the medium. This improvement of the quasiparticle picture due to the correlated medium accounted for by the consistent definition of the selfenergy as a derivative of the $\Phi-$ functional (\[Phi-A\]) is the reason the continuum correlations (\[ncont\]) are reduced by the factor $2\sin^2 \delta$ as compared to the traditional Beth-Uhlenbeck formula [@Beth:1936zz; @Beth:1937zz]. For details, see [@Zimmermann:1985ji; @SRS]. With the definition of the $\Phi-$ functional via the T-matrix in Fig. \[fig:2\] we were able to show the correspondence between the generalized Beth-Uhlenbeck approach and the $\Phi-$ derivable approach for the nonrelativistic potential model approach to two-particle correlations in a warm, dense Fermion system [@SRS; @Ropke:2012qv]. Now we would like to discuss its application to a relativistic model for Mesons in quark matter.
Mesons in quark matter
----------------------
In order to describe the problem of mesons in quark matter within the $\Phi-$ derivable approach we define the $\Phi-$ functional and the corresponding selfenergy in Fig. \[fig:3\].
![The $\Phi-$ functional (left panel) for the case of mesons in quark matter, where the bosonic meson propagator is defined by the dashed line and the fermionic quark propagators are shown by the solid lines with arrows. The corresponding meson and quark selfenergies are shown in the middle and right panels, respectively.[]{data-label="fig:3"}](Fig5C "fig:"){width="23.00000%"} ![The $\Phi-$ functional (left panel) for the case of mesons in quark matter, where the bosonic meson propagator is defined by the dashed line and the fermionic quark propagators are shown by the solid lines with arrows. The corresponding meson and quark selfenergies are shown in the middle and right panels, respectively.[]{data-label="fig:3"}](Fig6I "fig:"){width="35.00000%"} ![The $\Phi-$ functional (left panel) for the case of mesons in quark matter, where the bosonic meson propagator is defined by the dashed line and the fermionic quark propagators are shown by the solid lines with arrows. The corresponding meson and quark selfenergies are shown in the middle and right panels, respectively.[]{data-label="fig:3"}](Fig6D "fig:"){width="35.00000%"}
The meson polarization loop $\Pi_M(q,z)$ in the middle panel of Fig. \[fig:3\] enters the definition of the meson T-matrix (often called propagator) $$T^{-1}_M(q,\omega+i\eta) = G_S^{-1} - \Pi_M(q,\omega+i\eta)
=|T_M(q,\omega)|^{-1}{\rm e}^{- i \delta_M(q,\omega)}~,$$ which in the polar representation introduces a phase shift $\delta_M(q,\omega)=\arctan (\Im T_M/\Re T_M)$, that results in a generalized Beth-Uhlenbeck equation of state for the thermodynamics of the consistently coupled quark-meson system [@Blaschke:2013zaa]. $$\Omega = \Omega_{\rm MF} + \Omega_M ~,$$ where the selfconsistent quark meanfield contribution is $$\Omega_{\rm MF} =
\frac{\sigma^2_{\rm MF}}{4G_S} - 2N_c N_f \int\frac{d^3p}{(2\pi)^3}
\left[ E_p %+ \frac{\Sigma_++\Sigma_-}{2}
+T \ln\left(1+{\rm e}^{-(E_p-\Sigma_+-\mu)/T} \right)
+T \ln\left(1+{\rm e}^{-(E_p+\Sigma_-+\mu)/T} \right)
\right]
~,$$ with the quasiparticle energy shift for quarks (antiquarks) due to mesonic correlations given by $\Sigma_\pm=\sum_{M=\pi,\sigma} {\rm tr}_D \left[ \Sigma_M\Lambda_\pm \gamma_0\right]/2$ and the positive (negative) energy projection operators $\Lambda_\pm=(1\pm \gamma_0)/2$. The mesonic contribution to the thermodynamics is $$\Omega_M = d_M \int \frac{d^3 k}{(2\pi)^3} \int\frac{d\omega}{2\pi}
\left\{
\omega + 2 T \ln \left[1 - {\rm e}^{-\omega/T} \right] 2\sin^2\delta_M(k,\omega) ~
\frac{\delta_M(k,\omega)}{d\omega}
\right\}~,$$ where similar to the case of deuterons in nuclear matter the factor $2\sin^2\delta_M$ accounts for the fact that mesonic correlations in the continuum are partly already accounted for by the selfenergies $\Sigma_M$ defining the improved selfconsistent quasiparticle picture. In the previous works of Refs. [@Hufner:1994ma; @Zhuang:1994dw; @Blaschke:2013zaa; @Yamazaki:2012ux; @Wergieluk] on this topic, however, the effect of the backreaction from mesonic correlations on the quark meanfield thermodynamics had been disregarded. Here we note from the $\Phi-$ derivable approach that for consistency the quark propagator in the quark meanfield thermodynamic potential shall contain effects from the selfenergy $\Sigma_M$ due to the coupling to the mesonic correlations as in the right panel of Fig. \[fig:3\]. This total quark selfenergy is the given by $\Sigma({\bf p},p_0)=\sigma_{\rm MF} + \Sigma_M({\bf p},p_0)$, where for a local NJL model with scalar coupling constant $G_S$ the meanfield contribution is $$\begin{aligned}
\sigma_{\rm MF}= 2 N_f N_c G_S \int \frac{d^3 p}{(2\pi)^3} \frac{m}{E_p}[1-f_-(E_p)-f_+(E_p)] ~,\end{aligned}$$ and the contribution due to scalar/pseudoscalar mesons (corresponding to the diagram shown in the rightmost panel of Fig. \[fig:3\]) is given by [@Kitazawa:2014sga] $$\begin{aligned}
\label{SigmaM}
\Sigma_M({\bf 0},p_0)= d_M\int \frac{d^4q}{(2\pi)^4}\pi \varrho_M({\bf q},q_0)
\left\{\frac{(\gamma_0 +m/E_q)[1+g(q_0)-f_-(E_q)]}{q_0-p_0+E_q-\mu-i\eta}
+\frac{(\gamma_0 -m/E_q)[g(q_0)+f_+(E_q)]}{q_0-p_0-E_q-\mu-i\eta}
\right\}~,
\nonumber\\\end{aligned}$$ where $\varrho_M=(-1/\pi)\Im T_M({\bf q}, \omega+i\eta)$ is the meson spectral density and $E_q=\sqrt{q^2+m^2}$ is the quark dispersion law with the quark mass $m=m_0+\sigma_{\rm MF}$. One can observe the similarity of this result (\[SigmaM\]) with that for a Dirac fermion coupled to a pointlike scalar meson, as given in [@Blaizot:1991kh].
Hadrons in quark matter
-----------------------
Finally, we would like to sketch how the $\Phi-$ derivable approach can be employed to define a cluster virial expansion for quark-hadron matter consisting of quarks (Q), mesons (M), diquarks (D) and baryons (B). The thermodynamical potential for this system obtains the form very similar to the case of clustered nuclear matter, i.e. $$\begin{aligned}
\Omega &=& \sum_{i=Q,M,D,B} (-1)^{c_i} \left[{\rm Tr} \ln \left(-G_i^{-1}\right)
+ {\rm Tr} \left(\Sigma_i~G_i \right) \right]+\Phi\left[G_Q,G_M,G_D,G_B\right]~,\end{aligned}$$ where $c_i=0$ ($c_i=1$) for bosonic (fermionic) states and the minimal set of two-particle irreducible diagrams defining the $\Phi-$ functional is given in Fig. \[fig:4\].
![The contributions to the $\Phi-$ functional for the quark-meson-diquark-baryon system.[]{data-label="fig:4"}](Fig5A "fig:"){width="19.00000%"} ![The contributions to the $\Phi-$ functional for the quark-meson-diquark-baryon system.[]{data-label="fig:4"}](Fig5B "fig:"){width="19.00000%"} ![The contributions to the $\Phi-$ functional for the quark-meson-diquark-baryon system.[]{data-label="fig:4"}](Fig5C "fig:"){width="19.00000%"} ![The contributions to the $\Phi-$ functional for the quark-meson-diquark-baryon system.[]{data-label="fig:4"}](Fig5D "fig:"){width="19.00000%"} ![The contributions to the $\Phi-$ functional for the quark-meson-diquark-baryon system.[]{data-label="fig:4"}](Fig5E "fig:"){width="19.00000%"}
From this $\Phi-$ functional follow the selfenergies defining the full Greens functions of the system by functional derivation $$\Sigma_i = \frac{\delta~ \Phi\left[G_Q,G_M,G_D,G_B\right]}{\delta~G_i }~.$$ The resulting Feynman diagrams for the selfenergy contributions are given in Fig. \[fig:5\].
![The selfenergy contributions for the Greens functions of the quark-meson-diquark-baryon system, defined by the $\Phi-$ functional contributions shown in Fig. 4. From top to down the four rows of diagrams show the selfenergies for the full propagators of mesons, quarks, diquarks and baryons, respectively. \[fig:5\]](Fig6I "fig:"){width="19.00000%"} ![The selfenergy contributions for the Greens functions of the quark-meson-diquark-baryon system, defined by the $\Phi-$ functional contributions shown in Fig. 4. From top to down the four rows of diagrams show the selfenergies for the full propagators of mesons, quarks, diquarks and baryons, respectively. \[fig:5\]](Fig6J "fig:"){width="19.00000%"} ![The selfenergy contributions for the Greens functions of the quark-meson-diquark-baryon system, defined by the $\Phi-$ functional contributions shown in Fig. 4. From top to down the four rows of diagrams show the selfenergies for the full propagators of mesons, quarks, diquarks and baryons, respectively. \[fig:5\]](Fig6K "fig:"){width="19.00000%"}\
![The selfenergy contributions for the Greens functions of the quark-meson-diquark-baryon system, defined by the $\Phi-$ functional contributions shown in Fig. 4. From top to down the four rows of diagrams show the selfenergies for the full propagators of mesons, quarks, diquarks and baryons, respectively. \[fig:5\]](Fig6D "fig:"){width="19.00000%"} ![The selfenergy contributions for the Greens functions of the quark-meson-diquark-baryon system, defined by the $\Phi-$ functional contributions shown in Fig. 4. From top to down the four rows of diagrams show the selfenergies for the full propagators of mesons, quarks, diquarks and baryons, respectively. \[fig:5\]](Fig6C "fig:"){width="19.00000%"} ![The selfenergy contributions for the Greens functions of the quark-meson-diquark-baryon system, defined by the $\Phi-$ functional contributions shown in Fig. 4. From top to down the four rows of diagrams show the selfenergies for the full propagators of mesons, quarks, diquarks and baryons, respectively. \[fig:5\]](Fig6E "fig:"){width="19.00000%"}\
![The selfenergy contributions for the Greens functions of the quark-meson-diquark-baryon system, defined by the $\Phi-$ functional contributions shown in Fig. 4. From top to down the four rows of diagrams show the selfenergies for the full propagators of mesons, quarks, diquarks and baryons, respectively. \[fig:5\]](Fig6F "fig:"){width="19.00000%"} ![The selfenergy contributions for the Greens functions of the quark-meson-diquark-baryon system, defined by the $\Phi-$ functional contributions shown in Fig. 4. From top to down the four rows of diagrams show the selfenergies for the full propagators of mesons, quarks, diquarks and baryons, respectively. \[fig:5\]](Fig6A "fig:"){width="19.00000%"} ![The selfenergy contributions for the Greens functions of the quark-meson-diquark-baryon system, defined by the $\Phi-$ functional contributions shown in Fig. 4. From top to down the four rows of diagrams show the selfenergies for the full propagators of mesons, quarks, diquarks and baryons, respectively. \[fig:5\]](Fig6B "fig:"){width="19.00000%"}\
![The selfenergy contributions for the Greens functions of the quark-meson-diquark-baryon system, defined by the $\Phi-$ functional contributions shown in Fig. 4. From top to down the four rows of diagrams show the selfenergies for the full propagators of mesons, quarks, diquarks and baryons, respectively. \[fig:5\]](Fig6G "fig:"){width="19.00000%"} ![The selfenergy contributions for the Greens functions of the quark-meson-diquark-baryon system, defined by the $\Phi-$ functional contributions shown in Fig. 4. From top to down the four rows of diagrams show the selfenergies for the full propagators of mesons, quarks, diquarks and baryons, respectively. \[fig:5\]](Fig6H "fig:"){width="19.00000%"}
Note that it is immediately plain from this formulation that in the situation of confinement, when the propagators belonging to colored excitations (quarks and diquarks) and thus to states that could not be populated, the system simplifies considerably. When all closed loop diagrams containing quarks and diquarks are neglected, this system reduces to a meson-baryon system as, e.g., described in selfconsistent relativistic meanfield theories of nuclear matter. On the other hand, in the case of deconfined quarks when also chiral symmetry is restored, the meson and baryon states become unbound (Mott effect) and their contribution to the thermodynamics as captured in the corresponding phase shift functions is gradually vanishing at high temperatures and chemical potentials with just chiral quark matter remaining asymptotically.
Conclusion
==========
In this contribution the $\Phi-$ derivable approach is generalized to describe the formation of bound states (clusters) of different size in many particle systems with strong interactions. A generic form of $\Phi-$ functionals is presented that turns out to be fully equivalent to a selfconsistent cluster virial expansion up to the second virial coefficient for interactions among the clusters. As examples are considered: nuclei in nuclear matter and hadrons in quark matter, with particular attention to the case of the deuterons in nuclear matter and mesons in quark matter. Generalized Beth-Uhlenbeck equations of state are derived, where the quasiparticle virial expansion is extended to include arbitrary clusters. The approach is applicable to nonrelativistic potential models of nuclear matter as well as to relativistic field theoretic models of quark matter. It is particularly suited for a description of cluster formation and dissociation in hot, dense matter.
Acknowledgements {#acknowledgements .unnumbered}
----------------
Many ideas discussed in this contribution have emerged from a long standing collaboration with H. Grigorian, T. Klähn, G. Röpke, S. Typel and H. Wolter who are gratefully acknowledged for insightful discussions. I thank N. T. Gevorgyan for excellent assistance in preparing this manuscript. This work was supported in part by the Polish National Science Center (NCN) under grant number UMO-2011/02/A/ST2/00306.
[99]{}
G. Baym, L. P. Kadanoff, Phys. Rev. [**124**]{}, 287 (1961); G. Baym, Phys. Rev. [**127**]{}, 1391 (1962).
R. Dashen, S. Ma, W. J. Bernstein, Phys. Rev. [**187**]{}, 345 (1969).
J.M. Cornwall, R. Jackiw, E. Tomboulis, Phys. Rev. D [**10**]{}, 2428 (1974).
G.M. Carneiro, C.J. Pethick, Phys. Rev. D [**11**]{}, 1106 (1975).
T.D. Lee, M. Margulies, Phys. Rev. D [**11**]{}, 1591 (1975).
E.M. Nyman, M. Rho, Nucl. Phys. A [**268**]{}, 408 (1976).
W. Weinhold, B. Friman, W. Nörenberg, Phys. Lett. B [**433**]{}, 236 (1998). G. Röpke, M. Schmidt, L. Münchow, and H. Schulz, Nucl. Phys. A [**379**]{}, 536 (1982); Nucl. Phys. A [**399**]{}, 587 (1983); Phys. Lett. [**B 110**]{}, 21 (1982).
R. Zimmermann, H. Stolz, physica status solidi (b) [**131**]{}, 151 (1985).
M. Schmidt, G. Röpke, and H. Schulz, Ann. Phys. [**202**]{}, 57 (1990).
G. Uhlenbeck, E. Beth, Physica [**3**]{}, 729 (1936).
E. Beth, G. Uhlenbeck, Physica [**4**]{}, 915 (1937).
S. Typel, G. Röpke, T. Klähn, D. Blaschke and H. H. Wolter, Phys. Rev. C [**81**]{}, 015803 (2010). G. Röpke, arXiv:1411.4593 \[nucl-th\]. G. Röpke, N.-U. Bastian, D. Blaschke, T. Klähn, S. Typel and H. H. Wolter, Nucl. Phys. A [**897**]{}, 70 (2013). J. P. Blaizot, J. Korean Phys. Soc. [**25**]{}, S65 (1992). M. Kitazawa, T. Kunihiro and Y. Nemoto, Phys. Rev. D [**90**]{}, 116008 (2014). J. Hüfner, S. P. Klevansky, P. Zhuang and H. Voss, Annals Phys. [**234**]{}, 225 (1994). P. Zhuang, J. Hüfner and S. P. Klevansky, Nucl. Phys. A [**576**]{}, 525 (1994). D. Blaschke, D. Zablocki, M. Buballa, A. Dubinin and G. Röpke, Annals Phys. [**348**]{}, 228 (2014). K. Yamazaki and T. Matsui, Nucl. Phys. A [**913**]{}, 19 (2013). A. Wergieluk, D. Blaschke, Y. L. Kalinovsky and A. Friesen, Phys. Part. Nucl. Lett. [**10**]{}, 660 (2013).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we study commuting families of holomorphic mappings in $\mathbb{C}^n$ which form abelian semigroups with respect to their real parameter. Linearization models for holomorphic mappings are been used in the spirit of Schröder’s classical functional equation.'
address:
- |
Dipartimento Di Matematica\
Università di Roma Tor Vergata \
Via Della Ricerca Scientifica 1, 00133\
Roma, Italy
- |
Department of Mathematics\
ORT Braude College\
21982 Karmiel, Israel
author:
- Filippo Bracci
- Mark Elin
- David Shoikhet
title: Normal forms and linearization of holomorphic dilation type semigroups in several variables
---
The one-dimensional linearization models for holomorphic mappings and semigroups, based on Schröder’s and Abel’s functional equation have been studied by many mathematicians for more than a century.
These models are powerful tools in investigations of asymptotic behavior of semigroups, geometric properties of holomorphic mappings and their applications to Markov’s stochastic branching processes.
It turns out that solvability as well as constructions of the solution of Schröder’s or Abel’s functional equations properly, depend on the location of the so-called Denjoy–Wolff point of the given mappings or semigroups. In particular, recently many efforts were directed to the study of semigroups with a boundary Denjoy–Wolff point [@BE-PH; @SD; @AM-92; @RS-SD-96].
Multidimensional cases are more delicate even when the Denjoy–Wolff point is inside of the underlined domain. It appears that the existence of the solution (the so-called K[œ]{}nigs’ function) of a multidimensional Schröder’s equation depends also on the resonant properties of the linear part of a given mapping (or generator), and its relation to homogeneous polynomials of higher degrees.
In parallel, the study of commuting mappings (or semigroups) is of interest to many mathematicians and goes back to the classical theory of linear operators, differential equations and evolution problems.
In this paper we consider, in particular, the rigidity property of two commuting semigroups. Namely, the question we study is whether those semigroups coincide whenever the linear parts of their generators at their common null point are the same.
Let $D$ be a domain in ${\mathbb C}^n$. We denote the set of holomorphic mappings on $D$ which take values in a set $\Omega\subset{\mathbb C}^m$ by ${\mathop{\rm Hol}\nolimits}(D,\Omega)$. For each $f\in{\mathop{\rm Hol}\nolimits}(D,{\mathbb C}^m)$, the Frechét derivative of $f$ at a point $z\in D$ (which is understood as a linear operator acting from ${\mathbb C}^n$ to ${\mathbb C}^m$ or $n\times
m$-matrix) will be denoted by $df_z$.
For brevity, we write ${\mathop{\rm Hol}\nolimits}(D)$ for ${\mathop{\rm Hol}\nolimits}(D,D)$. The set ${\mathop{\rm Hol}\nolimits}(D)$ is a semigroup with respect to composition operation.
A family $\mathcal{S}=\{\v_t\}_{t\ge0}\subset{\mathop{\rm Hol}\nolimits}(D)$ of holomorphic self-mappings of $D$ is called a [*one-parameter continuous semigroup*]{} if the following conditions are satisfied:
\(i) $\v_{t+s}=\v_t\circ\v_s$ for all $s,\,t\ge0$;
\(ii) $\lim\limits_{t\to0^+}\v_t(z)=z$ for all $z\in D$.
It is more or less known that condition (ii) (the right continuity of a semigroup at zero) actually implies its continuity (right and left) on all of ${\mathbb R}^{+}=[0,\infty)$. Moreover, in this case the semigroup is differentiable on ${\mathbb R}^{+}$ with respect to the parameter $t\ge0$ (see [@BE-PH; @SD; @AM-92; @RS-SD-96]). Thus, for each $z\in D$ there exists the limit $$\label{gen}
\lim_{t\rightarrow 0^{+}}\frac{\v_t(z)-z}t=f(z),$$ which belongs to ${\mathop{\rm Hol}\nolimits}(D,{\mathbb C}^n)$. The mapping $f\in{\mathop{\rm Hol}\nolimits}(D,{\mathbb C}^n)$ defined by (\[gen\]) is called the [*(infinitesimal) generator*]{} of $\mathcal{S}=\left\{\v_t\right\}_{t\ge0}$.
Furthermore, the semigroup $\mathcal{S}$ can be defined as a (unique) solution of the Cauchy problem: $$\label{cauchy}
\left\{
\begin{array}{l}
{\displaystyle\frac{\partial \v_t(z)}{\partial
t}}=f(\v_t(z)),\quad t\ge0, \vspace{3mm}\\ \v_0(z)=z,\quad z\in D.
\end{array}
\right.$$
We say that a semigroup $\left\{\v_t\right\}_{t\ge0}$ is [*linearizable*]{} if there is a biholomorphic mapping $h\in{\mathop{\rm Hol}\nolimits}(D,{\mathbb C}^n)$ and a linear semigroup $\left\{\psi_t\right\}_{t\ge0}$ such that $\left\{\v_t\right\}_{t\ge0}$ conjugates with $\left\{\psi_t\right\}_{t\ge0}$ by $h$, namely, $h\circ\v_t=\psi_t\circ h$ for all $t\ge0$.
Linearization methods for semigroups on the open unit disk in ${\mathbb C}\left(={\mathbb C}^1\right)$ have been studied by many mathematicians (see, for example, [@Sis; @SD-07; @E-S-Y]). At the same time, little is known about multi-dimensional cases. For example, in [@Fab] and [@E-S-2004] the problem has been studied for some special class of the so-called one-dimensional type semigroups.
In this paper, we will concentrate on the case when a semigroup has a (unique) interior attractive fixed point, i.e., $\lim\limits_{t\to\infty}\v_t(z)=\tau\in D\subset{\mathbb C}^n$ for all $z\in D$. It is well known that this condition is equivalent to that fact that the spectrum $\sigma(A)$ of the linear operator (matrix) $A$ defined by $A:=df_\tau$ lies in the open left half-plane (see [@AM-88] and [@RS-SD-96]) and $d(\v_t)_\tau=e^{At}$. Usually, such semigroups are named [*of dilation type*]{}. Thus, for the one-dimensional case, it is possible to linearize the semigroup by solving Schröder’s functional equation: $$h\left(\v_t(z)\right)=e^{f'(\tau)t}h(z)$$ (see, for example, [@Sis; @SD]).
It should be noted that the latter equation involves the eigenvalue problem for the linear semigroup $\left\{C_t\right\}_{t\ge0}$ of composition operators on the space ${\mathop{\rm Hol}\nolimits}(D,{\mathbb C})$ defined by $C_t:\,h\mapsto h\circ\v_t$.
It is easy to show that the solvability of a higher dimensional analog of Schröder’s functional equation $$\label{schroder}
h\left(\v_t(z)\right)=e^{At}h(z),\quad A=df_\tau,$$ is equivalent to a generalized differential equation: $$\label{dif-schroder}
dh_z f(z)=A h(z).$$
It seems that in general useful criteria (necessary and sufficient conditions) for solvability of (\[dif-schroder\]) are unknown.
Without loss of generality, let us assume that $\tau=0$.
\[prop1\] Equation (\[schroder\]), or equivalently, (\[dif-schroder\]) is solvable if and only if there is a polynomial mapping $Q:\,{\mathbb C}^n\mapsto{\mathbb C}^n$ with $Q(O)=O$ and $dQ_O=\id$, such that the limit $$\lim_{t\to\infty}e^{-At}Q(\v_t(z))=:h(z),\quad z\in D,$$ exists.
This proposition is based on the following notation and lemma.
By $\lambda(A)$ we denote the spectrum distortion index of the matrix $A$, i.e., $$\lambda(A):=\frac{\max\limits_{\alpha\in\sigma(A)}|\Re
\alpha|}{\min\limits_{\alpha\in\sigma(A)}|\Re \alpha|}\,.$$
\[lem1\] Let $g\in{\mathop{\rm Hol}\nolimits}(D,{\mathbb C}^n)$ admit the expansion: $g(z)=\sum\limits_{\ell\ge m}Q_\ell(z)$, where $Q_\ell$ is a homogenous polynomial of order $\ell$ and $m>\lambda(A)$. Then $$\lim_{t\to\infty}e^{-At}g(\v_t(z))=O,\quad\mbox{for all }\ z\in D.$$
In many cases (and always — in the one dimensional case), a polynomial $Q$ in Proposition \[prop1\] can be chosen to be the identity mapping, $Q(z)=z$ for all $z$. Moreover, in this case $h\left(\v_t(z)\right)=e^{At}h(z)$, i.e., the mapping $h(z)=\lim\limits_{t\to\infty}e^{At}\v(z)$ forms a conjugation of a given semigroup $\left\{\v_t\right\}_{t\ge0}$ with the linear semigroup $\left\{e^{At}\right\}_{t\ge0}$.
Let $\mathcal{S}=\{\v_t\}_{t\ge0}$ be a continuous one-parameter semigroup of holomorphic self-mappings on a domain $D\subset{\mathbb C}^n$. We say that $\mathcal{S}$ is [*normally linearizable*]{} if the limit $$h(z)=\lim_{t\to\infty}e^{-At}\v_t(z),\quad z\in D,$$ exists.
A consequence of Lemma \[lem1\] is the following assertion.
\[prop-normal1\] Let $\mathcal{S}=\{\v_t\}_{t\ge0}$ be a one-parameter semigroup of holomorphic self-mappings on a domain $D\subset{\mathbb C}^n$ generated by $f\in{\mathop{\rm Hol}\nolimits}(D,{\mathbb C}^n)$. If $f$ admits the expansion on the series of homogenous polynomials: $f(z)=Az+\sum\limits_{\ell\ge m}Q_\ell(z)$, where $Q_\ell$ is a homogenous polynomial of order $\ell$ and $m>\lambda(A)$, then the semigroup $\mathcal{S}$ is normally linearizable.
In contrast with the one-dimensional case, for $n>1$ there are semigroups which are not normally linearizable.
\[ex2\] Let $\{\v_t\}_{t\ge0}$ be a semigroup in $\mathbb{C}^2$ defined by $$\begin{aligned}
\v_t(z_1,z_2)=\left( \begin{array}{c} z_1\exp\left(-(1+i)t\right) \vspace{2mm}\\
\Bigl[a{z_1}^2i\left(e^{-it}-1\right)+z_2\Bigr] e^{-(2+i)t}
\end{array}\right).\end{aligned}$$
It is easy to see that $$\lim_{t\to\infty}e^{-At}\v_t(z)=\lim_{t\to\infty}\left(\begin{array}{c}
z_1 \vspace{2mm}\\ a{z_1}^2i(\exp(-it)-1)+z_2 \end{array}\right)$$ does not exist. Thus, this semigroup is not normally linearizable.
Just differentiating $\v_t$ at $t=0^+$ we find the semigroup generator: $$\begin{aligned}
f(z_1,z_2)=\left( \begin{array}{c} -(1+i)z_1 \vspace{2mm}\\
-(2+i)z_2+a{z_1}^2 \end{array}\right).\end{aligned}$$ For this generator we have $\lambda(A)=m=2$, i.e., $f$ does not satisfy the conditions of Proposition \[prop-normal1\].
\[prop-normal2\] Let $D\subset {\mathbb C}^n$ be a domain containing $O$. Let $\left\{\v_t\right\}$ be a continuous dilation semigroup which is normally linearizable. If for some $t_0>0$ the semigroup element $\v_{t_0}$ is a linear map, then all the elements $\v_t,\ t\ge0,$ are linear.
Denote $h(z):=\lim\limits_{t\to\infty}e^{-At}\v_t(z)$. Then for all $s>0$ obviously $$h(\v_s(z)):=e^{As}\lim\limits_{t\to\infty}e^{-A(t+s)}\v_t(\v_s(z))=e^{As}h(z),$$ i.e., $h$ is a linearizing conjugation for $\left\{\v_t\right\}_{t\ge0}$. Since $\v_{t_0}=e^{At_0}$, we have $\v_{t_0n}=e^{At_0n}$ and $$h(z):=\lim_{n\to\infty}e^{-At_0n}\v_{t_0n}(z)=z,$$ so $h$ is the identity mapping. Therefore, $\v_s(z)=
h^{-1}\left(e^{As}h(z)\right)=e^{As}z$ for all $s\ge0$.
Example \[ex2\] above shows that this fact is not generally true. Indeed, for each $t_{\ell}=2\pi\ell,\ \ell\in\mathbb{Z},$ the semigroup element $\v_{t_{\ell}}$ is a linear mapping. Yet all other elements $\v_t,\ t\not=2\pi\ell,$ are not linear.
An additional problem is that that with exception of the one-dimensional case, linearizing conjugations may not be unique.
Let $\mathcal{F}=\left\{\v_s\right\}_{s\in\mathcal{A}}$ be a family of holomorphic self-mappings of $D$. We say that $\mathcal{F}$ is [*uniquely linearizable*]{} if there is a unique mapping $h$ biholomorphic in $D$ and normalized by $h(O)=O,\
dh_O=\id$, such that $$h\circ\v_s =B_s\circ h ,\quad s\in\mathcal{A},$$ where $\left\{B_s\right\}_{s\in\mathcal{A}}$ is an appropriate family of linear operators on ${\mathbb C}^n$.
Actually, it follows by the chain rule that $B_s=d(\v_s)_O$.
A family $\mathcal{F}$ may consist of a single mapping $F\in{\mathop{\rm Hol}\nolimits}(D)$ as well as a discrete or continuous semigroup of holomorphic self-mappings on $D$.
Our next example shows that even linear diagonal mappings may not be uniquely linearizable.
\[ex3\] Consider a linear mapping $\psi=(\psi_1,\psi_2)$ with $$\begin{aligned}
\psi_1(z_1,z_2)= \frac{z_1}2, \qquad \psi_2(z_1,z_2)= \frac{z_2}4\end{aligned}$$ and a holomorphic normalized mapping defined by $$h(z_1,z_2)= \left(\begin{array}{c} z_1\vspace{2mm} \\
{z_1}^2+z_2 \end{array}\right).$$ Then $h\circ\psi=\psi\circ h$, i.e., $h$ and also the identity mapping $\id$ linearize $\psi$.
Actually, the question whether a linear mapping $\psi(z)=Bz$ is uniquely linearizable can be formulated as the following rigidity problem:
[*When do the conditions $$Q\circ B=B\circ Q\quad\mbox{and} \quad Q'(O)=O$$ on a holomorphic mapping $Q$ imply that $Q\equiv O\ $?*]{}
In fact, it can be seen that if a matrix $B$ is diagonalazable and $\sigma(B)=\{\beta_1,\ldots,\beta_n\}\subset\Delta$, then $\psi$ is uniquely linearizable if and only if ${\beta_1}^{k_1}\cdot{\beta_2}^{k_2}\cdot\ldots\cdot{\beta_n}^{k_n}\neq\beta_j$ for all $j=1,\ldots,n$ and $k\in{\mathbb N}^n$.
\[commut-th\] Let $D\subset {\mathbb C}^n$ be a domain containing $O$. Let $\mathcal{S}=\left\{\v_t\right\}_{t\ge0}$ be a continuous semigroup of dilation type, and let $\psi$ be a holomorphic self-mapping of $D$ commuting with $\mathcal{S}$ such that $$\label{commut}
\psi\circ\v_t=\v_t\circ\psi$$ for all $t\ge0$. If $\psi$ is uniquely linearizable by a biholomorphic mapping $h:\,D\mapsto{\mathbb C}^n$, then all of the elements of the semigroup $\mathcal{S}$ are linearizable by the same mapping $h$.
Let $B$ denote a linear operator on ${\mathbb C}^n$ defined by $B=d\psi_O$. Also we denote $A=df_O$, where $f$ is the infinitesimal generator of the semigroup $\mathcal{S}$. First, by differentiating (\[commut\]) at $O$ we obtain $\left(d\psi_O\right)\circ e^{At}=e^{At}\circ d\psi_O$, i.e., $B$ commutes with the linear semigroup $\left\{e^{At}\right\}_{t\ge0}\
$ (in fact, $B$ commutes with $A$).
By our assumption, $h\circ\psi=B\circ h$. Therefore, for all $t\ge0$ we have $$h\circ\psi\circ\v_t=B\circ h\circ\v_t.$$ On the other hand, $h\circ\psi\circ\v_t= h\circ\v_t\circ\psi$ by (\[commut\]). Thus, $$e^{-At}\circ h\circ\v_t\circ\psi=e^{-At}\circ B\circ h\circ\v_t=
B\circ e^{-At}\circ h\circ\v_t.$$
Denoting $h_1:=e^{-At}\circ h\circ\v_t$ one rewrites the latter equality in the form $$h_1\circ\psi = B\circ h_1 .$$ Since $h_1(O)=O,\ d(h_1)_O=\id$ and $\psi$ is uniquely linearizable by $h$, we conclude that $h_1=e^{-At}\circ
h\circ\v_t=h$, or $$h\circ\v_t=e^{At}\circ h.$$ The proof is complete.
\[cor4\] Let $D\subset {\mathbb C}^n$ be a domain containing $O$. Let $\mathcal{S}=\left\{\v_t\right\}_{t\ge0}$ be a continuous semigroup of dilation type. If there exists $t_0>0$ such that $\v_{t_0}$ is uniquely linearizable by a biholomorphic mapping $h:\,D\mapsto{\mathbb C}^n$, then all the elements of $\mathcal{S}$ are linearizable by the same mapping $h$ which is a unique solution of the differential equation (\[dif-schroder\]) $$dh_z f(z)=A h(z),$$ normalized by the conditions $h(O)=O,\ dh_O=\id$.
Let $D\subset {\mathbb C}^n$ be a domain containing $O$. Let $\mathcal{S}_1=\left\{\v_t\right\}_{t\ge0}$ and $\mathcal{S}_2=\left\{\psi_t\right\}_{t\ge0}$ be two continuous semigroups on $D$ generated by mappings $f_1$ and $f_2$, respectively. Suppose that $d(f_1)_O=d(f_2)_O=A$ with $\Re\sigma(A)<0$ and that there exists $s_0>0$ such that
\(i) $\psi_{s_0}$ is uniquely linearizable and
\(ii) $\psi_{s_0}$ commutes with the semigroup $\mathcal{S}_1$ such that $\psi_{s_0}\circ\v_t=\v_t\circ\psi_{s_0}$ for all $t\ge0$.
Then the semigroups coincide.
By our assumption, there is a unique biholomorphic mapping $h$ normalized by $h(O)=O,\ dh_O=\id$, such that $$h\circ\psi_{s_0} =e^{As_0}\circ h .$$ Then Theorem \[commut-th\] (or Corollary \[cor4\]) implies that $h\circ\psi_s =e^{As}\circ h$ for all $s\ge0$. Since the mapping $h$ is biholomorphic, we have: $$\psi_s=h^{-1}\circ\left(e^{As}\circ h\right).$$
The commutativity of the mapping $\psi_{s_0}$ and the semigroup $\mathcal{S}_1$ implies by the same Theorem \[commut-th\] that all of the elements of $\mathcal{S}_1$ are linearizable by the mapping $h$, that is, $h\circ\v_t =e^{At}\circ h$ for all $t\ge0$. Thus $$\v_t=h^{-1}\circ\left(e^{At}\circ h\right).$$
If the semigroups $\mathcal{S}_1=\left\{\v_t\right\}_{t\ge0}$ and $\mathcal{S}_2=\left\{\psi_t\right\}_{t\ge0}$ commute in the sense: $\v_t\circ\psi_s=\psi_s\circ\v_t$ for all $t,s\ge0$, then the conclusion that they coincide holds under a formally weaker than condition (i) requirement that differential equation (\[dif-schroder\]) has a unique solution normalized by ${h(O)=O,}\ dh_O=\id$.
Let $D\subset {\mathbb C}^n$ be a domain containing $O$. Let $\mathcal{S}_1=\left\{\v_t\right\}_{t\ge0}$ and $\mathcal{S}_2=\left\{\psi_t\right\}_{t\ge0}$ be two commuting semigroups on $D$ generated by mappings $f_1$ and $f_2$, respectively. Suppose that $d(f_1)_O=d(f_2)_O=A$ with $\Re\sigma(A)<0$. If $\lambda(A)<2$ then the semigroups coincide.
The use of the Poincaré–Dulac theorem (see, for example, [@Arn]) is another approach to solve a linearization problem.
For simplicity, we assume in the sequel that $A$ is a diagonal matrix, $A=\diag(\alpha_1\ldots,\alpha_n)$ with $\Re\alpha_n\le\ldots\le\Re\alpha_1<0$.
Let $k:=(k_1,\ldots, k_n)\in {\mathbb N}^n$ be such that $|k|:=\sum k_j\ge
2$.
\[resonant\] We say that $A$ is [*resonant*]{} (or the $n$-tuple $(\alpha_1,\ldots\alpha_n)$ of the eigenvalues of $A$ is resonant) if for some $\ell=1,\ldots, n$ $$(\alpha,k):=\sum_{j=1}^n k_j\alpha_j= \alpha_\ell.$$ Such a relation is called a [*resonance*]{}. The number $|k|$ is called the [*order*]{} of the resonance.
If $\alpha_\ell=(\alpha,k)$, we call any map $G:{\mathbb C}^n\mapsto{\mathbb C}^n$ [*resonant monomial*]{} if it has the form $G(z)=(g_1(z),\ldots,
g_n(z))$ with $g_j\equiv 0$ for $j\neq \ell$ and $g_\ell(z)=a
z^k$.
\[resonances\] If $\,\Re\alpha_n\le\ldots\le\Re\alpha_1<0$ then there is at most a finite number of resonances for $\alpha$. Moreover, if $\alpha_j=(k,\alpha)$ then $k_j=\ldots=k_n=0$.
Both statements follow from the simple observation that if $\alpha_j=(k,\alpha)$, then $\Re \alpha_j=(k,\Re \alpha)$, and by the ordering of $\alpha_j$.
For simplicity of notation, let $$M_j:=\left\{
\begin{array}{l}
0, \quad \mbox{if there is no } k \mbox{ with } \alpha_j=(k,\alpha),\vspace{2mm}\\
\max\{|k|: \alpha_j=(\alpha,k)\}\quad \mbox{otherwise}.
\end{array}\right.$$ and $M(\alpha):=\max\{M_j:j=1,\ldots, n\}.$
A vector polynomial map $R:{\mathbb C}^n\mapsto {\mathbb C}^n,\ R(O)=O,$ is [*triangular*]{} if by switching coordinates $R(z)=(R_1(z),\ldots,
R_n(z))$ assumes the form $$R_j(z)=a_j z_j + r_j(z_1,\ldots, z_{j-1}),\quad j=1,\ldots, n$$ where $r_j$ is a polynomial.
Let $D\subset {\mathbb C}^n$ be a domain containing $O$. Let $\{\v_t\}_{t\ge0}$ be a continuous dilation type semigroup generated by $f\in{\mathop{\rm Hol}\nolimits}(D,{\mathbb C}^n)$ with $df_O=A$. Then there exists an injective holomorphic map $h:\,D\mapsto {\mathbb C}^n$ (independent of $\ t$) such that $h(O)=O$, $dh_O=\id$ and $$h \circ \v_t = P_t\circ h,$$ where $P_t(z)=e^{At}z+R_t(z)$ is a triangular polynomial group of automorphisms of ${\mathbb C}^n$ whose degree is less than or equal to $M(\alpha)$, and $R_t(z)$ containing only resonant monomials. In particular, if there are no resonances then $\{\v_t\}_{t\ge0}$ is linearizable.
Let $\v_t(z)=e^{At}z+\sum\limits_{|m|\ge2} P_{m,t}(z)$ be the homogeneous expansion at $O$ (which is defined on a small ball containing $O$ and contained in $D$). It follows from the theory of semigroups of holomorphic maps that each $P_{m,t}(z)$ is real analytic in $t$.
By our assumption, $A$ is diagonal and the convex hull in ${\mathbb C}$ of its eigenvalues does not contain $0$. Therefore by the classical Poincaré–Dulac theorem, there exist an open neighborhood $U$ of $O$ and a holomorphic map $h:\,U\mapsto {\mathbb C}^n$ normalized by $h(O)=O$ and $dh_O=\id$ such that $dh_{z}(f(z))=\widehat{f}(h(z))$, where $\widehat{f}(z)=Az+T(z)$ with $T$ being a polynomial vector field containing only resonant monomials.
The semigroup $\{\v_t\}_{t\ge0}$ is (locally around $O$) conjugated to the semigroup $\{\psi_t\}_{t\ge0},\ \psi_t=h \circ
\v_t\circ h^{-1}$, generated by $\widehat{f}=A+T$. Since $T$ contains only resonant monomials and $\Re \alpha_n\leq \ldots\leq
\Re \alpha_1<0$, Lemma \[resonances\] implies that $\widehat{f}$ is triangular, i.e., $\{\psi_t\}_{t\ge0}$ satisfies the following system: $$\begin{cases}
\stackrel{\cdot}{x_1}=\alpha_1 x_1\\
\stackrel{\cdot}{x_2}=\alpha_2 x_2+r_2(x_1)\\\ldots\\
\stackrel{\cdot}{x_n}=\alpha_n x_n+r_n(x_1,x_2,\ldots, x_{n-1}),
\end{cases}$$ where the $r_j$’s are polynomials in $x_1,\ldots, x_{j-1}$ containing only resonant monomials. Such a system can be integrated directly by first solving $\stackrel{\cdot}{x_1}=\alpha_1 x_1$, then substituting such solution into $\stackrel{\cdot}{x_2}=\alpha_2 x_2+r_2(x_1)$, and so on. In the end, $\psi_t$ is of the form $$\psi_t(z)=(e^{t\alpha_1}z_1, e^{t\alpha_2}(z_2+R_{2,t}(z_1)),
\ldots, e^{t\alpha_n}(z_n+R_{n,t}(z_1,z_2,\ldots, z_{n-1}))),$$ with $R_{j,t}$ a polynomial in $z_1,\ldots, z_{j-1}$ of (at most) degree $M_j$ containing with only resonant monomials. Moreover, $R_{j,t}$ depends also polynomially on $t$. It can be shown by induction. It is true for $j=1$, so assume it is true for $j-1$. Then the $l$-th component of $(\psi_t)$ for $l=1,\ldots, j-1$ is of the form $\psi_{t,l}(z)=
e^{t\alpha_l}(z_l+R_{l,t}(z_1,z_2,\ldots, z_{l-1}))$ with $R_{l,t}$ a polynomial in $z_1,\ldots, z_{l-1}$ of degree at most $M_l$ and depending polynomially on $t$. Substituting these into the differential equation $\stackrel{\cdot}{x_j}=\alpha_j
x_j+r_j(x_1,x_2,\ldots, x_{j-1})$, one obtains $$\begin{aligned}
\stackrel{\cdot}{x_j}=\alpha_j
x_j+r_j(e^{\alpha_1t}z_1,e^{t\alpha_2}(z_2+R_{2,t}(z_1)),\ldots\\
\ldots, e^{t\alpha_{j-1}}(z_{j-1}+R_{j-1,t}(z_1,z_2,\ldots,
z_{j-2}))).\end{aligned}$$ Therefore the solution is of the form $e^{\alpha_j t} g(t)$ for some function $g$ such that $g(0)=z_j$ and $$\stackrel{\cdot}{g}(t)=e^{-\alpha_j t}r_j(x_1,x_2,\ldots,
x_{j-1}).$$ Now, $r_j$ contains only resonant monomials for $\alpha_j$. Let $z^m$ be such a resonant monomial. Then, taking into account that $m_j=\ldots=m_n=0$ by Lemma \[resonances\], it follows $$z^m=a z_1^{m_1}\cdots
z_{j-1}^{m_{j-1}}=e^{(m,\alpha)t}[z_1^{m_1}\cdots (z_{j-1}+R_{j-1,t}(z_1,z_2,\ldots,
z_{j-2}))^{m_{j-1}}].$$ Hence $$\stackrel{\cdot}{g}(t)=e^{(-\alpha_j+(m,\alpha))
t}[z_1^{m_1}\cdots (z_{j-1}+R_{j-1,t}(z_1,z_2,\ldots,
z_{j-2}))^{m_{j-1}}],$$ and, being $\alpha_j=(m,\alpha)$, then actually $$\stackrel{\cdot}{g}(t)= z_1^{m_1}\cdots
(z_{j-1}+R_{j-1,t}(z_1,z_2,\ldots,
z_{j-2}))^{m_{j-1}}.$$ Since this holds for all resonant monomials in $r_j$, this proves that $R_{t,j}(z)$ is a polynomial in both $z_1,\ldots,
z_{j-1}$ and $t$. The degree of $R_{t,j}$ is at most $M_j$ because it contains only resonant monomials for $\alpha_j$. This proves the induction and the claim about the $R_{j,t}$’s.
This fact implies that $\psi_{-t}(z)$ is well defined for all $t\geq 0$ and $z\in {\mathbb C}^n$. Therefore, $\{\psi_t\}_{t\in{\mathbb R}}$ is a group of polynomial automorphisms of ${\mathbb C}^n$.
Finally, since $O$ is an attracting fixed point by hypothesis, then $h$ can be extended to all $D$ by imposing $h(w)=\psi_{-t}(h(\v_t(w)))$ for all $w\in D$.
\[ex1\] For $n=2$ there is only one possible resonance, namely, $\alpha_2=m\alpha_1$. Hence, up to conjugation, the dilation semigroups in ${\mathbb C}^2$ are of the form: $$\v_t(z)=(e^{\alpha_1 t}z_1, e^{\alpha_2 t}(z_2+atz_1^m))$$ for some $a\in {\mathbb C}$.
So, if the matrix $A=df_\tau$ is resonant, it may happen that all elements of the semigroup generated by $f$ are not linearizable. In this connection the following question arises naturally. Suppose that one of the elements of the semigroup $\mathcal{S}=\{\v_t\}_{t\ge0}$ (say, $\v_{t_0}$) is linearizable. Find conditions which ensure that all other elements $\v_t,\
t\not=t_0,$ are linearizable too.
To answer this question we need the following notion.
\[star\] We say that the matrix $A=\diag(\alpha_1\ldots\alpha_n)$ has [*pure real resonance*]{} if there are $j=1,\ldots,n$ and $k\in{\mathbb N}^n$ such that $\Re\alpha_j=\Re(\alpha,k)$ but $\alpha_j\not=(\alpha,k)$.
In particular, if all eigenvalues $\alpha_j$ have the same argument, then $A$ has not pure real resonance.
\[linear\] Let $D\subset {\mathbb C}^n$ be a domain containing $O$. Let $\{\v_t\}_{t\ge0}$ be a continuous dilation semigroup generated by $f\in{\mathop{\rm Hol}\nolimits}(D,{\mathbb C}^n)$ with $df_O=A$, where $A$ has not pure real resonance. If there exists $t_0>0$ such that $\v_{t_0}$ is linearizable by biholomorphic mapping $h:D\mapsto {\mathbb C}^n,\ h(O)=O$. Then the semigroup $\{\v_t\}_{t\ge0}$ is linearizable by $h$.
Not that even for the non-resonant case Theorem 3 completes Theorem 2 since it asserts the following fact: if $h\in{\mathop{\rm Hol}\nolimits}(D,
{\mathbb C}^n)$ is a linearizing mapping for $\v_{t_0}$, it also can serve as a linearizing mapping for all $\v_t,\ t\ge0$.
Let us define $\psi_t:=h\circ \v_t \circ h^{-1}$. Then $\psi_t$ is a semigroup on $h(D)$.
Let $\psi_t(z)=e^{At}z+\sum_m P_{m,t}(z)$ be the homogeneous expansion at $O$ (which is defined on a small ball containing $O$ and contained in $g(D)$), where $m\geq 2$ is the least positive integer such that $P_{m,t}\not\equiv 0$ for all $t,z$. If the theorem holds then $m=+\infty$ (namely, $(\psi_t)$ is linear). Seeking a contradiction, we assume that $m<+\infty$.
It follows from the theory of semigroups of holomorphic maps that each $P_{m,t}(z)$ is real analytic in $t$.
Since by hypothesis $\psi_{t_0}=h\circ \v_{t_0} \circ h^{-1}$ is linear, then $P_{m,t_0}\equiv 0$.
Now, from $\psi_{t+s}=\psi_t\circ \psi_s$ it follows that $$\label{uno}
P_{m,t+s}(z)=e^{At}P_{m,s}(z)+P_{m,t}(e^{As}z).$$ Write $P_{m,t}(z)=(\sum_{|k|=m}p^1_k(t)z^k,\ldots,
\sum_{|k|=m}p^n_k(t)z^k)$, where, as usual, $z^k=z_1^{k_1}\cdots z_n^{k_n}$. From it follows that for $j=1,\ldots, n$ $$p^j_{k}(t+s)=e^{\alpha_j t}p_k^j(s)+p_k^j(t) e^{(\alpha,k)s}.$$ Differentiating such an expression with respect to $t$ and setting $t=0$, we obtain the following differential equation: $$\label{diff1}
\frac{d}{dt}p^j_{k}(s)=\alpha_jp_k^j(s)+a_k^je^{(\alpha,k)s},$$ where we set $a_k^j=\frac{dp_k^j(t)}{dt}|_{t=0}$. There are two cases:
\(1) if $\Re\alpha_j\not=\Re(\alpha,k)$, then imposing the condition $p_k^j(0)=0$, equation has the solution $$\label{due}
p^j_k(t)=a^j_k \frac{e^{(\alpha,k)t}-e^{\alpha_j
t}}{(\alpha,h)-\alpha_j}.$$
\(2) if $\Re\alpha_j=\Re(\alpha,k)$, then by our assumption $\alpha_j=(\alpha,k)$. In this case, imposing the condition $p_k^j(0)=0$, equation has the solution $$\label{due-res}
p^j_k(t)=a^j_k e^{t\alpha_j}t.$$
By and it follows that $p_k^j(t_0)=0$ if and only if $p_k^j(t)=0$ for all $t\geq 0$, and hence $P_{m,t_0}\equiv 0$ if and only if $P_{m,t}\equiv 0$ for all $t\geq 0$, reaching a contradiction with our hypothesis.
Example \[ex2\] above shows that if $A$ has pure real resonance, Theorem \[linear\] fails.
\[cor1\] Let $\mathcal{S}=\{\v_t\}_{t\ge0}$ be a continuous semigroup of dilation type generated by $f\in{\mathop{\rm Hol}\nolimits}(D,{\mathbb C}^n)$ with $df_O=A=\diag(\alpha_1,\ldots,\alpha_n)$. Suppose that there is $t_0>0$ such that $\v_{t_0}$ is a linear mapping. Assume that one of the following conditions holds:
\(i) $A$ has not pure real resonance;
\(ii) $e^{(\alpha,k)t_0}\not=e^{\alpha_jt_0}$ for all $j=1,\ldots,n$ and $k\in{\mathbb N}^n$.
Then all elements of $\mathcal{S}$ are linear mappings.
If condition (i) holds, the assertion follows immediately by Theorem \[linear\].
Assume that condition (ii) holds. First, we show that $\v_{t_0}$ is uniquely linearizable. Indeed, let $h(z)=z+\ldots$ be a linearizing mapping different from $\id$. This means that $h\circ\v_{t_0}=\v_{t_0}\circ h$ and for some $j=1,\ldots,n,\,$ the $j$-th coordinate of $h$ contains a non-zero monomial $a_k{z_1}^{k_1}\ldots{z_n}^{k_n}$ with $|k|\ge2.$ Therefore, $$h_j\left(e^{\alpha_1t_0}z_1,\ldots,e^{\alpha_nt_0}\right)
=e^{\alpha_jt_0}h_j(z),$$ and so $$a_ke^{(\alpha,k)t_0}z^k=a_ke^{\alpha_jt_0}z^k.$$ The contradiction provides that $\v_{t_0}$ is uniquely linearizable by the identity mapping $\id$.
Now, Corollary \[cor4\] implies that the all mappings $\v_t,\
t\ge0,$ are linearizable by the identity mapping. Hence, they are linear.
Combining Corollary \[cor1\] with Proposition \[prop-normal2\], we get the following result.
Let ${\mathbb B}^n$ be the unit ball of ${\mathbb C}^n$ and let $\mathcal{S}=\{\v_t\}_{t\ge0}$ be a continuous semigroup of dilation type generated by $f\in{\mathop{\rm Hol}\nolimits}({\mathbb B},{\mathbb C}^n)$ with $df_O=A=\diag(\alpha_1,\ldots,\alpha_n)$. Suppose that there is $t_0>0$ such that $\v_{t_0}$ is a linear fractional self-mapping of ${\mathbb B}^n$. Assume that one of the following conditions holds:
\(i) $A$ has not pure real resonance;
\(ii) $\mathcal{S}$ is normally linearizable;
\(iii) $e^{(\alpha,k)t_0}\not=e^{\alpha_jt_0}$ for all $j=1,\ldots,n$ and $k\in{\mathbb N}^n$.
Then for all $t\geq 0$ the mapping $\v_{t}$ is a linear fractional self-map of ${\mathbb B}^n$.
According to [@B-C-DM Thm. 3.2 and Rmk. 3.4] (and its proof) there exists $h:\,{\mathbb B}^n\mapsto{\mathbb C}^n$ a linear fractional mapping fixing $O$ such that $h\circ \v_{t_0} \circ h^{-1}$ is linear. By Corollary \[cor1\] and Proposition \[prop-normal2\], it follows that $h\circ \v_{t} \circ h^{-1}$ is linear for all $t\ge0$. Therefore, $\v_t$ is the composition of linear fractional maps and hence linear fractional for all $t\ge0$.
Let $\mathcal{S}=\{\v_t\}_{t\ge0}$ be a continuous semigroup of dilation type generated by $f\in{\mathop{\rm Hol}\nolimits}({\mathbb B},{\mathbb C}^n),\
f(z)=Az+\sum\limits_{\ell\ge m}Q_\ell(z)$, where $Q_\ell$ is a homogenous polynomial of order $\ell$ and $m>\lambda(A)$. If for some $t_0>0$, the semigroup element $\v_{t_0}$ is a linear (respectively, linear fractional) mapping, then all the elements of $\mathcal{S}$ are linear (respectively, linear fractional) mappings.
A direct consequence of our Theorems 2 and 3 and a recent Forelli type extension theorem (see [@K-P-S Theorem 6.2]) is the following assertion.
Let $\mathcal{S}=\{\v_t\}_{t\ge0}$ be a continuous semigroup of dilation type generated by $f\in{\mathop{\rm Hol}\nolimits}(D,{\mathbb C}^n)$ with $df_O=A=\diag(\alpha_1,\ldots,\alpha_n)$, where all eigenvalues $\alpha_j$ have the same argument. Suppose that a function $F$ defined on $D$ is real analytic at $O$, and that its restrictions to the integral curves of the vector field $f$ are holomorphic. If at least one of the following conditions holds:
\(i) $A$ is not resonant,
or
\(ii) there is $t_0$ such that $\v_{t_0}$ is linearizable,
then $F$ is holomorphic on $D$.
[ABT2]{}
M. Abate, Converging semigroups of holomorphic maps, [*Atti Accad. Naz. Lincei Rend. Cl. Sci. Fis. Mat. Natur.*]{} [**82**]{} (1988), 223–227.
M. Abate, The infinitesimal generators of semigroups of holomorphic maps, [*Ann. Mat. Pura Appl.*]{} **161** (1992), 167–180.
V. I. Arnold, [Geometrical Methods in the Theory of Ordinary Differential Equations]{}, Springer, NY, 1988.
E. Berkson and H. Porta, Semigroups of analytic functions and composition operators, [*Michigan Math. J.*]{} [**25**]{} (1978), 101–115.
F. Bracci, M. D. Contreras, S. Diaz-Madrigal, [*Classification of semigroups of linear fractional maps in the unit ball*]{}. Adv. Math., [**208**]{} (2007), 318–350.
M. Elin, S. Reich and D. Shoikhet, Complex Dynamical Systems and the Geometry of Domains in Banach Spaces, [*Dissertationes Math. (Rozprawy Mat.)*]{} [**427**]{} (2004), 62 pp.
M. Elin and D. Shoikhet, Semigroups with boundary fixed points on the unit Hilbert ball and spirallike mappings, in: [*Geometric Function Theory in Several Complex Variables*]{}, 82–117, World Sci. Publishing, River Edge, NJ, 2004.
M. Elin, D. Shoikhet, F. Yacobzon, Linearization models for parabolic type semigroups, [*J. Nonlinear Convex Anal.*]{} 9 (2008), 205–214.
C. de Fabritiis, On the linearization of a class of semigroups on the unit ball of ${\mathbb C}^n$, [*Ann. Mat. Pura Appl.*]{} [**166**]{} (1994), 363–379.
K.-T. Kim, E. Poletsky, G. Schmalz, Functions holomorphic along holomorphic vector fields, in arXiv: math.CV/0811.1093v1, 2008.
S. Reich and D. Shoikhet, Generation theory for semigroups of holomorphic mappings in Banach spaces, [*Abstr. Appl. Anal.*]{} [**1**]{} (1996), 1–44.
D. Shoikhet, [*Semigroups in Geometrical Function Theory*]{}, Kluwer, Dordrecht, 2001.
D. Shoikhet, Koenigs type linearization models and asymptotic behavior of one-parameter semigroups, [*Contemporary Mathematics. Fundamental Directions*]{} [**21**]{} (2007), 149–166
A. Siskais, Semigroups of composition operators on spaces of analytic functions, a review. [*Studies on composition operators (Laramie, WY, 1996)*]{}, 229–252, [*Contemp. Math.*]{} [**213**]{} [*Amer. Math. Soc., Providence, RI*]{}, 1998.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Disks in the 6 Myr old cluster $\eta$ Chamaeleontis were searched for emission from hot [H$_2$ ]{}. Around the M3 star [ECHA J0843.4-7905 ]{}we detect circumstellar gas orbiting at $\sim$2 AU. If the gas is UV-excited, the ro-vibrational line traces a hot gas layer supported by a disk of mass $\sim$0.03[$~{\rm M}_{\odot}$ ]{}, similar to the minimum mass solar nebula. Such a gas reservoir at 6 Myr would promote the formation and inwards migration of gas giant planets.'
author:
- |
S.K. Ramsay Howat$^{1}$[^1] and J.S. Greaves$^{2}$[^2]\
$^{1}$ UK Astronomy Technology Centre, Blackford Hill, Edinburgh EH9 3HJ\
$^{2}$School of Physics and Astronomy, University of St. Andrews, North Haugh, St Andrews, Fife, KY16 9SS, UK.
date: 'Accepted; Received; in original form ...'
title: Molecular Hydrogen emission from disks in the eta Chamaeleontis cluster
---
\[firstpage\]
circumstellar matter, infrared:stars, planets:formation,
Introduction
============
Understanding the longevity of protoplanetary disks around stars, and their dust and gas content, is an important step in understanding the formation of planetary systems. The exploration of the dust content of disks has been carried out from near-infrared to millimetre wavelengths, probing from hot disks near the stellar surface out to hundreds of AU. The dust mass and temperature can be deduced from the spectral energy distribution, although shortwards of the far-infrared the dust emission is optically thick and so detailed models are needed to understand the distribution of material [@wood02]. Observations of carbon monoxide in disks are used to estimate the mass of molecular gas, however this is subject to a number of uncertainties. The CO may be photodissociated; it freezes out on grains and is only a small amount of the total gas. The H$_2$/CO ratio is highly variable due to these effects, and CO only effectively traces layers of the disk above and below the mid-plane. Thus a direct tracer of the bulk of the gas is to be preferred.
Molecular hydrogen is the principal gas component in disks and so direct observations of H$_2$ have been undertaken by several groups, using the rotational-vibrational transitions observable in the near- and mid-infrared. Mid-infrared observations of the pure rotational lines offer the most direct measure of the disk mass, as the lines are emitted by low temperature gas ($\sim$100 K). The line intensity should give a direct measure of the disk mass for an optically thin disk. To date, detection of the pure rotational lines has proved difficult. Thi and co-workers reported detections of copious quantities of [H$_2$ ]{}from a number of different types of source (debris disks, Herbig AeBe stars and T Tauri’s) using ISO SWS (@thi2001Naturea, @thi2001bb). However, attempts to replicate these detections in more sensitive observations with ground-based instruments (@richter02, @sheret03) have not detected the lines. The most likely explanation for this is that the H$_2$ detected by ISO was elsewhere in the large beam of the spectrometer, rather than in the circumstellar disk. Both Richter et al. and Sheret et al. explain their non-detections by pointing out that the disks are optically thick at mid-infrared wavelengths and that the [H$_2$ ]{}may be therefore be self-shielded. A ground based detection of the v=0-0 rotational lines from the disk around AB Aurigae has been reported by @bitner07, who detect the v=0-0 S(1), S(2) and S(4) lines from gas at T$\sim$670 K in a disk of mass 0.5[$~{\rm M}_{\odot}$ ]{}. Observations with Spitzer are proving effective for probing the pure rotational lines of H$_2$ in the circumstellar environment. @lahuis07 have detected the v=0-0 S(2) and v=0-0 S(3) lines of H$_2$ towards 6 of a total of 76 T Tauri and Herbig AeBe sources in their sample using the Spitzer InfraRed Spectrograph. These spectra show the presence of H$_2$ gas at T$\geq$500 K.
H$_2$ in disks may also be excited to higher vibration levels and emit in the near-infrared. The principle excitation mechanisms are by collisions in warm, dense gas or due to the passage of a shock, by absorption of a UV photon and subsequent radiative decay or by X-rays. All of these mechanisms may be in action in a circumstellar disk. The first detection of emission from the v=1-0 S(1) line at 2.1218$~\mu$m was been reported by @wkb00 in TW Hya. Following this, the line has been detected by @bwk03 and @bwk05 for a selection of T Tauri stars and by by @itoh03 for LkH$\alpha$ 264. In these sources, the authors conclude that the line emission arises from a quiescent disk. The line width of the [H$_2$ ]{}emission from LkH$\alpha$ 264 suggests emission from a radius of 1AU from the parent star. For the sources observed by Bary, Weintraub & Kastner (2003), the measured widths of the lines are from $9-14~{\rm km~s}^{-1}$ and the H$_2$ molecules are orbiting from 10–30 AU. The detection of excited [H$_2$ ]{}at these radii offers the possibility of a large reservoir of cold [H$_2$ ]{}associated with the zone of giant planet formation.
Estimating the total H$_2$ mass is a complex task, as each of the possible excitation mechanisms may be at work in a particular disk. Work so far [@bwk03] suggests that the v=1-0 S(1) line is tracing only a thin hot zone at the top surfaces of the gas disk, and so is not ideal as a bulk mass tracer. However, the presence of this line is very useful as an indicator that gas is still present in a disk; as a diagnostic of the orbital radius (with high resolution spectroscopy); and for comparison to dust excesses from the inner disk. These data may address the vital question of whether a gas supply still exists at a few Myr, when giant planet cores have grown massive enough to begin to accrete their thick atmospheres.
The discovery of a new young star cluster, $\eta$ Chamaeleontis at a distance of just 97pc by @mlf99 presents an exciting opportunity for studying disk formation at close range. The discovery was made in a deep ROSAT image. @lyo03 found that an unusually high proportion of the stars have an L-band excess indicative of the presence of a dust disk. For their derived age of the stars in the cluster (9 Myr), such a high disk fraction (60 %$\pm$13 %) implied that disk lifetimes may be higher than previously thought. @hll01 measured the disk fraction in clusters of different ages, finding a linear relationship between cluster age and disk fraction. From this, the disk fraction should drop below 60 % for an age of 2-3 Myr. Recent work on the $\eta$ Chamaeleontis cluster has challenged the idea that it has an unusually high disk fraction for the age of the cluster. The age has been revised downwards slightly to 6 (–1, +2) Myr [@ls04] and new L-band data [@hja05] cast doubt on some of the earlier disk excesses. The @hja05 results reduce the disk fraction to 17 %$\pm$11 %, which is consistent with the decline in disk fraction with age measured by @hll01.Longer wavelength observations with the IRAC camera on SPITZER confirm that two stars from the cluster (RECX11 and [ECHA J0843.4-7905 ]{}) show excesses in the $K_s$-\[3.6\] colour, closest to the K-L colour determined by @lyo04, but that more dust disks are revealed from the \[3.6\]-\[4.5\] versus \[5.8\]-\[8.0\] colour-colour diagram [@megeath05]. In total, @megeath05 find infared excesses from 6/15 of the stars in $\eta$ Cham.
Motivated by these discoveries and the successes of Bary et al. in detecting excited H$_2$, we elected to attempt detections of the v=1-0 S(1) line of H$_2$ in the seven $\eta$ Cham stars identified by Lyo et al. as having infrared excesses using high spectral resolution infrared spectroscopy. Four of these sources (RECX5, RECX11, RECX9 and [ECHA J0843.4-7905 ]{}) are identified as having disks by @megeath05. The presence of significant quantities of gas in these 5–8 Myr-old disks would extend the possible timescales for the formation of gas giant planets.
We report here on the search for H$_2$ in these disks and discuss the connection of gas to the dust content, excitation mechanisms and age of the stars.
Observations and Data Reduction
===============================
The observations were carried out on 25 February 2005 using the Phoenix high resolution infrared spectrograph on the southern Gemini telescope at Cerro Pachon, Chile. The observations were carried out in conditions of thin cirrus; the seeing, as measured from the full width half maxima of the stars, varied between 0.5 arcsec and 0.8 arcsec during the night. Phoenix is a single order spectrograph containing a 1024x1024 InSb Aladdin II array [@phx03]. With the K4748 filter and the grating centred at the wavelength of the H$_2$ 1-0 S(1) line (2.1218 $\mu$m) the wavelength coverage obtained was from 2.11787 $\mu$m to 2.12703 $\mu$m. The slit width selected was 0.35arcsec, matched to four detector pixels. The spectral resolving power close to the wavelength of the H$_2$ line was measured to be 62,400, using a telluric OH line at 2.121774 $\mu$m. This corresponds to a wavelength range per pixel of 0.34$\times10^{-4}~\mu$m. Each source was acquired through the Phoenix imaging channel. Observations of the science targets were ratioed by the spectrum of the bright star Beta Car (V=1.68, A2IV) to correct for atmospheric features. Three observations of the star were made throughout the night. Wavelength calibration was obtained using the six brightest telluric absorption features present in the spectrum. Using this calibration, the wavelengths of the two brightest OH emission lines present are obtained to better than 0.15$\times10^{-4}~\mu$m. We take this as a measure of the accuracy of the wavelength calibration. A flat-field frame was obtained from observations of a Quartz Halogen lamp in the Gemini Facility Calibration Unit [@gcal]. The flux calibration for the individual observations was carried out using the known K-band magnitude of the sources themselves (see Table 1).
The on-chip exposure time for the observations of the targets was 240 seconds, with exposures on the sky obtained by offsetting the source 5 arcsec along the 14arsec slit. Multiple 240 second exposures were taken and coadded to produce the total on-source observation time listed in Table 1.
------------------ ---------------- ------------- ----------- ------------------- ------------------
Name RA (J2000)[^3] Dec (J2000) Magnitude Spectral type[^4] Total on-source
observation time
RECX3 8 43 37.2 -79 03 31 K=9.61 M3.25 48 mins
RECX5 8 42 27.30 -78 57 48 K=9.96 M4 48 mins
RECX6 8 42 39 -78 54 43 K=9.46 M3 48 mins
RECX9 8 44 16.6 -78 59 34.4 K=9.50 M4.5 48 mins
RECX11 8 47 01.64 -78 59 34.4 K=7.71 K5.5 32 mins
RECX12 8 47 56.77 -78 54 53.1 K=8.51 M3.25 32 mins
ECHAJ0843.3-7905 8 41 30.6 -78 53 07 K=9.45 M3.25 48 mins
------------------ ---------------- ------------- ----------- ------------------- ------------------
{width="120mm"}
{width="120mm"}
{width="120mm"}
------------------ ---------------------------- ---------------------------------- -------------------------- -- -- -- -- --
Name Line Flux Hot H$_2$ mass Total disk mass
$\times10^{-18}$ Wm$^{-2}$ 10$^{-8}$[$~{\rm M}_{\odot}$ ]{} [$~{\rm M}_{\odot}$ ]{}
RECX3 $<$1.8 $<$0.36 $<$0.04–4
RECX5 $<$1.2 $<$0.24 $<$0.024–2.4
RECX6 $<$1.3 $<$0.26 $<$0.03–3
RECX9 $<$1.5 $<$0.30 $<$0.03–3
RECX11 $<$2.4 $<$0.48 $<$0.05–5
RECX12 $<$2.1 $<$0.42 $<$0.04–4
ECHAJ0843.3-7905 $2.5\pm0.1$ 0.5 0.05–5
GG Tau $6.9\pm0.5$ 2.8 0.28–28
DoAr21 $15\pm9$ 8.1 0.81–81
LkCa15 $1.7\pm0.2$ 0.7 0.07–7
TW Hya $1.0\pm0.1$ 0.06 6.4$\times 10^{-3}$–0.64
------------------ ---------------------------- ---------------------------------- -------------------------- -- -- -- -- --
------------------ -------------------------- ------------ ----------------- ----------------- --------------- --------------------- -- --
Name Total disk mass (K-L) \[3.6\]-\[4.5\] \[5.8\]-\[8.0\] logL$_x$ cluster or star age
[$~{\rm M}_{\odot}$ ]{} ergs s$^{-1}$ Myr
RECX3 $<$0.04–4 0.14 [^5] 0.06 [^6] -0.06 29.1 [^7] 6
RECX5 $<$0.024–2.4 0.32 0.09 0.48 29.0 6
RECX6 $<$0.03–3 0.19 0.06 0.0 29.5 6
RECX9 $<$0.06–6 0.39 0.19 0.6 28.4 6
RECX11 $<$0.1–10 0.66 0.23 0.6 30.1 6
RECX12 $<$0.04–4 0.39 0.04 0.03 30.1 6
ECHAJ0843.3-7905 0.05-5 1.32 0.47 0.91 $<$28.5 [^8] 6
GG Tau 0.28–28 0.87 [^9] 29.4 [^10] 0.8 [^11]
DoAr21 0.81–81 0.4 [^12] 31.2 1.6
LkCa15 0.07–7 0.57 8
TW Hya 6.4$\times 10^{-3}$–0.64 0.14 [^13] 30.3 10
------------------ -------------------------- ------------ ----------------- ----------------- --------------- --------------------- -- --
Results and Analysis
====================
Molecular hydrogen has been detected here in one of the seven sources observed in the $\eta$ Chamaeleontis cluster. The spectra of all the observed sources are shown in Figures 1–3. The H$_2$ v=1-0 S(1) line is detected in only ECHAJ0843.3-7905, one of the sources known to have a circumstellar disk [@megeath05]. The spectral types for the stars range from K5.5 to M4.5 and are listed in Table 1. The error on the spectral type is $\pm$0.25 subclasses for the M-stars and $\pm$0.5 subclasses for the K stars [@ls04].
After removal of the telluric features using Beta Car, the observed stellar continuum from the targets is seen to contain numerous absorption features that vary with spectral type. To obtain the best estimate of the continuum level, and hence the optimum measurement of the H$_2$ line, ECHAJ0843.3-7905 (spectral type M3.25) was ratioed by the spectrum of stars of the same spectral type (within the errors in spectral type). There are three such stars: RECX3 (M3.25), RECX6 (M3) and RECX12 (M3.25). The RECX3 spectrum has poor signal/noise and was not used for the ongoing analysis. The RECX12 and RECX6 spectra were ratioed by Beta Car, as above, and then normalised to the flux level at $2.1218~\mu$m to provide a template of the M3 stellar absorption features.
The [ECHA J0843.4-7905 ]{}spectrum was ratioed by the RECX6 and RECX12 templates separately, providing independent estimates of the [H$_2$ ]{}spectrum. Excellent cancellation of the stellar absorption features is obtained in this way and comparing the two [ECHA J0843.4-7905 ]{}spectra provides an estimate of the systematic errors. The spectrum ratioed by the RECX12 template is presented in Figure 4.
In both spectra, the H$_2$ line is well fitted by a single Gaussian. For the spectrum ratioed by RECX6 the line width is $1.40\pm0.08\times10^{-4}~\mu$m centred at $2.121894\pm0.03\times10^{-4}~\mu$m and, in the spectrum divided by RECX12, $1.24\pm0.06\times10^{-4}~\mu$m centred at $2.121903\pm0.02\times10^{-4}~\mu$m. Deconvolving the line widths with that measured for the unresolved telluric OH lines ($0.35\times10^{-4}~\mu$m) we obtain a line width in the range $1.35-1.18\times10^{-4}~\mu$m corresponding to a velocity of $18\pm1.2~{\rm km~s}^{-1}$. The measured flux in the line is $(2.5\pm0.1)\times10^{-18}{\rm~Wm}^{-2}$.
The presence of [H$_2$ ]{}in the cirumstellar environment may be attributed to molecular gas excited in a disk or in an outflow. For example, @takami04 concluded that excited [H$_2$ ]{}detected in the vicinity of DG Tauri arises from an outflow within 100AU of the star. This conclusion is based on a significant blueshift of the line relative to the systemic velocity of the star and to a spatial offset of the emission from the star. Moreover, DG Tau is known to have an energetic outflow. Similarly, @deming04 detected [H$_2$ ]{}v=1-0 S(1) emission from a bipolar outflow associated with KH 15D.
We have considered the possibility that the [H$_2$ ]{}emission is from an outflow associated with [ECHA J0843.4-7905 ]{}. We fit Gaussian functions to the the spatial profile of the star along the spectrograph slit at the wavelength of the [H$_2$ ]{}line and the adjacent continuum. Subtracting the continuum from the line emission, and fitting a Gaussian to the results we find that the peak of the line is centred on the same pixel as the peak of the continuum (to within 0.5 pixels or 0.04 arcsec) and there is no evidence for residual [H$_2$ ]{}emission in the wings of the line. At the distance of the $\eta$ Chamaeleontis cluster (97 pc), 0.04 arcsec corresponds to a scale of $\sim$4 AU. The seeing during the observations of [ECHA J0843.4-7905 ]{}was 0.7 arcsec (70 AU). Since emission from outflows is observed at these radii, this is consistent with either an outflow or disk origin of the [H$_2$ ]{}emission. Adopting the radial velocity of the cluster from @mlf99 for [ECHA J0843.4-7905 ]{}(+16 km s$^{-1}$) and correcting for the relative motion of the Earth on the night in question, there is no evidence for relative motion of the [H$_2$ ]{}and the star (to $\sim 1{\rm~km~s}^{-1}$). In addition, the luminosity of the [H$_2$ ]{}line (7.2$\times
10^{-7}$[$~{\rm L}_{\odot}$ ]{}) is 20–1000 times fainter than the [H$_2$ ]{}luminosity seen in a range of outflows from Class I and Class II sources [@davis01] or from DG Tau, though we note that there are many factors, including age of the source, that will affect the luminosity.
We take these diagnostics to mean that it is more likely that the [H$_2$ ]{}emission from [ECHA J0843.4-7905 ]{}arises in a disk and pursue this interpretation in the following section.
{width="120mm"}
Discussion
==========
Assuming Keplerian rotation around a star of the mass of [ECHA J0843.4-7905 ]{}(taken as 0.4[$~{\rm M}_{\odot}$ ]{}), the observed 1-0 S(1) line width implies that the gas is emitted from a region around a few AU (e.g. 2 AU for a disk inclination of 45 degrees). The [ECHA J0843.4-7905 ]{}line profile and width is similar to those observed in the other 4 disk detections, with the exception of LkCa 15 in which a clear double peaked signature is seen [@bwk03]. For comparison, the L-band excesses arise at smaller radii of $\leq 0.1$ AU [@hja05]. The correlation suggests that both trace properties of the inner disk. The prevalence of single peaked line profiles in [H$_2$ ]{}is perhaps suprising, given that a Keplerian rotation pattern tends to produce a double-peaked line profile (if the disk is moderately edge-on and optically thin in the gas line). A line broadening mechanism may be at work in which case, a narrower Keplerian velocity would imply that the hot gas is at $>$2 AU. The broader line width may also be due to the contribution of emission from slower moving gas at larger radii from the star. @nomura05 show the [H$_2$ ]{}emission in the 1-0 S(1) line can extend over 10 AU with flux greater than 0.1 of the peak flux. Thus a contribution for an extended disk is consistent with this model and with a single peaked line.
The v=1-0 S(1) traces regions of hot gas, comprising a small fraction of the total [H$_2$ ]{}reservoir of the disk. Estimating this bulk mass by extrapolation from the [H$_2$ ]{}line strength is difficult, and would require detailed modelling of the line excitation mechanism(s) to obtain a robust estimate. We follow previous authors in using the empirical scaling mechanism derived by @bwk03 to estimate the disk mass. Calculating the mass of hot H$_2$ from the line strength, we get 0.5$\times10^{-8}$[$~{\rm M}_{\odot}$ ]{}for ECHA J0843.3-7905. This calculation assumes that the excited gas is optically thin, that ro-vibrational levels of [H$_2$ ]{}are populated as for a gas in thermal equilibrium at 1500K and that the hot [H$_2$ ]{}is located at 10 AU to 30 AU. Using the scaling factors given in @bwk03 gives a mass in the range 0.05 to 5[$~{\rm M}_{\odot}$ ]{}for the [ECHA J0843.4-7905 ]{}disk. The estimated disk masses for our objects are listed along with those of the other systems with [H$_2$ ]{}detections in Table 2. A disk of the mass implied by the upper range of this scale would be very suprising for the end of the planet formation timescale and we consider it to be unfeasible. Indications of the likely disk masses may be judged from observations by e.g. Rodriguez et al. (2005) who measure a disk of 0.3-0.4[$~{\rm M}_{\odot}$ ]{}for the 0.8[$~{\rm M}_{\odot}$ ]{}Class 0 star IRAS16293-2422B and Andrews & Williams (2005) who obtain a mass of 0.5[$~{\rm M}_{\odot}$ ]{}for the Class I source L1551-IRS5, the latter being the largest disk observed to date. The assumptions inherent in our calculation will tend to produce an overestimate of the disk mass. At radii closer to the star, the gas will be hotter, the mass fraction in the excited state relatively higher and so the total mass reduced.
Although the mass estimate is very uncertain, important conclusions can be reached from the detection of gas in this and other systems. A moderate mass of H$_2$ must still persist around [ECHA J0843.4-7905 ]{}at an age of 6 Myr. It seems very unlikely that a hot component could exist unsupported by a large reservoir of cool gas in the bulk of the disk volume. The result for ECHA J0843.3-7905, in conjunction with H$_2$ 1-0 S(1) detections towards LkCa 15 (8 Myr) and TW Hya (10 Myr), indicates that a gas reservoir does persist to the ages when gas giant planets are presumed to form. For example, in the models of @hbl05, Jupiter and Saturn accreted their envelopes after 2–5 Myr. The mass of hot [H$_2$ ]{}against age for all the published [H$_2$ ]{}detections is shown in Figure 5. With the small number of objects so far studied in [H$_2$ ]{}, and the wide range of parameters that affects the observability of excited [H$_2$ ]{}, we can say only that there is a tendency for the older objects to contain a reduced mass of hot gas. Considering the sources cited, we also sought correlations between the detection of [H$_2$ ]{}line emission and various other star or disk parameters (listed in Table 3). In particular, [ECHA J0843.4-7905 ]{}is the star in $\eta$ Cham with the largest excess at both near and mid-infrared wavelengths, indicating a possible correlation with the mass of hot gas. However, we find no correlation between the mass of hot [H$_2$ ]{}and L-band excess. Since excitation of the [H$_2$ ]{}may be by UV radation or X-radiation, we also considered these but again found no correlation. In the following section, we discuss these two excitation mechanisms.
{width="120mm"}
Comparison to models
--------------------
Detailed models of H$_2$ excitation in protoplanetary disks have been published by @nomura05, @nomura07. They calculate the gas and dust temperature and density profiles self-consistently and predict line fluxes for H$_2$ in a UV-excited scenario, for comparison with the TW Hya disk. Their model disk contains 0.006 M$_{\odot}$ of gas and dust and has an outer radius of 100 AU. It may not directly apply to ECHA J0843.3-7905 but some interesting comparisons may be made.
@nomura05 model the UV and IR emission from [H$_2$ ]{}assuming that the UV radiation field is described either by the blackbody temperature of the star or, as with many T Tauri stars, that there is an excess of UV emission characterised by bremsstrahlung emission at higher temperature than the effective temperature of the star and emission from Ly $\alpha$. The latter model is appropriate for TW Hydrae. The ‘UV excess’ model predicts line strengths of $\sim10^3$ those of the model without a UV excess. Scaling the predicted 1-0 S(1) flux for TW Hydrae (at 56 pc) to the distance of [ECHA J0843.4-7905 ]{}(97 pc), the ‘UV excess’ model gives 1.14$\times10^{-18}\rm{W~m}^{-2}$, around five times lower than observed from [ECHA J0843.4-7905 ]{}. Making the simplifying assumption that the [H$_2$ ]{}flux would scale with disk mass, a disk of 0.03[$~{\rm M}_{\odot}$ ]{}would produce the observed [ECHA J0843.4-7905 ]{}flux, approximately consistent with the lowest mass derived for the [ECHA J0843.4-7905 ]{}disk by empirical methods. We therefore consider that the observed flux is consistent with UV excitation of the 0.03–0.05[$~{\rm M}_{\odot}$ ]{}disk by a star with significant UV excess radiation. The authors were unable to obtain any information on the UV spectrum of [ECHA J0843.4-7905 ]{}. One anomaly, compared with the model, is that the 1-0 S(1) emission is expected to peak at around 20AU, rather than around $\sim$2 AU. Since the radial dependence of the observable flux in the line depends on the disk geometry, we expect the difference to be attributable to the differences in disk mass.
@mht96 treat the excitation of [H$_2$ ]{}due to irradiation by X-rays. [ECHA J0843.4-7905 ]{}is not detected by ROSAT in the survey of the $\eta$ Chamaelontis cluster by @mlf99 giving an upper limit of log(L$_x$)=28.5 ergs s$^{-1}$ [@mlf00]. Therefore we can predict only an upper limit for the 1-0 S(1) emission due to X-ray excitation. Based on the upper limit for log (L$_x$), the rate of X-ray energy deposition at the radius of the emitting gas ($\sim$2 AU) and assuming an H$_2$ column density of 10$^{22}{\rm~cm}^{-3}$ is 2.5$\times
10^{-23}{\rm ergs~s}^{-1}$. Given a distance to the source of 97pc, assuming a gas density of 10$^5{\rm~cm}^{-3}$ and an emitting region stretching from 2 AU to 22 AU, we expect the emission in the 1-0 S(1) to be 2$\times 10^{-19}{\rm~W~m}^{-2}$ [@mht96]. Given the large emitting annulus that we have taken and the X-ray upper limit, this will most likely be an overestimate and yet is siginificantly smaller than the observed flux. Thus we conclude that X-ray excitation is not an important mechanism in this source.
Conclusions
===========
We have presented a survey of the $\eta$ Chamaelontis cluster in the v=1-0 S(1) line of [H$_2$ ]{}. Of four stars surveyed that are currently believed to have circumstellar disks, [H$_2$ ]{}was detected in a single source ([ECHA J0843.4-7905 ]{}). The line kinematics and flux have been interpreted as arising in a disk, illuminated by UV radiation from the central star. This result, combined with those in other clusters, shows that a significant amount of molecular gas is available for the formation of giant planets on timescales of 6-10 Myr. The gas around [ECHA J0843.4-7905 ]{}is located at $\sim$2 AU, or beyond if the line width is not all Keplerian. The total mass of gas is estimated at 0.03-0.05[$~{\rm M}_{\odot}$ ]{}, which is slightly greater than the Minimum Mass Solar Nebula (MMSN) of 0.02[$~{\rm M}_{\odot}$ ]{}.
This gas disk meets many of the conditions thought to be necessary for the formation of giant planets, including long duration, high surface density and suitable size. For example, the gas reservoir is still substantial at 6 Myr, so a thick gas atmosphere could be added to a Jupiter-analogue core that has taken $\geq 2$ Myr to form [@hbl05]. Further, the surface density is comparable to the conditions in which Jupiter formed: for a standard MMSN gas-plus-dust profile of $\Sigma = 1700
r_{AU}^{-3/2}$ g cm$^{-2}$ [@d05], $\Sigma$(5 AU) is 150 g cm$^{-2}$, while for [ECHA J0843.4-7905 ]{}, a gas disk of 0.03-0.05[$~{\rm M}_{\odot}$ ]{}exceeds an average $\Sigma =
200$ g cm$^{-2}$ even if spread over 20 AU radius. The gas also lies at suitable distances where giant planets should start to form, in particular in relation to the ’snowline’ outside which $\Sigma$ of solid particles is favourably enhanced because volatiles freeze out into icy mantles on grains. @lpsc06 show that the snowline in an MMSN disk lies at around 2 AU, and further in for a low stellar accretion rate such as the $10^{-9}$[$~{\rm M}_{\odot}$ ]{}per year of [ECHA J0843.4-7905 ]{}[@llm04]. The Keplerian estimate shows that the [ECHA J0843.4-7905 ]{}disk extends to at least 2 AU, and so beyond the snowline. Finally, in these conditions the Type I migration rate for an Earth-mass proto-planet core is long: the model of @mg04 suggest a migration timescale $> 2 \times
10^7$ years, so a planet in the [ECHA J0843.4-7905 ]{}disk is unlikely to fall into the star. Thus this star is a likely site for giant planet formation, but the three other stars surveyed without gas detections are less promising; this result is roughly in line with observations of main-sequence Sun-like stars, where in long-term monitoring programs $\sim 15$ % host gas giants [@fischer03].
Acknowledgments {#acknowledgments .unnumbered}
===============
This paper is based on observations obtained at the Gemini Observatory for proposal GS-2005B-C-15. The Gemini Observatory is operated by the Association of Universities for Research in Astronomy, Inc., under a cooperative agreement with the NSF on behalf of the Gemini partnership: the National Science Foundation (United States), the Particle Physics and Astronomy Research Council (United Kingdom), the National Research Council (Canada), CONICYT (Chile), the Australian Research Council (Australia), CNPq (Brazil) and CONICET (Argentina). Phoenix was designed and built at the National Optical Astronomical Observatories and is made available to Gemini as a visiting instrument. The authors are pleased to acknowledge the excellent support provided during the observing run by Verne Smith (NOAO), Claudia Winge (Gemini) and Sybil Adams (Gemini). We are also grateful to the anonymous referee and to David Weintraub for their helpful comments.
[99]{}
Bary, J.S., Weintraub, D.A. and Kastner, J.H. 2003, ApJ, 586, 1136.
Bary, J.S., Weintraub, D.A., Kastner, J.H., Shukla, S.J. and Chynoweth,K. 2005, Protostars and Planets V, 8197.
Bitner, M., Richter, M.J., Lacy, J.H., Greathouse, T.K., Jaffe, D.T. and Blake, G.A. 2007, astro-ph/0704.1481v1.
Calvet, N., D’Alessio, P., Hartmann, L., Wilner, D., Walsh, A., & Sitko, M. 2002, ApJ, 568, 1008
Davis, C. J., Ray, T. P., Desroches, L., & Aspin, C. 2001, MNRAS, 326, 524
Davis, S. S. 2005, ApJ 627, L153
Deming, D., Charbonneau, D., & Harrington, J. 2004, ApJL, 601, L87
Fischer, D. A., Butler, R. P., Marcy, G. W., Vogt, S. S, & Henry, G. W. 2003, ApJ 590, 1081
Haisch, K. E., Lada, E. A., & Lada, C. J. 2001, ApJL, 553, L153
Haisch, K. E., Jayawardhana, R., Alves, Jo. 2005, ApJ,627, L57
Hubickyj, O., Bodenheimer, P., & Lissauer, J. J. 2005, Icarus, 179, 415
Hinkle, K. H., Blum, R., Joyce, R.R., Ridgway, S.T., Rodgers, B., Sharp, N., Smith, V., Valenti, J., and van der Bliek, N., “The Phoenix Spectrograph at Gemini South,” 2003 Proc. SPIE 4834, 353
Itoh, Y., Sugitani, K., Ogura, K., & Tamura, M. 2003, PASJ, 55, L77
Lahuis, F., van Dischoeck, E.F., Blake, G.A., Evans, N.J., Kessler-Silacci, J.E. and Pontoppidan, K.M. 2007, astro-ph/0704.2305v2.
Lawson, W. A., Crause, L. A., Mamajek, E. E., & Feigelson, E. D. 2002, MNRAS, 329, L29
Lawson, W. A., Lyo, A.-R., & Muzerolle, J. 2004, MNRAS, 351, L39
Lecar, M., Podolak, M., Sasselov, D., & Chiang, E. 2006, ApJ 640, 1115
Luhman, K. L., & Steeghs, D. 2004, ApJ, 609, 917
A.-R. Lyo, W.A.Lawson, E.E.Mamajek, E.D. Fiegelson, E.-C. Sung and Crause, L.A. 2003, MNRAS, 338, 616.
Lyo, A.-R., Lawson, W. A., & Bessell, M. S. 2004, MNRAS, 355, 363
Mamajek, E.E., Lawson, W.A. and Feigelson, E.D. 1999, ApJ, 516, L77.
Mamajek, E.E., Lawson, W.A. and Feigelson, E.D. 2000, ApJ, 544, 356.
Maloney, P. R., Hollenbach, D. J., & Tielens, A. G. G. M. 1996, ApJ, 466, 561
Megeath, S. T., Hartmann, L., Luhman, K. L., & Fazio, G. G. 2005, ApJLZ, 634, L113
Menou, K., & Goodman, J. 2004, ApJ 606, 520
Nomura H., Millar T. J. 2005, A&A, 438, 923
Nomura H., Aikawa, Y., Tsujimoto, M. and Millar T. J. 2007, astro-ph/0702030v1.
Pollack J. B., Hubickyj O., Bodenheimer P., Lissauer J. J., Podolak M., Greenzweig Y. 1996, Icarus, 124, 62
Ramsay Howat, S. K., Harris, J.W., Gostick, D.C., Laidlaw,K., Kidd, N., Strachan, M. and Wilson, K. “The Gemini facility calibration unit”, 2000, Proc. SPIE Vol. 4008, p. 1351-1360, Optical and IR Telescope Instrumentation and Detectors, Masanori Iye; Alan F. Moorwood; Eds.
Richter, M. J., Jaffe, D. T., Blake, G. A., & Lacy, J. H. 2002, ApJL, 572, L161
Sheret, I., Ramsay Howat, S. K., & Dent, W. R. F. 2003, MNRAS, 343, L65
Siess, L., Forestini, M., & Bertout, C. 1999, A&A, 342, 480
Simon, M. Ghez, A.M., Leinert, C., Cassar, L., Chen, W. P., Howell, R. R., Jameson, R. F., Matthews, K., Neugebauer, G. & Richichi, A. 1995, ApJ, 443, 625
Takami, M., Chrysostomou, A., Ray, T. P., Davis, C., Dent, W. R. F., Bailey, J., Tamura, M., & Terada, H. 2004, A&A, 416, 213
Thi, W. F., et al. 2001, Nature, 409, 60
Thi, W. F., et al. 2001, ApJ, 561, 1074
Weintraub, D.A., Kastner, J.H. & Bary, J.S. 2000, ApJ, 541, 767
White, R.J. & Ghez, A.M. 2001, ApJ, 556, 265
Wood, K., Lada, C. J., Bjorkman, J. E., Kenyon, S. J., Whitney, B., Wolff, M. J. 2002, ApJ, 567, 1183
\[lastpage\]
[^1]: Email: [email protected]
[^2]: E-mail: [email protected](JSG)
[^3]: Coordinates from @mlf99.
[^4]: from @ls04
[^5]: [@hja05]
[^6]: @megeath05
[^7]: @mlf99
[^8]: @lcmf02
[^9]: @wg01
[^10]: @bwk03
[^11]: @siess99
[^12]: @simon95
[^13]: [@calvet02]
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'D. Eckert'
- 'S. Ettori'
- 'S. Molendi'
- 'F. Vazza'
- 'S. Paltani'
bibliography:
- 'aa20403.bib'
title: 'The X-ray/SZ view of the virial region\'
---
Introduction
============
The outskirts of galaxy clusters are the regions where the current activity of structure formation is taking place through the accretion of smaller structures onto massive clusters. In addition, they are the regions where the transition between virialized gas from clusters and accreting material from the large-scale structure occurs. Non-gravitational effects such as turbulence [e.g., @dolag05; @vazza11a] and cosmic rays [e.g., @vazza12a; @pfrommer07] are expected to play a more important role in these regions than in cluster cores, and the infall of smaller structures along large-scale filaments may cause the material in these regions to be clumpy [@nagai; @simionescu] and asymmetric [@vazzascat; @e12]. These effects may bias the measurements of cluster masses using X-ray and SZ proxies [@rasia; @nagai07; @piffaretti], thus setting limitations on the use of galaxy clusters as high-precision cosmological probes [@allen11].
Baryons in galaxy clusters, mainly in the form of hot X-ray emitting plasma with some contribution from stars in member galaxies, are subjected to gravitational and radiative processes that, in a cold dark matter (CDM) scenario, affect their distribution in the DM halo accordingly. By forming in the highest peaks of the primordial gravitational fluctuations, massive galaxy clusters are relatively well-isolated gravity-dominated structures, suggesting that their mass function and relative baryon budget are highly sensitive tests of the geometry and matter content of the Universe. In particular, estimates of the gas and total mass content of relaxed hot clusters in hydrostatic equilibrium can be obtained using X-ray observations, allowing one both to place a lower limit on the cluster baryon mass fraction, which is expected to match the cosmic value $\Omega_{\rm b} / \Omega_{\rm m}$ [e.g., @white93], and to constrain the cosmic dark energy by imposing the gas mass fraction as standard ruler [see e.g. @ettori03; @ettori09; @allen08].
In this paper, we combine the SZ pressure profiles from the sample of 62 galaxy clusters observed with the *Planck* satellite [@planck5 hereafter P12] with the *ROSAT* gas density profiles presented in @e12 [hereafter E12] to reconstruct gravitating mass profiles assuming hydrostatic equilibrium. This allows us to provide the first measurements of the gas fraction in cluster outskirts on relatively large samples, probing a volume several times larger than the typical *XMM* and *Chandra* limits and the entire azimuth (as opposed to most of the *Suzaku* results). We study the importance of additional effects (non-thermal pressure support, gas clumping, asymmetry) on the quantities recovered by assuming that the gas is in hydrostatic equilibrium. Similar results are also presented for a subset of 18 individual clusters (6 CC and 12 NCC) that are in common between the samples of P12 and E12. In a companion paper (hereafter Paper I), we demonstrate the validity of the method and apply it to infer the average thermodynamic properties of galaxy clusters.
Throughout the paper, we assume a $\Lambda$CDM cosmology with $\Omega_m=0.3$, $\Omega_\Lambda=0.7$, and $H_0=70$ km s$^{-1}$ Mpc$^{-1}$. We also refer to the best-fit results on the $\Lambda$CDM model provided from WMAP-7 years [see e.g. Table 8 in @wmap7], which imply a cosmic baryon fraction $\Omega_{\rm b} / \Omega_{\rm m} = 0.167 \pm 0.011$. Because our cluster sample is located at redshift below 0.2, the differences between these cosmological models affect the estimates of the gas mass fraction presented in this work by less than 1 per cent, through the variation of the angular diameter distance.
Basic formalism
===============
For this work, the two main observables are the pressure from SZ measurements and the gas number density from X-ray observations. In this case, the hydrostatic mass is simply given by
$$M_{{\rm hyd}}(<r)=-\frac{r^2}{G \mu m_{\rm p} n_{\rm gas}}\frac{{\rm d}P_{\rm gas}}{{\rm d}r}, \label{hydmass}$$
where $\mu=0.6$ is the mean molecular weight, while the gas mass profiles can be computed directly by integrating the density profiles,
$$M_{\rm gas}(<r) = \mu m_{\rm p} \int_0^r n_{\rm gas}(r^\prime) 4\pi r^{\prime 2}\, dr^\prime . \label{gasmass}$$
An estimate of $f_{\rm gas}(r)=M_{\rm gas}(<r)/M_{\rm hyd}(<r)$ can thus be obtained in a straightforward way. To describe the pressure and the gas density, we fit the data point using general parametric functions (see paper I for details). Following P12, the pressure is described by a generalized Navarro-Frenk-White profile [GNFW, @gnfw; @nagai07b]. The pressure gradient can be expressed analytically by differentiating the GNFW profile. The pressure profiles are taken from P12, for the average population as well as for the 18 individual clusters in common between the two samples.
For the density, we used a simplified version of the functional form introduced by @vikhlinin06. This functional form was projected along the line-of-sight and fit to the *ROSAT* emission-measure profiles. The error budget was estimated through a Monte-Carlo Markov chain (MCMC) method, which allows us to compute the envelopes of the curves derived from the fitting procedure. Alternatively, we also used deprojected density profiles as computed in E12, and interpolated them along a common grid through a cubic spline method. For more details and a thorough validation of the method, we refer the reader to Paper I, also for the details of the error calculation. In Fig. \[fgas\_all\] we show the gas fraction profiles derived for the 18 individual systems and the average pressure and density profiles. Since we focus our work on the outer regions, the MCMC procedure was performed only on the data beyond $0.2R_{500}$, which causes the bumps at low radii. Beyond $0.5R_{500}$, the most strongly deviating profile is that of A2163, a violently merging cluster that is likely out of hydrostatic equilibrium (see Appendix C of Paper I for a discussion of this system).
Results
=======
For a fiducial cluster with a virial temperature $kT=5$ keV, we used the universal pressure profiles from P12 together with the median gas density profiles from E12 to extract gas fraction profiles using the hydrostatic equilibrium equation (Eq. \[hydmass\]). Unlike the gas fraction profiles presented in Sect. 7.3 of P12, no extrapolation is needed to compute the gas fraction. In Fig. \[fig:fgas\], we show the average gas fraction profile obtained by combining the average profiles from P12 and E12, computed through the parametric forward-fitting method (green-shaded area) and the geometrical deprojection method (red points). Excellent agreement is found between the results obtained between the two methods, and hence the results presented here are stable against the assumed deprojection technique. The profiles obtained from the sample-averaged profiles are also compared to the median of the 18 individual objects for both deprojection methods. The uncertainties in the median were estimated using 1000 bootstrap reshufflings of the sample (see Paper I). Below $1.7R_{500}$[^1], all methods show a remarkable level of agreement, demonstrating the reliability of the measurements. In the following and for the sake of simplicity we focus on the results obtained with the parametric form, as in Paper I. However, we stress that the geometrical deprojection method always leads to consistent results.
At $R_{500}$, we find $f_{{\rm gas}, 500}=0.132\pm0.005$, which agrees well with previous estimates from [*Chandra*]{} [e.g., @vikhlinin06] and [*XMM-Newton*]{} [e.g., @ettori10]. The measured gas fraction reaches the cosmic baryon fraction [$\Omega_b/\Omega_m=0.167$, @wmap7] at $R=1.4R_{500}$ and slightly exceeds it beyond this radius, although it remains consistent within the error bars. At $R_{200}$, we measure $0.176\pm0.009$ ($0.160\pm0.019$ in the subsample of 18 objects) with the forward-fitting method, $0.158\pm0.019$ ($0.160\pm0.028$) with the deprojection technique. Given that the stellar mass contributes another $10-20\%$ of the baryon fraction [e.g., @gonzalez; @lagana], the gas fraction recovered assuming hydrostatic equilibrium exceeds the expected value in cluster outskirts at the $2\sigma$ level. Although the confidence level is low, this may indicate that either the mass recovered through hydrostatic equilibrium is biased low, as suggested from present hydrodynamical numerical simulations [e.g., @rasia06; @rasia12; @nelson; @burns10], or that inhomogeneities in the gas distribution bias high the estimated $M_{\rm gas}$ [@nagai; @simionescu], or a combination of the two effects.
[lcccccccc]{} Dataset & $f_{{\rm gas},2500}$ & $\sigma_{f,2500}$ & $f_{{\rm gas},1000}$ & $\sigma_{f,1000}$ & $f_{{\rm gas},500}$ & $\sigma_{f,500}$ & $f_{{\rm gas},200}$ & $\sigma_{f,200}$\
18 obj & $10.1\pm0.5$ & $1.5\pm0.4$ & $12.7\pm0.7$ & $2.4_{-0.6}^{+0.7}$ & $14.9\pm1.0$ & $3.7_{-0.8}^{+1.0}$ & $16.6\pm1.2$ & $4.2_{-0.9}^{+1.1}$\
6 CC obj & $8.7\pm0.5$ & $0.9\pm0.4$ & $10.7\pm0.5$ & $0.6\pm0.6$ & $12.3\pm0.6$ & $<1.9$ & $14.3_{-1.4}^{+1.5}$ & $2.8_{-1.0}^{+1.4}$\
12 NCC obj & $10.9\pm0.5$ & $1.2\pm0.6$ & $13.8\pm2.8$ & $2.8_{-0.8}^{+1.1}$ & $16.3\pm1.4$ & $4.4_{-1.0}^{+1.4}$ & $17.8\pm1.6$ & $4.7_{-1.1}^{+1.5}$\
\
18 obj, T-scaled & $9.9\pm0.3$ & $0.9\pm0.4$ & $12.6\pm0.5$ & $1.3\pm0.5$ & $14.6\pm0.7$ & $2.3_{-0.6}^{+0.7}$ & $16.4\pm1.1$ & $3.7_{-0.8}^{+1.0}$\
6 CC obj, T-scaled & $9.0\pm0.4$ & $<0.8$ & $11.3\pm0.5$ & $<1.0$ & $12.9\pm0.6$ & $<1.3$ & $14.9_{-1.5}^{+1.6}$ & $2.8_{-1.0}^{+1.5}$\
12 NCC obj, T-scaled & $10.9\pm0.5$ & $1.9_{-0.5}^{+0.6}$ & $13.4\pm0.7$ & $1.6_{-0.6}^{+0.7}$ & $15.6\pm1.0$ & $2.8_{-0.7}^{+0.9}$ & $17.2_{-1.4}^{+1.5}$ & $4.0_{-1.0}^{+1.4}$\
Similar to the entropy (see Paper I), we observe differences in the gas fraction profiles computed for the 6 CC and 12 NCC clusters independently (see Fig. \[fig:fgas\]). The NCC profiles tend to increase more steeply in cluster outskirts compared to CC profiles. From the sample-averaged profiles, at $R_{200}$ NCC clusters slightly exceed the cosmic baryon fraction ($f_{{\rm gas},NCC}=0.169\pm0.010$ using the average profiles; $0.178\pm0.016$ for the 12 individual NCC systems), while the gas fraction in CC systems is close to the expected gas fraction around $R_{200}$ ($f_{{\rm gas},CC}=0.134\pm0.011$; $0.143 \pm 0.015$ for the subsample of 6 CC objects). Despite its low statistical significance (less than 2$\sigma$), this result likely indicates that hydrostatic equilibrium breaks down in the outskirts of NCC systems, while in CC (relaxed) clusters the behavior is closer to self-similar expectations, in agreement with our results for the entropy (see Paper I).
In all cases, we can see that the enclosed gas fraction increases steadily with radius and gas temperature, as expected from previous work [see e.g., @arnaudlxt; @ettori99; @vikhlinin06]. To quantify the dependence on the gas temperature at different mass overdensities (200, 500, 1000, and 2500, see Fig. \[fgasTdelta\]), we fit a multi-linear function of the form $$\frac{f_{\rm gas}}{f_{\rm b, WMAP7}} = b_{{\rm gas },500}\left(\frac{\Delta}{500}\right)^\alpha \left(\frac{T_{vir}}{7\mbox{ keV}}\right)^\beta , \label{funcover}$$
where $b_{\rm gas}= f_{\rm gas}/f_{\rm b, WMAP7}$ is the depletion factor. The best-fitting values for the parameters are provided in Table \[tableover\]. Only the normalization $b_{500}$ is found to change significantly between the CC and NCC populations, in agreement with Fig. \[fig:fgas\]. The quality of the fit, however, is found to change significantly when comparing the CC and NCC populations, with only the fit to CC objects being formally acceptable. This indicates a significant cluster-to-cluster scatter in the latter population.
Dataset $b_{{\rm gas},500}$ $\alpha$ $\beta$ $\chi^2$/d.o.f
--------- --------------------- ------------------ ----------------- ----------------
Total $0.854 \pm 0.016$ $-0.20 \pm 0.02$ $0.54 \pm 0.05$ $154.4/69$
CC $0.757 \pm 0.024$ $-0.19 \pm 0.03$ $0.52 \pm 0.14$ $21.2/21$
NCC $0.919 \pm 0.021$ $-0.20 \pm 0.02$ $0.46 \pm 0.06$ $100.7/45$
: \[tableover\]Best-fitting parameters for the function given in Eq. \[funcover\].
Discussion
==========
We discuss here two important implications of our results on (i) the use of gas mass fraction as cosmological probe and (ii) the comparison with predictions from hydrodynamically simulated objects.
Implications on the use of clusters for cosmology
-------------------------------------------------
The use of $f_{\rm gas}$ as a standard ruler for cosmology [@ettori03; @allen08] implies that galaxy clusters are representative of the cosmic baryonic budget. In the previous section, we quantified how the gas mass fraction depends on the gas temperature and matter overdensity, providing for the first time an observationally constrained calibration of the cluster baryon budget in the hot X-ray emitting phase. We assess here the cluster-to-cluster scatter of $f_{\rm gas}$ at different overdensity radii. To compute the intrinsic scatter of a given population, we used a maximum-likelihood estimator as introduced by @maccacaro (see also Appendix A of E12). This method allows the mean value and the intrinsic scatter to be computed jointly. In Table \[tabscat\] we give the mean $f_{\rm gas}$ values and the intrinsic scatter $\sigma_f$ at overdensities of 2500, 1000, 500, and 200, for the entire population as well as for the CC and NCC clusters separately. We also report the scatter obtained when correcting the $f_{\rm gas}$ values for their dependence on the cluster’s virial temperature, i.e., by rescaling each value by the quantity $(T_{vir}/\mbox{7 keV})^\beta$ (see Eq. \[funcover\] and Table \[tableover\]).
In Table \[tabscat\], we can see that the level of scatter increases with radius, from $\sim15\%$ of the measured value at $R_{2500}$ to $\sim25\%$ at $R_{200}$. As expected, when a temperature scaling is performed, a lower scatter is measured in all cases. Interestingly, we note that the level of scatter is substantially higher for the NCC population than for the more relaxed, CC objects. When performing a temperature scaling, the level of intrinsic scatter is even consistent with 0 in CC systems out to $R_{500}$, while at $R_{200}$ some scatter at the level of $\sim18\%$ is measured, indicating that at the outermost radii even the most relaxed systems are affected by non-gravitational physics and/or inhomogeneous gas distributions (see E12). In the case of NCC systems, substantial cluster-to-cluster variations are found at all radii.
We remark that the trend of increasing scatter with radius is opposite to what we would expect when probing a larger volume, which shows that the the trend is likely caused by a physical effect. Both non-gravitational effects (turbulence, cosmic rays) and inhomogeneous gas densities can qualitatively explain this result. Indeed, numerical simulations show that non-thermal effects can be several times stronger in NCC systems [e.g., @lau09; @pfrommer07; @vazza11a; @vazza12a]. Gas clumping and large-scale asymmetries are also expected to be more prominent along large-scale filaments [@nagai; @vazza12c], and thus the effect should be stronger in merging systems. For instance, if clumping biases the observed density high by $\sim20\%$, as found in Paper I for NCC clusters, the actual gas fraction for this class of objects would be consistent with the expected hot gas fraction. For a more detailed discussion of the implications of our results on the physics of cluster outskirts, we refer the reader to Paper I.
To summarize, our findings (both on the global baryon budget and the variations in $f_{\rm gas}$ from one system to another) demonstrate that the gas fraction of CC systems can be used efficiently as a standard ruler, while in NCC systems non-gravitational effects and/or inhomogeneities in the gas distribution introduce significant systematics in the computation of the gas fraction. A very similar conclusion was recently obtained by @mahdavi12 by comparing X-ray and weak-lensing mass estimates. Indeed, X-ray mass estimates appear to be underestimated by $\sim15\%$ in NCC clusters, while good agreement is found for CC systems. The fact that a qualitatively similar result was obtained using different samples and techniques thus reinforces our conclusion.
Comparison with numerical simulations
-------------------------------------
Using the shock-capturing Eulerian adaptive mesh refinement N-body+gas-dynamical adaptive refinement tree code, @kravtsov05 measure in the most massive system of $M_{200} \approx 9.6 \times 10^{14} M_{\odot}$ a value of $b_{\rm gas}$ that rises from 0.86 (at $\Delta=2500$) to 1.04 (at the virial radius) in the run without dissipation and from 0.56 to 0.76 when radiative cooling, star formation, metal enrichment, and stellar feedback are considered. The constraints we obtain lie between these predictions, suggesting that both radiative processes are required and the feedback provided has to be less incisive in its action on distributing the ICM.
In a set of hydrodynamical simulations of massive galaxy clusters performed using the Tree+SPH code `GADGET-2`, @ettori06 [see also @deboni] estimate at $z=0$ a $b_{\rm gas}$ that increases from 0.82 to 0.89 from $\Delta=2500$ to $200$ in runs with gravitational heating only. On the other extreme, introducing cooling and star formation but no winds drastically reduces the expected relative gas fraction (0.29 at $\Delta=2500$; 0.59 at $\Delta=200$). Overall, however, none of the feedback models explored in that work are able to reproduce the high $b_{\rm gas}$ value of $0.74-0.9$ we observe at $\Delta < 500$, apart from the case where the alternative artificial low-viscosity scheme from @morris is considered, allowing about 30 per cent of the ICM thermal energy in the form of turbulent motion.
For massive halos extracted from a non-radiative gas-dynamical realization of the [*Millennium Simulations*]{}, @crain measured $b_{\rm bar}$ of about 0.9 (and rms of 6 per cent) with almost no dependence on the radius between $R_{500}$ and $R_{200}$, implying a relative contribution from stars, with respect to the gas component, of 0.22 and 0.19 at the two radii, respectively, slightly higher than the present observational constraints [e.g., $<0.17$ for $M_{500} > 3 \times 10^{14} M_{\odot}$ in @lagana]. In the [*Millennium Gas Project*]{}, @millenniumgas studied the baryon distribution in 170 objects, as SPH re-simulation of the Millennium simulation, with gas temperature higher than 3 keV. They considered models where the ICM is heated (i) solely by gravitational process, (ii) by pre-heating in the order of 200 keV cm$^2$ occuring at $z=4$ combined with cooling, (iii) by feedback from supernovae (SNe) and active galactic nuclei (AGN) as expected from the semi-analytic predictions on galaxy formation but not including radiative cooling. For the most massive systems, they measured $b_{\rm gas}$ in the range 0.88–0.91, 0.57–0.83, 0.70–0.82 between $\Delta=2500$ and $200$ for cases (i), (ii), (iii), respectively, suggesting that (a) some cooling is needed to lower the gas fraction predicted in the inner part and (b) the implemented action of either pre-heating or feedback from SNe and AGN is probably too strong to fully explain the observed gas fraction in the clusters’ outskirts.
Conclusion
==========
Combining X-ray (*ROSAT*/PSPC) and SZ (*Planck*) data, we analyzed the properties of the ICM at large radii using both sample-averaged profiles and the data for 18 individual systems. This allowed us to measure for the first time hydrostatic mass and gas fraction profiles out to the virial radius in a relatively large sample of objects. The results on the thermodynamic properties of galaxy clusters close to the virial radius are presented in a companion paper. Our results can be summarized as follows:
- The gas fraction reconstructed through the hydrostatic equilibrium equation increases with radius and reaches the value of $f_{\rm gas,200}=0.176\pm0.009$ around $R_{200}$. Different input data and deprojection techniques yield results that excellently agree. This value is consistent with the cosmic baryon fraction $\Omega_b/\Omega_m=0.167$ [@wmap7]. Given that the stellar content is expected to account for $10-20\%$ of the baryons in galaxy clusters, the measured gas fraction slightly exceeds the expected value. A 15% bias either in raising the hydrostatic masses, as expected from numerical simulations, or in lowering the gas mass allows us to reconcile our measurement with the expected gas fraction.
- We observe differences between the gas fraction measured in relaxed, CC systems compared to dynamically active, NCC systems. In NCC clusters, the gas fraction exceeds the cosmic baryon fraction significantly, indicating that hydrostatic equilibrium breaks down in the outer regions of perturbed systems, and/or that the gas is distributed in an inhomogeneous way. Conversely, in CC systems the gas fraction converges to the expected hot gas fraction, confirming that they correspond to more relaxed systems, which can be used more efficiently for cosmological purposes.
- From the 18 objects comprising our sample, we measured the scatter around the average $f_{\rm gas}$ value using a maximum-likelihood estimator. We find that the scatter increases with radius, from 15% at $R_{2500}$ to 25% at $R_{200}$. In contrast to NCC systems, in which we measure substantial cluster-to-cluster variations at all radii, CC systems exhibit a negligible level of scatter out to $R_{500}$, confirming that they are better suited for cosmological studies.
- We quantified the $f_{\rm gas}$ dependence upon gas temperature and matter overdensity with the relation $f_{\rm gas}/f_{\rm b, WMAP7} = b_{500} \left(\Delta/500\right)^\alpha \left(T_{\rm gas}/7\mbox{ keV}\right)^\beta$ and measured $b_{500} = 0.76 \pm 0.02$ in CC systems and $0.92 \pm 0.02$ in NCC clusters, $\alpha=-0.2$ and $\beta=0.5$. The gas bias (or depletion) factor $b_{500}$, and its variation with temperature and radii, disagrees with the results obtained for clusters extracted from recent hydrodynamical simulations, implying that the action of cooling and AGN and/or SN feedback is indeed needed but has to be regulated differently from what has been used so far.
[^1]: For a given overdensity $\Delta$, $R_\Delta$ is the radius for which $M_\Delta/(4/3\pi R_\Delta^3)=\Delta\rho_c$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The accumulation of atoms in the lowest energy level of a trap and the subsequent out-coupling of these atoms is a realization of a matter-wave analog of a conventional optical laser. Optical random lasers require materials that provide optical gain but, contrary to conventional lasers, the modes are determined by multiple scattering and not a cavity. We show that a Bose-Einstein condensate can be loaded in a spatially correlated disorder potential prepared in such a way that the Anderson localization phenomenon operates as a band-pass filter. A multiple scattering process selects atoms with certain momenta and determines laser modes which represents a matter-wave analog of an optical random laser.'
author:
- Marcin Płodzień
- Krzysztof Sacha
title: 'Matter-wave analog of an optical random laser'
---
Conventional optical lasers require two ingredients: material that provides optical gain and an optical cavity responsible for coherent feedback and selection of resonant laser modes. However, it is also possible to achieve laser action without the optical cavity provided the gain material is an active medium with disorder [@wiersma08]. Forty years ago Letokhov analyzed a light diffusion process with amplification and predicted that gain could overcome loss if the volume of a system exceeded a critical value [@letokhov67]. Random lasing (i.e. light amplification in disordered gain media), achieved in a laboratory in the 1990’s, attracts much experimental attention and offers possibilities for interesting applications [@lawandy94; @wiersma96; @cao98; @frolov99; @wiersma00; @wiersma08]. Theoretical understanding of this phenomenon is still imperfect. Although the Letokhov model of diffusion with gain is useful in predicting certain properties of random lasers, it neglects coherent phenomena. There are various theoretical models of random lasing but it is widely accepted that interference in a multiple scattering process determines the spatial and spectral mode structure of a random laser [@wiersma08].
Bose-Einstein condensation (BEC) of dilute atomic gases is a macroscopic accumulation of atoms in the lowest energy level of a trap when the temperature of the gas decreases [@anderson95]. This tendency of occupying a single state through the mechanism of stimulated scattering of bosons is an analog of mode selection in optical lasers due to the stimulated emission of photons. Gradual release of atoms from a trapped BEC allows for the realization of a matter-wave analog of a conventional optical laser [@mewes97; @hagley99; @bloch99; @cennini03; @guerin06; @tomek08]. The atom trap is an analog of the optical cavity. The lowest mode of the trap is a counterpart of an optical resonant mode. In conventional optical lasers the output coupler is usually a partially transmitting mirror. In atom lasers it involves, for example, a change of the internal state of the atom by means of a radio-frequency transition. In the present letter we propose the realization of a matter-wave analog of an optical random laser. Suppose the BEC of a dilute atomic gas has been achieved in a trapping potential. That is, we begin with the accumulation of atoms in a single mode of the resonator (i.e. the lowest eigenstate of the trap). Then, let us turn off the trap and turn on a weak disorder potential. Starting with a BEC we have a guarantee that the disorder medium is [*pumped*]{} with coherent matter-waves. We would like to raise the question of whether it is possible to prepare spatially correlated disorder potential in such a way that narrow peaks can be observed in the spectrum of atoms that are able to leave the area of the disorder potential? In other words: if the multiple scattering of atoms in a disorder medium can lead to a selective spectral emission of matter-waves from the medium?
In cold atom physics a disorder potential can be realized by means of an optical speckle potential [@schulte05; @clement06]. Transmission of coherent light through a diffusing plate leads to a random intensity pattern in the far field. Atoms experience the presence of the radiation as an external potential $V({\mathbf{r}})\propto \chi|E({\mathbf{r}})|^2$ proportional to the intensity of the light field $E({\mathbf{r}})$ and atomic polarizabitlity $\chi$ whose sign depends on the detuning of the light frequency from the atomic resonance. Diffraction from the diffusive plate onto the location of atoms determines correlation functions of the speckle potential. We assume that the origin of the energy is shifted so that $\overline{V({\mathbf{r}})}=0$ where the overbar denotes an ensemble average over disorder realizations. Standard deviation of the speckle potential $V_0$ measures the strength of the disorder.
Let us begin with a one-dimensional (1D) problem. In a weak disorder potential atoms with $k$-momentum undergo multiple scattering, diffusive motion and finally localize with an exponentially decaying density profile due to the Anderson localization process, provided that the system size exceeds the localization length [@anderson58; @lee85; @tiggelen99]. Taking the Born approximation to the second order in the potential strength, the inverse of the localization length is [@lifshits88] $l_{loc}^{-1}=(mV_0/\hbar^2k)^2{\cal P}(2k)$, where the Fourier transform of the pair correlation function of the speckle potential is (k)=(q)(k-q), \[pk\] and $\gamma(k)=\int dz \tilde\gamma(z)e^{-ikz}$ is the Fourier transform of the complex degree of coherence $\tilde\gamma(z)=\overline{E^*(z+z')E(z')}/\overline{|E(z)|^2}=\int dy {\cal A}(y)e^{izy/\alpha}/\int dy {\cal A}(y)$. ${\cal A}(y)$ describes the aperture of the optics and $\alpha$ is a constant dependant on the wavelength of the laser radiation and the distance of the diffusive plate from the atomic trap. In Ref. [@billy08], where the experimental realization of the Anderson localization of matter-waves is reported (see also [@roati08]), a simple Heaviside step function ${\cal A}(z)=\Theta(R-|z|)$ describes the aperture. The corresponding $\gamma(k)=\pi\sigma_R\Theta(1-|k\sigma_R|)$, where $\sigma_R=\alpha/R$ is the correlation length of the speckle potential. Consequently the power spectrum (\[pk\]) decreases linearly and becomes zero for $|k|\ge 2/\sigma_R$. Thus, the Born approximation predicts an effective mobility edge at $|k|=1/\sigma_R$, i.e. atoms with larger momenta do not localize [@sanchez07; @billy08] (actually higher order calculations [@lugan09] show they do localize but with very large localization lengths, much larger than the system size in the experiment). Hence, neglecting atom interactions, if the width of the initial atom momentum distribution exceeds the mobility edge particles at the tail of the distribution avoid the Anderson localization and may leave the disorder area.
![(Color online) Examples of the speckle potential (top panels) and the corresponding localization length (bottom panels) obtained within the Born approximation (dashed black curves) and numerically in the transfer-matrix calculations (solid red curves). Panels (a) and (c) show the results for the single obstacle in the diffusive plate where $\sigma_R=0.066$ (0.31 $\mu$m), $\sigma_R/\sigma_\rho=0.4$ and $V_0=3.5$. Panels (b) and (d) correspond to the case of two obstacles with the same $\sigma_R$ and $V_0$ but $\sigma_R/\sigma_\rho=0.7$ and $\sigma_R/\sigma_\zeta=0.1$. The results are shown for rubidium-87 atoms. Red detuning of the laser radiation from the atomic resonance is assumed. All values are presented in the harmonic oscillator units, i.e. energy $E_0=\hbar\omega$ and length $l_0=\sqrt{\hbar/m\omega}$ where $\omega/2\pi=5.4$ Hz. []{data-label="one"}](Fig1.eps){width="0.95\linewidth"}
Let us modify the experiment reported in Ref. [@billy08] by introducing an obstacle at the center of the diffusive plate so that the aperture is now described by ${\cal A}(z)=\Theta(R-|z|)-\Theta(\rho-|z|)$ where $\rho<R$. It implies that (k)=(-)\^[-1]{}\[ (1-|k\_R|)-(1-|k\_|)\], where $\sigma_\rho=\alpha/\rho$. If the size of the obstacle $\rho>R/3$, interference of light passing through such a [*double-slit* ]{} diffusive plate creates a peculiar speckle potential. That is, the power spectrum (\[pk\]) disappears for $|k|\ge 2/\sigma_R$ as previously but it is also zero for $\frac{1}{\sigma_R}-\frac{1}{\sigma_\rho}<|k|<\frac{2}{\sigma_\rho}$. Thus, according to the Born approximation there is a momentum interval where the localization length diverges. It implies that the Anderson localization process is able to operate as a band-pass filter letting particles with specific momenta leave the region of the disorder. Detection of escaping atoms should reveal a peak in the momentum spectrum corresponding to the interval where the localization length diverges.
![(Color online) Momentum distributions of the atoms localized in the disorder potential (dashed black lines) and the atoms which escaped from the disorder area (solid red lines). Panels (a) and (c) show the results for a single experimental realization of the disorder with parameters as in Fig. \[one\]a,c while panels (b) and (d) are related to the parameters as in Fig. \[one\]b,d. Panel (a) corresponds to the evolution time $t=100$ (2.9 s), panel (b) to $t=70$ (2 s), panels (c) and (d) to $t=200$ (5.7 s). Fraction of atoms that escaped the disorder region is about 9% in (a) and (b), and 20% in (c) and (d). In order to take into account experimental resolution all data have been convoluted with Gaussian of $\Delta k=0.3$ width. All values are presented in the harmonic oscillator units, i.e. energy $E_0=\hbar\omega$, length $l_0=\sqrt{\hbar/m\omega}$ and time $t_0=1/\omega$ where $\omega/2\pi=5.4$ Hz.[]{data-label="two"}](Fig2.eps){width="0.9\linewidth"}
Introducing two (or more) obstacles in the diffusive plate we can increase the number of momentum intervals with diverging $l_{loc}$. In Fig. \[one\] we present examples of the speckle potentials in the single obstacle case and the case of two obstacles located symmetrically around the plate center. In the latter case the aperture is described by ${\cal A}(z)=\Theta(R-|z|)-\Theta(\rho-|z|)+\Theta(\zeta-|z|)$ where $\zeta<\rho$ and (k)&=& (-+)\^[-1]{} \[(1-|k\_R|) &&-(1-|k\_|) +(1-|k\_|)\], with $\sigma_\zeta=\alpha/\zeta$. The figure also presents localization lengths obtained numerically in the transfer-matrix calculation [@lugan09] that confirms the Born predictions.
To simulate an experiment, we follow the parameters used in Ref. [@billy08] where Anderson localization of matter-waves has been observed. We assume that a BEC of $N=1.7\cdot 10^4$ rubidium-87 atoms is initially prepared in a quasi-1D harmonic trap with longitudal and transverse frequencies $\omega/2\pi=5.4$ Hz and $\omega_\perp/2\pi=70$ Hz, respectively. In the following we adopt the harmonic oscillator units: $E_0=\hbar\omega$, $l_0=\sqrt{\hbar/m\omega}$ and $t_0=1/\omega$ for energy, length and time, respectively. When the trapping potential is turned off and the speckle potential is turned on the expansion of the atomic cloud is initially dominated by the particle interactions until the density drops significantly and the atoms start feeling only the disorder potential. This initial stage of the gas expansion sets the momentum distribution of the atoms which may be approximated by an inverted parabola with an upper cut-off $k_{max}=2\sqrt{\mu}=12.7$ where $\mu$ is the initial chemical potential of the system [@sanchez07; @miniatura09].
The disorder potentials we choose are attainable in the experiment reported in Ref. [@billy08], i.e. they extend 862 units (4 mm) along the $z$ direction with the correlation length $\sigma_R=0.066$ (0.31 $\mu$m). We consider the potentials obtained by introducing one or two obstacles in the diffusive plate that are presented in Fig. \[one\]. If the atomic cloud starts at the center of the disorder potentials and if the cut-off of the momentum distribution is $k_{max}\lesssim 13$ then we may expect that time evolution leads to emission of atoms with $|k|\approx 5.5$ in the single obstacle case (cf. Fig. \[one\]a,c) and with $|k|\approx 9$ in the case of two obstacles (cf. Fig. \[one\]b,d). In the latter case we may expect also a small leakage of atoms with $|k|\approx 3.5$. For $|k|\approx 3.5$ the localization length shown in Fig \[one\]d reaches locally a maximum value of $l_{loc}\approx 120$. Before the particle interactions become negligible the gas spreads over a significant range of the disorder region. Therefore the Anderson localization is not able to diminish completely the leakage of atoms with $|k|\approx 3.5$. Starting with the ground state of the stationary Gross-Pitaevskii equation [@mewes97; @hagley99; @bloch99; @cennini03; @guerin06; @tomek08; @schulte05] in the presence of the harmonic trap, we integrate the time-dependent Gross-Pitaevskii equation when the trap is turned off and a disorder is turned on. Figure \[two\] shows momentum distributions of atoms that escaped from the disorder region and those that are localized for the disorder potentials corresponding to Fig. \[one\] at different moments in time. The expected selective spectral emissions of atoms are apparent in the figure. Interestingly in Fig. \[two\]d, i.e. for longer evolution time, small peaks around $|k|\approx 3.5$ become visible.
![(Color online) The Boltzmann transport mean-free path (a) and the localization length (b) for atoms in the 2D speckle potential created by transmission of a laser beam through the circularly shaped diffusive plate with the obstacle in the form of a ring, i.e. the aperture of the optics is described by $\Theta(R-|{\mathbf{r}}|)-\Theta(\rho-|{\mathbf{r}}|)+\Theta(\zeta-|{\mathbf{r}}|)$ with $\sigma_R/\sigma_\rho=0.99$ and $\sigma_R/\sigma_\zeta=0.15$. The potential strength corresponds to $\eta=0.15$. The quantities presented in the figure are dimensionless. []{data-label="three"}](Fig3.eps){width="0.9\linewidth"}
Finally let us consider a possibility of the realization of an atom analog of an optical random laser in 2D. The Boltzmann transport mean-free path $l_B$ is the characteristic spatial scale beyond which memory of the initial direction of the particle momentum is lost. In 2D $l_{loc}= l_Be^{\pi kl_B/2}$ and thus the localization length is much larger than $l_B$ which is in contrast to the 1D case where these two quantities are nearly identical $l_{loc}=2l_B$ [@tiggelen99]. For the circularly shaped diffusing plate with the radius $R$ the classical transport mean-free path [@kuhn05; @miniatura09], to the second order in the potential strength, reads = (1-)[P]{}(2k\_R), \[2dlb\] where $\eta=V_0/E_\sigma$ is the ratio of the potential strength and correlation energy $E_\sigma=\hbar^2/(m\sigma_R^2)$ with $\sigma_R=\alpha/R$. The power spectrum ${\cal P}(k)$ of the optical speckle potential disappears for $k\ge 2/\sigma_R$. Nevertheless, the $l_B$ (and consequently also $l_{loc}$) is always finite. In the bulk 2D system, an initially prepared atomic wave-packet follows a diffusive motion at short time but eventually the dynamics slow down and freeze due to the Anderson localization process [@kuhn05; @miniatura09; @vincent10].
By introducing obstacles in the diffusive plate we are able to shape the power spectrum of the speckle potential. On one hand the fact that ${\cal P}(k)$ may disappear at certain momentum intervals does not mean divergence of the corresponding transport mean-free path (\[2dlb\]). On the other hand any non-monotonic behaviour of $l_B(k)$ is dramatically amplified in the behaviour of $l_{loc}(k)$ because the localization length is an exponential function of $l_B$. In Fig. \[three\] we present an example related to the obstacle in the form of a ring, i.e. the aperture of the optics is described by ${\cal A}({\mathbf{r}})=\Theta(R-|{\mathbf{r}}|)-\Theta(\rho-|{\mathbf{r}}|)+\Theta(\zeta-|{\mathbf{r}}|)$. At $k\sigma_R\approx 0.4$ both $l_B$ and $l_{loc}$ shows a maximum. However, while the transport mean-free path changes by only a few in the neighboring region the localization length changes by four orders of magnitude. If the width of the momentum distribution of a BEC loaded in such a disorder potential is smaller than $0.6/\sigma_R$ and the radius of the disorder medium is greater than $10^3\sigma_R$ but less than $10^5\sigma_R$ the multiple scattering process leads to an isotropic emission of atoms with $k\approx 0.4/\sigma_R$.
We have outlined a proposal for the realization of a matter-wave analog of an optical random laser. Spatially correlated disorder potential for atoms with a peculiar pair correlation function can be created by transmitting a laser beam through a diffusive plate with obstacles. The resulting Anderson localization length reveals non-monotonic behaviour as a function of particle momentum. It allows for filtering momenta of particles that leave the area of the disorder, if the size of the disorder medium is suitably chosen. The disorder medium is assumed to be initially loaded with a BEC which guarantees that the matter-waves emitted from the medium are coherent. We have restricted ourselves to the 1D and 2D cases but the atom analog of an optical random laser can be also anticipated in 3D. In 3D the Ioffe-Regel criterion discriminates between waves that are Anderson localized ($kl_B\lesssim 1$) or not [@tiggelen99; @kuhn07]. Thus, a spatially correlated disorder potential for which the Ioffe-Regel criterion is not fulfilled for specific momenta should allow for selective emission of matter-waves in 3D.
Our proposal is directly applicable to atomic matter-wave experiments. From the point of view of the optical random lasers our analysis is not complete because it is restricted to passive random materials without gain. There is an interesting question whether a disorder with properties similar to those analyzed here play a role in optical random lasers and which modes are important when the gain is included in a system.
The non-monotonic behaviour of the localization length results in the appearance of a multiple effective mobility edge if a disorder system is finite. Wave transport is then unusual and interesting on its own. A shallow non-monotonical bahaviour of the Anderson localization length versus energy has been observed also in a classical wave system, see Ref. [@chorwat].
We are grateful to D. Delande for encouraging discussion. and to R. Marcinek and J. Zakrzewski for critical reading of the manuscript. This work is supported by the Polish Government within research projects 2009-2012 (MP) and 2008-2011 (KS).
[*Note added:*]{} After submission of this article, we became aware of a related theoretical study [@laurent11].
[99]{}
D. S. Wiersma, Nature Physics [**4**]{}, 359 (2008).
V. S. Letokhov, Zh. Eksp. Teor. Fiz. [**53**]{}, 1442 (1967); Sov. Phys. JETP [**26**]{}, 835 (1968).
N. M. Lawandy [*et al.*]{}, Nature [**368**]{}, 436 (1994).
D. S. Wiersma and A. Lagendijk, Phys. Rev. E [**54**]{}, 4256 (1996).
H. Cao [*et al.*]{}, Appl. Phys. Lett. [**73**]{}, 3656 (1998); H. Cao [*et al.*]{}, Phys. Rev. Lett. [**82**]{}, 2278 (1999).
S.V. Frolov [*et al.*]{}, Phys. Rev. B [**57**]{}, 9141 (1999); T.V. Shahbazyan [*et al.*]{}, Phys. Rev. B [**61**]{}, 13 266 (2000).
D. S. Wiersma, Nature [**406**]{}, 135 (2000).
M. H. Anderson [*et al.*]{}, Science [**269**]{}, 198 (1995).
M.-O. Mewes [*et al.*]{}, Phys. Rev. Lett. [**78**]{}, 582 (1997).
E. W. Hagley [*et al.*]{}, Science [**283**]{}, 1706 (1999).
I. Bloch [*et al.*]{}, Phys. Rev. Lett. [**82**]{}, 3008 (1999).
G. Cennini [*et al.*]{}, Phys. Rev. Lett. [**91**]{}, 240408 (2003)
W. Guerin [*et al.*]{}, Phys. Rev. Lett. [**97**]{}, 200402 (2006)
A. Couvert [*et al.*]{}, Europhys. Lett. [**83**]{}, 50001 (2008).
J. T. Schulte [*et al.*]{}, Phys. Rev. Lett. [**95**]{}, 170411 (2005).
D Clément [*et al.*]{}, New J. Phys. [**8**]{}, 166 (2006).
P.W. Anderson, Phys. Rev. [**109**]{}, 1492 (1958).
P. A. Lee and T.V. Ramakrishnan, Rev. Mod. Phys. [**57**]{}, 287 (1985).
B. van Tiggelen, in [*Diffuse Waves in Complex Media*]{}, edited by J.-P. Fouque, NATO Advanced Study Institutes, Ser. C, Vol. 531 (Kluwer, Dordrecht, 1999).
I. M. Lifshits, S. Gredeskul, and L. A. Pastur, Introduction to the Theory of Disordered Systems (Wiley, New York, 1988).
J. Billy [*et al.*]{}, Nature [**453**]{}, 891 (2008).
G. Roati [*et al.*]{}, Nature [**453**]{}, 895 (2008).
L. Sanchez-Palencia [*et al.*]{}, Phys. Rev. Lett. [**98**]{}, 210401 (2007)
E. Gurevich and O. Kenneth, Phys. Rev. A [**79**]{}, 063617 (2009); P. Lugan [*et al.*]{}, Phys. Rev. A [**80**]{}, 023605 (2009).
R. C. Kuhn [*et al.*]{}, Phys. Rev. Lett. [**95**]{}, 250403 (2005)
C. Miniatura [*et al.*]{}, Eur. Phys. J. B [**68**]{}, 353 (2009).
M. Robert-de-Saint-Vincent [*et al.*]{}, Phys. Rev. Lett. [**104**]{}, 220602 (2010).
R. C. Kuhn [*et al.*]{}, New J. Phys. [**9**]{}, 161 (2007).
D. Čapeta [*et al.*]{}, Phys. Rev. A [**84**]{}, 011801(R) (2011).
M. Piraud, A. Aspect, L. Sanchez-Palencia, arXiv:1104.2314.
| {
"pile_set_name": "ArXiv"
} |
=-1cm =3.cm
\
\
\
\
Wuppertal, Juni 1995\
WU B 95-07\
DELPHI 95-80 PHYS 515\
hep-ex/951127\
[**Abstract**]{}\
\
Event shape and charged particle inclusive distributions determined from 750 000 hadronic Z events measured with the DELPHI detector at LEP are presented. The statistical and systematic precision of this data allows for a decisive confrontation with Monte Carlo models of the hadronization process and a better understanding of the structure of the Z hadronic final state.
Improved tunings of the JETSET, ARIADNE and HERWIG parton shower models and the JETSET matrix element model are obtained by fitting the models to identified particle distributions from all LEP experiments and the DELPHI data presented. The description of the data distributions by the models is critically reviewed with special importance attributed to identified particles. \
\
\
\
Wuppertal, Juni 1995\
WU B 95-07\
DELPHI 95-80 PHYS 515\
Introduction {#intro}
============
Precision measurements at LEP using the hadronic final state, such as determinations of the strong coupling constant $\alpha_s$ from event shapes, the measurement of the $Z$ mass and width, the forward backward asymmetries for quarks or at higher energies the $W^\pm$ mass require a precise modelling of the properties of the corresponding final states. Perturbative QCD cannot provide full theoretical insight to the transition of the primary quarks to observable hadrons, the so-called fragmentation or hadronization process. Only part of this transition including large momentum transfer, mainly the radiation of hard gluons or the evolution of a parton shower are calculable perturbatively. The final formation of hadrons is hidden due to the increase of the strong coupling constant $\alpha_s$ at small momentum transfer and hence the failure of perturbation theory.
Guidance to a better understanding of the hadronization process must therefore come from detailed experimental investigations of the hadronic final state including attempts to describe this process by phenomenological models inspired by QCD. LEP I is a unique unrivaled place to pursue these studies. The clean well defined initial state in $e^+e^-$ annihilation provides an excellent testing field, since the event rate at the $Z$ is huge, the energy is large, and the capabilities of the experimental apparatus are much improved with respect to previous experiments.
This paper attempts to determine parameters for the most frequently used hadronization models which lead to an optimal description of the observed hadronic event shapes and charged particle inclusive distributions, as measured with the DELPHI experiment at LEP, as well as the available information on identified particles from all LEP experiments. The latter allows one to precisely determine more model parameters than from event shapes only and also to check the internal consistency of the models. The performance of these models is compared and critically reviewed.
This paper is organized as follows: Section \[exp\] gives a brief overview to the relevant detector components and describes the experimental procedure applied to determine event shape and inclusive distributions and the related systematic errors. Section \[models\] discusses the models employed and the relevant parameters. Section \[fit\] describes the optimization strategy applied to obtain best parameters for the fragmentation models and justifies the choice of distributions used in the fit. The fits are discussed in detail and the resulting optimized parameters and their errors are presented. In Section \[results\] a confrontation of the models to the event shape, charge particle spectra and identified particle data is presented. Finally we summarize in Section \[summary\]. The appendices contain the definitions of the variables used throughout this paper, tables of model parameter settings and tables and plots of the relevant data distributions and model comparisons.
Detector and Data Analysis {#exp}
==========================
The data used in this analysis are from the 1991, 1992 and 1993 data taking with the DELPHI detector at LEP. For the determination of the event shape distributions we use charged particles measured in the solenoidal 1.2 T magnetic field of DELPHI and neutral clusters measured with the electromagnetic or hadronic calorimeters. The following detectors are relevant to the analysis [@detector_paper]:
- The Vertex Detector VD, measuring the $R\varphi$ coordinate with up to three layers of silicon micro strips at radii between 6.3 cm and 11 cm. It covers polar angles $\theta$ between $37^\circ$ and $143^\circ$ .
- The Inner Detector ID, a 24 layer cylindrical jet chamber ($\theta$ coverage $17^\circ$ to $163^\circ$ ).
- The Time Projection Chamber TPC, the principal tracker of DELPHI, with twice 6 sector plates, 24 pad rows and 192 sense wires each. Inner and outer radius of the TPC are $30$ cm and $122$ cm, the polar coverage is $20^\circ$ to $160^\circ$.
- The Outer Detector OD, a five layer drift chamber at $192$ cm radius covering polar angles between $43^\circ$ and $137^\circ$.
- Two sets of forward planar drift chambers FCA and FCB with 6 or 12 layers respectively with a overall polar angle coverage between $11^\circ$ to $35^\circ$ and $145^\circ$ to $169^\circ$.
- The High density Projection Chamber HPC, a lead-gas electromagnetic calorimeter with a very good spatial resolution inside the DELPHI coil between $208$ cm to $260$ cm radius. It measures the showers using the TPC principle and covers polar angles between $43^\circ$ and $137^\circ$. The precision on the energy can be described by $\Delta E/E \approx 29\%/\sqrt{(E)}+4\%$.
- The Forward Electromagnetic Calorimeter FEMC, a lead-glass array in both endcaps each consisting of 4500 lead glass blocks with a pointing geometry. The angular acceptance is between $10^\circ$ and $36.5^\circ$ to the beam. The precision on the energy is $\Delta E/E \approx \sqrt{(0.35+5/\sqrt{E})^2 + (6/E)^2}~\%$.
- The Hadron Calorimeter HAC, a iron-gas hadronic calorimeter outside the coil consisting of at least 19 layers of streamer tubes and $5$ cm thick iron plates also used as flux return. The overall angular covarage is $11.2^\circ$ to $168.8^\circ$. The precision on the energy is $120\%/\sqrt{E}$.
Depending on the polar angle and the detectors included in the track fit the precision on the momentum, $p$, for charged particles in hadronic events is $\Delta p / p^2 = 0.001$ to $0.01 GeV^{-1}$. Charged particles were accepted in the analysis if they satisfy the following criteria:
- $p \geq 200 MeV$
- $\Delta p /p \leq 1$
- $20^\circ \leq \theta_{track} \leq 160^\circ$
- Measured track length $ \geq 50$ cm.
- Impact parameter with respect to the nominal interaction point within $2$ cm perpendicular or $5$ cm along the beam.
Furthermore we require that charged particles with large momenta ($p \geq 25$ GeV$)$ within the geometrical acceptance of the OD or FCB have indeed been measured by these detectors as well as by the ID or VD. This requirement assures a good momentum resolution for high momentum particles.
Neutral clusters or photons reconstructed from conversions were accepted if:
- $1\, GeV \leq E_{cluster}$ for clusters measured with the HAC,
- $0.5\, GeV \leq E_{cluster/ \gamma}$ for electromagnetic clusters or photons.
Because the available statistics is very large, the final experimental error is dominated by the systematic error. Therefore cuts were applied to the event kinematics to assure that the major components of the event were measured in DELPHI with optimal efficiency and resolution, as well as to minimize secondary interactions in the detector material. However care has been taken not too strongly bias the measured distributions by these cuts. Events were selected if they fulfil the following conditions:
- there were at least 5 charged particles selected,
- the total energy of charged particles exceeds 15 GeV and 3 GeV for each half of the detector, defined by the (x,y) plane,
- the polar angle of the thrust axis is between $50^\circ$ and $85^\circ$,
- the momentum imbalance of the event along the beam direction is $|\sum p_z| / \sqrt{s} \leq 0.15$
In total about $750.000$ events satisfy these cuts. The contamination of beam gas events, $\gamma\gamma$-events, and leptonic events other than $\tau^+\tau^-$, is expected to be less than 0.1% and has been neglected. The influence of $\tau^+\tau^-$ events which have a pronounced 2-jet topology and contain high momentum particles has been determined by a simulation study using events generated by the KORALZ model [@koralz] treated by the full simulation of the DELPHI detector DELSIM [@delsim] and the standard data reconstruction chain. The $\tau^+\tau^-$ contributions have been subtracted from the measured data according to the relative abundance of $\tau^+\tau^-$ ($0.16\pm 0.03$%) and hadronic events.
We present differential event shape and inclusive single particle distributions as a function of the physical observables defined in appendix A normalized to the number of hadronic Z decays. Three different types of results are given, cross-sections measured from charged particles only corrected to the initial charged final state, or to the charged plus neutral final state and measured charged plus neutral particles corrected to the full final state. The observed data distributions were corrected for kinematic cuts, limited acceptance, and resolution of the detector as well as effects due to reinteractions of particles inside the detector material. This correction has been calculated using simulated events generated by JETSET 7.3 [@jetset] treated by the full simulation and analysis chain as described above.
For each bin a correction factor $C_i$ has been calculated as the ratio of the generated and final distributions. Particles with a lifetime bigger than 1 ns were considered as stable in the generated distributions. The correction for initial state photon radiation has been determined separately using events generated by JETSET 7.3 PS with and without initial state radiation as predicted by DYMU3 [@dymu3].
The simple unfolding by correction factors in general leads to biases of the final results when the detector smearing is bigger than the bin width used, and when the model does not describe the data well [@unfolding]. In our case the model has been tuned to DELPHI data as described in [@fuerstenau] and in this paper and is in good agreement for all distributions considered. The relevant parameter settings are given in appendix \[delphi\_par\]. To keep residual biases small with respect to other systematic errors we ensure that the bin width of all presented distributions is at least as big as the detector resolution.
To account for a possible imperfect representation of the DELPHI detector or secondary processes in the simulation program DELSIM, the cuts given above have been varied over a wide range, including small polar angles for the event axis, demanding events with more than 7 tracks etc. Further cuts have been imposed to exclude the boundaries of the TPC sectors for high momentum particles where the agreement between data and simulation is less good. From the stability of the measured distributions a systematic uncertainty as function of the physical observables has been deduced as the R.M.S. of the deviation with respect to the published central value. As the systematic error is expected to grow proportional to the deviation of the overall correction factor from unity an additional relative systematic uncertainty of 10% of this deviation has been added quadratically to the above value. A further systematic error has been added in quadrature for a few bins where the results of the individual data sets, corresponding to the different years of data-taking were found to be statistically incompatible. This error has been calculated such that the $\chi^2$ per degree of freedom for the merging was 1 when this error was considered in addition. The individual estimates of the systematic uncertainty show similar trends. This a posteriori justifies the methods used, however also indicates correlations between the error estimates which could lead to an overestimate of the systematic uncertainty. The final systematic uncertainty has been smoothened in dependence of the individual variables.
The uncertainties shown in the graphs in Appendix \[plots\] comparing data and models are final experimental uncertainties adding statistical and systematic uncertainties in quadrature. With the exception of the point-to-point scatter they should be interpreted like a statistical uncertainty. The individual errors are given in the tables in appendix \[tables\] together with the cross-sections.
Monte Carlo Models {#models}
==================
In this paper we consider the most frequently used fragmentation models, namely JETSET 7.3/7.4 PS and ME [@jetset], ARIADNE 4.06 [@ariadne] and HERWIG 5.8 c [@herwig]. HERWIG and JETSET are complete models describing the parton shower evolution (or matrix element calculation), the hadronization of partons into hadrons, and the subsequent decays of short lived particles. ARIADNE only models the parton shower, the consequent hadronization and decays are treated by the corresponding JETSET routines.
JETSET 7.3 / 7.4 Parton Shower {#jetsetps}
------------------------------
JETSET with the parton shower option is used with the settings as given in table \[jetset\_switch\].
Variable Value Variable Value Variable Value
----------- ------- ----------- ------- ---------- -------
MSTJ(11) 3 MSTJ(12) 3 MSTJ(41) 2
MSTJ(45) 5 MSTJ(46) 3 MSTJ(51) 0
MSTJ(101) 5 MSTJ(107) 0
: \[jetset\_switch\] Setting of JETSET PS Switches
The JETSET parton shower algorithm is a coherent leading log algorithm (LLA) with angular ordering. The shower evolves in the CM frame of the partons obeying energy momentum conservation at each step of the shower. The lowest order 3-jet cross-section is reproduced by rejecting part of the first branchings of the initial $q\bar{q}$ system as predicted by the LLA formalism. Angular ordering of the branchings is explicitely imposed and gluon helicity effects can be included. $\alpha_s$ is running with a scale given by the transverse momentum squared of the branching. The shower evolution is stopped at a mass scale $Q_0$, then fragmentation takes over. $Q_0$ and $\alpha_s$ (i.e. $\Lambda_{QCD}$) are parameters of the parton shower part of JETSET.
The fragmentation is performed using the Lund string scheme which can be formulated as an iterative procedure. A string stretches in between the oppositely coloured quark and antiquark via the gluon colour charges. Two close by gluons act similary to a single gluon with equal momentum. Therefore the string model is infrared safe. The longitudinal momentum fraction $z$ of a hadron is determined using the Lund symmetric fragmentation function: $$f(z) = \frac{(1-z)}{z}^a \cdot \exp{-\frac{b \cdot m_t^2}{z}}$$ $m_t^2 = m^2 + p_t^2$ is the transverse mass squared of the hadron. $a$ and $b$ are parameters of the fragmentation function. $b$ is universal, $a$ in principle can depend on the quark flavour. For heavy quarks we instead use the Peterson fragmentation function [@pet83]: $$f(z) = \frac{1} {z \left ( 1 - \frac{1}{z} -\frac{\epsilon_q}{1-z} \right )^2}$$ with parameters $\epsilon_{b(c)}$ which gives a better description of heavy quark fragmentation. The transverse momenta of hadrons are determined from the $p_t$ of its constituent quarks which in turn is chosen from a tunneling process. This leads to a exponential $p_t$ distribution, thus a Gaussian $p_t^2$ behaviour. The relevant parameter is the standard deviation of this distribution, $\sigma_q$.
The tunneling also determines the quark flavour generated in the string breakup, leading to a dependence and thus negligible heavy quark production in the fragmentation. Already strangeness production is strongly suppressed: $P(u\bar{u}):P(d\bar{d}):P(s\bar{s}) = 1:1:\gamma_s \approx 0.3$. Mesons are produced according to their quark contents in the six multiplets with smallest mass, i.e. in the states: $^1S_0$, $^3S_1$, $^1P_1$, $^3P_0$, $^3P_1$ and $^3P_2$. Contrary to the formulation implemented in JETSET we define individual production probabilities for these multiplets and light, strange or (both) heavy flavours. Except for the light $^3P_0$ multiplet the probabilities for the P-multiplets are taken $\propto (1-P(^1S_0)-P(^3S_1)) \cdot (2s+1)$. In a (2 dimensional) string picture production of particles with angular momentum (i.e. $P$-states) is suppressed and expected to be small (10% [@lund_rev]).
Using additional mass relations the tunneling mechanism is also applied to baryon production (replacing a quark by a diquark). Parameters related to baryon production are the relative diquark production rate $P(qq)/P(q)$, an extra strange diquark suppression $P(us)/P(ud)/\gamma_s$ and an extra suppression $P(ud1)/P(ud0)$ of spin 1 diquark relative to spin 0 ones leading to $s=3/2$ and $s=1/2$ baryons respectively. Furthermore it turned out to be necessary to include an extra suppression of leading baryons. This extra suppression is not used in heavy quark fragmentation and in the simulation of heavy particle decays by fragmentation because here it leads to soft baryon spectra and a strong overall suppression of baryons.
The parameter related to baryon-meson-baryon production, the so-called Popcorn parameter we leave at its default value, also because so far experimental determinations of this parameter are not fully conclusive (see [@ak0] and references therein). Furthermore we do not include simulation of Bose Einstein interference (by LUBOEI) although with properly chosen parameters this procedure results in good representation of the correlation functions for identical particles as well as in a strongly improved description of light meson resonance lineshapes. The energy momentum rescaling performed in this routine strongly influences angular distributions between particles and multi jet rates. Inclusion of this routine leads to widely different model parameters (compare e.g. the resulting parameters given in section \[fit\] (tables \[terg73d\],\[terg74d\]) and the parameters of the DELPHI simulation (see table \[delphi\_tuning\] ) which includes Bose Einstein interference).
For JETSET 7.3 and a tuning of ARIADNE 4.06 heavy particles decays have been modified to obtain a better description of the heavy particle branching fractions. The modifications act similarly to those performed in JETSET 7.4.
JETSET 7.4 Matrix Elements
--------------------------
The parton shower simulation is replaced in the historically older matrix element version of JETSET by the exact second order matrix element calculation which provides up to 4 partons. Two calculations GKS [@GKS] and ERT [@ERT] are available in JETSET. In this paper we consider only the default ERT option because it is expected to be more exact. At PETRA/PEP the predicted 4-jet rate turned out to be too small [@sj_yb] for a given 3-jet rate. This has been connected with higher order terms missing in the second order calculation. These terms can be partially accounted for by choosing a suitable renormalization scale according to the “optimal perturbation theory” description [@stevenson] by choosing a new scale $Q^2 = \mu \cdot s, \mu \leq 1$. To obtain an optimal description we fit the parameter $\mu$ together with the $\Lambda_{QCD}$ parameter and the fragmentation parameters as described in the JETSET PS section \[jetsetps\]. Formally the scale parameter $\mu$ replaces the cutoff parameter $Q_0$ in the ME fit. JETSET 7.4 ME parameter settings are given in table \[jetset\_me\_switch\].
Variable Value Variable Value Variable Value
----------- ------- ----------- ------- ----------- -------
MSTJ(11) 3 MSTJ(12) 3 MSTJ(41) 2
MSTJ(45) 5 MSTJ(46) 3 MSTJ(51) 0
MSTJ(101) 2 MSTJ(107) 0 MSTJ(111) 1
: \[jetset\_me\_switch\] Setting of JETSET ME Switches
ARIADNE 4.06
------------
ARIADNE is a particulary elegant formulation of a parton shower based on colour dipoles [@leningrad; @ariadne]. The emission of a gluon from a colour dipole, i.e. the initial quark antiquark pair creates two new dipoles, one stretched from the quark to the gluon and one from the gluon to the antiquark. Both in turn can independently radiate further gluons. This ansatz automatically includes ordering in angle (or transverse momentum) as well as azimuthal dependencies in a proper way. The dipole chain resembles the Lund string picture. Parameters are the QCD scale parameter $\Lambda_{QCD}$ and the cut off scale $p_t^{min}$ which corresponds to $Q_0$ for JETSET. The evolution variable is the transverse momentum squared. As for JETSET the first order 3 jet cross section is reproduced in ARIADNE. The ARIADNE running parameters are set as in table \[ariadne\_switch\].
Variable Value Variable Value Variable Value
----------- ------- ----------- ------- ----------- -------
MSTA(1) 1 MSTA(2) 1 MSTA(3) 0
MSTA(5) 1 MSTA(12) 1
MSTJ(11) 3 MSTJ(12) 3 MSTJ(41) 0
MSTJ(45) 5 MSTJ(46) 3 MSTJ(51) 0
MSTJ(101) 5 MSTJ(105) 0 MSTJ(107) 0
: \[ariadne\_switch\] Setting of ARIADNE Switches
HERWIG 5.8c
-----------
The evolution of the parton shower in HERWIG is based on the Coherent Parton Braching formalism, an extension of the LLA. It accounts for the leading and sub-leading logarithmic terms arising from soft and (or) collinear gluon emission. HERWIG pays special attention to the simulation of QCD interference phenomena [@herwig2]. Most important parameters of the parton shower algorithm are $\Lambda_{QCD}$ (QCDLAM), the quark masses (RMASS(1-6)) and the effective gluon mass (RMASS(13)). The parton shower in HERWIG 5.8c is matched with the first order 3 jet cross section. At the end of the parton shower evolution gluons are split nonperturbatively into $q\overline{q}$-pairs.
The hadronization in HERWIG proceeds via the so-called “cluster algorithm” based on the preconfinement characteristic of QCD. The colour charge of a parton is compensated to leading order by a anticolour object which is close by in phase space. Combining colour and anticolour objects low mass clusters with no colour are formed. Higher mass clusters are further split into two lighter ones. The splitting is controlled by the parameters CLMAS and CLPOW. In the decay of a cluster containing a quark from the perturbative phase the direction of this quark is remembered. A gaussian smearing is applied controlled by the parameter CLSMR.
Hadrons then are formed in two body cluster decays according to phase space and spin factors. The particle and hadron transverse momentum are thus dynamically produced as a consequence of the cluster mass spectrum. Particle production in cluster decays is modified by changing the a priori weights for the individual hadron types. These are VECWT, TENWT [^1] and DECWT for vector or tensor mesons and decuplet baryons respectively. The a priory weights for quarks are PWT(6).
Light particle decays are simulated in HERWIG using decay tables. Particles including heavy quarks decay via quark decay and subsequent fragmentation.
Relevant parameter settings for the tuning of HERWIG are given in table \[herwig\_switch\].
Variable Value Variable Value Variable Value
---------- ------- ---------- ------- ---------- -------
IPROC 100 SUDORD 1 CLDIR 1
: \[herwig\_switch\] Setting of HERWIG parameters
Fit of Monte Carlo Models to Experimental Data {#fit}
==============================================
Classical optimization strategies like hill climbing methods in general fail to converge if they are directly applied to the optimization of a Monte Carlo model because the physical observables predicted by the models are defined only on a statistical basis (and thus only known within the statistical errors). Moreover, this straight forward strategy requires much computer time and cannot be easily repeated to change the input data to be fitted, or check the influence of systematic errors of the data etc. Therefore, similar to previous work [@a_fit; @t_fit; @l_fit] we chose to approximate the dependence of the physical observables on the model parameters analytically. For [*each bin of each distribution*]{} we use the quadratic form: $$\begin{aligned}
\label{expans}
MC(\vec{p}_0 + \delta\vec{p})(x) \approx f(\vec{p}_0 + \delta\vec{p})(x)
&=& a_0^{(1)}(x) + \sum\limits^{n}_{i=1} a^{(2)}_i(x) \delta p_i +
\sum\limits^{n}_{i,j=1} a^{(3)}_{ij}(x) \delta p_i \delta p_j\end{aligned}$$
Here $f(x)$ denotes the predicted distribution of a physical observable $x$, $\vec{p}_0$ is a central parameter setting in the $n$-dimensional parameter space and $\delta p_i$ the deviation of parameter i from this setting. The last term includes correlating terms between the model parameters. These terms are not present in a linear approximation as used in [@l_fit].
The optimal $m=\frac{n}{2}\cdot(n+3)+1$ parameters $a^{(k)}$ of the expansion are determined from a fit of ansatz \[expans\] to $l > m$ model reference distributions with different parameter settings. This fit is equivalent to the solution of a system of linear equations: $$\begin{aligned}
A \cdot \vec{c} = \vec{f} = \vec{MC}
\label{linear_eq}
$$ where column i of the matrix A contains the parameter variations of model set i: $$\begin{aligned}
A_{i,1..m} = ( 1, \delta p^{(i)}_1, ~...~ ,\delta p^{(i)}_n,
\delta p^{(i)2}_1, ~...~ , \delta p^{(i)2}_n,
\delta p^{(i)}_1 \cdot \delta p^{(i)}_2, ~...~,
\delta p^{(i)}_n \cdot \delta p^{(i)}_{n-1} )
\nonumber\end{aligned}$$ $\vec{c}$ is the vector of coefficients $a^{(k)}_{i(j)}(x)$, and $\vec{MC}$ the vector of model predictions corresponding to the parameters $\vec{p_0}+\delta\vec{p}^{(i)}$. The optimal solution of this, for $l \geq m$ overconstrained linear system is obtained using a standard singular value decomposition method [@svd; @nag].
The parameters of the $l$ reference models (generated with equal statistic) are randomly chosen in parameter space around the central point $\vec{p}_0$. Assuming the (a priori chosen) parameter intervals to be renormalized to $ \pm 1$ it turns out to be unimportant, except for minor differences in the statistical precision, whether this volume is a cube or a sphere or when the points are placed on the surface of the sphere. Our choice should ensure that the precision of the fitted linear function is roughly constant within the hypersphere.
Statistical errors of the simulated data sets have been neglected because all sets are generated with fixed statistics, and the variation of the distributions for the different parameter choices is relatively small. Consequently the exact statistical uncertainty of the linear approximation is unknown. However the model statistic is chosen such that the overall statistical error is small compared to the experimental uncertainty of the data.
The optimal parameters $p_i$, their errors $\sigma_i$, and correlation coefficients $\varrho_{ij}$ are then determined from a standard $\chi^2$-fit using MINUIT [@minuit] of the analytic approximations (\[expans\]) for [*all distributions and bins*]{} considered to the corresponding data. Note that the overall analytic approximation contains $n_{bins}\cdot m$ coefficients $a^{(k)}_{i(j)}(x)$.
If the sensitivity of a distribution to a given model parameter $p_i$ is negligible compared to the sensitivity to other parameters (see also next section), the dependence of the approximation on this parameter has been suppressed by omitting this parameter from the linear system (\[linear\_eq\]). This leads to a better convergence of the minimization and more robust results.
The method has been tested by generating $l+1=51$ simulated sets of 50.000 events each with 6 independent parameters and simultaneously fitting the parameters of one set using the other $l$ sets as reference. Here, the statistical error of the approximation is negligible compared to the statistical error of the test set. The pull distributions $(p_i^{fit} - p_i^{true})/\sigma_i$ are approximately standard normal distributions, i.e. the fitting method is unbiased and produces correct errors.
Choice of Distributions
-----------------------
The parameters of a fragmentation model have a well-defined physical meaning. However some parameters are directly coupled like $a$, $b$, and $\sigma_q$ in the Lund fragmentation function, or the effect of the parameter on physically observed quantities is obscured by other processes like decays. Therefore the choice of distributions to tune the model (parameters) to is not always evident. So far among the many possible distributions some have been choosen in an ad hoc way, and alternative changes of distributions have been used to estimate systematic errors of parameters [@fuerstenau],[@a_fit],[@l_fit],[@o_fit].
In practice, to keep the influence of statistical errors as small as possible it is clear that the models should be fitted to those distributions which show the strongest dependence on the parameter under consideration and least correlations with others. For each distribution $MC(x)$ we therefore calculated its sensitivity to a given model parameter, i.e. the quantity: $$\begin{aligned}
S_i(x) = \frac{\delta MC(x)}{MC(x)}
\Bigl|_{p_i}
\Bigl/ \frac{\delta p_i}{p_i} \approx
\frac{\partial \ln{MC(x)}}{\partial \ln{p_i}} \Bigr|_{p_i}\end{aligned}$$ where $\delta MC(x)$ is the change of the distribution $MC$ when changing the model parameter $p_i$ by $\delta p_i$ around its central value. The fraction $\frac{\delta p_i}{p_i}$ gives all parameters the same normalization.
It is observed, that the sensitivity of single particle inclusive distributions to all model parameters is in general larger than for event shape distributions. Compare for example the sensitivities of inclusive spectra and event shapes to JETSET parameters (figures \[sens3\] and \[sens1\]).
It is seen that almost no sensitivity to the parton shower cut off $Q_0$ is present in the event shape distributions. For this parameter the $x_p$-momentum spectrum at large $x_p$ is most important. Correlations among the individual parameters in the inclusive spectra are very strong as can be seen by similar (or opposite) behaviour of the sensitivities for different parameters. Especially the opposite, almost symmetric behaviour of the sensitivities on $a$ and $b$ explains the strong correlation between these parameters and why it is possible to find good descriptions for many different choices of $a$ and $b$. As has been expected, $\sigma_Q$ is best determined by the $p_t^{out}$ spectrum or related quantities. The variation of the sensitivities are smallest for the rapidity $y$. Another important quantity is the charged multiplicity which is know to high precision. Obviously it depends on very many model parameters and details of particle decays.
Event shape distributions measuring the overall shape like $T$, $S$ or $M_{high}^2/E^2_{vis}$ mainly depend on $\alpha_s$ i.e. $\Lambda_{QCD}$ (see \[sens1\]) except in the 2-jet region where also fragmentation effects are relevant.
The 3-jet rate as measured by $D_2^{Durham}$ or $D_2^{Jade}$ is best suited to determine $\Lambda_{QCD}$ because the sensitivities on other parameters is negligible. Thus the $\Lambda_{QCD}$-determination can almost be decoupled from the determination of other parameters. This also underlines the reliability of $\alpha_s$-determinations from jet-rates.
Distributions measuring the aplanarity of events like $A$, $m$ and also $M_{low}^2/E^2_{vis}$ again show increased sensitivities especially on $\alpha_s$ but also strong correlations between the parameters. This explains why quantities measuring differences of event shape variables like $O$, $M_{diff}^2/E^2_{vis}$ and the $AEEC$, contrary to popular belief, depend both on $\Lambda_{QCD}$ and also on many other fragmentation parameters.
For a comprehensive overview tables \[sensjt74\] and \[sensh58\] show the sensitivities to JETSET 7.4 and HERWIG 5.8 model parameters averaged geometrically over the bins of the distributions. The averaging dilutes part of the sensitivities (compare e.g. fig. \[sens3\] and table \[sensjt74\]).
From the above discussion we conclude that for a determination of the general parton-shower and fragmentation parameters the models should be fitted to:
- the inclusive (charged) particle distributions as function of $x_p$, $p_t^{in}$ and $p_t^{out}$ with respect to the thrust or sphericity axis,
- the differential three jet rate $D_2$ or alternatively $R_3$,
- a combination of event shape distributions like $T$, $M$, $m$, or $S$, $A$, $P$, or $M_{high}^2/E^2_{vis}$ and $M_{low}^2/E^2_{vis}$.
Any combination of event shape distributions or event axes for the $p_t$-spectra is in principle equivalent and can be used to estimate the stability of the fit and systematic errors of the parameters determined or can be viewed as check of the “predictive power” of the models. However because of the strong correlations among the model parameters correlations need to be considered.
The action of the model parameters linked directly to identified particle production (like the strange quark suppression, $\gamma_s$, or the relative probability to form a qq pair to make a baryon, $P(qq)/P(q)$, or others) usually follows the physical interpretation more directly. However identified particle spectra have also high sensitivity to fragmentation parameters (see tables \[sensjt74\] and \[sensh58\]). Therefore the concept described above to determine the optimal dependencies is used.
Strategy of the Fit
-------------------
### JETSET & ARIADNE
First we perform fits of the general fragmentation parameters ( $\Lambda_{QCD}$,$Q_0$ and $p_t^{cut}$ or $\mu$, $a$, $b$ and $\sigma_q$) to the charged particle inclusive and global event shape distributions. For each fit 50 simulated sets with 100.000 events are used. Parameters related to identified particles are set to values similar to those used for the DELPHI simulation (see table \[delphi\_tuning\]) which are already known to describe identified particle rates and spectra well.
The average scaled momentum $<x_E>$ of charm and beauty particles in the models only depends on $\epsilon_{c(b)}$ and $\Lambda_{QCD}$. Stable hadron spectra only weakly depend on heavy quark fragmentation parameters. To further reduce dependencies of stable particle distributions on heavy quark parameters and at the same time assure correct $<x_E>$ values for $D^*$ and $B$ mesons we adopt the following procedure: the $<x_E>$ values for $D^*$ and $B$ mesons for several choices of $\Lambda_{QCD}$ and $\epsilon_{c(b)}$ are determined within the model. Then the dependence of $\epsilon_{c(b)}$ on $<x_E>_{D^* (B)}$ and $\Lambda_{QCD}$ is parametrized. This allows to chose for each model set and its given $\Lambda_{QCD}$ value a corresponding value of $\epsilon_{c(b)}$ such that the model reproduces the average heavy meson scaled momenta as measured by the LEP experiments [@mean_x_heavy] ($<x_E>_{D^*}=0.504,\,\, <x_E>_{B}=0.701$).
The models were then fit to several combinations of event shape and semi-inclusive distributions. The results for $\Lambda_{QCD}$ and $\sigma_q$ were relatively stable, the variation for $Q_0$ or $p_t^{cut}$ was bigger and there were many different solutions for $a$ and $b$. The latter is due to the strong correlation between these two parameters. Therefore from now on we take one central value of $b$ and treat only $a$ as variable. $a$ is prefered because it is less directly coupled to $\sigma_q$ in the Lund fragmentation function.
Parameters relevant only to the production of specific particles were then adjusted or fitted to the related data. In overview these are:
- The extra $\eta$ and $\eta '$ suppressions were fitted to data from [@leta],[@aeta].
- The probability for the different B meson multiplets were adjustet to agree with recent measurements [@oB**], [@dB**]. The corresponding D mesons probabilities were interpolated between the B and light meson values.
- $P(us)/P(ud)/\gamma_s$, The strange baryon suppression was adjusted to the ratio of proton- and $\Lambda^0$-production [@ak0],[@ap],[@dp],[@op].
- $P(ud1)/P(ud0)$ the spin 1 diquark suppression was fitted to the ratio of $\Sigma(1385)$ and $\Lambda^0$ or proton production [@ak0],[@ok0],[@ap],[@op],[@dp].
- The leading baryon suppression was adjusted to the large momentum tail of the proton and $\Lambda^0$ spectra [@dlambda],[@ak0] ,[@ostr_bary].
Then a simultaneous fit of 10 important parameters (see parameters where generated ranges are specified in tables \[terg73d\] – \[terg74m\]) was prepared generating 100 sets of 100.000 events each. The analytical approximation obtained from these simulated sets was fitted to different choices of event shape and particle distributions. For the event shapes we chose the combinations of variables shown in table \[evtfit1\]. The choices for identified particles are shown in table \[evtfit2\]. These were used to account for the discrepancies observed in the in the proton data from different experiments and the imperfect representation of the $K^{\pm}$ spectra in the models.
-------------------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- ---------------
distribution 0 1 2 3 4 5 6 7 8 9 10 11
$<N_{ch}>$ [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$x_p$ [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$y_T$ [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$y_S$ [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$p^{in}_t,p^{out}_t\,\,\, (T)$ [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$p^{in}_t,p^{out}_t\,\,\, (S)$ [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$S,\,\,A,\,\,P$ [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$T,\,\,M,\,\,m$ [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$D^{Durham}_2$ [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$D^{Jade}_2$ [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$D^{Durham}_3$ [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$D^{Jade}_3$ [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$D^{Durham}_4$ [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$D^{Jade}_4$ [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
-------------------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- ---------------
: \[evtfit1\] Overview of the combinations of inclusive distributions and shape distributions fits of the JETSET, ARIADNE and HERWIG models. Distributions used are marked\
by [$\bullet$]{}.
---------------- ----------------------------------------------- --------------- --------------- --------------- --------------- ---------------
0 1 2 3 4
$\rho^\circ$ DELPHI[@drho] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$\omega$ L3[@lomega] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$f_0$, $f_2$ DELPHI[@drho] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$K^0$ ALEPH[@ak0], OPAL[@ok0] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$K^\pm$ ALEPH[@ap], DELPHI[@dp], OPAL[@op] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$K^{\ast 0}$ OPAL[@okstar0] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$K^{\ast \pm}$ ALEPH[@akstar], DELPHI[@drho], OPAL[@okstar0] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$\Phi$ DELPHI[@dp], OPAL[@op] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$p$ ALEPH[@ap],DELPHI[@dp] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$p$ OPAL[@op] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
---------------- ----------------------------------------------- --------------- --------------- --------------- --------------- ---------------
: \[evtfit2\] Overview of the combinations of identified particle data used for the JETSET and ARIADNE fits. Data used are marked by [$\bullet$]{}.
The $\Lambda^0$ data has been excluded from the fit because at small momenta it is not described by the models. Inclusion of this distribution leads to ill defined fits and unstable results. The proton data was only used to determine the diquark suppression parameter $P(qq)/P(q)$ and was excluded from the $\chi^2$ calculation for the fits of the remaining parameters. Baryons at intermediate and large momenta are likely, within the models, to be primary particles. Therefore these data strongly influence the fragmentation function and the related parameters. Since proton and $\Lambda^0$ spectra appear poorly described by the models and because it is even necessary to modify the fragmentation function by an extra suppression at large momenta this strong impact on the fit results is felt to be unphysical and therefore has been excluded.
Separate fits were performed for all combinations of choices for shapes and identified particles. To check the stability of the fits the optimization has been started with 6 random start values of fragmentation parameters. The fits were stable and converged to the same solution for about 95% of the cases. Differences mainly showed up for two strongly correlated parameters. For those fits which show more than 1 solution we used the one with the better $\chi^2$. A full (MINOS [@minuit]) error estimate for all parameters was performed.
The central result corresponds to the combination S6/P0. This combination has been choosen because it contains event shape distributions linear and quadratic in the particle momenta, the charged multiplicity and all relevant identified particle information and thus should result in a comprehensive overall description. Results are given with the statistical and systematic errors in tables \[terg73d\] – \[terg74m\]. The results for the individual parameters are strongly correlated. The correlation coefficients for the JETSET 7.4 (default decays) and ARIADNE 4.06 (DELPHI decays) are given representatively in appendix \[correl\_tabs\]. Beside these statistical correlations further correlations exist due to the different possible choices of input data. The systematic error is the R.M.S obtained from all combinations of input data which turn out to give parameters smaller respectively bigger than the central fit.
The following observations were made comparing results of the different input data choices for the JETSET PS fits:\
As should be expected the results for $\Lambda_{QCD}$, $Q_0$ and $\sigma_q$ are almost independent of the choice of identified particle data. The resulting value of $\Lambda_{QCD}$ does not depend on the algorithm used (JADE or DURHAM) if only $D_2$ is included in the fit however if also the higher jet rates were used the $\Lambda_{QCD}$ value turns out to be somewhat bigger ($\approx 8\%$) in case of the JADE algorithm. Values of $\Lambda_{QCD}$ and $Q_0$ are positively correlated. This implies that the number of final partons is more stable within the models than might be expected from the error of $Q_0$ alone. $\Lambda_{QCD}$, and $a$ and $\sigma_q$ are anticorrelated. This is due to the a compensation of transverse momentum generated in the parton shower and fragmentation phase of the model.\
The results for $\gamma_s$ are higher if $K^{\pm}$ and $K^0$ data are included compared to $K^0$ only. $P(qq)/P(q)$ is larger if the ALEPH proton spectrum is included. Production parameters for strange vector and pseudoscalar mesons are anticorrelated. If the charged multiplicity is not included and therefore the predicted multiplicity is smaller than the measured result, the primary production probabilities for light pseudoscalar and vector mesons are $P(^1S_0)_{ud} \approx 0.40$ and $P(^3S_1)_{ud} \approx 0.26$ respectively. If the multiplicity is fixed to $N_{ch} \approx 20.9$ these values are $0.28$ and $0.29$. This implies substantial production probabilities for light p-wave mesons of $0.36$ to $0.43$.
Most of these observations similary also apply in case of ARIADNE. However here the different choices of $D_2, D_3$ etc. lead to stable results for $\Lambda_{QCD}$. There is a tendency to obtain slightly bigger values from the JADE algorithm (0.245) then from Durham (0.237). This goes again conform with bigger values for $p_t^{QCD}$ from JADE ($\approx 0.9$) then from DURHAM ($\approx 0.6$).
For JETSET ME $\Lambda_{QCD}$ is anticorrelated with the scale $\mu$. The fragmentation parameters have different values as in the PS case. Especially $\sigma_q \approx 0.48 $ is much bigger. This partially compensates missing higher orders in the ME ansatz. The values of $\gamma_s$ and $P(qq)/P(q)$ are smaller compared to the PS case and show a decreased dependence on the choice of the related input data. The probability for p-wave states is much bigger $\approx 0.36$ already for the fits with free multiplicity.
### HERWIG
HERWIG employs much less parameters than JETSET especially in the hadronization sector of the model. Therefore a simultaneous fit of all model parameters which are found to be important ( compare table \[sensh58\] ) is easily performed. These parameters are:
- the QCD scale parameter QCDLAM and the gluon mass RMASS(13) as major parameters of the PS phase,
- the cluster fragmentation parameters CLPOW, CLMAX and CLSMR (CLDIR=1) and
- the a priori weights DECWT for decuplet baryons, PWT(3) for strange quarks and PWT(7) for diquarks.
Beside the particle spectra the cluster parameters also strongly influence the stable charged particle and event shape distributions. Fits have therefore been performed to combinations of shape distributions (see table \[evtfit1\]) and identified particle data (see table \[evtfit3\]). Of special importance is the simultaneous usage of proton and $\Lambda^0$ data. The average scaled momenta of heavy mesons are important for the determination of CLPOW and CLMAX.
\[t\]
-------------------------------------- ----------------------------------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- ---------------
0 1 2 3 4 5 6 7 8 9
$\rho^\circ$ DELPHI[@drho] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$\omega$ L3[@lomega] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$f_2$ DELPHI[@drho] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$K^0$ ALEPH[@ak0], OPAL[@ok0] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$K^\pm$ ALEPH[@ap], DELPHI[@dp], OPAL[@op] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$K^{\ast 0}$ OPAL[@op] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$K^{\ast \pm}$ ALEPH[@akstar], DELPHI[@drho], OPAL[@okstar0] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$\Phi$ DELPHI[@dp], OPAL[@op] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$\Lambda^0$ ALEPH[@ak0],DELPHI[@dlambda] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$\Lambda^0$ OPAL[@okstar0] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$p$ ALEPH[@ap],DELPHI[@dp] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$p$ OPAL[@op] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$\eta$ ALEPH[@aeta] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$\eta$’ ALEPH[@aeta] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$<x_E>\,\, D^{\ast\pm}, D^{\ast 0}$ [@mean_x_heavy] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$<x_E>\,\, B^{0}, B^{\pm}$ [@mean_x_heavy] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$\Sigma^\pm (1385)$ DELPHI[@dstrange], OPAL [@okstar0] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$\Xi^-$ DELPHI[@dstrange], OPAL [@okstar0] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
$\Xi_0(1530)$ DELPHI[@dstrange], OPAL [@okstar0] [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{} [$\bullet$]{}
-------------------------------------- ----------------------------------------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- --------------- ---------------
: \[evtfit3\] Overview of the combinations of identified particle data used for the HERWIG fits. Data used are marked by [$\bullet$]{}.
The different choices of the identified particle data (see table \[evtfit3\]) are made such that systematic differences in the data are reflected in the systematic errors of the parameters. The central fit result for HERWIG corresponds to the combination $S_6P_5$.
Following observations are made comparing results of the different input data choices for the HERWIG fits:\
Values obtained for QCDLAM are in general bigger ($\approx 0.01$) when fitting DURHAM compared to JADE jet rates. Contrary to the JETSET case QCDLAM is smaller (by $\approx 0.005$) when differential 2-, 3-, and 4-jet rates are fitted compared to 2-jet only. QCDLAM is insensitive to the inclusion of the charged multiplicity in the fit. The gluon mass RMASS(13) shows some dependency on the selected identified particle information and on the multiplicity. The cluster parameters CLMAX and CLPOW depend on the identified particle spectra only in case that the multiplicity is not included in the fit. CLPOW is higher for JADE than for DURHAM jet rates. CLSMR depends only on the charged particle and event shape information. The inclusion of the baryon decuplet data tends to spoil the description of the octet sector.
Comparison of Models to Data {#results}
============================
The fits of the individual fragmentation models are compared to DELPHI data in figures \[bildxpp\] - \[bilddjb\] shown in appendix \[plots\]. The lower insets of the plots depict the relative deviation of the models to the data. Also shown as shaded area in these insets is the total experimental error obtained by adding quadratically the systematic and statistical error in each bin. Except for the point-to-point scatter the error should be interpreted like a statistical ”$1 \sigma$” uncertainty. For most shape distributions we show on the left side the data as measured from charged particles only, on the right measured from charged and neutral particles. The distributions are corrected to the corresponding final states.
We restrict to comparisons to the following models:
- JETSET 7.3 DELPHI decays labeled JT 7.3 PS
- JETSET 7.4 default decays labeled JT 7.4 PS
- ARIADNE 4.06 DELPHI decays labeled AR 4.06
- HERWIG 5.8 c default decays labeled H 5.8C
- JETSET 7.4 ME default decays labeled JT 7.4 ME
The different decay treatments lead only to negligible differences for the stable charged particle and event shape distributions.
[**Inclusive Charged Particle Spectra**]{}\
All models describe well the general trends of the data. Almost no discrepencies show up from a direct data model comparison. More quantitatively the comparison of model and data (lower insets) shows:\
The $x_p$-spectrum (fig. \[bildxpp\]) for $x_p < 0.4$ is almost perfectly described by the ARIADNE and JETSET PS models. At large $x_p$ these models slightly underestimate the data. This trend is reduced if the multiplicity is left free in the fit. HERWIG and JETSET ME exhibit a slight wave structure when compared to the data. This structure is also reflected in the rapidity distributions (see fig. \[bildyts\]).\
Although the $p^{in}_t$-distribution (fig. \[bildptin\]) is in agreement to data for all models the tail of the $p^{out}_t$-distribution (fig. \[bildptout\]) ($p^{out}_t > 1$) disagrees with the data for all models. This tail is almost unaffected by fragmentation but is sensitive to the partonic part of the models. For the ME model a discrepancy might be expected because of missing higher order terms in the second order matrix element calculation. The failure of the parton shower models can possibly be traced back to missing large angle terms in the LLA basic to the models. Accepting these reasons for the failure of the models it should be expected that a matching of the second order calculation and the LLA formalism should lead to an improved description of $p^{out}_t$ and related distributions.
[**Shape Distributions**]{}\
The general shape distributions $1-T$, $S$, $C$ and $B_{sum}$ (figs. \[bildthr\],\[bildsph\],\[bildcp\],\[bildtjb\]) measured from charged particles are described within the small experimental errors (typically 2-3%) by all parton shower models. HERWIG for large values of these observables tends to lie slightly above the data. The agreement to the corresponding charged plus neutral distributions is still satisfactory, although these distributions have not been used in the fit. This illustrates furher the quality of the models and of the data. The description of distributions sensitive to transverse momenta in the event plane like $M$, $B_{max}$ or $M_h^2/E^2_{vis}$ is only slightly less good (typically better than 5%). Turning to distributions sensitive to transverse momenta out of the event plane $m$, $O$, $A$, $M_l^2/E^2_{vis}$, or $B_{min}$ the following pattern is observed: ARIADNE generally describes the data well, HERWIG tends to overestimate the distributions and JETSET PS is usually below the data for higher values of the observables. The latter is to be expected from the underestimation of the $p_t^{out}$-distribution by all PS models. A similar pattern is also observed for the jet rates. The differential 2-jet rate $D_2$ (figs. \[bild2dd\],\[bild2dj\]) is well described by all models. ARIADNE describes well also the higher jet rates (figs. \[bild3dd\]–\[bild4dj\]), HERWIG overestimates and JETSET PS underestimates them. This behaviour is consistently observed for the JADE as well as for the DURHAM algorithm and for the charged and charged plus neutral data.\
The description of the shape distributions by the JETSET ME model is less perfect than for the PS models. There is a shortcome for the extreme 2 jet region, the multijet rates and observables sensitive to radiation out of the event plane. In general however the description by the JETSET ME is quite satisfactory.
[**Identified Particle Rates**]{}\
Table \[part\_rates\] compares the particle rates predicted from the models with the current measured LEP averages.
[lr@[.]{}lr@[.]{}lr@[.]{}lr@[.]{}lr@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l]{} & & & & & &\
\
$<N_{ch}>$ & 20&87 & 20&81 & 20&80 & 20&86 & 20&94 & 20&95 & 0&21\
\
$\pi^{\pm}$ & 17&19 & 17&09 & 17&13 & 17&36 & 17&66 & 17&1 & 0&4\
$\pi^{0}$ & 9&85 & 9&83 & 9&82 & 10&03 & 9&81 & 9&9 & 0&08\
$K^{\pm}$ & 2&20 & 2&23 & 2&19 & 2&15 & 2&11 & 2&42 & 0&13\
$K^{0}$ & 2&13 & 2&17 & 2&12 & 2&10 & 2&08 & 2&12 & 0&06\
$\eta$ & 1&07 & 1&10 & 1&09 & 1&16 & 1&02 & 0&73 & 0&07\
$\eta$’(958) & 0&10 & 0&09 & 0&10 & 0&10 & 0&14 & 0&17 & 0&05\
$D^{+}$ & 0&19 & 0&20 & 0&20 & 0&20 & 0&24 & 0&20 & 0&03\
$D^{0}$ & 0&46 & 0&49 & 0&48 & 0&49 & 0&53 & 0&40 & 0&06\
$B^{\pm}$,$B^{0}$ & 0&36 & 0&36 & 0&36 & 0&36 & 0&36 & 0&34 & 0&06\
\
$f_0(980)$ & 0&17 & 0&16 & 0&17 & 0&16 & & 0&14 & 0&06\
\
$\rho^{\circ}(770)$ & 1&29 & 1&27 & 1&26 & 1&29 & 1&43 & 1&40 & 0&1\
$K^{\ast\pm}(892)$ & 0&78 & 0&77 & 0&79 & 0&77 & 0&74 & 0&78 & 0&08\
$K^{\ast 0}(892)$ & 0&80 & 0&77 & 0&81 & 0&78 & 0&74 & 0&77 & 0&09\
$\phi(1020)$ & 0&109& 0&107& 0&107 & 0&102 & 0&099& 0&086 & 0&018\
$D^{\ast\pm}(2010)$ & 0&18 & 0&22 & 0&19 & 0&22 & 0&22 & 0&17 & 0&02\
$D^{\ast 0}(2007)$ & 0&20 & 0&22 & 0&20 & 0&22 & 0&23 &\
\
$f_2(1270)$ & 0&29 & 0&29 & 0&29 & 0&30 & 0&26 & 0&31 & 0&12\
\
$p$ & 0&97 & 0&97 & 0&96 & 0&90 & 0&78 & 0&92 & 0&11\
$\Lambda^0$ & 0&361& 0&349& 0&365& 0&309& 0&368& 0&348 & 0&013\
$\Xi^-$ & 0&0288& 0&0300& 0&0300& 0&0256& 0&0493 & 0&0238 & 0&0024\
$\Delta^{++}(1232)$ & 0&158& 0&160& 0&136& 0&158& 0&154& 0&077 &0&018\
$\Sigma^{\pm}(1385)$ & 0&037& 0&036& 0&032& 0&033& 0&065& 0&0380 & 0&0062\
$\Xi^0(1530)$ & 0&0073& 0&0069& 0&0063 & 0&0060& 0&0249 & 0&0063 & 0&0014\
$\Omega^-$ & 0&0013& 0&0019& 0&0021& 0&0010& 0&0077& 0&0051 & 0&0013\
$\Lambda_b^0$ & 0&032& 0&033& 0&032 & 0&029& 0&007& 0&031 & 0&016\
After inclusion of the charged multiplicity in the fits it is well described by the models. Neglecting this constraint ARIADNE and JETSET PS predict the multiplicity too low ( 20.2 – 20.4 ) and JETSET ME too high . HERWIG is correct without constraint.
All meson rates with the exception of the $K^{\pm}$ and the $\eta$ rate are well described within errors (2$\sigma$). The measured $\eta$ momentum spectrum is however well described. We therefore suspect a too small error quoted in the extrapolation needed to determine the total $\eta$ rate. The $K^{\pm}$ rate is sensitive to heavy quark decays (see below). The octet baryons also agree reasonably. Only HERWIG overestimates the $\Xi^-$ rate by about a factor 2. Some discrepancies show up in the decuplet baryon sector.
[**Meson Momentum Spectra**]{}\
In figures \[bildk0\],\[bildkpm\] we compare the models to the identified kaon spectra as function of $\xi_p=\log{\frac{1}{x_p}}$ from different LEP experiments. For a more quantitative comparison which also allows to judge the agreement among the different experiments we show on the right the relative deviation between the individual data sets and each model as lines. In the fragmentation region at large $\xi_p$ (i.e. small $x_p$) where the $K^{\pm}$ data is more precise and therefore dominates the fit the $K^0$’s are overestimated by $\approx 10-20\%$ by all models. At more central momenta all kaons are well described. $K^{\pm}$ are underestimated in the range ($0.8 < \xi_p < 2.5$). K’s from heavy particle decays tend to contribute mainly in this momentum range.
Wrong branching fractions of b-hadrons in the models presumably cause this discrepancy. Indeed a recent DELPHI measurement of the inclusive particle production in b-events supports this interpretation [@b-inclusive]. The difference of $K^{\pm}$ and $K^0$ production was found to be $0.60 \pm 0.50$. The corresponding value in the models is only $\approx 0.30$ for JETSET and ARIADNE and $0.04$ for HERWIG. This difference causes the relatively large systematic error on $\gamma_s$. An improvement is possible as soon as K measurements for identified b and light quark events become available. Finally at very small $\xi_p$ (i.e. large $x_p$) there is an indication that the models underestimate slightly the K production.
Fragmentation functions of vector mesons, which are likely to be primary particles, are compared to in figs. \[bildkstar0\],\[bildkstarpm\],\[bildphi\]. Within the large errors of these resonance measurements all models describe the spectra very well. There is only a tendency to predict a somewhat harder fragmentation than measured. This is most evident for the $\Phi (1020)$ spectrum.\
It is important to note that the strange pseudoscalar and vector meson data shown already imply large production rates of higher mass resonances with strangeness. The large $\xi_p$ part of the $K^{0, \pm}$ spectra can only be described if large resonance rates are assumed in the models. Due to the stronger boost (because of their high mass) kaons from resonance decays acquire in average higher momenta (i.e. smaller $\xi_p$). As the strange vector meson rates are relatively moderate higher mass strange resonances must therefore be present. This reflects in high production probabilities, of the order of 25%, for p-wave resonances with strangeness in JETSET and ARIADNE. HERWIG dynamically predicts similar rates.
Fig. \[bildf0\] depicts the the good agreement of the model predictions to the measurements of the $f_0(980)$ scalar and $f_2(1270)$ tensor meson resonance. With large production probabilities for p-wave resonances all models describe the measured data well. Note that high production of p-wave states ($>10\%$) is not expected in the string picture [@lund_rev].
[**Baryon Momentum Spectra**]{}\
The proton spectra of OPAL and ALEPH show a severe discrepancy at high momentum (or small $\xi_p$) (see fig. \[bildproton\]). The central model fits interpolate between these experimental results. Including only the ALEPH or OPAL data in the fit it turns out that the OPAL proton spectrum cannot be described by JETSET PS and ARIADNE. The ALEPH data can be reasonably well described. On the other hand the JETSET ME model gives a fair description of the OPAL proton data. The rate predicted by HERWIG is too high at small $\xi_p$ and too low at small $\xi_p$. At small $\xi_p$ the behaviour of JETSET PS and ARIADNE is similar to HERWIG if the extra baryon suppression is not used. The need for an extra suppression of baryon production may also indicate the presence of nucleon resonances. Although the mass splitting between angular excited baryons and the related ground states is smaller than in the mesons case these states are so far not foreseen in JETSET or HERWIG. Decays of these states would lead to a softening of the spectra of stable baryons and would also modify the predicted correlations between baryon antibaryon pairs.
The experimental situation for the $\Lambda^0$ spectrum also shows some discrepancies (see fig. \[bildlambda\]). JETSET PS and ARIADNE describe or slightly overestimate the small $\xi_p$ region. HERWIG here is again too high. At large $\xi_p$ where the measurements agree all models underestimate the $\Lambda^0$ production by $\approx 15\%$. This region is of special interest because here particle production is expected to be influenced by coherence phenomena.
Fig. \[bildbaryons\] shows in a comprehensive overview a comparison of the octet and decuplet baryon production and the model predictions. The discrepancies discussed in the last paragraphs are hidden here due to the large scales. JETSET and ARIADNE describe the gross features of the octet and decuplet baryon production well. The $\Delta^{++}$ rate [@deltapp] is however overestimated by about a factor 2. HERWIG predicts too hard baryon fragmentation. The relative production rates of the different multiplet states are less well predicted by HERWIG.
Summary
=======
From 750.000 $e^+e^- \rightarrow Z \rightarrow hadrons$ events measured by the DELPHI experiment precise fully corrected semi-inclusive charged particle and event shape distributions have been determined.
A systematic, quantitative study has been undertaken to determine the optimal choice of distributions to tune fragmentation models to. Semi-inclusive charged particle and identified particle distributions constrain the hadronization part of the models whereas 3-jet rate and most event shape distributions mainly control the parton shower parameters ( especially $\Lambda_{QCD}$) in the models.
Optimal parameters for the ARIADNE 4.06, HERWIG 5.8c and JETSET 7.3 and 7.4 parton shower models and for the JETSET 7.4 matrix element model have been determined. The models were fitted to the measured event properties and inclusive data and to identified particle data measured by the LEP experiments. The fit algorithm employed allowed for a simultaneous fit of up to 10 model parameters. Statistical and systematic errors as well as correlations of the model parameters have been determined.
All models reasonably describe the inclusive and the event shape distributions. The data measured form charged particles and charged plus neutral particles when compared with the corresponding model predictions yield consistent results.\
All models underestimate the tail of the $p_t^{out}$ distribution by more than $25\%$. With this exception the best overall description of event shapes is provided by the ARIADNE 4.06 model. HERWIG 5.8c tends to overestimate and JETSET 7.3/7.4 to underestimate the production of 4 and more jet events. Correspondingly the tails of shape distributions sensitive to particle production out of the event plane are overestimated (underestimated) by HERWIG (JETSET). The matrix element model JETSET 7.4 ME with optimized scale also provides reasonable predictions. It however shows the expected discrepancies due to missing higher orders in the extreme 2-jet and multijet regions.
Identified meson spectra are fairly well described by all models. It has been found that strong production of p-wave resonances ($25-40\%$) has to be considered. This is not expected in a string fragmentation picture.
The gross features of baryon production are described by JETSET and ARIADNE. Some discrepancies however show up in the momentum spectra and in the relative abundance of the individual multiplet states. HERWIG shows stronger discrepancies, especially the predicted fragmentation functions are too hard. In JETSET and ARIADNE a similar tendency has been corrected by an extra leading baryon suppression.
[**Acknowledgement**]{}\
We would like to thank L. Lönnblad, I. Knowles, M. Seymour and T. Sjöstrand for useful discussions.
Definition of Variables {#vars}
=======================
Throughout this paper we use the following definitions of event shape and inclusive particle variables:
Inclusive Single Particle Variables
-----------------------------------
Scaled Momentum, $x_{p}$
: \
The Scaled Momentum $x_{p}$ is the absolute momentum $|\vec{p}|$ of a particle scaled to the beam momentum.
Transverse Momenta, $p_t^{in}, p_t^{out}$
: \
We distinguish the transverse momentum of a particle in the event plane $p_t^{in}=\vec{p}\cdot\vec{n}_{Major}$ and out of the event plane $p_t^{out}=\vec{p}\cdot\vec{n}_{Minor}$. Alternatively we use the axes as defined by the eigenvectors of the quadratic momentum tensor.
Rapidity $y$
: \
The rapidity is given by: $$y=\frac{1}{2}\cdot\log\frac{E-p_{\parallel}}{E+p_{\parallel}}$$ where $p_{\parallel}$ is a particles momentum parallel to $\vec{n}_{Thrust}$ or $\vec{n}_{Sphericity}$.
Event Shapes Variables
----------------------
Thrust $T$, Major $M$, Minor $m$, Oblateness $O$
: \
Thrust $T$ [@thrust] and Thrust-axis are defined by: $$\nonumber
T = \max_{\vec{n}_{Thrust}}
\frac{\sum\limits_{i=1}^{N_{particle}} \left| \vec{p}_{i} \cdot
\vec{n}_{Thrust}
\right| } {\sum\limits_{i=1}^{N_{particle}} \left| \vec{p}_{i} \right| }$$ $\vec{p}$ is the 3-momentum of a particle, $\vec{n}_{Thrust}$ is a unit-vector along the Thrust-axis. Major and Minor are defined via the same expression, however replacing $\vec{n}_{Thrust}$ by $\vec{n}_{Major} \perp \vec{n}_{Thrust}$ or $ \vec{n}_{Minor} \, = \, \vec{n}_{Major} \times \vec{n}_{Thrust}$ respectively. The oblateness is $O=M-m$.
Sphericity $S$, Aplanarity $A$, Planarity $P$
: \
Ordering the eigenvalues $\lambda$ of the quadratic momentum tensor: $$M^{\alpha\beta}=\sum\limits_{i=1}^{N_{particle}} p^{\alpha}_{i} p^{\beta}_{j}
{}~~~~~~~~~~~( \alpha,\beta = 1,2,3 )$$ $$\lambda_{1}\ge\lambda_{2} \ge\lambda_{3}
{}~~~~~~~~~~~\lambda_{1}+\lambda_{2}+\lambda_{3}=1$$ The Spericity is $S=\frac{3}{2} (\lambda_{2}+\lambda_{3})$, the Aplanarity $A=\frac{3}{2} \lambda_{3}$ and the Planarity is $P=\frac{2}{3}(S-2A)$ [@sphericity]. The Sphericity-axis is parallel to the eigenvector corresponding to $\lambda_1$. As the momenta enter quadratically the Sphericity-axis is influenced more strongly by large momentum particles than the Thrust-axis.
$C$- and $D$-Parameter
: \
C- and D-Parameter are defined through the eigenvalues $\lambda$ of the linear momentum tensor [@c_d]: $$\Theta^{\alpha\beta} =
\frac{1}
{\sum\limits_{k=1}^{N_{particle}} \left|\vec{p}_{k}\right|}
\cdot
\sum\limits_{i=1}^{N_{particle}}
\frac{p^{i}_{k}p^{j}_{k}}{\left|\vec{p}_{k}\right|}$$ $$C=3\cdot(\lambda_{1}\lambda_{2}+\lambda_{2}\lambda_{3}+
\lambda_{3}\lambda_{1})~~~~~~~~~~
D=27\cdot\lambda_{1}\lambda_{2}\lambda_{3}$$
Jet Masses $M^2_{high}/E^{2}_{vis}, M^2_{low}/E^{2}_{vis},
M^2_{diff}/E^{2}_{vis} $
: \
Particles are ordered in the two hemispheres of an event separated by the plane normal to $\vec{n}_{Thrust}$. $$\frac{M^{2}_{high}}{E^{2}_{vis}} = \frac{1}{E^{2}_{vis}} \cdot
\max \left( \left(
\sum\limits_{ \vec{p}_{k} \cdot \vec{n}_{Thrust}>0} p_{k}
\right)^{2} \, , \,\, \left(
\sum\limits_{ \vec{p}_{k} \cdot \vec{n}_{Thrust}<0} p_{k}
\right)^{2}\right)$$ For $\frac{M^2_{low}}{E^{2}_{vis}}$ the maximum in the above formula is replaced by the minimum and $\frac{M^2_{diff}}{E^{2}_{vis}} =
\frac{M^2_{high}}{E^{2}_{vis}}-\frac{M^2_{low}}{E^{2}_{vis}}$. This definition differs from the original one by Clavelli [@clavelli] but is easier to calculate and therefore prefered by many experiments.
Jet Broadening $B_{max}, B_{min}, B_{sum}, B_{diff}$
: \
Similarly to the definition of the jetmasses the transverse momenta with respect to the Thrust-axis are added up [@bbw], i.e.: $$B_{\pm} = \frac{
\sum
\limits_{ \pm \vec{p}_{i} \cdot \vec{n}_{Thrust} > 0}
\left | \right. \vec{p}_{i} \times \vec{n}_{Thrust} \left | \right.
}
{2 \sum\limits_i \left | \vec{p}_i \right |}$$ Then: $$B_{max}=\max{B_+, B_-}~~~~B_{min}=\min{B_+, B_-}
{}~~~~B_{sum}=B_+ + B_-~~~~B_{diff}=|B_+ - B_-|$$
Differential Jet Rates $D_i(y)$
: \
Jets are reconstructed using cluster finding algorithms of the JADE typ [@jade]. For each event and each pair of particles i and j the scaled invariant mass or transverse momentum $y_{ij}$ for the JADE or Durham algorithm respectively are evaluated: $$y_{ij}^{Jade} =
2E_{i}E_{j}/E_{vis}^{2}\cdot(1-\cos\theta_{ij}) ~\\$$ $$y_{ij}^{Durham} =
2\min(E_{i}^{2},E_{j}^{2})/E_{vis}^{2}\cdot(1-\cos\theta_{ij}) \\$$ where $E_i$, $E_j$ are the energies and $\theta_{ij}$ the angle between the momentum vectors of the two particles. The particle pair with the lowest value $y_{ij}$ is selected and replaced by a pseudo-particle with four momentum $(p_i + p_j)$, hereby reducing the multiplicity by one. In successive steps the procedure is repeated until the scaled invariant masses of all pairs of (pseudo-)particles are larger than a given resolution $y$. The remaining (pseudo-)particles are called jets. The differential 2-jet rate $D_{2}$ is derived from the 2-jet rate $R_2=\frac{N_{2-jets}}{N}$ [@smolik]: $$D_{2}(y) = \frac {R_{2}(y + \Delta y)- R_{2}(y)}{\Delta y}$$ Higher differential rates follow from the recursion: $$D_{n}(y) = \frac {R_{n}(y + \Delta y)- R_{n}(y)}{\Delta y} + D_{n-1}(y)$$
Energy-Energy-Correlation $EEC$ and Asymmetry $AEEC$
: \
The Energy-Energy-Correlation $EEC$ is the histogram of angles $\chi_{ij}$ between all particles weighted by their scaled energies [@eec]: $$EEC(cos{\chi}) = \frac{1}{N}\frac{1}{\Delta\cos{\chi}} \cdot
\sum_{events}^N\sum_{i,j} \frac{E_i}{E_{vis}}\frac{E_j}{E_{vis}}
\Theta(\Delta\cos{\chi} - |cos{\chi}-cos{\chi_{ij}}| )$$ $cos{\chi}$ and $\Delta\cos{\chi}$ are the low edge and width of a bin and $\Theta$ is the step function. For $cos{\chi}\geq 0$ the asymmetry $AEEC$ of the $EEC$ then is: $$AEEC(cos{\chi}) = EEC(-cos{\chi}) - EEC(cos{\chi})$$
Parameter Settings of the DELPHI Monte Carlo {#delphi_par}
============================================
The DELPHI tuning of JETSET has been obtained by tuning the model to charged particle data from the 1991 and 1992 data taking. Care has been taken to describe especially well the observables relevant to standard precision analyses which are the charged multiplicity, the momentum spectrum, 2-jet rate and Thrust and Sphericity distribution. Also included is the simulation of Bose Einstein interference (by LUBOEI) to obtain a correct description of two particle correlations and light resonance line shapes. The BE parameters are taken from the DELPHI measurement [@bepaper] Particle spectra have been adjusted and partially fitted to available data [@ada]. Heavy particle decays have been adjusted. When using adjusted decay tables this is quoted as DELPHI decays.
Variable Dec. 93 Sept. 94
----------- ---------- ----------
MSTJ(11) 3 3
MSTJ(12) 3 3
MSTJ(41) 2 2
MSTJ(45) 5 5
MSTJ(46) 3 3
MSTJ(51) 2 2
MSTJ(52) 7 7
MSTJ(101) 5 5
MSTJ(107) 0 0
PARJ(1) 0.10 0.10
PARJ(2) 0.28 0.28
PARJ(3) 0.55 0.55
PARJ(4) 0.07 0.07
PARJ(5) 0.5 0.5
PARJ(11) 0.55
PARJ(12) 0.55
PARJ(13) 0.75
PARJ(14) 0.090
PARJ(15) 0.070
PARJ(16) 0.085
PARJ(17) 0.140
PARJ(19) 0.5 0.5
PARJ(21) 0.417 0.428
PARJ(25) 0.7 0.7
PARJ(26) 0.2 0.2
PARJ(41) 0.5 0.354
PARJ(42) 0.701 0.523
PARJ(54) -0.0631 -0.0305
PARJ(55) -0.00414 -0.00233
: DELPHI parameter setting of JETSET 7.3 PS (December 1993 tuning used for modelling detector effects for 1993 data and September 1994 tuning used for modelling detector effects for 1994 data )[]{data-label="delphi_tuning"}
Variable Dec. 93 Sept. 94
----------------- --------- ----------
PARJ(81) 0.297 0.346
PARJ(82) 1.732 2.25
PARJ(92) 1 1
PARJ(93) 0.394 0.394
$P(^1S_0)_{ud}$ .423
$P(^3S_1)_{ud}$ .275
$P(^1P_1)_{ud}$ .067
$P(^3P_0)_{ud}$ .056
$P(^3P_1)_{ud}$ .067
$P(^3P_2)_{ud}$ .112
$P(^1S_0)_{s }$ .388
$P(^3S_1)_{s }$ .296
$P(^1P_1)_{s }$ .079
$P(^3P_0)_{s }$ .026
$P(^3P_1)_{s }$ .079
$P(^3P_2)_{s }$ .132
$P(^1S_0)_{c }$ .250
$P(^3S_1)_{c }$ .400
$P(^1P_1)_{c }$ .087
$P(^3P_0)_{c }$ .030
$P(^3P_1)_{c }$ .087
$P(^3P_2)_{c }$ .146
$P(^1S_0)_{b }$ .1625
$P(^3S_1)_{b }$ .4875
$P(^1P_1)_{b }$ .087
$P(^3P_0)_{b }$ .030
$P(^3P_1)_{b }$ .087
$P(^3P_2)_{b }$ .146
: DELPHI parameter setting of JETSET 7.3 PS (December 1993 tuning used for modelling detector effects for 1993 data and September 1994 tuning used for modelling detector effects for 1994 data )[]{data-label="delphi_tuning"}
Tables of Sensitivities
=========================
[lcccccccccc]{} Parameter & & & &\
& & & & & & & &\
$\Lambda_{QCD}$ & 32& 30& 29& 26& 23& 18& 22& 23\
$\epsilon_C$ & 9& & 9& & 10& & 9&\
$\epsilon_B$ & & 8& & 8& & 9& & 10\
[lccccccccccc]{} & &\
Property & & & & & & & & & & &\
$S $& 1& 2& 5& 10& 1& 4& 2& 1& 1& 0& 0\
$A $& 2& 6& 12& 20& 1& 9& 12& 1& 2& 4& 1\
$P $& 1& 2& 2& 8& 1& 2& 1& 1& 1& 0& 1\
$1-T $& 2& 4& 6& 10& 1& 2& 2& 0& 1& 0& 0\
$M $& 6& 13& 11& 19& 6& 4& 4& 1& 1& 2& 0\
$m $& 5& 13& 11& 18& 4& 1& 5& 1& 1& 1& 1\
$O $& 2& 5& 6& 6& 1& 3& 2& 1& 0& 0& 1\
$D_2^D $& 1& 1& 6& 9& 2& 6& 2& 0& 2& 1& 1\
$D_3^D $& 1& 2& 9& 22& 6& 12& 11& 3& 4& 1& 1\
$D_4^D $& 4& 9& 16& 45& 5& 16& 17& 9& 6& 5& 3\
$D_2^J $& 1& 1& 2& 8& 2& 3& 2& 0& 1& 0& 1\
$D_3^J $& 2& 4& 6& 19& 1& 2& 1& 2& 0& 0& 1\
$D_4^J $& 3& 9& 11& 20& 3& 7& 4& 3& 1& 1& 1\
$C $& 2& 5& 5& 10& 1& 3& 2& 1& 1& 1& 1\
$D $& 2& 6& 6& 15& 2& 4& 5& 1& 0& 1& 1\
$M^2_h/E_{vis.}^2 $& 2& 5& 3& 9& 1& 4& 1& 1& 0& 1& 1\
$M^2_l/E_{vis.}^2 $& 5& 11& 2& 23& 3& 4& 4& 1& 1& 1& 1\
$M^2_d/E_{vis.}^2 $& 1& 4& 4& 6& 0& 6& 2& 1& 2& 2& 0\
$B_{max} $& 3& 8& 7& 14& 2& 2& 2& 1& 0& 1& 1\
$B_{min} $& 3& 7& 7& 20& 4& 2& 6& 3& 1& 2& 2\
$B_{sum} $& 5& 10& 8& 15& 4& 5& 4& 1& 1& 1& 0\
$B_{diff.} $& 1& 4& 5& 5& 1& 5& 3& 1& 2& 2& 1\
$p_t^{in} $& 5& 8& 9& 5& 3& 3& 3& 2& 2& 2& 1\
$p_t^{out}$& 4& 8& 27& 11& 2& 2& 2& 4& 2& 2& 0\
$y_T $& 2& 5& 6& 8& 2& 1& 2& 1& 1& 1& 0\
$x_p $& 3& 6& 7& 6& 4& 1& 1& 2& 1& 1& 0\
$p_t^{out} vs. x_p $& 3& 7& 7& 9& 7& 11& 6& 3& 2& 2& 0\
$p_t^{trans.} vs. x_p $& 3& 7& 7& 9& 7& 11& 6& 3& 2& 2& 0\
$EEC $& 1& 1& 1& 4& 0& 1& 0& 0& 0& 0& 0\
$AEEC $& 1& 6& 16& 9& 13& 32& 17& 6& 2& 7& 3\
$\rho^{\circ} $& 4& 7& 9& 8& 4& 4& 3& 3& 9& 1& 1\
$K^0 / K^{\pm} $& 4& 7& 9& 6& 4& 13& 4& 1& 0& 6& 2\
$K^{\ast 0}/ K^{\ast \pm} $& 4& 8& 10& 7& 4& 14& 4& 0& 1& 8& 10\
$\phi $& 5& 8& 7& 9& 4& 35& 9& 3& 3& 5& 23\
$p $& 5& 12& 13& 5& 6& 3& 22& 3& 1& 1& 3\
$\Lambda^0 $& 5& 11& 12& 12& 4& 17& 21& 2& 2& 2& 2\
$\Sigma^{\pm}(1193) $& 4& 15& 18& 11& 8& 9& 21& 4& 4& 5& 5\
[lccccccccc]{}Property & & & & & & & &\
$S $& 11& 3& 2& 3& 1& 1& 0& 1\
$A $& 25& 3& 7& 5& 2& 2& 1& 3\
$P $& 9& 2& 2& 2& 0& 0& 1& 1\
$1-T $& 12& 6& 1& 1& 0& 1& 0& 1\
$M $& 13& 7& 3& 7& 1& 1& 1& 2\
$m $& 20& 5& 3& 4& 1& 1& 1& 2\
$O $& 8& 1& 2& 1& 0& 1& 1& 1\
$D_2^D $& 12& 2& 3& 1& 1& 1& 1& 0\
$D_3^D $& 27& 6& 4& 3& 1& 1& 1& 2\
$D_4^D $& 39& 5& 8& 6& 3& 5& 3& 2\
$D_2^J $& 11& 2& 2& 1& 0& 0& 1& 1\
$D_3^J $& 21& 5& 3& 3& 1& 2& 1& 1\
$D_4^J $& 24& 6& 7& 5& 1& 1& 1& 2\
$C $& 12& 6& 1& 1& 0& 1& 0& 1\
$D $& 20& 5& 2& 1& 1& 1& 1& 2\
$M^2_h/E_{vis.}^2 $& 10& 7& 4& 2& 1& 1& 1& 1\
$M^2_l/E_{vis.}^2 $& 17& 14& 5& 2& 1& 2& 1& 1\
$M^2_d/E_{vis.}^2 $ & 10& 3& 4& 3& 1& 0& 1& 1\
$B_{max} $& 13& 8& 3& 4& 1& 1& 0& 1\
$B_{min} $& 19& 7& 3& 2& 1& 1& 1& 2\
$B_{sum} $& 12& 7& 1& 4& 1& 0& 1& 2\
$B_{diff.} $& 8& 1& 2& 1& 0& 1& 1& 1\
$p_t^{in} $& 6& 4& 5& 11& 3& 1& 0& 1\
$p_t^{out}$& 15& 5& 14& 7& 2& 1& 0& 2\
$y_T $& 6& 4& 7& 10& 2& 1& 1& 1\
$x_p $& 2& 3& 4& 9& 3& 1& 1& 0\
$p_t^{out} vs. x_p $& 1& 3& 3& 50& 5& 2& 2& 5\
$p_t^{trans.} vs. x_p $& 1& 3& 3& 50& 5& 2& 2& 5\
$EEC $& 5& 1& 1& 0& 0& 0& 0& 0\
$AEEC $& 7& 3& 3& 4& 1& 1& 2& 4\
$\rho^{\circ} $& 4& 6& 7& 12& 3& 4& 2& 2\
$K^0 / K^{\pm} $& 6& 11& 13& 23& 6& 13& 4& 2\
$K^{\ast 0}/ K^{\ast \pm} $& 8& 14& 8& 15& 4& 14& 5& 3\
$\phi $& 15& 24& 10& 21& 6& 33& 10& 3\
$p $& 5& 7& 48& 20& 7& 2& 3& 15\
$\Lambda^0 $& 9& 13& 65& 27& 8& 18& 8& 11\
$\Sigma^{\pm}(1193) $& 10& 18& 46& 16& 7& 13& 7& 7\
Tables of Inclusive Charge Particle and Event Shape Distributions {#tables}
===================================================================
The cross section has been determined from charged particles only and from charged and neutral particles. The tables show first the measurement from charged particles corrected to the charged final state, then corrected to the full final state. The third column contains the measurement from charged plus neutral particles corrected to the full final state. For inclusive charged particle distributions the relevant event axes have been evaluated for the corresponding final states. The cross section given in the table is followed by the statistical and systematic errors.
Tables of Inclusive Charged Particle Distributions
----------------------------------------------------
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.100 & 44&667 & 0&036 & 1&653 & 44&740 & 0&036 & 1&685 & 46&663 & 0&037 & 1&758\
0.100 & 0.200 & 39&331 & 0&033 & 1&062 & 39&475 & 0&033 & 1&078 & 39&823 & 0&033 & 1&092\
0.200 & 0.300 & 29&429 & 0&028 & 0&603 & 29&491 & 0&028 & 0&606 & 29&351 & 0&028 & 0&608\
0.300 & 0.400 & 21&326 & 0&024 & 0&353 & 21&338 & 0&024 & 0&352 & 21&034 & 0&024 & 0&350\
0.400 & 0.500 & 15&547 & 0&020 & 0&225 & 15&463 & 0&020 & 0&222 & 15&156 & 0&020 & 0&219\
0.500 & 0.600 & 11&463 & 0&017 & 0&156 & 11&437 & 0&017 & 0&154 & 11&149 & 0&017 & 0&150\
0.600 & 0.700 & 8&668 & 0&015 & 0&116 & 8&589 & 0&015 & 0&115 & 8&348 & 0&015 & 0&111\
0.700 & 0.800 & 6&664 & 0&013 & 0&090 & 6&608 & 0&013 & 0&091 & 6&430 & 0&013 & 0&087\
0.800 & 1.000 & 4&6916 & 0&0078 & 0&0646 & 4&6518 & 0&0078 & 0&0661 & 4&5131 & 0&0076 & 0&0624\
1.000 & 1.200 & 3&0694 & 0&0063 & 0&0430 & 3&0346 & 0&0063 & 0&0449 & 2&9522 & 0&0062 & 0&0420\
1.200 & 1.400 & 2&0913 & 0&0052 & 0&0298 & 2&0847 & 0&0052 & 0&0322 & 2&0401 & 0&0052 & 0&0299\
1.400 & 1.600 & 1&5030 & 0&0044 & 0&0218 & 1&4924 & 0&0044 & 0&0243 & 1&4597 & 0&0044 & 0&0222\
1.600 & 1.800 & 1&0916 & 0&0038 & 0&0163 & 1&0961 & 0&0038 & 0&0189 & 1&0796 & 0&0038 & 0&0171\
1.800 & 2.000 & 0&8306 & 0&0033 & 0&0130 & 0&8228 & 0&0033 & 0&0153 & 0&8155 & 0&0033 & 0&0136\
2.000 & 2.500 & 0&5311 & 0&0017 & 0&0089 & 0&5315 & 0&0017 & 0&0108 & 0&5326 & 0&0017 & 0&0095\
2.500 & 3.000 & 0&2946 & 0&0013 & 0&0055 & 0&2935 & 0&0013 & 0&0066 & 0&2988 & 0&0013 & 0&0057\
3.000 & 3.500 & 0&17405 & 0&00097 & 0&00372 & 0&17393 & 0&00097 & 0&00441 & 0&18067 & 0&00099 & 0&00383\
3.500 & 4.000 & 0&10909 & 0&00077 & 0&00276 & 0&10957 & 0&00077 & 0&00319 & 0&11471 & 0&00079 & 0&00273\
4.000 & 5.000 & 0&05911 & 0&00041 & 0&00182 & 0&05897 & 0&00040 & 0&00200 & 0&06305 & 0&00042 & 0&00171\
5.000 & 6.000 & 0&02669 & 0&00028 & 0&00101 & 0&02672 & 0&00027 & 0&00106 & 0&03040 & 0&00029 & 0&00095\
6.000 & 7.000 & 0&01369 & 0&00021 & 0&00064 & 0&01299 & 0&00019 & 0&00060 & 0&01501 & 0&00021 & 0&00054\
7.000 & 8.000 & 0&00675 & 0&00014 & 0&00039 & 0&00670 & 0&00013 & 0&00035 & 0&00858 & 0&00016 & 0&00035\
8.000 & 10.000 & 0&00271 & 0&00006 & 0&00019 & 0&00278 & 0&00006 & 0&00016 & 0&00376 & 0&00008 & 0&00017\
10.000 & 12.000 & 0&00095 & 0&00004 & 0&00008 & 0&00084 & 0&00003 & 0&00005 & 0&00123 & 0&00004 & 0&00006\
12.000 & 14.000 & 0&00031 & 0&00002 & 0&00003 & 0&00030 & 0&00002 & 0&00002 & 0&00044 & 0&00003 & 0&00002\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.100 & 67&512 & 0&043 & 1&956 & 67&673 & 0&044 & 1&818 & 66&160 & 0&043 & 1&822\
0.100 & 0.200 & 50&517 & 0&037 & 1&024 & 50&698 & 0&037 & 1&166 & 49&794 & 0&037 & 1&149\
0.200 & 0.300 & 33&618 & 0&030 & 0&593 & 33&803 & 0&030 & 0&692 & 33&544 & 0&030 & 0&678\
0.300 & 0.400 & 21&228 & 0&024 & 0&368 & 21&278 & 0&024 & 0&403 & 21&407 & 0&024 & 0&397\
0.400 & 0.500 & 13&160 & 0&019 & 0&235 & 13&171 & 0&019 & 0&240 & 13&466 & 0&019 & 0&239\
0.500 & 0.600 & 8&183 & 0&015 & 0&152 & 8&156 & 0&015 & 0&147 & 8&527 & 0&015 & 0&150\
0.600 & 0.700 & 5&178 & 0&012 & 0&101 & 5&126 & 0&011 & 0&094 & 5&448 & 0&012 & 0&097\
0.700 & 0.800 & 3&3455 & 0&0095 & 0&0684 & 3&3285 & 0&0093 & 0&0631 & 3&5845 & 0&0098 & 0&0658\
0.800 & 1.000 & 1&8640 & 0&0050 & 0&0407 & 1&8393 & 0&0049 & 0&0374 & 2&0309 & 0&0052 & 0&0398\
1.000 & 1.200 & 0&8873 & 0&0035 & 0&0210 & 0&8706 & 0&0034 & 0&0198 & 0&9959 & 0&0037 & 0&0216\
1.200 & 1.400 & 0&4600 & 0&0026 & 0&0117 & 0&4474 & 0&0024 & 0&0114 & 0&5288 & 0&0028 & 0&0127\
1.400 & 1.600 & 0&2500 & 0&0020 & 0&0068 & 0&2462 & 0&0018 & 0&0070 & 0&2987 & 0&0021 & 0&0079\
1.600 & 1.800 & 0&1429 & 0&0015 & 0&0042 & 0&1373 & 0&0014 & 0&0043 & 0&1755 & 0&0016 & 0&0051\
1.800 & 2.000 & 0&0841 & 0&0011 & 0&0026 & 0&0833 & 0&0011 & 0&0029 & 0&1086 & 0&0013 & 0&0034\
2.000 & 2.500 & 0&04069 & 0&00052 & 0&00140 & 0&03877 & 0&00047 & 0&00154 & 0&05266 & 0&00058 & 0&00189\
2.500 & 3.000 & 0&01468 & 0&00033 & 0&00057 & 0&01338 & 0&00028 & 0&00063 & 0&01885 & 0&00035 & 0&00080\
3.000 & 3.500 & 0&00501 & 0&00018 & 0&00022 & 0&00501 & 0&00017 & 0&00028 & 0&00814 & 0&00023 & 0&00040\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
/GeV & & &\
& & &\
0.000 & 0.100 & 49&063 & 0&038 & 1&068 & 48&833 & 0&038 & 1&696 & 49&206 & 0&038 & 1&672\
0.100 & 0.200 & 38&338 & 0&032 & 0&819 & 38&301 & 0&033 & 1&265 & 38&461 & 0&033 & 0&984\
0.200 & 0.300 & 28&259 & 0&027 & 0&593 & 28&231 & 0&027 & 0&886 & 28&203 & 0&027 & 0&571\
0.300 & 0.400 & 20&492 & 0&023 & 0&422 & 20&554 & 0&023 & 0&613 & 20&391 & 0&023 & 0&349\
0.400 & 0.500 & 15&073 & 0&020 & 0&306 & 15&123 & 0&020 & 0&428 & 14&926 & 0&020 & 0&233\
0.500 & 0.600 & 11&274 & 0&017 & 0&225 & 11&312 & 0&017 & 0&304 & 11&133 & 0&017 & 0&168\
0.600 & 0.700 & 8&566 & 0&015 & 0&169 & 8&601 & 0&015 & 0&219 & 8&458 & 0&015 & 0&129\
0.700 & 0.800 & 6&627 & 0&013 & 0&129 & 6&659 & 0&013 & 0&161 & 6&548 & 0&013 & 0&102\
0.800 & 1.000 & 4&7065 & 0&0078 & 0&0898 & 4&7326 & 0&0079 & 0&1055 & 4&6706 & 0&0078 & 0&0747\
1.000 & 1.200 & 3&0724 & 0&0063 & 0&0574 & 3&0960 & 0&0064 & 0&0620 & 3&0684 & 0&0064 & 0&0504\
1.200 & 1.400 & 2&1120 & 0&0052 & 0&0388 & 2&1289 & 0&0053 & 0&0383 & 2&1299 & 0&0053 & 0&0359\
1.400 & 1.600 & 1&4873 & 0&0044 & 0&0270 & 1&5068 & 0&0045 & 0&0244 & 1&5201 & 0&0045 & 0&0264\
1.600 & 1.800 & 1&0882 & 0&0038 & 0&0196 & 1&0925 & 0&0038 & 0&0160 & 1&1143 & 0&0039 & 0&0201\
1.800 & 2.000 & 0&8101 & 0&0033 & 0&0146 & 0&8173 & 0&0033 & 0&0109 & 0&8398 & 0&0034 & 0&0159\
2.000 & 2.500 & 0&5156 & 0&0017 & 0&0094 & 0&5172 & 0&0017 & 0&0060 & 0&5334 & 0&0017 & 0&0107\
2.500 & 3.000 & 0&2799 & 0&0012 & 0&0053 & 0&2793 & 0&0012 & 0&0029 & 0&2968 & 0&0013 & 0&0065\
3.000 & 3.500 & 0&16206 & 0&00095 & 0&00323 & 0&16110 & 0&00093 & 0&00168 & 0&17343 & 0&00098 & 0&00418\
3.500 & 4.000 & 0&09848 & 0&00074 & 0&00212 & 0&09814 & 0&00073 & 0&00116 & 0&10741 & 0&00078 & 0&00292\
4.000 & 5.000 & 0&04987 & 0&00038 & 0&00124 & 0&04903 & 0&00036 & 0&00080 & 0&05615 & 0&00040 & 0&00176\
5.000 & 6.000 & 0&02113 & 0&00025 & 0&00065 & 0&02055 & 0&00024 & 0&00054 & 0&02473 & 0&00027 & 0&00089\
6.000 & 7.000 & 0&00933 & 0&00017 & 0&00036 & 0&00907 & 0&00016 & 0&00037 & 0&01157 & 0&00019 & 0&00048\
7.000 & 8.000 & 0&00401 & 0&00011 & 0&00019 & 0&00400 & 0&00011 & 0&00024 & 0&00561 & 0&00013 & 0&00026\
8.000 & 10.000 & 0&00132 & 0&00005 & 0&00008 & 0&00135 & 0&00004 & 0&00013 & 0&00204 & 0&00006 & 0&00010\
10.000 & 12.000 & 0&00026 & 0&00002 & 0&00002 & 0&00030 & 0&00002 & 0&00005 & 0&00049 & 0&00003 & 0&00002\
12.000 & 14.000 & 0&00007 & 0&00001 & 0&00001 & 0&00007 & 0&00001 & 0&00002 & 0&00012 & 0&00001 & 0&00001\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.100 & 68&848 & 0&044 & 1&931 & 69&511 & 0&045 & 1&992 & 66&825 & 0&043 & 1&506\
0.100 & 0.200 & 51&541 & 0&037 & 1&039 & 51&918 & 0&038 & 1&141 & 50&556 & 0&037 & 1&102\
0.200 & 0.300 & 34&351 & 0&030 & 0&607 & 34&419 & 0&030 & 0&644 & 34&241 & 0&030 & 0&726\
0.300 & 0.400 & 21&343 & 0&024 & 0&373 & 21&274 & 0&024 & 0&372 & 21&708 & 0&024 & 0&451\
0.400 & 0.500 & 12&939 & 0&018 & 0&236 & 12&816 & 0&018 & 0&224 & 13&481 & 0&019 & 0&277\
0.500 & 0.600 & 7&765 & 0&014 & 0&150 & 7&641 & 0&014 & 0&140 & 8&314 & 0&015 & 0&170\
0.600 & 0.700 & 4&757 & 0&011 & 0&098 & 4&660 & 0&011 & 0&091 & 5&180 & 0&012 & 0&106\
0.700 & 0.800 & 2&9543 & 0&0090 & 0&0651 & 2&8799 & 0&0085 & 0&0605 & 3&2986 & 0&0094 & 0&0679\
0.800 & 1.000 & 1&5438 & 0&0047 & 0&0373 & 1&5015 & 0&0044 & 0&0354 & 1&7559 & 0&0049 & 0&0370\
1.000 & 1.200 & 0&6885 & 0&0032 & 0&0186 & 0&6683 & 0&0030 & 0&0182 & 0&8187 & 0&0034 & 0&0181\
1.200 & 1.400 & 0&3309 & 0&0023 & 0&0099 & 0&3218 & 0&0021 & 0&0100 & 0&4064 & 0&0024 & 0&0096\
1.400 & 1.600 & 0&1730 & 0&0017 & 0&0057 & 0&1678 & 0&0015 & 0&0058 & 0&2175 & 0&0018 & 0&0055\
1.600 & 1.800 & 0&0956 & 0&0013 & 0&0034 & 0&0909 & 0&0011 & 0&0035 & 0&1232 & 0&0014 & 0&0034\
1.800 & 2.000 & 0&05183 & 0&00093 & 0&00199 & 0&05046 & 0&00085 & 0&00215 & 0&0712 & 0&0011 & 0&0022\
2.000 & 2.500 & 0&02330 & 0&00041 & 0&00101 & 0&02209 & 0&00036 & 0&00109 & 0&03217 & 0&00047 & 0&00115\
2.500 & 3.000 & 0&00729 & 0&00023 & 0&00037 & 0&00708 & 0&00021 & 0&00042 & 0&01112 & 0&00029 & 0&00050\
3.000 & 3.500 & 0&00230 & 0&00013 & 0&00013 & 0&00219 & 0&00011 & 0&00015 & 0&00387 & 0&00017 & 0&00021\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.250 & 5&7889 & 0&0093 & 0&5509 & 5&7590 & 0&0093 & 0&5296 & 5&9517 & 0&0095 & 0&5628\
0.250 & 0.500 & 6&3749 & 0&0094 & 0&4306 & 6&4046 & 0&0095 & 0&4411 & 6&4291 & 0&0095 & 0&4417\
0.500 & 0.750 & 6&6630 & 0&0093 & 0&3197 & 6&6834 & 0&0094 & 0&3399 & 6&6831 & 0&0094 & 0&3319\
0.750 & 1.000 & 6&7572 & 0&0089 & 0&2316 & 6&8119 & 0&0090 & 0&2522 & 6&7763 & 0&0089 & 0&2429\
1.000 & 1.250 & 6&7638 & 0&0084 & 0&1677 & 6&8102 & 0&0085 & 0&1810 & 6&7650 & 0&0085 & 0&1755\
1.250 & 1.500 & 6&7463 & 0&0082 & 0&1238 & 6&7853 & 0&0083 & 0&1284 & 6&7230 & 0&0083 & 0&1277\
1.500 & 1.750 & 6&6488 & 0&0081 & 0&0933 & 6&6622 & 0&0082 & 0&0898 & 6&6085 & 0&0081 & 0&0939\
1.750 & 2.000 & 6&5006 & 0&0080 & 0&0729 & 6&4845 & 0&0081 & 0&0648 & 6&4346 & 0&0080 & 0&0710\
2.000 & 2.250 & 6&2303 & 0&0079 & 0&0623 & 6&1936 & 0&0079 & 0&0619 & 6&1697 & 0&0078 & 0&0617\
2.250 & 2.500 & 5&8228 & 0&0076 & 0&0582 & 5&7642 & 0&0076 & 0&0576 & 5&7692 & 0&0076 & 0&0577\
2.500 & 2.750 & 5&1930 & 0&0072 & 0&0519 & 5&1348 & 0&0072 & 0&0513 & 5&1450 & 0&0072 & 0&0514\
2.750 & 3.000 & 4&3686 & 0&0067 & 0&0437 & 4&3142 & 0&0066 & 0&0431 & 4&3511 & 0&0066 & 0&0435\
3.000 & 3.250 & 3&4544 & 0&0059 & 0&0345 & 3&4239 & 0&0059 & 0&0342 & 3&4481 & 0&0059 & 0&0345\
3.250 & 3.500 & 2&5583 & 0&0051 & 0&0256 & 2&5473 & 0&0051 & 0&0255 & 2&5852 & 0&0051 & 0&0259\
3.500 & 3.750 & 1&7687 & 0&0042 & 0&0177 & 1&7878 & 0&0043 & 0&0196 & 1&7999 & 0&0043 & 0&0180\
3.750 & 4.000 & 1&1330 & 0&0034 & 0&0113 & 1&1585 & 0&0034 & 0&0150 & 1&1669 & 0&0035 & 0&0117\
4.000 & 4.250 & 0&6770 & 0&0026 & 0&0068 & 0&7044 & 0&0027 & 0&0107 & 0&7054 & 0&0027 & 0&0071\
4.250 & 4.500 & 0&3814 & 0&0019 & 0&0038 & 0&4054 & 0&0020 & 0&0071 & 0&3997 & 0&0020 & 0&0041\
4.500 & 5.000 & 0&14699 & 0&00084 & 0&00147 & 0&16077 & 0&00091 & 0&00339 & 0&15673 & 0&00089 & 0&00177\
5.000 & 5.500 & 0&03148 & 0&00038 & 0&00031 & 0&03466 & 0&00042 & 0&00091 & 0&03374 & 0&00041 & 0&00043\
5.500 & 6.000 & 0&00486 & 0&00016 & 0&00005 & 0&00543 & 0&00017 & 0&00017 & 0&00502 & 0&00016 & 0&00007\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.250 & 6&5422 & 0&0096 & 0&5337 & 6&5778 & 0&0097 & 0&5082 & 6&5680 & 0&0097 & 0&5323\
0.250 & 0.500 & 6&5743 & 0&0096 & 0&4132 & 6&6009 & 0&0096 & 0&4157 & 6&5901 & 0&0096 & 0&4246\
0.500 & 0.750 & 6&5997 & 0&0092 & 0&3151 & 6&6327 & 0&0093 & 0&3348 & 6&6094 & 0&0093 & 0&3329\
0.750 & 1.000 & 6&5896 & 0&0087 & 0&2357 & 6&6311 & 0&0088 & 0&2630 & 6&6152 & 0&0088 & 0&2554\
1.000 & 1.250 & 6&5660 & 0&0083 & 0&1734 & 6&6019 & 0&0084 & 0&2011 & 6&5917 & 0&0084 & 0&1908\
1.250 & 1.500 & 6&5623 & 0&0081 & 0&1266 & 6&5787 & 0&0082 & 0&1499 & 6&5817 & 0&0082 & 0&1393\
1.500 & 1.750 & 6&4983 & 0&0080 & 0&0911 & 6&5086 & 0&0081 & 0&1079 & 6&5221 & 0&0081 & 0&0983\
1.750 & 2.000 & 6&3665 & 0&0080 & 0&0656 & 6&3719 & 0&0080 & 0&0747 & 6&4097 & 0&0080 & 0&0673\
2.000 & 2.250 & 6&1189 & 0&0078 & 0&0612 & 6&1073 & 0&0078 & 0&0611 & 6&1741 & 0&0079 & 0&0617\
2.250 & 2.500 & 5&6636 & 0&0076 & 0&0566 & 5&6636 & 0&0076 & 0&0566 & 5&7542 & 0&0076 & 0&0575\
2.500 & 2.750 & 5&0282 & 0&0071 & 0&0503 & 5&0161 & 0&0071 & 0&0502 & 5&1066 & 0&0072 & 0&0511\
2.750 & 3.000 & 4&2102 & 0&0066 & 0&0421 & 4&1948 & 0&0065 & 0&0419 & 4&2721 & 0&0066 & 0&0427\
3.000 & 3.250 & 3&3399 & 0&0058 & 0&0334 & 3&3206 & 0&0058 & 0&0332 & 3&3718 & 0&0059 & 0&0337\
3.250 & 3.500 & 2&5260 & 0&0051 & 0&0253 & 2&4999 & 0&0050 & 0&0250 & 2&5185 & 0&0051 & 0&0252\
3.500 & 3.750 & 1&7861 & 0&0042 & 0&0179 & 1&7772 & 0&0043 & 0&0178 & 1&7588 & 0&0042 & 0&0176\
3.750 & 4.000 & 1&2194 & 0&0035 & 0&0142 & 1&2049 & 0&0035 & 0&0155 & 1&1589 & 0&0034 & 0&0130\
4.000 & 4.250 & 0&8020 & 0&0028 & 0&0110 & 0&7860 & 0&0028 & 0&0133 & 0&7327 & 0&0027 & 0&0105\
4.250 & 4.500 & 0&4995 & 0&0022 & 0&0080 & 0&4933 & 0&0022 & 0&0106 & 0&4402 & 0&0021 & 0&0078\
4.500 & 5.000 & 0&2475 & 0&0011 & 0&0049 & 0&2406 & 0&0011 & 0&0071 & 0&1952 & 0&0010 & 0&0046\
5.000 & 5.500 & 0&08078 & 0&00060 & 0&00203 & 0&07713 & 0&00064 & 0&00326 & 0&05574 & 0&00050 & 0&00180\
5.500 & 6.000 & 0&02382 & 0&00032 & 0&00073 & 0&02004 & 0&00033 & 0&00113 & 0&01306 & 0&00024 & 0&00055\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l]{}&\
&\
&\
&\
0.000 & 0.010 & 389&18 & 0&37 & 28&68\
0.010 & 0.020 & 407&47 & 0&32 & 10&44\
0.020 & 0.030 & 266&10 & 0&26 & 3&67\
0.030 & 0.040 & 185&45 & 0&22 & 1&85\
0.040 & 0.050 & 137&73 & 0&19 & 1&38\
0.050 & 0.060 & 106&07 & 0&16 & 1&08\
0.060 & 0.070 & 84&57 & 0&15 & 0&95\
0.070 & 0.080 & 68&87 & 0&13 & 0&85\
0.080 & 0.090 & 57&09 & 0&12 & 0&78\
0.090 & 0.100 & 47&79 & 0&11 & 0&71\
0.100 & 0.120 & 37&756 & 0&071 & 0&619\
0.120 & 0.140 & 28&006 & 0&061 & 0&501\
0.140 & 0.160 & 21&330 & 0&054 & 0&409\
0.160 & 0.180 & 16&660 & 0&048 & 0&337\
0.180 & 0.200 & 13&043 & 0&042 & 0&273\
0.200 & 0.250 & 9&019 & 0&022 & 0&195\
0.250 & 0.300 & 5&464 & 0&018 & 0&124\
0.300 & 0.400 & 2&8036 & 0&0088 & 0&0683\
0.400 & 0.500 & 1&1938 & 0&0057 & 0&0327\
0.500 & 0.600 & 0&5322 & 0&0040 & 0&0169\
0.600 & 0.700 & 0&2287 & 0&0027 & 0&0084\
0.700 & 0.800 & 0&0936 & 0&0016 & 0&0039\
0.800 & 1.000 & 0&0201 & 0&0005 & 0&0023\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l]{}&\
&\
&\
&\
0.000 & 0.200 & 0&0151 & 0&0004 & 0&0022\
0.200 & 0.400 & 0&0774 & 0&0011 & 0&0051\
0.400 & 0.600 & 0&1987 & 0&0018 & 0&0083\
0.600 & 0.800 & 0&3970 & 0&0023 & 0&0116\
0.800 & 1.000 & 0&6805 & 0&0031 & 0&0161\
1.000 & 1.200 & 1&0520 & 0&0038 & 0&0226\
1.200 & 1.400 & 1&5128 & 0&0046 & 0&0311\
1.400 & 1.600 & 2&0270 & 0&0053 & 0&0398\
1.600 & 1.800 & 2&5997 & 0&0060 & 0&0481\
1.800 & 2.000 & 3&1993 & 0&0066 & 0&0553\
2.000 & 2.200 & 3&8069 & 0&0071 & 0&0610\
2.200 & 2.400 & 4&3851 & 0&0076 & 0&0651\
2.400 & 2.600 & 4&9235 & 0&0080 & 0&0680\
2.600 & 2.800 & 5&4160 & 0&0083 & 0&0702\
2.800 & 3.000 & 5&8079 & 0&0086 & 0&0716\
3.000 & 3.200 & 6&1608 & 0&0088 & 0&0734\
3.200 & 3.400 & 6&4026 & 0&0090 & 0&0753\
3.400 & 3.600 & 6&5371 & 0&0091 & 0&0779\
3.600 & 3.800 & 6&5822 & 0&0091 & 0&0819\
3.800 & 4.000 & 6&4738 & 0&0090 & 0&0869\
4.000 & 4.200 & 6&2383 & 0&0089 & 0&0933\
4.200 & 4.400 & 5&8490 & 0&0087 & 0&1003\
4.400 & 4.600 & 5&3231 & 0&0084 & 0&1068\
4.600 & 4.800 & 4&6603 & 0&0079 & 0&1098\
4.800 & 5.000 & 3&9029 & 0&0073 & 0&1064\
5.000 & 5.200 & 3&1150 & 0&0066 & 0&0952\
5.200 & 5.400 & 2&3781 & 0&0058 & 0&0777\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l]{}&\
&\
&\
&\
0.000 & 0.010 & 0&11124 & 0&00012 & 0&00097\
0.010 & 0.020 & 0&19506 & 0&00014 & 0&00113\
0.020 & 0.030 & 0&23694 & 0&00019 & 0&00118\
0.030 & 0.040 & 0&25986 & 0&00023 & 0&00130\
0.040 & 0.050 & 0&27485 & 0&00028 & 0&00137\
0.050 & 0.060 & 0&28678 & 0&00032 & 0&00143\
0.060 & 0.070 & 0&29812 & 0&00037 & 0&00149\
0.070 & 0.080 & 0&30653 & 0&00041 & 0&00153\
0.080 & 0.090 & 0&31324 & 0&00046 & 0&00157\
0.090 & 0.100 & 0&31969 & 0&00051 & 0&00160\
0.100 & 0.120 & 0&32899 & 0&00041 & 0&00164\
0.120 & 0.140 & 0&34140 & 0&00049 & 0&00171\
0.140 & 0.160 & 0&35107 & 0&00058 & 0&00176\
0.160 & 0.180 & 0&35833 & 0&00066 & 0&00179\
0.180 & 0.200 & 0&37165 & 0&00076 & 0&00186\
0.200 & 0.250 & 0&38189 & 0&00059 & 0&00191\
0.250 & 0.300 & 0&40047 & 0&00078 & 0&00200\
0.300 & 0.400 & 0&41720 & 0&00079 & 0&00223\
0.400 & 0.500 & 0&4348 & 0&0012 & 0&0031\
0.500 & 0.600 & 0&4436 & 0&0019 & 0&0044\
0.600 & 0.700 & 0&4412 & 0&0031 & 0&0062\
0.700 & 0.800 & 0&4525 & 0&0046 & 0&0093\
0.800 & 1.000 & 0&3778 & 0&0051 & 0&0112\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l]{}&\
&\
&\
&\
0.000 & 0.010 & 0&18791 & 0&00016 & 0&00171\
0.010 & 0.020 & 0&36450 & 0&00019 & 0&00214\
0.020 & 0.030 & 0&48985 & 0&00027 & 0&00245\
0.030 & 0.040 & 0&57682 & 0&00035 & 0&00288\
0.040 & 0.050 & 0&64402 & 0&00043 & 0&00322\
0.050 & 0.060 & 0&69974 & 0&00051 & 0&00350\
0.060 & 0.070 & 0&75195 & 0&00059 & 0&00376\
0.070 & 0.080 & 0&79370 & 0&00067 & 0&00397\
0.080 & 0.090 & 0&83170 & 0&00075 & 0&00416\
0.090 & 0.100 & 0&85958 & 0&00083 & 0&00430\
0.100 & 0.120 & 0&90697 & 0&00069 & 0&00453\
0.120 & 0.140 & 0&96796 & 0&00083 & 0&00484\
0.140 & 0.160 & 1&01237 & 0&00098 & 0&00506\
0.160 & 0.180 & 1&0555 & 0&0011 & 0&0053\
0.180 & 0.200 & 1&0903 & 0&0013 & 0&0055\
0.200 & 0.250 & 1&1594 & 0&0010 & 0&0061\
0.250 & 0.300 & 1&2363 & 0&0014 & 0&0076\
0.300 & 0.400 & 1&3226 & 0&0014 & 0&0100\
0.400 & 0.500 & 1&4043 & 0&0022 & 0&0135\
0.500 & 0.600 & 1&4327 & 0&0035 & 0&0183\
0.600 & 0.700 & 1&4210 & 0&0056 & 0&0246\
0.700 & 0.800 & 1&4134 & 0&0084 & 0&0336\
0.800 & 1.000 & 1&226 & 0&011 & 0&040\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l]{}&\
&\
&\
&\
-1.000 & -0.960 & 3&0850 & 0&0085 & 0&0154\
-0.960 & -0.920 & 0&7097 & 0&0040 & 0&0035\
-0.920 & -0.880 & 0&3794 & 0&0029 & 0&0019\
-0.880 & -0.840 & 0&2517 & 0&0024 & 0&0013\
-0.840 & -0.800 & 0&18476 & 0&00202 & 0&00092\
-0.800 & -0.760 & 0&14645 & 0&00179 & 0&00073\
-0.760 & -0.720 & 0&11995 & 0&00161 & 0&00060\
-0.720 & -0.680 & 0&10155 & 0&00149 & 0&00051\
-0.680 & -0.640 & 0&08866 & 0&00140 & 0&00044\
-0.640 & -0.600 & 0&07845 & 0&00132 & 0&00039\
-0.600 & -0.560 & 0&07092 & 0&00126 & 0&00036\
-0.560 & -0.520 & 0&06498 & 0&00122 & 0&00037\
-0.520 & -0.480 & 0&05975 & 0&00117 & 0&00037\
-0.480 & -0.440 & 0&05559 & 0&00113 & 0&00038\
-0.440 & -0.400 & 0&05211 & 0&00109 & 0&00038\
-0.400 & -0.360 & 0&04894 & 0&00106 & 0&00038\
-0.360 & -0.320 & 0&04681 & 0&00104 & 0&00038\
-0.320 & -0.280 & 0&04499 & 0&00102 & 0&00039\
-0.280 & -0.240 & 0&04383 & 0&00102 & 0&00039\
-0.240 & -0.200 & 0&04212 & 0&00100 & 0&00039\
-0.200 & -0.160 & 0&04078 & 0&00098 & 0&00039\
-0.160 & -0.120 & 0&03970 & 0&00096 & 0&00039\
-0.120 & -0.080 & 0&03900 & 0&00096 & 0&00039\
-0.080 & -0.040 & 0&03882 & 0&00096 & 0&00040\
-0.040 & 0.000 & 0&03849 & 0&00096 & 0&00040\
0.000 & 0.040 & 0&03825 & 0&00096 & 0&00040\
0.040 & 0.080 & 0&03803 & 0&00095 & 0&00040\
0.080 & 0.120 & 0&03816 & 0&00096 & 0&00040\
0.120 & 0.160 & 0&03812 & 0&00095 & 0&00040\
0.160 & 0.200 & 0&03874 & 0&00096 & 0&00041\
0.200 & 0.240 & 0&03943 & 0&00097 & 0&00041\
0.240 & 0.280 & 0&04023 & 0&00098 & 0&00041\
0.280 & 0.320 & 0&04090 & 0&00099 & 0&00041\
0.320 & 0.360 & 0&04233 & 0&00101 & 0&00041\
0.360 & 0.400 & 0&04361 & 0&00102 & 0&00041\
0.400 & 0.440 & 0&04521 & 0&00103 & 0&00040\
0.440 & 0.480 & 0&04759 & 0&00106 & 0&00040\
0.480 & 0.520 & 0&05050 & 0&00109 & 0&00040\
0.520 & 0.560 & 0&05331 & 0&00111 & 0&00039\
0.560 & 0.600 & 0&05702 & 0&00114 & 0&00038\
0.600 & 0.640 & 0&06238 & 0&00119 & 0&00038\
0.640 & 0.680 & 0&06943 & 0&00126 & 0&00038\
0.680 & 0.720 & 0&07782 & 0&00133 & 0&00039\
0.720 & 0.760 & 0&08949 & 0&00142 & 0&00045\
0.760 & 0.800 & 0&10526 & 0&00153 & 0&00053\
0.800 & 0.840 & 0&12980 & 0&00170 & 0&00065\
0.840 & 0.880 & 0&17143 & 0&00196 & 0&00086\
0.880 & 0.920 & 0&2539 & 0&0024 & 0&0013\
0.920 & 0.960 & 0&4860 & 0&0033 & 0&0024\
0.960 & 1.000 & 2&8965 & 0&0083 & 0&0166\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l]{}\
&\
&\
&\
&\
-1.000 & -0.960 & 0&18959 & 0&00051 & 0&00924\
-0.960 & -0.920 & 0&22480 & 0&00150 & 0&00220\
-0.920 & -0.880 & 0&12560 & 0&00120 & 0&00130\
-0.880 & -0.840 & 0&08023 & 0&00092 & 0&00080\
-0.840 & -0.800 & 0&05496 & 0&00073 & 0&00055\
-0.800 & -0.760 & 0&04122 & 0&00062 & 0&00049\
-0.760 & -0.720 & 0&03054 & 0&00050 & 0&00046\
-0.720 & -0.680 & 0&02383 & 0&00042 & 0&00044\
-0.680 & -0.640 & 0&01929 & 0&00036 & 0&00043\
-0.640 & -0.600 & 0&01613 & 0&00032 & 0&00042\
-0.600 & -0.560 & 0&01391 & 0&00029 & 0&00041\
-0.560 & -0.520 & 0&01182 & 0&00026 & 0&00079\
-0.520 & -0.480 & 0&00938 & 0&00021 & 0&00036\
-0.480 & -0.440 & 0&00803 & 0&00019 & 0&00035\
-0.440 & -0.400 & 0&00696 & 0&00017 & 0&00033\
-0.400 & -0.360 & 0&00546 & 0&00014 & 0&00029\
-0.360 & -0.320 & 0&00459 & 0&00012 & 0&00027\
-0.320 & -0.280 & 0&00408 & 0&00011 & 0&00026\
-0.280 & -0.240 & 0&00360 & 0&00010 & 0&00025\
-0.240 & -0.200 & 0&00268 & 0&00008 & 0&00020\
-0.200 & -0.160 & 0&00212 & 0&00006 & 0&00017\
-0.160 & -0.120 & 0&00148 & 0&00005 & 0&00013\
-0.120 & -0.080 & 0&00104 & 0&00003 & 0&00010\
-0.080 & -0.040 & 0&00069 & 0&00003 & 0&00010\
-0.040 & 0.000 & 0&00013 & 0&00002 & 0&00005\
Tables of Event Shape Distributions
-------------------------------------
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{} &\
& & &\
& & &\
& & &\
0.000 & 0.010 & 17&598 & 0&068 & 0&176 & 17&956 & 0&072 & 0&458 & 16&198 & 0&067 & 0&208\
0.010 & 0.020 & 18&570 & 0&068 & 0&186 & 19&635 & 0&071 & 0&267 & 20&008 & 0&072 & 0&246\
0.020 & 0.030 & 12&129 & 0&053 & 0&121 & 12&249 & 0&054 & 0&122 & 12&896 & 0&056 & 0&153\
0.030 & 0.040 & 8&225 & 0&044 & 0&082 & 7&985 & 0&042 & 0&080 & 8&237 & 0&043 & 0&094\
0.040 & 0.050 & 5&825 & 0&036 & 0&058 & 5&686 & 0&036 & 0&057 & 5&885 & 0&037 & 0&065\
0.050 & 0.060 & 4&600 & 0&033 & 0&046 & 4&345 & 0&031 & 0&046 & 4&458 & 0&032 & 0&048\
0.060 & 0.080 & 3&274 & 0&019 & 0&033 & 3&155 & 0&019 & 0&040 & 3&272 & 0&019 & 0&034\
0.080 & 0.100 & 2&309 & 0&016 & 0&023 & 2&256 & 0&016 & 0&032 & 2&290 & 0&016 & 0&023\
0.100 & 0.120 & 1&692 & 0&014 & 0&017 & 1&666 & 0&014 & 0&026 & 1&699 & 0&014 & 0&017\
0.120 & 0.160 & 1&2001 & 0&0082 & 0&0132 & 1&1914 & 0&0082 & 0&0203 & 1&2018 & 0&0082 & 0&0120\
0.160 & 0.200 & 0&8053 & 0&0067 & 0&0104 & 0&7994 & 0&0067 & 0&0145 & 0&7988 & 0&0067 & 0&0080\
0.200 & 0.250 & 0&5688 & 0&0050 & 0&0084 & 0&5758 & 0&0051 & 0&0107 & 0&5610 & 0&0050 & 0&0063\
0.250 & 0.300 & 0&4031 & 0&0043 & 0&0065 & 0&3972 & 0&0042 & 0&0071 & 0&3926 & 0&0042 & 0&0051\
0.300 & 0.350 & 0&2931 & 0&0036 & 0&0049 & 0&2908 & 0&0036 & 0&0044 & 0&2810 & 0&0035 & 0&0043\
0.350 & 0.400 & 0&2206 & 0&0031 & 0&0035 & 0&2244 & 0&0032 & 0&0023 & 0&2099 & 0&0030 & 0&0037\
0.400 & 0.500 & 0&1528 & 0&0018 & 0&0025 & 0&1500 & 0&0018 & 0&0015 & 0&1441 & 0&0018 & 0&0032\
0.500 & 0.600 & 0&0886 & 0&0014 & 0&0021 & 0&0900 & 0&0014 & 0&0026 & 0&0842 & 0&0013 & 0&0023\
0.600 & 0.700 & 0&04319 & 0&00094 & 0&00209 & 0&04351 & 0&00095 & 0&00394 & 0&04160 & 0&00092 & 0&00129\
0.700 & 0.850 & 0&00778 & 0&00032 & 0&00081 & 0&00809 & 0&00034 & 0&00184 & 0&00758 & 0&00032 & 0&00024\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{} &\
& & &\
& & &\
& & &\
0.000 & 0.005 & 81&41 & 0&20 & 0&81 & 81&83 & 0&21 & 1&03 & 75&10 & 0&19 & 0&75\
0.005 & 0.010 & 49&12 & 0&15 & 0&49 & 51&34 & 0&16 & 0&62 & 55&31 & 0&17 & 0&55\
0.010 & 0.015 & 24&05 & 0&11 & 0&24 & 24&15 & 0&10 & 0&33 & 26&03 & 0&11 & 0&28\
0.015 & 0.020 & 13&668 & 0&079 & 0&137 & 13&242 & 0&076 & 0&224 & 13&927 & 0&079 & 0&176\
0.020 & 0.030 & 6&893 & 0&039 & 0&088 & 6&477 & 0&037 & 0&142 & 6&768 & 0&038 & 0&098\
0.030 & 0.040 & 3&270 & 0&027 & 0&057 & 3&013 & 0&025 & 0&087 & 3&014 & 0&025 & 0&056\
0.040 & 0.060 & 1&443 & 0&013 & 0&032 & 1&295 & 0&012 & 0&048 & 1&281 & 0&012 & 0&035\
0.060 & 0.080 & 0&5841 & 0&0080 & 0&0153 & 0&5404 & 0&0077 & 0&0248 & 0&5181 & 0&0075 & 0&0188\
0.080 & 0.100 & 0&2825 & 0&0056 & 0&0085 & 0&2687 & 0&0055 & 0&0150 & 0&2619 & 0&0054 & 0&0118\
0.100 & 0.120 & 0&1515 & 0&0041 & 0&0051 & 0&1465 & 0&0041 & 0&0098 & 0&1461 & 0&0041 & 0&0079\
0.120 & 0.140 & 0&0831 & 0&0030 & 0&0031 & 0&0797 & 0&0030 & 0&0061 & 0&0758 & 0&0029 & 0&0043\
0.140 & 0.160 & 0&0535 & 0&0026 & 0&0023 & 0&0511 & 0&0025 & 0&0040 & 0&0467 & 0&0023 & 0&0027\
0.160 & 0.200 & 0&0262 & 0&0012 & 0&0014 & 0&0239 & 0&0012 & 0&0017 & 0&0234 & 0&0011 & 0&0014\
0.200 & 0.250 & 0&00969 & 0&00067 & 0&00062 & 0&01040 & 0&00071 & 0&00058 & 0&00884 & 0&00061 & 0&00052\
0.250 & 0.300 & 0&00320 & 0&00037 & 0&00075 & 0&00263 & 0&00035 & 0&00009 & 0&00310 & 0&00040 & 0&00018\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{} &\
& & &\
& & &\
& & &\
0.000 & 0.005 & 64&81 & 0&18 & 0&65 & 68&07 & 0&19 & 0&82 & 68&69 & 0&19 & 0&74\
0.005 & 0.010 & 32&31 & 0&12 & 0&32 & 31&44 & 0&12 & 0&39 & 31&66 & 0&12 & 0&35\
0.010 & 0.015 & 18&177 & 0&092 & 0&182 & 17&196 & 0&088 & 0&215 & 17&091 & 0&088 & 0&188\
0.015 & 0.020 & 12&001 & 0&075 & 0&120 & 11&362 & 0&072 & 0&145 & 11&370 & 0&072 & 0&127\
0.020 & 0.025 & 8&607 & 0&062 & 0&086 & 8&373 & 0&062 & 0&109 & 8&417 & 0&062 & 0&095\
0.025 & 0.030 & 6&792 & 0&056 & 0&068 & 6&542 & 0&055 & 0&087 & 6&578 & 0&055 & 0&075\
0.030 & 0.035 & 5&542 & 0&051 & 0&055 & 5&414 & 0&050 & 0&074 & 5&479 & 0&050 & 0&063\
0.035 & 0.040 & 4&519 & 0&045 & 0&045 & 4&461 & 0&045 & 0&062 & 4&493 & 0&045 & 0&052\
0.040 & 0.050 & 3&640 & 0&029 & 0&036 & 3&585 & 0&029 & 0&051 & 3&610 & 0&029 & 0&042\
0.050 & 0.060 & 2&757 & 0&025 & 0&028 & 2&747 & 0&025 & 0&041 & 2&749 & 0&025 & 0&033\
0.060 & 0.080 & 1&986 & 0&015 & 0&022 & 1&986 & 0&015 & 0&031 & 1&987 & 0&015 & 0&024\
0.080 & 0.100 & 1&362 & 0&012 & 0&017 & 1&394 & 0&013 & 0&023 & 1&362 & 0&012 & 0&017\
0.100 & 0.120 & 1&008 & 0&011 & 0&014 & 1&026 & 0&011 & 0&018 & 1&008 & 0&011 & 0&013\
0.120 & 0.160 & 0&6777 & 0&0061 & 0&0107 & 0&6898 & 0&0062 & 0&0133 & 0&6676 & 0&0061 & 0&0093\
0.160 & 0.200 & 0&4477 & 0&0050 & 0&0083 & 0&4502 & 0&0051 & 0&0096 & 0&4248 & 0&0048 & 0&0063\
0.200 & 0.250 & 0&2755 & 0&0034 & 0&0060 & 0&2807 & 0&0035 & 0&0067 & 0&2692 & 0&0034 & 0&0042\
0.250 & 0.300 & 0&1801 & 0&0028 & 0&0045 & 0&1833 & 0&0029 & 0&0048 & 0&1742 & 0&0028 & 0&0029\
0.300 & 0.350 & 0&1114 & 0&0022 & 0&0032 & 0&1121 & 0&0022 & 0&0033 & 0&1042 & 0&0021 & 0&0019\
0.350 & 0.400 & 0&0575 & 0&0015 & 0&0018 & 0&0598 & 0&0016 & 0&0019 & 0&0566 & 0&0015 & 0&0011\
0.400 & 0.500 & 0&0137 & 0&0005 & 0&0005 & 0&0144 & 0&0006 & 0&0005 & 0&0145 & 0&0006 & 0&0003\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{} &\
& & &\
& & &\
& & &\
0.000 & 0.010 & 3&386 & 0&028 & 0&132 & 1&366 & 0&025 & 0&125 & 1&030 & 0&019 & 0&076\
0.010 & 0.020 & 13&901 & 0&057 & 0&249 & 10&185 & 0&048 & 0&446 & 10&951 & 0&051 & 0&527\
0.020 & 0.030 & 15&733 & 0&061 & 0&157 & 16&184 & 0&062 & 0&288 & 17&645 & 0&066 & 0&547\
0.030 & 0.040 & 12&535 & 0&055 & 0&125 & 13&572 & 0&059 & 0&173 & 14&192 & 0&061 & 0&292\
0.040 & 0.050 & 9&494 & 0&048 & 0&095 & 10&004 & 0&050 & 0&134 & 10&009 & 0&050 & 0&152\
0.050 & 0.060 & 7&248 & 0&042 & 0&072 & 7&581 & 0&044 & 0&108 & 7&572 & 0&044 & 0&101\
0.060 & 0.070 & 5&665 & 0&037 & 0&057 & 5&933 & 0&039 & 0&090 & 5&760 & 0&038 & 0&076\
0.070 & 0.080 & 4&544 & 0&033 & 0&045 & 4&723 & 0&035 & 0&072 & 4&619 & 0&034 & 0&062\
0.080 & 0.090 & 3&722 & 0&030 & 0&040 & 3&929 & 0&032 & 0&056 & 3&792 & 0&031 & 0&051\
0.090 & 0.100 & 3&085 & 0&027 & 0&038 & 3&265 & 0&029 & 0&043 & 3&176 & 0&028 & 0&042\
0.100 & 0.120 & 2&413 & 0&017 & 0&027 & 2&578 & 0&018 & 0&035 & 2&456 & 0&018 & 0&032\
0.120 & 0.140 & 1&800 & 0&015 & 0&018 & 1&916 & 0&016 & 0&029 & 1&825 & 0&015 & 0&022\
0.140 & 0.160 & 1&387 & 0&013 & 0&014 & 1&477 & 0&014 & 0&025 & 1&401 & 0&013 & 0&016\
0.160 & 0.180 & 1&074 & 0&011 & 0&011 & 1&142 & 0&012 & 0&020 & 1&074 & 0&011 & 0&011\
0.180 & 0.200 & 0&8362 & 0&0099 & 0&0096 & 0&889 & 0&011 & 0&016 & 0&8262 & 0&0100 & 0&0083\
0.200 & 0.250 & 0&5683 & 0&0051 & 0&0101 & 0&6005 & 0&0055 & 0&0123 & 0&5525 & 0&0051 & 0&0065\
0.250 & 0.300 & 0&3138 & 0&0038 & 0&0077 & 0&3295 & 0&0040 & 0&0083 & 0&3030 & 0&0038 & 0&0058\
0.300 & 0.350 & 0&1286 & 0&0024 & 0&0045 & 0&1395 & 0&0026 & 0&0062 & 0&1312 & 0&0025 & 0&0044\
0.350 & 0.400 & 0&0180 & 0&0009 & 0&0011 & 0&0232 & 0&0012 & 0&0026 & 0&0238 & 0&0012 & 0&0014\
0.400 & 0.500 & 0&0004 & 0&0001 & 0&0001 & 0&0005 & 0&0002 & 0&0002 & 0&0007 & 0&0002 & 0&0001\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{} &\
& & &\
& & &\
& & &\
0.000 & 0.020 & 0&00207 & 0&00087 & 0&00313 & 0&0022 & 0&0023 & 0&0016 & 0&00040 & 0&00090 & 0&00005\
0.020 & 0.040 & 0&2165 & 0&0053 & 0&0114 & 0&1274 & 0&0058 & 0&0094 & 0&0590 & 0&0030 & 0&0058\
0.040 & 0.050 & 1&300 & 0&018 & 0&013 & 0&895 & 0&017 & 0&024 & 0&642 & 0&013 & 0&028\
0.050 & 0.060 & 2&734 & 0&027 & 0&027 & 2&337 & 0&026 & 0&059 & 2&178 & 0&024 & 0&086\
0.060 & 0.070 & 4&245 & 0&033 & 0&042 & 4&132 & 0&033 & 0&098 & 4&303 & 0&034 & 0&155\
0.070 & 0.080 & 5&280 & 0&037 & 0&053 & 5&469 & 0&038 & 0&122 & 5&849 & 0&039 & 0&192\
0.080 & 0.100 & 6&130 & 0&027 & 0&061 & 6&370 & 0&028 & 0&128 & 6&889 & 0&030 & 0&194\
0.100 & 0.120 & 5&861 & 0&027 & 0&059 & 5&870 & 0&026 & 0&103 & 6&342 & 0&028 & 0&143\
0.120 & 0.140 & 4&881 & 0&024 & 0&049 & 4&726 & 0&023 & 0&072 & 4&890 & 0&024 & 0&085\
0.140 & 0.160 & 3&966 & 0&022 & 0&040 & 3&876 & 0&021 & 0&052 & 3&900 & 0&021 & 0&050\
0.160 & 0.200 & 3&024 & 0&013 & 0&030 & 3&002 & 0&013 & 0&033 & 2&960 & 0&013 & 0&030\
0.200 & 0.240 & 2&166 & 0&011 & 0&023 & 2&169 & 0&011 & 0&022 & 2&124 & 0&011 & 0&021\
0.240 & 0.280 & 1&5714 & 0&0095 & 0&0177 & 1&6042 & 0&0097 & 0&0160 & 1&5562 & 0&0095 & 0&0156\
0.280 & 0.320 & 1&1782 & 0&0082 & 0&0146 & 1&2190 & 0&0085 & 0&0122 & 1&1807 & 0&0083 & 0&0118\
0.320 & 0.360 & 0&8962 & 0&0071 & 0&0125 & 0&9228 & 0&0074 & 0&0111 & 0&8693 & 0&0071 & 0&0087\
0.360 & 0.400 & 0&6771 & 0&0061 & 0&0106 & 0&6967 & 0&0064 & 0&0109 & 0&6493 & 0&0061 & 0&0065\
0.400 & 0.440 & 0&5164 & 0&0054 & 0&0092 & 0&5232 & 0&0055 & 0&0108 & 0&4820 & 0&0052 & 0&0048\
0.440 & 0.480 & 0&3940 & 0&0047 & 0&0079 & 0&3998 & 0&0048 & 0&0108 & 0&3493 & 0&0044 & 0&0055\
0.480 & 0.520 & 0&2796 & 0&0039 & 0&0064 & 0&2788 & 0&0040 & 0&0097 & 0&2497 & 0&0037 & 0&0065\
0.520 & 0.560 & 0&1835 & 0&0031 & 0&0047 & 0&1725 & 0&0031 & 0&0076 & 0&1489 & 0&0028 & 0&0058\
0.560 & 0.600 & 0&1048 & 0&0023 & 0&0030 & 0&0892 & 0&0022 & 0&0049 & 0&0714 & 0&0019 & 0&0038\
0.600 & 0.640 & 0&0453 & 0&0015 & 0&0015 & 0&0306 & 0&0014 & 0&0020 & 0&0203 & 0&0010 & 0&0014\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.020 & 0&1449 & 0&0043 & 0&0107 & 0&0689 & 0&0067 & 0&0117 & 0&0156 & 0&0017 & 0&0036\
0.020 & 0.040 & 3&109 & 0&019 & 0&031 & 1&968 & 0&019 & 0&100 & 1&236 & 0&013 & 0&066\
0.040 & 0.050 & 7&968 & 0&044 & 0&080 & 6&346 & 0&041 & 0&071 & 5&706 & 0&037 & 0&073\
0.050 & 0.060 & 10&353 & 0&050 & 0&104 & 9&491 & 0&048 & 0&116 & 9&714 & 0&048 & 0&125\
0.060 & 0.070 & 11&272 & 0&052 & 0&113 & 11&116 & 0&051 & 0&147 & 12&015 & 0&054 & 0&155\
0.070 & 0.080 & 10&890 & 0&051 & 0&109 & 11&163 & 0&051 & 0&159 & 12&437 & 0&055 & 0&161\
0.080 & 0.100 & 9&104 & 0&033 & 0&091 & 9&417 & 0&033 & 0&149 & 10&404 & 0&036 & 0&136\
0.100 & 0.120 & 6&292 & 0&027 & 0&063 & 6&569 & 0&028 & 0&118 & 6&918 & 0&029 & 0&092\
0.120 & 0.140 & 4&029 & 0&022 & 0&048 & 4&233 & 0&022 & 0&085 & 4&250 & 0&023 & 0&058\
0.140 & 0.160 & 2&519 & 0&017 & 0&037 & 2&612 & 0&018 & 0&058 & 2&517 & 0&017 & 0&035\
0.160 & 0.200 & 1&2832 & 0&0085 & 0&0243 & 1&3590 & 0&0092 & 0&0349 & 1&2561 & 0&0086 & 0&0187\
0.200 & 0.240 & 0&5255 & 0&0055 & 0&0130 & 0&5586 & 0&0059 & 0&0170 & 0&4895 & 0&0054 & 0&0080\
0.240 & 0.280 & 0&2160 & 0&0035 & 0&0067 & 0&2448 & 0&0040 & 0&0086 & 0&2112 & 0&0036 & 0&0039\
0.280 & 0.320 & 0&0928 & 0&0023 & 0&0066 & 0&1020 & 0&0026 & 0&0041 & 0&0879 & 0&0023 & 0&0018\
0.320 & 0.400 & 0&0248 & 0&0009 & 0&0012 & 0&0291 & 0&0010 & 0&0014 & 0&0250 & 0&0009 & 0&0006\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.020 & 7&415 & 0&031 & 0&106 & 8&103 & 0&033 & 0&193 & 9&357 & 0&036 & 0&178\
0.020 & 0.040 & 10&840 & 0&037 & 0&108 & 10&809 & 0&036 & 0&204 & 11&508 & 0&038 & 0&140\
0.040 & 0.060 & 7&800 & 0&030 & 0&078 & 7&294 & 0&029 & 0&109 & 7&215 & 0&029 & 0&072\
0.060 & 0.080 & 5&273 & 0&025 & 0&053 & 5&030 & 0&024 & 0&060 & 4&736 & 0&023 & 0&047\
0.080 & 0.100 & 3&797 & 0&021 & 0&038 & 3&697 & 0&021 & 0&037 & 3&477 & 0&020 & 0&035\
0.100 & 0.120 & 2&838 & 0&018 & 0&028 & 2&850 & 0&018 & 0&029 & 2&696 & 0&018 & 0&027\
0.120 & 0.140 & 2&218 & 0&016 & 0&022 & 2&208 & 0&016 & 0&022 & 2&106 & 0&016 & 0&021\
0.140 & 0.160 & 1&760 & 0&014 & 0&018 & 1&793 & 0&015 & 0&018 & 1&690 & 0&014 & 0&017\
0.160 & 0.200 & 1&3199 & 0&0087 & 0&0132 & 1&3247 & 0&0088 & 0&0132 & 1&2648 & 0&0085 & 0&0126\
0.200 & 0.240 & 0&8874 & 0&0070 & 0&0089 & 0&9033 & 0&0072 & 0&0091 & 0&8403 & 0&0069 & 0&0087\
0.240 & 0.280 & 0&6208 & 0&0058 & 0&0078 & 0&6179 & 0&0060 & 0&0065 & 0&5674 & 0&0056 & 0&0065\
0.280 & 0.320 & 0&4354 & 0&0049 & 0&0055 & 0&4268 & 0&0049 & 0&0049 & 0&3842 & 0&0046 & 0&0050\
0.320 & 0.360 & 0&2959 & 0&0040 & 0&0030 & 0&2855 & 0&0040 & 0&0046 & 0&2573 & 0&0037 & 0&0043\
0.360 & 0.400 & 0&1987 & 0&0033 & 0&0021 & 0&1841 & 0&0032 & 0&0045 & 0&1594 & 0&0029 & 0&0037\
0.400 & 0.440 & 0&1163 & 0&0025 & 0&0032 & 0&0977 & 0&0022 & 0&0032 & 0&0836 & 0&0020 & 0&0030\
0.440 & 0.520 & 0&0453 & 0&0011 & 0&0007 & 0&0293 & 0&0009 & 0&0014 & 0&0221 & 0&0007 & 0&0015\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.010 & 4&450 & 0&032 & 0&260 & 2&035 & 0&028 & 0&180 & 1&994 & 0&027 & 0&166\
0.010 & 0.020 & 17&310 & 0&062 & 0&173 & 16&492 & 0&059 & 0&165 & 18&580 & 0&065 & 0&709\
0.020 & 0.030 & 17&290 & 0&066 & 0&267 & 20&056 & 0&074 & 0&611 & 20&678 & 0&076 & 0&729\
0.030 & 0.040 & 12&996 & 0&058 & 0&206 & 13&772 & 0&061 & 0&377 & 13&377 & 0&060 & 0&412\
0.040 & 0.050 & 9&496 & 0&050 & 0&154 & 9&456 & 0&051 & 0&233 & 8&965 & 0&049 & 0&239\
0.050 & 0.060 & 6&957 & 0&043 & 0&116 & 6&862 & 0&043 & 0&153 & 6&558 & 0&041 & 0&151\
0.060 & 0.080 & 4&811 & 0&025 & 0&083 & 4&762 & 0&025 & 0&092 & 4&515 & 0&024 & 0&082\
0.080 & 0.100 & 3&064 & 0&020 & 0&055 & 3&071 & 0&020 & 0&051 & 2&914 & 0&019 & 0&037\
0.100 & 0.120 & 2&109 & 0&016 & 0&039 & 2&126 & 0&017 & 0&032 & 1&991 & 0&016 & 0&020\
0.120 & 0.140 & 1&468 & 0&014 & 0&028 & 1&478 & 0&014 & 0&023 & 1&406 & 0&013 & 0&014\
0.140 & 0.160 & 1&078 & 0&012 & 0&021 & 1&096 & 0&012 & 0&018 & 1&010 & 0&011 & 0&010\
0.160 & 0.200 & 0&6959 & 0&0067 & 0&0142 & 0&6983 & 0&0068 & 0&0149 & 0&6319 & 0&0063 & 0&0063\
0.200 & 0.250 & 0&3444 & 0&0042 & 0&0074 & 0&3461 & 0&0042 & 0&0118 & 0&3085 & 0&0039 & 0&0051\
0.250 & 0.300 & 0&1497 & 0&0027 & 0&0033 & 0&1457 & 0&0027 & 0&0082 & 0&1115 & 0&0022 & 0&0039\
0.300 & 0.350 & 0&0547 & 0&0015 & 0&0012 & 0&0309 & 0&0013 & 0&0027 & 0&0184 & 0&0008 & 0&0012\
0.350 & 0.400 & 0&0192 & 0&0009 & 0&0004 & 0&0022 & 0&0006 & 0&0003 & 0&0008 & 0&0002 & 0&0001\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.010 & 39&362 & 0&093 & 1&220 & 24&708 & 0&077 & 1&403 & 23&414 & 0&074 & 1&595\
0.010 & 0.020 & 31&177 & 0&092 & 1&074 & 35&91 & 0&10 & 2&35 & 39&12 & 0&11 & 2&65\
0.020 & 0.030 & 13&733 & 0&065 & 0&523 & 17&370 & 0&079 & 1&196 & 18&080 & 0&081 & 1&215\
0.030 & 0.040 & 6&486 & 0&044 & 0&270 & 7&628 & 0&051 & 0&516 & 7&704 & 0&052 & 0&514\
0.040 & 0.050 & 3&358 & 0&032 & 0&147 & 3&829 & 0&036 & 0&251 & 3&922 & 0&036 & 0&260\
0.050 & 0.060 & 1&879 & 0&024 & 0&082 & 2&085 & 0&026 & 0&132 & 2&128 & 0&026 & 0&140\
0.060 & 0.080 & 0&901 & 0&012 & 0&040 & 0&989 & 0&013 & 0&063 & 1&013 & 0&013 & 0&066\
0.080 & 0.100 & 0&3212 & 0&0072 & 0&0175 & 0&3326 & 0&0073 & 0&0234 & 0&3748 & 0&0079 & 0&0241\
0.100 & 0.120 & 0&1087 & 0&0044 & 0&0082 & 0&1135 & 0&0044 & 0&0096 & 0&1412 & 0&0050 & 0&0089\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.010 & 29&229 & 0&082 & 0&292 & 31&955 & 0&086 & 0&332 & 35&393 & 0&092 & 0&354\
0.010 & 0.020 & 19&648 & 0&068 & 0&196 & 20&327 & 0&071 & 0&203 & 20&745 & 0&071 & 0&207\
0.020 & 0.030 & 12&462 & 0&055 & 0&125 & 11&909 & 0&054 & 0&119 & 11&426 & 0&052 & 0&114\
0.030 & 0.040 & 8&332 & 0&045 & 0&083 & 7&654 & 0&043 & 0&077 & 7&170 & 0&041 & 0&072\
0.040 & 0.060 & 5&164 & 0&025 & 0&052 & 4&706 & 0&024 & 0&047 & 4&344 & 0&023 & 0&043\
0.060 & 0.080 & 2&987 & 0&019 & 0&034 & 2&816 & 0&019 & 0&034 & 2&605 & 0&017 & 0&026\
0.080 & 0.120 & 1&646 & 0&010 & 0&022 & 1&5632 & 0&0098 & 0&0267 & 1&4238 & 0&0092 & 0&0142\
0.120 & 0.160 & 0&8298 & 0&0070 & 0&0136 & 0&8043 & 0&0070 & 0&0186 & 0&7061 & 0&0064 & 0&0071\
0.160 & 0.200 & 0&4606 & 0&0053 & 0&0091 & 0&4374 & 0&0051 & 0&0132 & 0&3831 & 0&0046 & 0&0044\
0.200 & 0.250 & 0&2358 & 0&0033 & 0&0060 & 0&2223 & 0&0032 & 0&0087 & 0&1836 & 0&0028 & 0&0032\
0.250 & 0.300 & 0&1043 & 0&0022 & 0&0036 & 0&0838 & 0&0020 & 0&0042 & 0&0579 & 0&0015 & 0&0018\
0.300 & 0.350 & 0&0417 & 0&0013 & 0&0020 & 0&0144 & 0&0010 & 0&0009 & 0&0075 & 0&0006 & 0&0006\
0.350 & 0.400 & 0&0151 & 0&0008 & 0&0010 & 0&0008 & 0&0005 & 0&0001 & 0&0003 & 0&0002 & 0&0001\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.010 & 63&52 & 0&12 & 0&64 & 62&64 & 0&12 & 0&63 & 63&81 & 0&12 & 0&64\
0.010 & 0.020 & 12&230 & 0&054 & 0&122 & 12&326 & 0&054 & 0&126 & 11&589 & 0&052 & 0&116\
0.020 & 0.030 & 6&128 & 0&038 & 0&061 & 6&324 & 0&039 & 0&080 & 6&040 & 0&038 & 0&060\
0.030 & 0.040 & 3&828 & 0&030 & 0&038 & 3&993 & 0&031 & 0&058 & 3&823 & 0&030 & 0&038\
0.040 & 0.050 & 2&756 & 0&025 & 0&030 & 2&888 & 0&027 & 0&046 & 2&697 & 0&026 & 0&027\
0.050 & 0.060 & 1&995 & 0&022 & 0&028 & 2&095 & 0&023 & 0&036 & 2&041 & 0&022 & 0&020\
0.060 & 0.080 & 1&470 & 0&013 & 0&025 & 1&484 & 0&014 & 0&027 & 1&410 & 0&013 & 0&014\
0.080 & 0.100 & 0&932 & 0&011 & 0&019 & 0&963 & 0&011 & 0&020 & 0&952 & 0&011 & 0&011\
0.100 & 0.120 & 0&6551 & 0&0091 & 0&0162 & 0&6537 & 0&0089 & 0&0160 & 0&6708 & 0&0090 & 0&0102\
0.120 & 0.140 & 0&4985 & 0&0081 & 0&0144 & 0&4723 & 0&0076 & 0&0138 & 0&4831 & 0&0077 & 0&0098\
0.140 & 0.160 & 0&3442 & 0&0065 & 0&0114 & 0&3431 & 0&0065 & 0&0120 & 0&3660 & 0&0068 & 0&0098\
0.160 & 0.180 & 0&2619 & 0&0059 & 0&0103 & 0&2614 & 0&0057 & 0&0112 & 0&2830 & 0&0061 & 0&0102\
0.180 & 0.200 & 0&2023 & 0&0055 & 0&0099 & 0&1841 & 0&0048 & 0&0100 & 0&2190 & 0&0054 & 0&0109\
0.200 & 0.250 & 0&0978 & 0&0024 & 0&0060 & 0&1001 & 0&0023 & 0&0069 & 0&1357 & 0&0028 & 0&0092\
0.250 & 0.300 & 0&0248 & 0&0013 & 0&0019 & 0&0309 & 0&0015 & 0&0027 & 0&0605 & 0&0022 & 0&0055\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.002 & 318&65 & 0&62 & 3&19 & 318&41 & 0&62 & 3&18 & 333&97 & 0&64 & 3&34\
0.002 & 0.004 & 91&82 & 0&33 & 0&92 & 90&83 & 0&32 & 1&07 & 85&51 & 0&31 & 0&86\
0.004 & 0.006 & 33&87 & 0&20 & 0&34 & 33&23 & 0&20 & 0&49 & 29&49 & 0&18 & 0&29\
0.006 & 0.008 & 16&86 & 0&14 & 0&17 & 16&99 & 0&14 & 0&31 & 14&78 & 0&13 & 0&15\
0.008 & 0.012 & 8&383 & 0&070 & 0&120 & 8&546 & 0&073 & 0&189 & 7&583 & 0&067 & 0&076\
0.012 & 0.016 & 3&967 & 0&048 & 0&079 & 4&119 & 0&051 & 0&112 & 3&707 & 0&047 & 0&037\
0.016 & 0.020 & 2&218 & 0&037 & 0&058 & 2&325 & 0&039 & 0&077 & 2&119 & 0&037 & 0&042\
0.020 & 0.024 & 1&302 & 0&028 & 0&041 & 1&479 & 0&032 & 0&058 & 1&348 & 0&030 & 0&042\
0.024 & 0.028 & 0&906 & 0&024 & 0&033 & 0&961 & 0&026 & 0&043 & 0&906 & 0&024 & 0&038\
0.028 & 0.032 & 0&621 & 0&020 & 0&025 & 0&669 & 0&022 & 0&032 & 0&646 & 0&021 & 0&034\
0.032 & 0.040 & 0&344 & 0&010 & 0&016 & 0&387 & 0&012 & 0&019 & 0&366 & 0&011 & 0&025\
0.040 & 0.050 & 0&1892 & 0&0071 & 0&0104 & 0&2045 & 0&0078 & 0&0097 & 0&1965 & 0&0076 & 0&0177\
0.050 & 0.060 & 0&0867 & 0&0050 & 0&0060 & 0&0869 & 0&0050 & 0&0041 & 0&0877 & 0&0050 & 0&0100\
0.060 & 0.080 & 0&0331 & 0&0024 & 0&0030 & 0&0346 & 0&0023 & 0&0021 & 0&0367 & 0&0024 & 0&0054\
0.080 & 0.100 & 0&00742 & 0&00136 & 0&00088 & 0&00615 & 0&00096 & 0&00066 & 0&0073 & 0&0010 & 0&0014\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.001 & 688&3 & 1&3 & 6&9 & 684&2 & 1&3 & 6&8 & 711&0 & 1&3 & 7&1\
0.001 & 0.002 & 190&40 & 0&67 & 1&90 & 193&30 & 0&67 & 1&93 & 184&32 & 0&66 & 1&84\
0.002 & 0.003 & 60&98 & 0&37 & 0&61 & 61&50 & 0&38 & 2&31 & 53&35 & 0&35 & 1&06\
0.003 & 0.004 & 25&69 & 0&24 & 0&26 & 25&32 & 0&25 & 0&98 & 21&18 & 0&22 & 0&53\
0.004 & 0.005 & 12&63 & 0&17 & 0&18 & 12&43 & 0&18 & 0&49 & 10&15 & 0&15 & 0&30\
0.005 & 0.006 & 6&89 & 0&13 & 0&15 & 6&90 & 0&13 & 0&28 & 5&67 & 0&11 & 0&20\
0.006 & 0.008 & 3&442 & 0&063 & 0&112 & 3&377 & 0&066 & 0&145 & 2&768 & 0&057 & 0&114\
0.008 & 0.010 & 1&454 & 0&040 & 0&068 & 1&457 & 0&044 & 0&067 & 1&203 & 0&039 & 0&060\
0.010 & 0.012 & 0&750 & 0&031 & 0&045 & 0&744 & 0&032 & 0&037 & 0&617 & 0&028 & 0&036\
0.012 & 0.014 & 0&348 & 0&020 & 0&025 & 0&405 & 0&024 & 0&021 & 0&357 & 0&022 & 0&024\
0.014 & 0.016 & 0&204 & 0&016 & 0&017 & 0&258 & 0&020 & 0&015 & 0&215 & 0&017 & 0&016\
0.016 & 0.018 & 0&130 & 0&013 & 0&013 & 0&1551 & 0&0156 & 0&0095 & 0&144 & 0&015 & 0&011\
0.018 & 0.020 & 0&0712 & 0&0092 & 0&0078 & 0&0901 & 0&0118 & 0&0059 & 0&0745 & 0&0103 & 0&0064\
0.020 & 0.025 & 0&0342 & 0&0041 & 0&0044 & 0&0483 & 0&0057 & 0&0036 & 0&0434 & 0&0052 & 0&0042\
0.025 & 0.030 & 0&0110 & 0&0023 & 0&0017 & 0&0151 & 0&0028 & 0&0014 & 0&0133 & 0&0028 & 0&0014\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.010 & 26&014 & 0&077 & 0&260 & 24&267 & 0&073 & 0&442 & 26&317 & 0&078 & 0&289\
0.010 & 0.020 & 18&243 & 0&065 & 0&182 & 18&064 & 0&066 & 0&269 & 17&670 & 0&064 & 0&197\
0.020 & 0.030 & 11&945 & 0&054 & 0&119 & 12&277 & 0&056 & 0&157 & 11&622 & 0&054 & 0&134\
0.030 & 0.040 & 8&378 & 0&046 & 0&084 & 8&677 & 0&047 & 0&103 & 8&177 & 0&045 & 0&099\
0.040 & 0.050 & 6&218 & 0&039 & 0&062 & 6&442 & 0&041 & 0&079 & 6&053 & 0&039 & 0&078\
0.050 & 0.060 & 4&817 & 0&035 & 0&048 & 4&983 & 0&036 & 0&069 & 4&731 & 0&035 & 0&065\
0.060 & 0.080 & 3&490 & 0&021 & 0&035 & 3&691 & 0&022 & 0&059 & 3&479 & 0&021 & 0&051\
0.080 & 0.100 & 2&358 & 0&017 & 0&024 & 2&525 & 0&018 & 0&048 & 2&434 & 0&018 & 0&038\
0.100 & 0.120 & 1&680 & 0&014 & 0&020 & 1&797 & 0&015 & 0&039 & 1&740 & 0&015 & 0&029\
0.120 & 0.140 & 1&247 & 0&012 & 0&017 & 1&326 & 0&013 & 0&034 & 1&314 & 0&013 & 0&023\
0.140 & 0.160 & 0&941 & 0&011 & 0&016 & 0&993 & 0&011 & 0&029 & 0&981 & 0&011 & 0&018\
0.160 & 0.180 & 0&7392 & 0&0097 & 0&0154 & 0&7562 & 0&0098 & 0&0250 & 0&7495 & 0&0098 & 0&0150\
0.180 & 0.200 & 0&5596 & 0&0084 & 0&0153 & 0&5766 & 0&0086 & 0&0207 & 0&5711 & 0&0086 & 0&0126\
0.200 & 0.250 & 0&3330 & 0&0042 & 0&0120 & 0&3416 & 0&0042 & 0&0130 & 0&3633 & 0&0044 & 0&0091\
0.250 & 0.300 & 0&1317 & 0&0028 & 0&0062 & 0&1296 & 0&0026 & 0&0051 & 0&1621 & 0&0030 & 0&0046\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.002 & 107&87 & 0&36 & 1&08 & 103&55 & 0&35 & 1&04 & 113&89 & 0&37 & 1&14\
0.002 & 0.004 & 121&68 & 0&38 & 1&22 & 122&74 & 0&37 & 1&41 & 130&22 & 0&39 & 1&30\
0.004 & 0.006 & 74&79 & 0&29 & 0&75 & 73&19 & 0&29 & 0&95 & 71&16 & 0&28 & 0&80\
0.006 & 0.008 & 46&84 & 0&23 & 0&47 & 47&04 & 0&24 & 0&69 & 43&53 & 0&22 & 0&51\
0.008 & 0.012 & 27&71 & 0&13 & 0&28 & 28&13 & 0&13 & 0&48 & 25&56 & 0&12 & 0&32\
0.012 & 0.016 & 15&164 & 0&097 & 0&152 & 15&73 & 0&10 & 0&32 & 14&035 & 0&095 & 0&191\
0.016 & 0.020 & 9&124 & 0&075 & 0&091 & 9&798 & 0&082 & 0&228 & 8&816 & 0&075 & 0&133\
0.020 & 0.024 & 5&939 & 0&060 & 0&070 & 6&363 & 0&066 & 0&168 & 5&733 & 0&061 & 0&095\
0.024 & 0.028 & 4&101 & 0&051 & 0&064 & 4&417 & 0&056 & 0&130 & 4&009 & 0&051 & 0&074\
0.028 & 0.032 & 2&920 & 0&043 & 0&057 & 3&163 & 0&047 & 0&103 & 2&897 & 0&044 & 0&059\
0.032 & 0.040 & 1&857 & 0&024 & 0&047 & 2&026 & 0&027 & 0&076 & 1&849 & 0&025 & 0&044\
0.040 & 0.050 & 0&970 & 0&016 & 0&033 & 1&058 & 0&018 & 0&047 & 0&988 & 0&017 & 0&029\
0.050 & 0.060 & 0&509 & 0&012 & 0&022 & 0&542 & 0&013 & 0&028 & 0&513 & 0&012 & 0&019\
0.060 & 0.080 & 0&1952 & 0&0051 & 0&0113 & 0&2241 & 0&0057 & 0&0144 & 0&2203 & 0&0057 & 0&0114\
0.080 & 0.100 & 0&0554 & 0&0029 & 0&0043 & 0&0594 & 0&0031 & 0&0047 & 0&0654 & 0&0033 & 0&0049\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.001 & 248&15 & 0&77 & 2&48 & 227&18 & 0&75 & 3&97 & 240&63 & 0&78 & 4&58\
0.001 & 0.002 & 286&76 & 0&82 & 2&87 & 289&93 & 0&82 & 5&54 & 309&26 & 0&86 & 4&96\
0.002 & 0.003 & 167&03 & 0&63 & 1&67 & 173&10 & 0&64 & 3&59 & 171&69 & 0&64 & 2&39\
0.003 & 0.004 & 97&00 & 0&48 & 1&04 & 99&50 & 0&49 & 2&22 & 92&65 & 0&47 & 1&18\
0.004 & 0.005 & 59&58 & 0&38 & 0&71 & 60&08 & 0&39 & 1&44 & 52&87 & 0&35 & 0&65\
0.005 & 0.006 & 37&48 & 0&30 & 0&49 & 38&54 & 0&31 & 0&98 & 33&41 & 0&28 & 0&43\
0.006 & 0.008 & 21&61 & 0&16 & 0&32 & 22&33 & 0&17 & 0&62 & 18&83 & 0&15 & 0&27\
0.008 & 0.010 & 11&10 & 0&12 & 0&19 & 11&66 & 0&13 & 0&36 & 9&45 & 0&11 & 0&17\
0.010 & 0.012 & 6&235 & 0&087 & 0&124 & 6&717 & 0&098 & 0&231 & 5&567 & 0&086 & 0&122\
0.012 & 0.014 & 3&602 & 0&065 & 0&080 & 4&068 & 0&077 & 0&153 & 3&262 & 0&066 & 0&089\
0.014 & 0.016 & 2&205 & 0&052 & 0&055 & 2&454 & 0&060 & 0&100 & 2&047 & 0&052 & 0&070\
0.016 & 0.018 & 1&373 & 0&041 & 0&037 & 1&556 & 0&048 & 0&069 & 1&330 & 0&042 & 0&055\
0.018 & 0.020 & 0&874 & 0&032 & 0&026 & 1&035 & 0&040 & 0&049 & 0&912 & 0&036 & 0&044\
0.020 & 0.025 & 0&460 & 0&015 & 0&016 & 0&538 & 0&018 & 0&028 & 0&456 & 0&016 & 0&025\
0.025 & 0.030 & 0&1735 & 0&0096 & 0&0070 & 0&202 & 0&011 & 0&012 & 0&195 & 0&011 & 0&012\
0.030 & 0.050 & 0&0303 & 0&0021 & 0&0017 & 0&037 & 0&002 & 0&003 & 0&034 & 0&003 & 0&002\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.010 & 0.020 & 2&027 & 0&023 & 0&020 & 1&396 & 0&023 & 0&014 & 0&6707 & 0&0096 & 0&1077\
0.020 & 0.030 & 9&923 & 0&051 & 0&232 & 9&815 & 0&051 & 0&747 & 7&538 & 0&038 & 0&809\
0.030 & 0.040 & 14&758 & 0&061 & 0&245 & 15&520 & 0&062 & 0&819 & 14&690 & 0&058 & 0&745\
0.040 & 0.050 & 13&374 & 0&056 & 0&140 & 12&966 & 0&054 & 0&413 & 13&942 & 0&057 & 0&592\
0.050 & 0.060 & 10&620 & 0&050 & 0&106 & 10&186 & 0&048 & 0&136 & 11&298 & 0&053 & 0&379\
0.060 & 0.070 & 8&369 & 0&044 & 0&084 & 8&211 & 0&044 & 0&082 & 9&065 & 0&048 & 0&266\
0.070 & 0.080 & 6&811 & 0&040 & 0&068 & 6&790 & 0&040 & 0&068 & 7&387 & 0&043 & 0&222\
0.080 & 0.100 & 5&146 & 0&025 & 0&051 & 5&166 & 0&025 & 0&052 & 5&445 & 0&026 & 0&176\
0.100 & 0.120 & 3&584 & 0&020 & 0&036 & 3&685 & 0&021 & 0&037 & 3&796 & 0&022 & 0&127\
0.120 & 0.140 & 2&577 & 0&017 & 0&026 & 2&651 & 0&018 & 0&027 & 2&670 & 0&018 & 0&087\
0.140 & 0.170 & 1&733 & 0&012 & 0&017 & 1&780 & 0&012 & 0&018 & 1&756 & 0&012 & 0&051\
0.170 & 0.200 & 1&0540 & 0&0090 & 0&0105 & 1&0990 & 0&0094 & 0&0110 & 1&0580 & 0&0092 & 0&0218\
0.200 & 0.240 & 0&5388 & 0&0056 & 0&0135 & 0&5548 & 0&0058 & 0&0321 & 0&5288 & 0&0056 & 0&0053\
0.240 & 0.280 & 0&1710 & 0&0031 & 0&0104 & 0&1687 & 0&0031 & 0&0284 & 0&1460 & 0&0028 & 0&0071\
0.280 & 0.320 & 0&0262 & 0&0011 & 0&0028 & 0&0057 & 0&0008 & 0&0018 & 0&0029 & 0&0004 & 0&0003\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.010 & 2&502 & 0&025 & 0&025 & 1&552 & 0&026 & 0&115 & 0&645 & 0&010 & 0&096\
0.010 & 0.020 & 18&823 & 0&068 & 0&188 & 15&484 & 0&064 & 0&243 & 11&169 & 0&045 & 1&006\
0.020 & 0.030 & 28&683 & 0&084 & 0&287 & 29&092 & 0&084 & 0&291 & 28&908 & 0&082 & 1&823\
0.030 & 0.040 & 20&703 & 0&071 & 0&207 & 22&082 & 0&073 & 0&275 & 25&972 & 0&083 & 1&478\
0.040 & 0.050 & 11&654 & 0&052 & 0&117 & 11&600 & 0&052 & 0&139 & 14&119 & 0&061 & 0&860\
0.050 & 0.060 & 6&558 & 0&039 & 0&066 & 6&423 & 0&038 & 0&089 & 7&500 & 0&044 & 0&494\
0.060 & 0.080 & 3&191 & 0&019 & 0&032 & 3&205 & 0&019 & 0&052 & 3&405 & 0&021 & 0&233\
0.080 & 0.100 & 1&295 & 0&012 & 0&026 & 1&376 & 0&013 & 0&034 & 1&320 & 0&013 & 0&089\
0.100 & 0.120 & 0&5561 & 0&0078 & 0&0185 & 0&5970 & 0&0086 & 0&0222 & 0&5448 & 0&0082 & 0&0328\
0.120 & 0.150 & 0&1921 & 0&0037 & 0&0054 & 0&2114 & 0&0042 & 0&0070 & 0&1916 & 0&0040 & 0&0104\
0.150 & 0.180 & 0&0319 & 0&0015 & 0&0012 & 0&0356 & 0&0017 & 0&0016 & 0&0366 & 0&0017 & 0&0034\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.020 & 0.030 & 0&956 & 0&016 & 0&010 & 0&550 & 0&017 & 0&030 & 0&2030 & 0&0055 & 0&0383\
0.030 & 0.040 & 3&765 & 0&031 & 0&038 & 2&804 & 0&029 & 0&033 & 1&628 & 0&015 & 0&183\
0.040 & 0.050 & 7&117 & 0&043 & 0&071 & 6&646 & 0&042 & 0&066 & 4&999 & 0&031 & 0&463\
0.050 & 0.060 & 9&301 & 0&048 & 0&093 & 9&342 & 0&048 & 0&093 & 8&190 & 0&041 & 0&644\
0.060 & 0.070 & 10&059 & 0&050 & 0&101 & 10&265 & 0&050 & 0&103 & 9&887 & 0&047 & 0&661\
0.070 & 0.080 & 9&212 & 0&047 & 0&092 & 9&272 & 0&047 & 0&093 & 9&883 & 0&049 & 0&564\
0.080 & 0.090 & 7&958 & 0&043 & 0&080 & 7&915 & 0&043 & 0&079 & 9&007 & 0&047 & 0&443\
0.090 & 0.100 & 6&815 & 0&040 & 0&068 & 6&788 & 0&040 & 0&068 & 7&746 & 0&044 & 0&332\
0.100 & 0.110 & 5&898 & 0&037 & 0&059 & 5&899 & 0&037 & 0&059 & 6&714 & 0&041 & 0&255\
0.110 & 0.130 & 4&785 & 0&024 & 0&048 & 4&812 & 0&024 & 0&048 & 5&393 & 0&026 & 0&180\
0.130 & 0.150 & 3&619 & 0&021 & 0&036 & 3&667 & 0&021 & 0&038 & 3&998 & 0&023 & 0&125\
0.150 & 0.170 & 2&737 & 0&018 & 0&027 & 2&799 & 0&018 & 0&033 & 2&980 & 0&019 & 0&098\
0.170 & 0.190 & 2&143 & 0&016 & 0&021 & 2&203 & 0&016 & 0&031 & 2&294 & 0&017 & 0&085\
0.190 & 0.210 & 1&670 & 0&014 & 0&017 & 1&728 & 0&014 & 0&029 & 1&747 & 0&015 & 0&075\
0.210 & 0.240 & 1&2151 & 0&0096 & 0&0122 & 1&255 & 0&010 & 0&024 & 1&242 & 0&010 & 0&063\
0.240 & 0.270 & 0&8189 & 0&0078 & 0&0082 & 0&8504 & 0&0082 & 0&0194 & 0&8125 & 0&0080 & 0&0469\
0.270 & 0.300 & 0&5167 & 0&0061 & 0&0065 & 0&5243 & 0&0063 & 0&0141 & 0&4974 & 0&0062 & 0&0296\
0.300 & 0.330 & 0&2701 & 0&0044 & 0&0054 & 0&2599 & 0&0045 & 0&0087 & 0&2285 & 0&0041 & 0&0119\
0.330 & 0.360 & 0&0856 & 0&0024 & 0&0028 & 0&0831 & 0&0027 & 0&0042 & 0&0732 & 0&0024 & 0&0007\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.010 & 24&778 & 0&078 & 0&248 & 26&513 & 0&082 & 0&265 & 26&630 & 0&081 & 0&459\
0.010 & 0.020 & 18&957 & 0&068 & 0&190 & 18&385 & 0&066 & 0&184 & 18&684 & 0&066 & 0&292\
0.020 & 0.030 & 13&050 & 0&056 & 0&131 & 12&151 & 0&053 & 0&122 & 12&343 & 0&054 & 0&186\
0.030 & 0.040 & 9&129 & 0&046 & 0&091 & 8&722 & 0&045 & 0&087 & 8&819 & 0&046 & 0&134\
0.040 & 0.050 & 6&689 & 0&040 & 0&067 & 6&562 & 0&039 & 0&066 & 6&688 & 0&040 & 0&106\
0.050 & 0.060 & 5&220 & 0&035 & 0&052 & 5&134 & 0&035 & 0&051 & 5&111 & 0&035 & 0&084\
0.060 & 0.070 & 4&087 & 0&031 & 0&041 & 4&083 & 0&031 & 0&043 & 4&071 & 0&031 & 0&068\
0.070 & 0.080 & 3&314 & 0&028 & 0&033 & 3&293 & 0&028 & 0&043 & 3&271 & 0&028 & 0&054\
0.080 & 0.090 & 2&677 & 0&025 & 0&027 & 2&674 & 0&025 & 0&042 & 2&681 & 0&025 & 0&043\
0.090 & 0.100 & 2&218 & 0&023 & 0&022 & 2&240 & 0&023 & 0&040 & 2&233 & 0&023 & 0&035\
0.100 & 0.120 & 1&662 & 0&014 & 0&017 & 1&688 & 0&014 & 0&033 & 1&647 & 0&014 & 0&026\
0.120 & 0.140 & 1&145 & 0&012 & 0&011 & 1&162 & 0&012 & 0&025 & 1&111 & 0&011 & 0&019\
0.140 & 0.160 & 0&8022 & 0&0098 & 0&0108 & 0&8134 & 0&0099 & 0&0218 & 0&7618 & 0&0095 & 0&0144\
0.160 & 0.180 & 0&5135 & 0&0078 & 0&0109 & 0&5384 & 0&0081 & 0&0181 & 0&5138 & 0&0078 & 0&0119\
0.180 & 0.200 & 0&3325 & 0&0063 & 0&0091 & 0&3481 & 0&0067 & 0&0143 & 0&3167 & 0&0062 & 0&0098\
0.200 & 0.240 & 0&1540 & 0&0029 & 0&0053 & 0&1605 & 0&0031 & 0&0097 & 0&1265 & 0&0026 & 0&0056\
0.240 & 0.280 & 0&0389 & 0&0014 & 0&0023 & 0&0206 & 0&0013 & 0&0025 & 0&0117 & 0&0008 & 0&0008\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.040 & 0&3806 & 0&0047 & 0&0140 & 0&1375 & 0&0047 & 0&0138 & 0&0881 & 0&0030 & 0&0067\
0.040 & 0.080 & 2&388 & 0&012 & 0&050 & 1&5046 & 0&0098 & 0&0858 & 1&5383 & 0&0100 & 0&0831\
0.080 & 0.120 & 3&653 & 0&015 & 0&042 & 3&514 & 0&014 & 0&108 & 3&909 & 0&016 & 0&142\
0.120 & 0.160 & 3&323 & 0&014 & 0&033 & 3&570 & 0&015 & 0&058 & 3&833 & 0&016 & 0&088\
0.160 & 0.200 & 2&637 & 0&013 & 0&026 & 2&748 & 0&013 & 0&027 & 2&835 & 0&013 & 0&040\
0.200 & 0.240 & 2&091 & 0&011 & 0&021 & 2&149 & 0&011 & 0&021 & 2&164 & 0&012 & 0&022\
0.240 & 0.280 & 1&680 & 0&010 & 0&017 & 1&732 & 0&010 & 0&017 & 1&716 & 0&010 & 0&017\
0.280 & 0.320 & 1&3757 & 0&0092 & 0&0138 & 1&4127 & 0&0093 & 0&0141 & 1&3860 & 0&0092 & 0&0139\
0.320 & 0.360 & 1&1545 & 0&0083 & 0&0129 & 1&1920 & 0&0086 & 0&0122 & 1&1623 & 0&0084 & 0&0116\
0.360 & 0.400 & 0&9709 & 0&0076 & 0&0127 & 1&0093 & 0&0079 & 0&0115 & 0&9720 & 0&0077 & 0&0097\
0.400 & 0.440 & 0&8100 & 0&0068 & 0&0121 & 0&8629 & 0&0074 & 0&0108 & 0&8349 & 0&0072 & 0&0083\
0.440 & 0.480 & 0&7162 & 0&0065 & 0&0120 & 0&7545 & 0&0068 & 0&0105 & 0&7161 & 0&0066 & 0&0072\
0.480 & 0.520 & 0&6247 & 0&0060 & 0&0116 & 0&6544 & 0&0064 & 0&0101 & 0&6205 & 0&0061 & 0&0062\
0.520 & 0.560 & 0&5573 & 0&0057 & 0&0113 & 0&5914 & 0&0061 & 0&0103 & 0&5441 & 0&0057 & 0&0054\
0.560 & 0.600 & 0&4939 & 0&0053 & 0&0108 & 0&5187 & 0&0057 & 0&0101 & 0&4844 & 0&0054 & 0&0050\
0.600 & 0.640 & 0&4245 & 0&0049 & 0&0100 & 0&4558 & 0&0053 & 0&0101 & 0&4209 & 0&0050 & 0&0063\
0.640 & 0.680 & 0&3860 & 0&0048 & 0&0097 & 0&3968 & 0&0049 & 0&0101 & 0&3699 & 0&0046 & 0&0079\
0.680 & 0.720 & 0&3589 & 0&0046 & 0&0095 & 0&3697 & 0&0047 & 0&0106 & 0&3286 & 0&0044 & 0&0099\
0.720 & 0.760 & 0&3063 & 0&0041 & 0&0084 & 0&3250 & 0&0044 & 0&0104 & 0&2813 & 0&0040 & 0&0129\
0.760 & 0.800 & 0&2672 & 0&0039 & 0&0073 & 0&2807 & 0&0042 & 0&0096 & 0&2178 & 0&0033 & 0&0151\
0.800 & 0.840 & 0&1485 & 0&0028 & 0&0037 & 0&1803 & 0&0035 & 0&0062 & 0&1287 & 0&0026 & 0&0130\
0.840 & 0.880 & 0&0716 & 0&0020 & 0&0015 & 0&0831 & 0&0023 & 0&0026 & 0&0542 & 0&0016 & 0&0076\
0.880 & 0.920 & 0&0307 & 0&0013 & 0&0004 & 0&0361 & 0&0016 & 0&0008 & 0&0212 & 0&0009 & 0&0040\
[r@[ - ]{}rr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}lr@[.]{}l@[ $\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l r@[.]{}l@[$\pm$ ]{}r@[.]{}l@[ $\pm$ ]{}r@[.]{}l ]{}&\
& & &\
& & &\
& & &\
0.000 & 0.008 & 29&731 & 0&093 & 0&308 & 22&567 & 0&083 & 0&645 & 22&228 & 0&082 & 0&868\
0.008 & 0.016 & 19&922 & 0&077 & 0&199 & 20&963 & 0&080 & 0&401 & 22&766 & 0&085 & 0&440\
0.016 & 0.030 & 11&026 & 0&044 & 0&110 & 11&522 & 0&045 & 0&169 & 12&107 & 0&047 & 0&150\
0.030 & 0.044 & 6&513 & 0&034 & 0&065 & 6&763 & 0&035 & 0&093 & 6&879 & 0&035 & 0&079\
0.044 & 0.066 & 4&066 & 0&021 & 0&045 & 4&262 & 0&022 & 0&064 & 4&284 & 0&022 & 0&053\
0.066 & 0.088 & 2&640 & 0&017 & 0&033 & 2&790 & 0&018 & 0&048 & 2&727 & 0&018 & 0&036\
0.088 & 0.112 & 1&864 & 0&014 & 0&025 & 1&964 & 0&014 & 0&039 & 1&909 & 0&014 & 0&028\
0.112 & 0.136 & 1&352 & 0&012 & 0&019 & 1&469 & 0&013 & 0&033 & 1&415 & 0&012 & 0&022\
0.136 & 0.162 & 1&0013 & 0&0095 & 0&0149 & 1&113 & 0&011 & 0&028 & 1&051 & 0&010 & 0&018\
0.162 & 0.188 & 0&7829 & 0&0085 & 0&0123 & 0&8654 & 0&0094 & 0&0239 & 0&7977 & 0&0089 & 0&0145\
0.188 & 0.218 & 0&6187 & 0&0071 & 0&0103 & 0&6835 & 0&0078 & 0&0204 & 0&6155 & 0&0073 & 0&0117\
0.218 & 0.248 & 0&4490 & 0&0059 & 0&0081 & 0&5253 & 0&0069 & 0&0168 & 0&4566 & 0&0063 & 0&0089\
0.248 & 0.284 & 0&3425 & 0&0048 & 0&0068 & 0&3915 & 0&0055 & 0&0134 & 0&3341 & 0&0049 & 0&0065\
0.284 & 0.320 & 0&2575 & 0&0042 & 0&0057 & 0&2834 & 0&0046 & 0&0103 & 0&2452 & 0&0042 & 0&0049\
0.320 & 0.360 & 0&1884 & 0&0034 & 0&0047 & 0&2142 & 0&0039 & 0&0083 & 0&1774 & 0&0033 & 0&0037\
0.360 & 0.400 & 0&1363 & 0&0028 & 0&0038 & 0&1629 & 0&0034 & 0&0067 & 0&1234 & 0&0028 & 0&0028\
0.400 & 0.450 & 0&0962 & 0&0021 & 0&0029 & 0&1095 & 0&0025 & 0&0049 & 0&0902 & 0&0021 & 0&0023\
0.450 & 0.500 & 0&0644 & 0&0017 & 0&0021 & 0&0758 & 0&0020 & 0&0036 & 0&0603 & 0&0017 & 0&0018\
0.500 & 0.560 & 0&0394 & 0&0012 & 0&0014 & 0&0466 & 0&0015 & 0&0023 & 0&0368 & 0&0012 & 0&0013\
0.560 & 0.620 & 0&0277 & 0&0010 & 0&0010 & 0&0314 & 0&0012 & 0&0016 & 0&0222 & 0&0009 & 0&0009\
0.620 & 0.710 & 0&0151 & 0&0006 & 0&0006 & 0&0182 & 0&0007 & 0&0009 & 0&0128 & 0&0006 & 0&0006\
0.710 & 0.800 & 0&0071 & 0&0004 & 0&0003 & 0&0084 & 0&0005 & 0&0040 & 0&0052 & 0&0004 & 0&0003\
Tables of Results of the Fits
=============================
\[fit\_tab\]
----------- --- ---- --- ----- --- ----- --- ----- --- ----- -- --
Parameter
QCDLAM 0 18 0 155 0 205 0 163 0 001
RMASS(13) 0 75 0 64 0 96 0 65 0 01
CLMAX 3 35 2 35 4 35 3 48 0 04
CLPOW 2 0 1 0 2 0 1 49 0 04
CLSMR 0 0 0 1 0 0 36 0 04
DECWT 1 0 0 0 1 0 0 77 0 08
PWT(3) 1 0 0 0 1 0 0 83 0 02
PWT(7) 1 0 0 0 1 0 0 74 0 09
----------- --- ---- --- ----- --- ----- --- ----- --- ----- -- --
: \[th58c\] Parameter setting and fit results for HERWIG 5.8 C
------------------------- ---------- --- ---- ---- ------- --- ------- --- ----- --- ----- -- --
Parameter Name
$\lambda_{QCD}$ PARJ(81) 0 4 0 25 0 35 0 297 0 005
$Q_0 $ PARJ(82) 1 0 1 0 2 0 1 34 0 08
$a$ PARJ(41) 0 5 0 1 0 5 0 409 0 020
$b$ PARJ(42) 0 9
$\sigma_q$ PARJ(21) 0 35 0 36 0 44 0 401 0 006
$P(^1S_0)_{ud}$ - 0 5 0 3 0 5 0 272 0 030
$P(^3S_1)_{ud}$ - 0 5 0 2 0 4 0 306 0 049
$P(^1P_1)_{ud}$ - 0
$P(oth. P-states)_{ud}$ - 0
$\gamma_s$ PARJ(2) 0 30 0 27 0 31 0 307 0 007
$P(^1S_0)_{s}$ - 0 4 0 3 0 5 0 416 0 029
$P(^3S_1)_{s}$ - 0 6 0 2 0 4 0 304 0 023
$P(P-states)_{s}$ - 0
$\epsilon_c $ PARJ(54) -0 0372 0 0008
$P(^1S_0)_{c}$ - 0 25
$P(^3S_1)_{c}$ - 0 75
$P(P-states)_{c}$ - 0
$\epsilon_b $ PARJ(55) -0 00284 0 00006
$P(^1S_0)_{b}$ - 0 25
$P(^3S_1)_{b}$ - 0 75
$P(P-states)_{b}$ - 0
$P(qq)/P(q)$ PARJ(1) 0 1 0 08 0 11 0 095 0 001
$P(us)/P(ud)/\gamma_s$ PARJ(3) 0 4
$P(ud1)/P(ud0)$ PARJ(4) 0 05
extra baryon supp. PARJ(19) 0
extra $\eta$ supp. PARJ(25) 1 0
extra $\eta'$ supp. PARJ(26) 1 0
------------------------- ---------- --- ---- ---- ------- --- ------- --- ----- --- ----- -- --
: \[terg73d\] Parameter setting and fit results for JETSET 7.3 PS with DELPHI decays
------------------------- ---------- --- ---- ---- ------- --- ------- --- ----- --- ----- -- --
Parameter Name
$\lambda_{QCD}$ PARJ(81) 0 4 0 25 0 35 0 297 0 004
$Q_0 $ PARJ(82) 1 0 1 0 2 0 1 56 0 11
$a$ PARJ(41) 0 5 0 1 0 5 0 417 0 022
$b$ PARJ(42) 0 9
$\sigma_q$ PARJ(21) 0 35 0 36 0 44 0 408 0 005
$P(^1S_0)_{ud}$ - 0 5 0 3 0 5 0 297 0 021
$P(^3S_1)_{ud}$ - 0 5 0 2 0 4 0 289 0 038
$P(^1P_1)_{ud}$ - 0
$P(oth. P-states)_{ud}$ - 0
$\gamma_s$ PARJ(2) 0 30 0 27 0 31 0 308 0 007
$P(^1S_0)_{s}$ - 0 4 0 3 0 5 0 410 0 038
$P(^3S_1)_{s}$ - 0 6 0 2 0 4 0 297 0 021
$P(P-states)_{s}$ - 0
$\epsilon_c $ PARJ(54) -0 0372 0 0007
$P(^1S_0)_{c}$ - 0 25
$P(^3S_1)_{c}$ - 0 75
$P(P-states)_{c}$ - 0
$\epsilon_b $ PARJ(55) -0 00284 0 00005
$P(^1S_0)_{b}$ - 0 25
$P(^3S_1)_{b}$ - 0 75
$P(P-states)_{b}$ - 0
$P(qq)/P(q)$ PARJ(1) 0 1 0 08 0 11 0 099 0 001
$P(us)/P(ud)/\gamma_s$ PARJ(3) 0 4
$P(ud1)/P(ud0)$ PARJ(4) 0 05
extra baryon supp. PARJ(19) 0
extra $\eta$ supp. PARJ(25) 1 0
extra $\eta'$ supp. PARJ(26) 1 0
------------------------- ---------- --- ---- ---- ------- --- ------- --- ----- --- ----- -- --
: \[terg74d\] Parameter setting and fit results for JETSET 7.4 PS with default decays
------------------------- ---------- --- ---- ---- ------- --- ------- --- ----- --- ----- -- --
Parameter Name
$\lambda_{QCD}$ PARA(1) 0 22 0 20 0 30 0 237 0 003
$p_t^{QCD}$ PARA(3) 0 6 0 4 1 2 0 64 0 08
$a$ PARJ(41) 0 5 0 1 0 5 0 391 0 047
$b$ PARJ(42) 0 9
$\sigma_q$ PARJ(21) 0 35 0 36 0 44 0 390 0 005
$P(^1S_0)_{ud}$ - 0 5 0 3 0 5 0 276 0 017
$P(^3S_1)_{ud}$ - 0 5 0 2 0 4 0 292 0 056
$P(^1P_1)_{ud}$ - 0
$P(oth. P-states)_{ud}$ - 0
$\gamma_s$ PARJ(2) 0 30 0 27 0 31 0 302 0 004
$P(^1S_0)_{s}$ - 0 4 0 3 0 5 0 389 0 026
$P(^3S_1)_{s}$ - 0 6 0 2 0 4 0 301 0 019
$P(P-states)_{s}$ - 0
$\epsilon_c $ PARJ(54) -0 0378 0 0005
$P(^1S_0)_{c}$ - 0 25
$P(^3S_1)_{c}$ - 0 75
$P(P-states)_{c}$ - 0
$\epsilon_b $ PARJ(55) -0 00255 0 00004
$P(^1S_0)_{b}$ - 0 25
$P(^3S_1)_{b}$ - 0 75
$P(P-states)_{b}$ - 0
$P(qq)/P(q)$ PARJ(1) 0 1 0 08 0 11 0 096 0 001
$P(us)/P(ud)/\gamma_s$ PARJ(3) 0 4
$P(ud1)/P(ud0)$ PARJ(4) 0 05
extra baryon supp. PARJ(19) 0
extra $\eta$ supp. PARJ(25) 1 0
extra $\eta'$ supp. PARJ(26) 1 0
------------------------- ---------- --- ---- ---- ------- --- ------- --- ----- --- ----- -- --
: \[tar46d\] Parameter setting and fit results for ARIADNE 4.06 with DELPHI decays
------------------------- ---------- --- ---- ---- ------- --- ------- --- ----- --- ----- -- --
Parameter name
$\lambda_{QCD}$ PARA(1) 0 22 0 20 0 30 0 237 0 004
$p_t^{QCD}$ PARA(3) 0 6 0 4 1 2 0 68 0 06
$a$ PARJ(41) 0 5 0 1 0 5 0 388 0 030
$b$ PARJ(42) 0 9
$\sigma_q$ PARJ(21) 0 35 0 36 0 44 0 387 0 008
$P(^1S_0)_{ud}$ - 0 5 0 3 0 5 0 326 0 017
$P(^3S_1)_{ud}$ - 0 5 0 2 0 4 0 263 0 039
$P(^1P_1)_{ud}$ - 0
$P(oth. P-states)_{ud}$ - 0
$\gamma_s$ PARJ(2) 0 30 0 27 0 31 0 282 0 005
$P(^1S_0)_{s}$ - 0 4 0 3 0 5 0 395 0 023
$P(^3S_1)_{s}$ - 0 6 0 2 0 4 0 325 0 015
$P(P-states)_{s}$ - 0
$\epsilon_c $ PARJ(54) -0 0378 0 0005
$P(^1S_0)_{c}$ - 0 25
$P(^3S_1)_{c}$ - 0 75
$P(P-states)_{c}$ - 0
$\epsilon_b $ PARJ(55) -0 00255 0 00004
$P(^1S_0)_{b}$ - 0 25
$P(^3S_1)_{b}$ - 0 75
$P(P-states)_{b}$ - 0
$P(qq)/P(q)$ PARJ(1) 0 1 0 08 0 11 0 101 0 001
$P(us)/P(ud)/\gamma_s$ PARJ(3) 0 4
$P(ud1)/P(ud0)$ PARJ(4) 0 05
extra baryon supp. PARJ(19) 0
extra $\eta$ supp. PARJ(25) 1 0
extra $\eta'$ supp. PARJ(26) 1 0
------------------------- ---------- --- ---- ---- ------- --- ------- --- ----- --- ----- -- --
: \[tar46dd\] Parameter setting and fit results for ARIADNE 4.06 with default decays
------------------------- ----------- --- ----- --- ------ --- ------ --- ------- --- ------- -- --
Parameter Name
$\lambda_{QCD}$ PARJ(122) 0 22 0 20 0 30 0 163 0 002
$Scale$ $\mu$ PARJ(129) 0 002 0 0005 0 0025 0 00250 0 00006
$a$ PARJ(41) 0 5 0 1 0 5 0 903 0 014
$b$ PARJ(42) 0 9
$\sigma_q$ PARJ(21) 0 35 0 36 0 44 0 477 0 004
$P(^1S_0)_{ud}$ - 0 5 0 3 0 5 0 327 0 023
$P(^3S_1)_{ud}$ - 0 5 0 2 0 4 0 274 0 061
$P(^1P_1)_{ud}$ - 0
$P(oth. P-states)_{ud}$ - 0
$\gamma_s$ PARJ(2) 0 30 0 27 0 31 0 277 0 006
$P(^1S_0)_{s}$ - 0 4 0 3 0 5 0 413 0 028
$P(^3S_1)_{s}$ - 0 6 0 2 0 4 0 330 0 021
$P(P-states)_{s}$ - 0
$\epsilon_c $ PARJ(54)
$P(^1S_0)_{c}$ - 0 25
$P(^3S_1)_{c}$ - 0 75
$P(P-states)_{c}$ - 0
$\epsilon_b $ PARJ(55)
$P(^1S_0)_{b}$ - 0 25
$P(^3S_1)_{b}$ - 0 75
$P(P-states)_{b}$ - 0
$P(qq)/P(q)$ PARJ(1) 0 1 0 08 0 11 0 087 0 002
$P(us)/P(ud)/\gamma_s$ PARJ(3) 0 4
$P(ud1)/P(ud0)$ PARJ(4) 0 05
extra baryon supp. PARJ(19) 0
extra $\eta$ supp. PARJ(25) 1 0
extra $\eta'$ supp. PARJ(26) 1 0
------------------------- ----------- --- ----- --- ------ --- ------ --- ------- --- ------- -- --
: \[terg74m\] Parameter setting and fit results for JETSET 7.4 ME with default decays
Tables of Correlation Coefficients
==================================
\[correl\_tabs\]
\[h\]
----------------- -- --- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- --
$a$ 0 33 -0 60 0 01 -0 10 0 27 -0 11 0 06 0 03
$\sigma_q$ 1 00 -0 69 -0 08 0 43 -0 41 0 11 -0 26 -0 29
$\lambda_{QCD}$ 1 00 0 18 -0 26 0 19 -0 10 0 15 0 20
$Q_0$ 1 00 0 20 0 06 -0 38 -0 35 -0 13
$\gamma_s$ 1 00 -0 04 0 19 -0 73 -0 65
$P(^1S_0)_{ud}$ 1 00 0 37 -0 29 0 12
$P(^3S_1)_{ud}$ 1 00 -0 10 -0 26
$P(^1S_0)_{s}$ 1 00 0 41
$P(^3S_1)_{s}$ 1 00
----------------- -- --- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- --
: \[covjt74\] Table of Correlation Coefficients for JETSET 7.4 PS Fit
\[h\]
----------------- -- --- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- --- ---- --
$a$ 0 11 -0 18 0 90 -0 29 0 71 -0 04 0 29 0 07
$\sigma_q$ 1 00 -0 60 -0 07 0 12 -0 47 -0 06 0 02 0 05
$\lambda_{QCD}$ 1 00 0 16 0 02 0 18 0 03 -0 03 -0 06
$p_t^{QCD}$ 1 00 -0 21 0 65 -0 02 0 26 0 05
$\gamma_s$ 1 00 -0 27 -0 02 -0 41 -0 47
$P(^1S_0)_{ud}$ 1 00 -0 11 -0 03 0 13
$P(^3S_1)_{ud}$ 1 00 0 08 -0 03
$P(^1S_0)_{s}$ 1 00 0 02
$P(^3S_1)_{s}$ 1 00
----------------- -- --- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- ---- --- ---- --
: \[covar46d\] Table of Correlation Coefficients for ARIADNE 4.06 PS Fit
Figures of Data and Model Comparisons {#plots}
=====================================
The figures in this section compare DELPHI data and identified particle spectra to the following fragmentation models:
- JETSET 7.3 DELPHI decays labeled JT 7.3 PS
- JETSET 7.4 default decays labeled JT 7.4 PS
- ARIADNE 4.06 DELPHI decays labeled AR 4.06
- HERWIG 5.8 c default decays labeled H 5.8C
- JETSET 7.4 ME default decays labeled JT 7.4 ME
The references for the identified particle data can be found in table \[evtfit3\]. For most shape distributions we show on the left side the data as measured from charged particles only, on the right as measured from charged plus neutral particles. The distributions are corrected to the corresponding final states. The total correction factor for each distribution is shown in the upper inset. The lower insets of the plots decipt the relative deviation of the models to the data. Also shown as shaded area in these insets is the total experimental error obtained by adding quadratically the systematic and statistical error in each bin. Except for the point-to-point scatter the error should be interpreted like a statistical ”$1 \sigma$” uncertainty.
Inclusive Distributions
-----------------------
Shape Distributions
-------------------
\[h\]
Identified Particle Distributions
---------------------------------
\[t\]
\[t\]
[99]{} DELPHI Coll., P. Aarnio et al., Nucl. Inst. Methods A303 (1991) 187 F. Jadach, B.F.L. Ward, Z. Was, Comput. Phys. Commun. 40 (1986) 285; Nucl. Phys. B253 (1985) 441 DELPHI Coll., DELPHI 89-67(1989); DELPHI 89-68(1989). T. Sjöstrand, Comput. Phys. Commun. 39 (1986) 347; T. Sjöstrand and M. Bengtsson, Comput. Phys. Commun. 46 (1987) 367 J.E. Campagne and R. Zitoun, Z. Phys. C 43 (1989) 469 V. Blobel, Unfolding methods in high energy physics experiments, Lecture given at the 1984 School of Computing, DESY 84-118; V.B. Anikeev, V.P. Zhigunov, Phys. of Part. and Nucl. 24 Nr. 4 (1993) 424 H. Fürstenau, Ph. D. Thesis, IEKP-KA 92-16 L. Lönnblad, Comp. Phys. Comm. 71 (1992) 15. G. Marchesini et al., Comp. Phys. Comm. 67 (1992) 465. C. Peterson, D. Schlatter, I. Schmitt and P. Zerwas, Phys. Rev. D27 (1983) 105 B. Andersson, G. Gustafson, G. Ingelman, T. Sjöstrand, Phys. Rep. 97 (1983) 31. G. Gustavson, U. Petterson, Nucl. Phys. B308 (1988) 746 F. Gutbrod, G. Kramer, G. Schierholz, Z. Phys. C21(1984) 235\
K. Fabricius et al., Z. Phys C11(1981) 315 R.K. Ellis, D.A. Ross, E.A. Terrano, Phys. Rev. Lett. 45 (1980) 1225,\
Nucl. Phys. B178(1981) 421 compare discussion and references in: QCD Generators, B. Bambah et al. CERN 89-08 Vol. 3 p. 143 ff. P.M. Stevenson, Phys. Rev. D23 (1981) 2916 compare discussion and references in: QCD Generators, B. Bambah et al. CERN 89-08 Vol. 3 p. 235 ff. ALEPH Coll., D. Busculic et al., Z. Phys. C 64, 361 (1994) ALEPH Coll., D. Decamp et al., Z. Phys. C 55, 209 (1992) TASSO Coll., M. Althoff et al., Z. Phys. C 26 (1984) 157; TASSO Coll., W. Braunschweig et al. Z. Phys. C 41 (1988) 359 L3 Coll., B. Adeva et al., Z. Phys. C 55 (1992) 39 W.H. Press, B.P. Flannery, S.A. Teukolsky, W.T. Vetterling, Numerical Recipes, [*The Art of Scientific Computing*]{}, Cambridge University Press NAG library MK16, NAG ltd., Oxford, GBR. F. James, M. Goossens, [*MIUNIT, Function Minimization and Error Analysis*]{}, Reference Manual, CERN Program Library Long Writeup D506 (1992). M.Z. Akrawy et al., Z. Phys. C 47 (1991) 505 V. Gibson, Charm and Beauty Hadron Production at $\sqrt{s} \approx M_Z$, Proc. XXVII ICHEP, Glasgow 1994, P.J. Bussey and I. G. Knowles (ed.) ALEPH Collaboration, Production of $K^{\circ}$ and $\Lambda$ in Hadronic Z Decays, CERN-PPE/94-74, submitted to Z. Phys. C. OPAL Coll., R. Akers et al. CERN-PPE/94-206, submitted to Z. Phys. C. DELPHI Coll., P. Abreu et al., Phys. Lett 345B (1995) 598 ALEPH Coll., D. Busculic et al., CERN-PPE/94-201, submitted to Z. Phys. C. OPAL Coll., R. Akers et al., CERN-PPE/95-24 (1995), submitted to Z. Phys. C. OPAL Coll., R. Akers et al., Z. Phys. C.63 (1994) 181 DELPHI Coll., P. Abreu et al., Nucl. Phys. B444 (1995) 3 OPAL Coll., Inclusive strange vector and tensor meson production in hadronic $Z^{\circ}$ decays, preliminary results presented at ICHEP94 Glasgow. ALEPH Coll., preliminary results presented at ICHEP94 Glasgow, GLS0548. Holger Hepp (ALEPH Coll.),\
Inklusive Produktion von geladenen $K^*$-Mesonen in hadronischen Z-Zerfällen, Diplomarbeit, HD-IHEP 93-04. OPAL Collaboration, Strange Baryon Production and Correlations in Hadronic $Z^{\circ}$ Decays, preliminary results presented at ICHEP94 Glasgow. DELPHI Coll., P. Abreu et al., Phys. Lett. 318B (1993) 249 DELPHI Collaboration, Strange Baryon Production in $Z^{\circ}$ Hadronic Decays. preliminary results presented at ICHEP94 Glasgow DELPHI Coll., P. Abreu et al., Z. Phys. C65 (1995) 587 L3 Coll., Y. Pei, talk given at the CERN PPE Seminar L3 Coll., O. Adriani et al., Phys. Lett. 286 B (1992) 403. L3 Coll., M. Acciari et al., Phys. Lett. 328 B (1994) 223 ALEPH Coll., D. Buskulic et al., Phys. Lett. 292 B (1992) 210. ALEPH Coll., Production Rates of $\eta$ and $\eta'$ in Hadronic Z Decays, preliminary results presented at ICHEP94 Glasgow. DELPHI Coll., P. Abreu et al., CERN-PPE/95-39, submitted to Z. Phys. C. Review of Particle Properties 1994, Phys. Rev. D50 (1994) 3 DELPHI Coll., P. Abreu et al., Z. Phys. C63 (1994)17, Phys. Lett. 323B (1994) 242, CERN-PPE/95-77, submitted to Phys. Lett. A. De Angelis, J. Phys. G19, 1233 (1993) OPAL Collaboration, M.Z. Akrawy Z. Phys. C 47 (1990) 505,\
ALEPH Collaboration, D. Decamp et al., Phys. Lett. 273B(1991)181,\
DELPHI Collaboration, A. Abreu et al., Z. Phys. C 50(1991)185,\
OPAL Collaboration, P.D. Acton et al., Z. Phys. C 53 (1992) 539. DELPHI Coll. P. Abreu et al., Phys. Lett B347 (1995) 447 Delphi Coll., P. Abreu et al. DELPHI 95-50 PHYS 488, contribution to the eps conference on High Energy Physics, Brussels 1995 S. Brandt et al., Phys. Lett. 12 (64) 57; E. Fahri, Phys. Rev. Lett. 39 (1977) 1587 J.D. Bjorken, S. Brodsky: Phys. Rev. D1 (1970) 1416 R.K. Ellis, D.A. Ross, A.E. Terrano: Nucl. Phys. B178 (1981) 421 G. Parisi: Phys. Lett 74B (1978) 65. J.F. Donohue, F.E. Low, S.Y. Pi: Phys. Rev. D20 (1979) 2759 L. Clavelli: Phys. Lett. B85 (1979) 111. JADE Coll., W. Bartel et al.: Z. Phys. C33 (1986) 23 JADE Coll., S. Bethke et al.: Phys. Lett. B213 (1988) 235 L. Smolik, DESY 90-089 C.L. Basham, L.S. Brown, S.D. Ellis, S.T. Love: Phys. Rev. D17 (1978) 2298, Phys. Rev. Lett. 41 (1978) 1585, Phys. Rev. D19 (1979) 2018 S. Catani, G. Turnock, B.R. Webber, CERN-TH.6640/92 (1992)
[^1]: The $1^{+-}$ and $0^{++}$ meson multiplets are not included in HERWIG.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider one-dimensional random walks in random environment which are transient to the right. Our main interest is in the study of the sub-ballistic regime, where at time $n$ the particle is typically at a distance of order $O(n^\kappa)$ from the origin, $\kappa\in(0,1)$. We investigate the probabilities of moderate deviations from this behaviour. Specifically, we are interested in quenched and annealed probabilities of slowdown (at time $n$, the particle is at a distance of order $O(n^{\nu_0})$ from the origin, $\nu_0\in (0,\kappa)$), and speedup (at time $n$, the particle is at a distance of order $n^{\nu_1}$ from the origin, $\nu_1\in (\kappa,1)$), for the current location of the particle and for the hitting times. Also, we study probabilities of backtracking: at time $n$, the particle is located around $(-n^\nu)$, thus making an unusual excursion to the left. For the slowdown, our results are valid in the ballistic case as well.'
address:
- 'Université de Lyon; Université Lyon 1; INSA de Lyon, F-69621; Ecole Centrale de Lyon; CNRS, UMR5208, Institut Camille Jordan, 43 blvd du 11 novembre 1918, F-69622 Villeurbanne-Cedex, France'
- 'CeNos, Center for Nonlinear Science, and Institut für Mathematische Statistik, Fachbereich Mathematik und Informatik, Universität Münster, Einsteinstr. 62, D–48149 Münster, Germany'
- 'Instituto de Matemática e Estatística, Universidade de São Paulo, rua do Matão 1010, CEP 05508–090, São Paulo SP, Brasil'
author:
- Alexander FRIBERGH
- Nina GANTERT
- Serguei POPOV
title: On slowdown and speedup of transient random walks in random environment
---
Introduction and results {#s_introres}
========================
Let $\omega:=(\omega_i, \, i \in {\mathbb{Z}})$ be a family of i.i.d. random variables taking values in $(0,1)$. Denote by ${\mathbf{P}}$ the distribution of $\omega$ and by ${\mathbf{E}}$ the corresponding expectation. After choosing an environment $\omega$ at random according to the law ${\mathbf{P}}$, we define the random walk in random environment (usually abbreviated as RWRE) as a nearest-neighbour random walk on ${\mathbb{Z}}$ with transition probabilities given by $\omega$: $(X_n, \, n \geq 0)$ is the Markov chain satisfying $X_0=z$ and for $n \geq 0,$ $$\begin{aligned}
{P_{\omega}}^z[ X_{n+1} = x+1 \mid X_n =x] &= \omega_x, \nonumber\\
{P_{\omega}}^z[ X_{n+1} = x-1 \mid X_n =x] &= 1-\omega_x.\end{aligned}$$ As usual, ${P_{\omega}}^z$ is called the *quenched* law of $(X_n, \, n \geq 0)$ starting from $X_0=z$, and we denote by ${E_{\omega}}^z$ the corresponding quenched expectation. Also, we denote by ${\mathbb{P}}^z$ the semi-direct product ${\mathbf{P}}\times {P_{\omega}}^z$ and by ${\mathbb{E}}^z$ the expectation with respect to ${\mathbb{P}}^z$; ${\mathbb{P}}^z$ and ${\mathbb{E}}^z$ are called the *annealed* probability and expectation. When $z=0$, we write simply ${P_{\omega}}$, ${E_{\omega}}$, ${\mathbb{P}}$, ${\mathbb{E}}$.
In this paper we will also consider RWRE on ${\mathbb{Z}}_+$, with reflection to the right at the origin. This RWRE can be defined as above, in the environment ${{\tilde\omega}}$ given by $${{\tilde\omega}}_i = \begin{cases}
\omega_i, & i\neq 0,\\
1, & i=0
\end{cases}$$ (provided, of course, that the starting point is nonnegative). We then write ${P_{{{\tilde\omega}}}}^z$, ${E_{{{\tilde\omega}}}}^z$ for the quenched probability and expectation in the case of RWRE reflected at the origin, ${{\tilde{\mathbb{P}}}}^z$ and ${{\tilde{\mathbb{E}}}}^z$ for the annealed probability and expectation, keeping the simplified notation ${P_{{{\tilde\omega}}}}$, ${E_{{{\tilde\omega}}}}$, ${{\tilde{\mathbb{P}}}}$, ${{\tilde{\mathbb{E}}}}$ for the RWRE starting at the origin.
For all $i \in {\mathbb{Z}}$, let us introduce $$\rho_i:=\frac{1-\omega_i}{\omega_i}.$$ Throughout this paper, we assume that $$\label{logrho}
{\mathbf{E}}[\ln\rho_0]<0,$$ which implies (cf. [@Solomon]) that $\lim_{n\to\infty} X_n=+\infty$ ${P_{\omega}}$-a.s. for ${\mathbf{P}}$-a.a. $\omega$, so that the RWRE is transient to the right (or simply transient, in the case of RWRE with reflection at the origin).
We refer to [@Zeitouni] for a general overview of results on RWRE. In the following we always work under the assumption that $$\label{def_kappa}
\text{there exists a unique $\kappa>0$, such that } {\mathbf{E}}[\rho_0^{\kappa}]=1 \text{ and } {\mathbf{E}}[\rho_0^{\kappa}\ln^+ \rho_0]<\infty.$$
This constant plays a central role for RWRE, in particular when it exists, its value separates the *ballistic* from the *sub-ballistic* regime: $$\kappa >1 \text{ if and only if } \frac{X_n}n \to v>0,\quad \text{${\mathbb{P}}$-a.s.}$$ We refer to the case $\kappa > 1$ as the *ballistic* regime and to the case $\kappa \leq 1$ as the *sub-ballistic* regime. In this paper we mainly consider the case where the RWRE is transient (to the right) and *sub-ballistic*, i.e. the asymptotic speed is equal to $0$. The following result was proved in [@KKS] and partially refined in [@ESZ]:
\[kks\] Let $\omega:=(\omega_i, \, i \in {\mathbb{Z}})$ be a family of independent and identically distributed random variables such that
- $-\infty \leq {\mathbf{E}}[\ln \rho_0]<0$,
- there exists $0<\kappa\leq1$ for which ${\mathbf{E}}\left[\rho_0^{\kappa}\right] = 1$ and ${\mathbf{E}}\left[ \rho_0^{\kappa} \ln^+\rho_0 \right]<\infty,$
- the distribution of $\ln \rho_0$ is non-lattice.
Then, if $\kappa<1$, we have $$\frac{X_n}{n^{\kappa}} \,
\stackrel{law}{\longrightarrow}\, C_1 \left(\frac 1{\mathcal{S}_{\kappa}^{ca}}\right)^{\kappa},$$ where $\stackrel{law}{\longrightarrow}$ stands for convergence in distribution with respect to the annealed law ${\mathbb{P}}$, $C_1$ is a positive constant and $\mathcal{S}_{\kappa}^{ca}$ is the completely asymmetric stable law of index $\kappa$. If $\kappa=1$, we have $$\frac{X_n}{n/\ln n} \,
\stackrel{law}{\longrightarrow}\,
C_2 \frac 1{\mathcal{S}_{1}^{ca}}.$$
In the quenched case, the limiting behaviour is more complicated, as discussed in [@PZ]. However, one still can say that at time $n$ the particle is “typically” at distance roughly $n^\kappa$ from the origin, since the weaker result $\lim_{n\to\infty}\ln X_n/\ln n = \kappa$, ${\mathbb{P}}$-a.s., is still valid[^1].
Besides the results about the location of the particle at time $n$, we are interested also in the first hitting times of certain regions in space. For any set $A\subset{\mathbb{Z}}$, define: $$T_A := \min\{n\geq 0 : X_n\in A\}.$$ To simplify the notations, for one-point sets we write $T_a:=T_{\{a\}}$. In the case where $a$ is not an integer, the notation $T_a$ will correspond to $T_{\lfloor a \rfloor}$.
In this paper we investigate the following types of unusual behaviour of the random walk:
- *slowdown*, which means that at time $n$ the particle is around $n^{\nu_0}$, $\nu_0<1 \wedge \kappa$, so that the particle goes to the right much slower than it typically does;
- *backtracking*, that is, at time $n$ the particle is found around $(-n^\nu)$, thus performing an unlikely excursion to the left instead of going to the right (this is, of course, only for RWRE without reflection);
- *speedup*, which means that the particle is going to the right faster than it should (but still with sublinear speed): at time $n$ the particle is around $n^{\nu_1}$, $\kappa<\nu_1<1$ (this is possible only for $\kappa <1$).
We refer to all of the above as *moderate deviations*, even for the slowdown in the ballistic case $\kappa>1$. Indeed, in the latter case the deviation from the typical position is linear in time, but we have that the large deviation rate function $I$ satisfies $I(0) = 0$, and the known large deviation results only tell us that slowdown probabilities decay slower than exponentially in $n$ (see, for instance, [@CGZ]).
We mention here that in the literature one can find some results on moderate deviations for the case of recurrent RWRE (often referred to as RWRE in “Sinai’s regime”), see [@CP; @CP04], and also [@HS] for the continuous space and time version.
Now, we state the results we are going to prove in this paper. In addition to (\[def\_kappa\]), we will use the following weak integrability hypothesis: $$\label{integ_hyp}
\text{there exists ${\varepsilon}_0>0$ such that } {\mathbf{E}}[\rho_0^{-{\varepsilon}_0}]<\infty.$$
First, we discuss the results about quenched slowdown probabilities. It turns out that the quenched slowdown probabilities behave differently depending on whether one considers RWRE with or without reflection at the origin. Also, it matters which of the following two events is considered: (i) the position of the particle at time $n$ is at most $n^\nu$, $\nu<\kappa$ (i.e., the event $\{X_n<n^\nu\}$), or (ii) the hitting time of $n^\nu$ is greater than $n$ (i.e., the event $\{T_{n^\nu}>n\}$). Here we prove that in all these cases the quenched probability of slowdown is roughly $e^{-n^\beta}$, where $\beta=1-\frac{\nu}{\kappa}$ for the “hitting time slowdown” in the reflected case, and $\beta=(1-\frac{\nu}{\kappa})\wedge\frac{\kappa}{\kappa+1}$ in the other cases. More precisely, we have
\[t\_q\_slow\] [**Slowdown, quenched**]{} Suppose that (\[logrho\]), (\[def\_kappa\]) and (\[integ\_hyp\]) hold. For $\nu \in (0, 1\wedge \kappa)$ the quenched slowdown probabilities behave in the following way. For the reflected RWRE, $$\begin{aligned}
\lim_{n\to\infty} \frac{\ln(-\ln {P_{{{\tilde\omega}}}}[T_{n^{\nu}}>n])}{\ln n} &= 1-\frac{\nu}{\kappa},
\qquad \text{${\mathbf{P}}$-a.s.},\label{refl_hit_q_sl}\\
\lim_{n\to\infty} \frac{\ln(-\ln {P_{{{\tilde\omega}}}}[X_n<n^{\nu}])}{\ln n}
&= \Bigl(1-\frac{\nu}{\kappa}\Bigr) \wedge \frac{\kappa}{\kappa+1},
\qquad \text{${\mathbf{P}}$-a.s.} \label{refl_loc_q_sl}\end{aligned}$$ For the RWRE without reflection, we obtain $$\begin{aligned}
\lim_{n\to\infty} \frac{\ln(-\ln {P_{\omega}}[T_{n^{\nu}}>n])}{\ln n}
&= \Bigl(1-\frac{\nu}{\kappa}\Bigr) \wedge \frac{\kappa}{\kappa+1},
\qquad \text{${\mathbf{P}}$-a.s.},\label{nonrefl_hit_q_sl}\\
\lim_{n\to\infty} \frac{\ln(-\ln {P_{\omega}}[X_n<n^{\nu}])}{\ln n}
&= \Bigl(1-\frac{\nu}{\kappa}\Bigr) \wedge \frac{\kappa}{\kappa+1},
\qquad \text{${\mathbf{P}}$-a.s.} \label{nonrefl_loc_q_sl}\end{aligned}$$
For a heuristical explanation of the reason for the different behaviours of the quenched slowdown probabilities we refer to the beginning of Section \[s\_q\_slowdown\].
For the annealed slowdown probabilities, we obtain that there is no difference between reflecting/nonreflecting cases (at least on the level of precision we are working here) and also it does not matter which one of the slowdown events $\{T_{n^{\nu}}>n\}$, $\{X_n<n^{\nu}\}$ one considers. In all these cases, the annealed probability of slowdown decays polynomially, roughly as $n^{-(\kappa-\nu)}$:
\[t\_a\_slow\] [**Slowdown, annealed**]{} Suppose that (\[logrho\]), (\[def\_kappa\]) and (\[integ\_hyp\]) hold. For $\nu \in (0, 1\wedge \kappa)$, $$\label{a_slowdown}
\lim_{n\to\infty} \frac{\ln {\mathbb{P}}[X_n<n^{\nu}]}{\ln n}
=\lim_{n\to\infty} \frac{\ln {\mathbb{P}}[T_{n^{\nu}}>n]}{\ln n} = -(\kappa-\nu).$$ The same result holds if one changes ${\mathbb{P}}$ to ${{\tilde{\mathbb{P}}}}$ in (\[a\_slowdown\]).
In the case of RWRE on ${\mathbb{Z}}$ (i.e., without reflection at the origin) there is another kind of untypically slow escape to the right. Namely, before going to $+\infty$, the particle can make an untypically big excursion to the left of the origin. While it is easy to control the distribution of the leftmost site touched by this excursion (e.g., by means of the formula (\[exit\_probs\]) below), it is interesting to study the probability that at time $n$ the particle is far away to the left of the origin:
\[t\_backtrack\] [**Backtracking**]{} Suppose that (\[logrho\]), (\[def\_kappa\]) and (\[integ\_hyp\]) hold. For $\nu \in (0, 1)$, we have $$\begin{aligned}
\lim_{n\to\infty} \frac{\ln(-\ln {P_{\omega}}[X_n<-n^{\nu}])}{\ln n} &= \nu \vee \frac{\kappa}{\kappa+1},
\qquad \text{${\mathbf{P}}$-a.s.}
\label{q_backtrack}\\
\lim_{n\to\infty} \frac{\ln(-\ln {\mathbb{P}}[X_n<-n^{\nu}])}{\ln n} &= \nu,
\label{a_backtrack}\end{aligned}$$ and $$\label{backtrack_T}
\lim_{n\to\infty} \frac{\ln(-\ln {\mathbb{P}}[T_{-n^{\nu}}<n])}{\ln n}
= \lim_{n\to\infty} \frac{\ln(-\ln {P_{\omega}}[T_{-n^{\nu}}<n])}{\ln n} = \nu
\qquad \text{${\mathbf{P}}$-a.s.}$$
Another kind of deviation from the typical behaviour is the speedup of the particle, i.e., at time $n$ the particle is at a distance larger than $n^\kappa$ from the origin (here we of course assume that $\kappa<1$). There are results in the literature that cover the *large deviations* case, i.e., the case when at time $n$ the particle is at distance $O(n)$ from the origin, see e.g. Section 2.3 of [@Zeitouni], or [@CGZ]. In this paper we are interested in the probabilities of moderate speedup: the displacement of the particle is sublinear, but still bigger than in the typical case. Namely, we show that the quenched probability that $X_n$ is of order $n^\nu$, $\kappa<\nu<1$, is roughly $e^{-n^\beta}$, where $\beta=\frac{\nu-\kappa}{1-\kappa}$. It is remarkable that the annealed probability is roughly of the same order. More precisely, we are able to prove the following result:
\[t\_speedup\] [**Speedup**]{} Suppose that (\[logrho\]), (\[def\_kappa\]) and (\[integ\_hyp\]) hold. For $\nu \in (\kappa,1)$ we can control the probabilities of the moderate speedup in the following way: $$\label{eq_q_speed}
\lim_{n\to\infty} \frac{\ln(-\ln {P_{\omega}}[X_n> n^{\nu}])}{\ln n}
= \lim_{n\to\infty} \frac{\ln(-\ln {P_{\omega}}[T_{n^{\nu}}< n])}{\ln n}
= \frac{\nu-\kappa}{1-\kappa},\qquad \text{${\mathbf{P}}$-a.s.},$$ and $$\label{eq_a_speed}
\lim_{n\to\infty} \frac{\ln(-\ln {\mathbb{P}}[X_n > n^{\nu}])}{\ln n}
= \lim_{n\to\infty} \frac{\ln(-\ln {\mathbb{P}}[T_{n^{\nu}} < n])}{\ln n} = \frac{\nu-\kappa}{1-\kappa}.$$ The same result holds for the RWRE with reflection at the origin.
For the case $\kappa\in (0,1)$, the quenched moderate deviations for the random walk on ${\mathbb{Z}}$ are well summed up by the plot of the following function on Figure \[f\_exponent\]: $$f(\nu)=\left\{
\begin{array}{ll}
\lim_{n\to\infty} \ln(-\ln {P_{\omega}}[X_n<-n^{-\nu}])/\ln n, & \text{ if } \nu\in (-1,0],\\
\lim_{n\to\infty} \ln(-\ln {P_{\omega}}[X_n<n^{\nu}])/\ln n, & \text{ if } \nu\in (0,\kappa), \\
\lim_{n\to\infty} \ln(-\ln {P_{\omega}}[X_n>n^{\nu}])/\ln n, & \text{ if } \nu\in [\kappa, 1). \\
\end{array}
\right.$$
![The plot of $f(\nu)$, $-1<\nu< 1$[]{data-label="f_exponent"}](exponent){width="12cm"}
The rest of this paper is organized in the following way. In Section \[s\_notations\] we give the (standard) definition of the potential and the reversible measure for the RWRE. We then decompose the environment into a sequence of valleys. In this decomposition the valleys do not only depend on the environment but the construction is time-dependent. Also, we derive some basic facts about the valleys needed later. In Section \[s\_estimates\_env\] we mainly study the properties of that sequence of valleys. In Section \[s\_confinement\], we recall some results concerning the spectral properties of RWRE restricted to a finite interval, and then obtain some bounds on the probability of confinement in a valley. In Section \[s\_induced\_rw\] we define the induced random walk whose state is the current valley (more precisely, the last visited boundary between two neighbouring valleys) where the particle is located. Theorems \[t\_q\_slow\], \[t\_a\_slow\], \[t\_backtrack\], \[t\_speedup\] are proved in Sections \[s\_q\_slowdown\], \[s\_a\_sl\], \[s\_backtracking\], \[s\_speedup\] respectively. We denote by $\gamma, {\gamma}_0, {\gamma}_1, {\gamma}_2, {\gamma}_3,\ldots$ the “important” constants (those that can be used far away from the place where they appear for the first time), and by $C_1,C_2,C_3,\ldots$ the “local” ones (those that are used only in a small neighbourhood of the place where they appear for the first time), restarting the numeration at the beginning of each section in the latter case. All these constants are either universal or depend only on the law of the environment.
More notations and some basic facts {#s_notations}
===================================
An important ingredient of our proofs is the analysis of the [*potential*]{} associated with the environment, which was introduced by Sinai in [@Sinai]. The potential, denoted by $V= (V(x), \; x\in {\mathbb{Z}}),$ is a function of the environment $\omega$. It is defined in the following way: $$V(x) :=\left\{\begin{array}{ll}
\sum_{i=1}^x \ln \rho_i, & \text{if } x \geq 1, \\
0\vphantom{\sum^N}, & \text{if } x=0, \\
-\sum_{i=x+1}^0 \ln \rho_i, &\text{if } x\leq -1,
\end{array}\right.$$ so it is a random walk with negative drift, because ${\bf E}[\ln \rho_0]<0$. This notation is extended on ${\mathbb{R}}$ by $V(x):=V(\lfloor x \rfloor)$. We also define a reversible measure $$\label{pidef}
\pi(x) :=e^{-V(x)}+e^{-V(x-1)},\qquad \text{for $x\in {\mathbb{Z}}$,}$$ (one easily verifies that $\omega_x\pi(x)=(1-\omega_{x+1})\pi(x+1)$ for all $x$). We will also use the notation $\pi([x,y])=\sum_{i=\lfloor x \rfloor -1}^{\lfloor y \rfloor } \pi (i)$, for $x<y$ two real numbers.
The function $V(\cdot)$ enables us to define the valleys, parts of the environment which acts as traps for the random walk. The valleys are responsible for the sub-ballistic behaviour and hence play a central role for slowdown and speedup phenomena.
We define by induction the following environment dependent sequence $(K_i(n))_{i \geq 0}$ by $$\begin{aligned}
K_0(n) =& -n,\\
K_{i+1}(n) =& \min \Bigl\{ j \geq K_i(n): V(K_i(n))-\min_{k\in[K_i(n),j]}V(k)\geq \frac 3{1\wedge \kappa}\ln n,\\
& \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad V(j)=\max_{k\geq j} V(k) \Bigr\}.\end{aligned}$$ The dependence with respect to $n$ will be frequently omitted to ease the notations. The portion of the environment $[K_i,K_{i+1})$ is called the $i$-th valley, and we will prove that for $n$ large enough the valleys are descending in the sense that $V(K_{i+1})<V(K_i)$ for all $i\in[0,n]$. We associate to the $i$-th valley the bottom point $$b_i=\inf\Bigl\{x \in [K_{i},K_{i+1}): V(x)= \min_{y\in [K_{i},K_{i+1})} V(y)\Bigr\},$$ and the depth $$\begin{aligned}
H_i&=\max_{x\in [K_{i},K_{i+1})}\Bigl(\max_{y\in [x,K_{i+1})} V(y)-\min_{y\in [K_{i},x)} V(y)\Bigr)\\
& = \max_{K_i(n)\leq j<k<K_{i+1}(n)} \bigl(V(k)-V(j)\bigr),\end{aligned}$$ see Figure \[f\_valleys\].
![On the definition of the sequence of valleys[]{data-label="f_valleys"}](valleys){width="12cm"}
Let us denote $$\label{Ndef}
N_n(m,m')=\{ i\geq 1: \ [K_{i},K_{i+1})\cap [\lfloor m\rfloor, \lfloor m'\rfloor) \neq \emptyset\}$$ and again we will often omit the index $n$. Let us emphasize that we do not include the valley of index 0, which is different from the others because of border issues.
The valleys for $i\geq 1$ are non-overlapping parts of ${\mathbb{Z}}$, for any value of $n$. Moreover the potential in the valleys are i.i.d. up to space-shift, in the sense that for any $n$ and $i\geq 1$ the sequence of vectors of random length $ \bigl(V(j)-V(K_{i+1}(n)),j=K_i(n),\ldots,K_{i+1}(n)-1\bigr)$, $i\geq 1$, is i.i.d.
We introduce the two following indices which will be used regularly $$\label{indices}
i_0={\mathop{\mathrm{card}}}N(-n,0) \ \text{ and }\ i_1={\mathop{\mathrm{card}}}N(-n,n^{\nu}).$$
To carry over the proofs easily to the reflected case, we introduce the following notation $$\label{tilddef}
\widetilde{K}_{i_0}=0 \ \text{ and }\ \widetilde{K}_i=K_i \text{ for $i\geq i_0$}.$$
We can estimate the depth of the valleys using a result of renewal theory which concerns the maximum of random walks with negative drift. We refer to [@Feller] for a detailed introduction to renewal theory. Denoting $S=\max_{i\geq 0} V(i)$, under assumptions (\[logrho\]), (\[def\_kappa\]) and $(iii)$ of Theorem \[kks\], we have $$\label{fellerthm_nlattice}
{\mathbf{P}}[S>h] \sim C_F \, e^{-\kappa h},\qquad \qquad h\to \infty,$$ which is a result due to Feller which can be found in this form in [@igle].
If $(iii)$ in Theorem \[kks\] fails, $\ln \rho_0$ is concentrated on $\lambda {\mathbb{Z}}$ for some $\lambda>0$, so that $V(\cdot)$ is a Markov chain with i.i.d. increments of law $\ln \rho_i$. In this case, under our assumptions (\[logrho\]) and (\[def\_kappa\]) we can use a result in [@Spitzer] (p. 218) stating the discrete version of the previous equation. In the case of an aperiodic Markov chain we have $$\label{fellerthm_lattice}
{\mathbf{P}}[S\geq n \lambda] \sim C_F' \, e^{-\kappa \lambda n}, \qquad \qquad n\to \infty,$$ and in the general case we obtain similar asymptotics by noticing that $(V(nd+k))_{n\geq 0}$ is aperiodic for $k\in [0,d-1]$ and $d$ the period of $V(\cdot)$ (which is well defined and finite by $(i)$ and $(ii)$).
Hence we can easily deduce from the assumptions (\[logrho\]) and (\[def\_kappa\]) and equations (\[fellerthm\_nlattice\]) and (\[fellerthm\_lattice\]) that $$\label{fellerthm}
{\mathbf{P}}[S>h]=\Theta(e^{-\kappa h}),$$ where $f(n)=\Theta(g(n))$ means that $f(n)=O(g(n))$ and $g(n)=O(f(n))$.
Let us recall also the following basic fact. For any integers $a<x<b$, the (quenched) probability for RWRE to reach $b$ before $a$ starting from $x$ can be easily computed: $$\label{exit_probs}
{P_{\omega}}^x[T_b < T_a] = \frac{ \sum_{y=a}^{x-1}
e^{V(y)}}{ \sum_{y=a}^{b-1} e^{V(y)}},$$ see e.g. Lemma 1 in [@Sinai] or formula (2.1.4) in [@Zeitouni].
Estimates on the environment {#s_estimates_env}
============================
Let us introduce the event $$\label{Adef}
A(n)= \Bigl\{ \max_{i\leq 2n} (K_{i+1}-K_i) \leq (\ln n)^2\Bigr\}.$$
The following lemma shows that the valleys are not very wide.
\[widthvalley\] We have $${\mathbf{P}}[A(n)^c]=O\Bigl(\frac 1 {n^2}\Bigr).$$
We have $$\begin{aligned}
{\mathbf{P}}[A(n)^c] &= {\mathbf{P}}\Bigl[\max_{i\leq 2n} (K_{i+1}-K_i) > (\ln n)^2\Bigr]\nonumber\\
&\leq 2n {\mathbf{P}}[K_2-K_1 > (\ln n)^2]+{\mathbf{P}}[\overline K_1 > (\ln n)^2],\label{Astep1}\end{aligned}$$ where $$\overline K_1(n) = \min \Bigl\{ j \geq 0: -\min_{k\in[0,j]}V(k)\geq \frac 3{1\wedge \kappa}\ln n, \quad V(j)=\max_{k\geq j} V(k) \Bigr\}.$$ Now $$\begin{aligned}
{\mathbf{P}}[K_2-K_1 > (\ln n)^2]&={\mathbf{P}}[\overline K_1> (\ln n)^2\mid \max_{i\geq 0} V(i) \leq 0]\\
&\leq \frac{{\mathbf{P}}[\overline K_1 > (\ln n)^2]}{{\mathbf{P}}[\max_{i\geq 0} V(i)\leq 0]},\end{aligned}$$ where ${\mathbf{P}}[\max_{i\geq 0} V(i)\leq 0]>0$ since ${\mathbf{E}}[\ln \rho_0]<0$. Choose $\ell$ such that $\varepsilon_0\ell > 3(1\wedge \kappa)$, with $\varepsilon_0$ from (\[integ\_hyp\]). Note that if $V((\ln n)^2) \leq -\frac {3+3\ell}{1\wedge\kappa} \ln n$, $ \min_{j\leq (\ln n)^2} \left(V(j)-V(j-1)\right)\geq - \frac{\ell}{1\wedge\kappa} \ln n$ and $\max_{j\geq (\ln n)^2} V(j)-V((\ln n)^2)\leq \frac 3{1\wedge\kappa} \ln n$, then the set $$\left\{i \in [0, (\ln n)^2],~V(i)\in \left(-\frac 3{1\wedge\kappa} \ln n,-\frac{3+2\ell}{1\wedge\kappa} \ln n\right)\right\}$$ is non-empty. Moreover its largest element $m$ is such that $\max_{j\geq m} V(j)=V(m)$, hence we have $\overline K_1 \leq (\ln n)^2$. This yields $$\begin{aligned}
\label{Astep2}
{\mathbf{P}}[\overline K_1 > (\ln n)^2] & \leq {\mathbf{P}}\Bigl[V((\ln n)^2) >-\frac {3+3\ell}{1\wedge\kappa} \ln n
\\ & \qquad \nonumber \qquad \text{ or } \min_{j\leq (\ln n)^2} \left(V(j)-V(j-1)\right)< -\frac{\ell}{1\wedge\kappa} \ln n
\\ & \qquad \nonumber \qquad \text{ or } \max_{j\geq (\ln n)^2} V(j)-V((\ln n)^2)> \frac 3{1\wedge\kappa} \ln n\Bigr].\end{aligned}$$
Using (\[fellerthm\]), we obtain $$\label{Astep3}
{\mathbf{P}}\Bigl[\max_{j\geq (\ln n)^2} V(j)-V((\ln n)^2)> \frac 3{1\wedge\kappa} \ln n\Bigr] =O(n^{-3}).$$
Furthermore, using Chebyshev’s inequality and (\[integ\_hyp\]) we get $$\begin{aligned}
\lefteqn{{\mathbf{P}}\Bigl[ \min_{j\leq (\ln n)^2}\left( V(j)-V(j-1)\right)< - \frac{\ell}{1\wedge\kappa} \ln n\Bigr]} \nonumber\\
&\leq (\ln n)^2 {\mathbf{P}}\left[\ln \rho_0 <
-\frac{\ell}{1\wedge\kappa} \ln n\right] \nonumber\\
&\leq (\ln n)^2 {\mathbf{P}}\left[\rho_0^{-\varepsilon_0} > \exp\left(\varepsilon_0\frac{\ell}{1\wedge\kappa}\ln n\right) \right]\nonumber\\
& \leq (\ln n)^2 {\mathbf{E}}[\rho_0^{-{\varepsilon}_0}] n^{-\varepsilon_0\ell/(1\wedge \kappa)}\nonumber\\
&= o(n^{-3})\, .
\label{Astep5}\end{aligned}$$
Now, since $V(\cdot)$ is a sum of i.i.d. random variables with exponential moments by the assumptions (\[def\_kappa\]) and (\[integ\_hyp\]), we can use large deviations techniques to get $$\begin{aligned}
{\mathbf{P}}[V((\ln n)^2) >-C_1 \ln n] &\leq {\mathbf{P}}\bigl[{\left\vert V((\ln n)^2)-{\mathbf{E}}[V(1)] (\ln n)^2\right\vert} > C_2(\ln n)^2\bigr] \label{Astep4} \\
&\leq \exp(- C_3(\ln n)^2)\nonumber\\
&= o(n^{-3}),\nonumber\end{aligned}$$ since ${\mathbf{E}}[V(1)]={\mathbf{E}}[\ln \rho_0]\in(-\infty,0)$. Putting together (\[Astep1\]), (\[Astep2\]), (\[Astep3\]), (\[Astep5\]) and (\[Astep4\]) we obtain the result.
Consider $a\in [0, \nu)$, and define the event $$B(n,\nu,a)^c=\Bigl\{ {\mathop{\mathrm{card}}}\Bigl\{i \in N_n(-n^{\nu},n^{\nu}): H_i \geq \frac a {\kappa} \ln n + \ln \ln n\Bigr\} \geq n^{\nu -a} \Bigr\}.$$
The following lemma will tell us that asymptotically, between levels $-n^{\nu}$ and $n^{\nu}$ there are at most $n^{\nu-a}$ valleys of depth greater than $(a/ \kappa) \ln n + \ln \ln n$.
\[depthvalley1\] For any $a\in [0,\nu)$, we have $${\mathbf{P}}[B(n,\nu,a)^c]=O(n^{-2}).$$
We have easily that (“$\prec$” means “stochastically dominated”) $$\begin{aligned}
\lefteqn{{\mathop{\mathrm{card}}}\Bigl\{i \leq N_n(-n^{\nu},n^{\nu}): H_i \geq \frac a {\kappa} \ln n + \ln \ln n\Bigr\}}\\
& \prec \text{Bin}\Bigl(2\lfloor n^{\nu}\rfloor+2,{\mathbf{P}}\Bigl[S\geq \frac a {\kappa} \ln n + \ln \ln n\Bigr]\Bigr),\end{aligned}$$ since we have at most $2\lfloor n^{\nu}\rfloor+2$ integers on the right of which we need an increase of potential of $(a/\kappa)\ln n+\ln \ln n$ to create a valley of sufficient depth.
Using (\[fellerthm\]), we have $${\mathbf{P}}\Bigl[S\geq \frac a {\kappa} \ln n + \ln \ln n\Bigr]=O\Bigl( \frac{n^{-a}}{(\ln n)^{\kappa}}\Bigr).$$ Now, using Chebyshev’s exponential inequality, we can write $$\begin{aligned}
\lefteqn{{\mathbf{P}}\Bigl[ \text{Bin}\Bigl(2 \lfloor n^{\nu}\rfloor+2,{\mathbf{P}}\Bigl[S\geq \frac a {\kappa} \ln n + \ln \ln n\Bigr]\Bigr)
\geq n^{\nu -a}\Bigr]}\\
&\leq C_4 \exp(-n^{\nu -a})\exp(C_5 n^{\nu-a}(\ln n)^{-\kappa}),\end{aligned}$$ and, since $\nu>a$, the result follows.
We introduce for $m\in {\mathbb{Z}}^+$ the following event, which, by Lemma \[depthvalley1\], has probability converging to 1, $$\label{def_b_prime}
B'(n,\nu,m)=\bigcap_{k=1}^{m-1}B(n,\nu, k \nu/m).$$
Also, set $$\begin{aligned}
G(n)^c=&\Bigl\{\max_{ k \geq n} (V(k)-V(n)) \geq \frac 1 {\kappa}(\ln n + 2 \ln \ln n) \Bigr\}
\nonumber\\
&\bigcup \Bigl\{\max_{ k \geq -n} (V(k)-V(-n)) \geq \frac 1 {\kappa}(\ln n + 2 \ln \ln n) \Bigr\}.\end{aligned}$$
\[depthdiff\] We have $${\mathbf{P}}[G(n)^c] =O \Bigl(\frac 1 {n (\ln n)^2}\Bigr).$$
This is a direct consequence of (\[fellerthm\]).
We now show that Lemma \[depthdiff\] implies that asymptotically, in the interval $[-n,n]$, the deepest valley we can find has depth lower than $ \frac 1 {\kappa}(\ln n + 2 \ln \ln n)$. Let $$\label{Gdef}
G_1(n)=\Bigl\{\ \max_{ i\in [-n, n]} \max_{k\geq i} ( V(k)-V(i) ) \leq \frac {1} {\kappa}(\ln n+ 2\ln \ln n) \Bigr\}.$$
\[depthvalley\] For ${\mathbf{P}}$-almost all $\omega$, there is $N= N(\omega)$ such that $\omega \in G_1(n)$ for $n \geq N$.
By symmetry, it suffices to give the proof for $$G_2(n)=\Bigl\{\ \max_{ i\in[0,n]} \max_{k\geq i} ( V(k)-V(i) ) \leq \frac {1} {\kappa}(\ln n+ 2\ln \ln n) \Bigr\}$$ instead of $G_1(n)$. Let $$n_0 : = \min\Bigl\{j \geq 0: \max_{k\geq i} ( V(k)-V(i) ) \leq \frac {1} {\kappa}(\ln i + 2\ln \ln i), \, \forall i\geq j \Bigr\}$$ and $$K = \max_{0 \leq i\leq n_0}\max_{k\geq i} ( V(k)-V(i) ) .$$ Due to Lemma \[depthdiff\], $n_0$ is finite ${\mathbf{P}}$-almost surely. Now, take $N$ large enough such that $N \geq n_0$ and $$\frac {1} {\kappa}(\ln N +2\ln \ln N)\geq K.$$ Then for $n\geq N$, let $\ell\in[0,n]$ be such that $ \max_{ i\in[0,n]} \max_{k\geq i} ( V(k)-V(i) ) = \max_{k\geq \ell} ( V(k)-V(\ell) )$. We have either $\ell \leq n_0$ and then $\max_{k\geq \ell} ( V(k)-V(\ell) ) \leq K$ by the definition of $K$, or $\ell > n_0$ and then, by the definition of $n_0$, $\max_{k\geq \ell} ( V(k)-V(\ell) ) \leq \frac {1} {\kappa}(\ln \ell + 2\ln \ln \ell)\leq \frac {1} {\kappa}(\ln n + 2\ln \ln n)$.
Let us define $$\begin{aligned}
D(n)^c=&\Bigl\{\ \max_{ i\in[0,n]} \max_{k\geq i} ( V(k)-V(i) ) \leq \frac {1} {\kappa}(\ln n-4\ln \ln n) \Bigr\}
\nonumber\\
&\bigcup \Bigl\{\ \max_{ i\in[-n,0]} \max_{k\geq i} ( V(k)-V(i) ) \leq \frac {1} {\kappa}(\ln n-4\ln \ln n) \Bigr\}.\end{aligned}$$
\[depthvalley2\] We have $${\mathbf{P}}[D(n)^c]=O(n^{-2}).$$
First, we notice that $$\begin{aligned}
{\mathbf{P}}[D(n)^c] & \leq 2{\mathbf{P}}\Bigl[\max_{ i \in [0,\frac{n}{\lfloor (\ln n)^2\rfloor}]} \max_{k \leq (\ln n)^2} V(i(\ln n)^2+k)-V(i(\ln n)^2) \\
& \qquad\qquad\qquad\qquad\qquad \leq \frac {1} {\kappa}(\ln n-4\ln \ln n)\Bigr]+{\mathbf{P}}[A(n)^c],\end{aligned}$$ where ${\mathbf{P}}[A(n)^c]=O(n^{-2})$ by Lemma \[widthvalley\].
Let us introduce $$D^{(1)}(n)=\Bigl\{\max_{ k >\lfloor (\ln n)^2 \rfloor} V(k)-V(0) \geq \frac {1} {\kappa}(\ln n-4\ln \ln n)\Bigr\},$$ then we have $$\begin{aligned}
{\mathbf{P}}[D^{(1)}(n)]\leq &{\mathbf{P}}\Bigl[\max_{ k \geq 0} V(k)-V(0) > \frac {1} {\kappa}(\ln n-4\ln \ln n)\Bigr]
\nonumber\\
&+{\mathbf{P}}\Bigl[\max_{ k \geq 0} V(k)-V(0) \neq \max_{ k \leq (\ln n)^2} V(k)-V(0)\Bigr] =\Theta\Bigl(\frac {(\ln n)^4} {n}\Bigr),\end{aligned}$$ using a reasoning similar to the proof of Lemma \[widthvalley\] (cf. equations (\[Astep2\]) and (\[Astep3\])) to show that the second term is at most $O(n^{-2})$.
So, we obtain for $n$ large enough $${\mathbf{P}}[D(n)^c] \leq 2\Bigl(1-\frac{C_6(\ln n)^4} {n}\Bigr)^{n/(\ln n)^2}\leq 2\exp\left(-C_7 (\ln n)^2\right),$$ hence the result.
Finally, let us introduce $$F(n)=\bigl\{ \min_{i \in [-n,n]} (1-\omega_i) >n^{-3/{\varepsilon}_0}\bigr\}.$$
\[easy\_backtrack\] We have $${\mathbf{P}}[F(n)^c]=O\Bigl(\frac 1 {n^2}\Bigr).$$
We notice that $1-\omega_i\geq \min(1/2, \rho_i/2)$, so that it is enough to prove that ${\mathbf{P}}[\rho_i<2n^{-3/{\varepsilon}_0}]=O(n^{-3})$ which is a consequence of (\[integ\_hyp\]), since by Chebyshev’s inequality $${\mathbf{P}}\Bigl[\rho_i^{-1}>\frac{n^{3/{\varepsilon}_0}}{2}\Bigr]\leq \frac {2^{{\varepsilon}_0}{\mathbf{E}}[ \rho_0^{-{\varepsilon}_0}]}{n^3}.$$
Using the Borel-Cantelli Lemma one can obtain that for ${\mathbf{P}}$-almost all $\omega$ and $n$ large enough, we have $\omega \in A(n)\cap B'(n,\nu,m) \cap G_1(n) \cap D(n)\cap F(n)$. That is, the width of the valleys is lower than $(\ln n)^2$, their depth lower than $(\ln n + 2 \ln \ln n)/\kappa$, we can control the number of valleys deeper than $\frac {a} {\kappa}\ln n-\ln \ln n$, and there is at least one valley of depth $(\ln n -4 \ln \ln n)/\kappa$.
Due to the definition of the valleys, the potential goes down at least by $\frac{3}{1\wedge \kappa}\ln n$ in a valley and on $G_1(n)$ the biggest increase of potential is lower than $\frac 1{\kappa} (\ln n+ 2 \ln \ln n)$ for all valleys in $[-n,n]$. In particular, on $G_1(n)$, $(V(K_i))_{i\leq 2n}$ is a decreasing sequence and we have $$\begin{aligned}
V(b_{i+1}) & \leq V(b_i) - \frac{3}{1\wedge \kappa}\ln n + \frac 1{\kappa} (\ln n+ 2 \ln \ln n)
\nonumber\\
& \leq V(b_i) - \frac{2}{1\wedge \kappa}\ln n +
\frac 2{\kappa} \ln \ln n\end{aligned}$$ implying using (\[pidef\]) that for all valleys in $[-n, n]$, $$\label{weightbottom}
\qquad \pi(b_i)\leq 2e^{-V(b_i)}\leq \frac {2(\ln n)^{2/\kappa}}{n^{2/(1\wedge \kappa)}} \pi(b_{i+1})\leq \frac 12 \pi(b_{i+1}).$$
In a similar fashion, we can give an upper bound for $V(K_{i})-V(b_i)$ on $G_1(n)\cap F(n)$. We claim that on $G_1(n)\cap F(n)$, for a constant $\gamma_0$, $$\label{bound_pot}
V(K_{i})-V(K_{i+1}) \leq V(K_{i})-V(b_i)\leq \gamma_0 \ln n.$$ To show (\[bound\_pot\]), let $x$ be the smallest integer larger than $K_{i}$ such that $V(x)\leq V(K_{i})-(3/(1\wedge \kappa)) \ln n$. By definition of $K_{i+1}$ it satisfies $V(x)\leq V(K_{i+1})$. But on $F(n)$ we know that $V(x)\geq V(K_{i}) -(3/(1\wedge \kappa)+3/{\varepsilon}_0)\ln n$. Recalling that on $G_1(n)$ we have $V(b_i)\geq V(K_{i+1})-(2/\kappa)\ln n$, we get for $n$ large enough $$\begin{aligned}
V(K_i)-V(b_i)& \leq V(K_i)-(V(K_{i+1})-\frac{2}{\kappa} \ln n)\nonumber\\
& \leq V(K_i)-(V(x)-\frac{2}{\kappa} \ln n)\nonumber\\
& \leq \Bigl(\frac{3}{1 \wedge \kappa} +\frac{3}{{\varepsilon}_0} +\frac{2}{\kappa}\Bigr) \ln n\, .\end{aligned}$$
Bounds on the probability of confinement {#s_confinement}
========================================
In this section, let $I=[a,c]$ be a finite interval of ${\mathbb{Z}}$ containing at least four points and let the potential $V(x)$ be an arbitrary function defined for $x \in [a-1, c]$, with $V(a-1)=0$. This potential defines transition probabilities given by $\omega_x = e^{-V(x)}/\pi(x)$, $x \in [a,c]$ where $\pi(x)$ is defined as in (\[pidef\]) (taking $V(a-1)=0$ is no loss of generality since the transition probabilities remain the same if we replace $V(x)$ by $V(x) + c$, $\forall x$). We denote by $X$ the Markov chain restricted on $I$ in the following way: the transition probability $\omega_a$ from $a$ to $a+1$ is defined as above, and with probability $1-\omega_a$ the walk just stays in $a$; in the same way, we define the reflection at the other border $c$. We denote $$\begin{aligned}
H_+&=\max_{x\in[a, c]} \Bigl(\max_{y \in [x,c]} V(y) -\min_{y\in [a,x)} V(y)\Bigr) ,\\
H_-&=\max_{x\in[a, c]} \Bigl(\max_{y \in [a,x]} V(y) -\min_{y\in (x,c]} V(y)\Bigr),\end{aligned}$$ and $$H=H_+\wedge H_-.$$ Let us denote also by $${\tilde M} = \max_{y \in [a,c]} V(y) -\min_{y\in [a,c]} V(y)$$ the maximal difference between the values of the potential in the interval $[a,c]$. Also, we set $$\begin{aligned}
f&= \begin{cases}
c, & \text{ if } H=H_+, \\
a, & \text{ otherwise. }
\end{cases}\end{aligned}$$
To avoid confusion, let us mention that the results of this section (Propositions \[spectral\], \[cost\_to\_climb\], \[Lboundconf\]) hold for both the unrestricted and restricted random walks (as long as the starting point belongs to $I$). First, we prove the following
\[spectral\] There exists ${\gamma}_1>0$, such that for all $u\geq 1$ $$\begin{aligned}
\lefteqn{\max_{x\in I} {P_{\omega}}^x\Bigl[\frac{T_{\{a,c\}}}{{\gamma}_1(c-a)^3((c-a)+{\tilde M})e^H}>u \Bigr]}\\
&\leq \max_{x\in I} {P_{\omega}}^x\Bigl[\frac{T_{f}}{{\gamma}_1(c-a)^3((c-a)+{\tilde M})e^H}>u \Bigr] \\ &\leq e^{-u}.\end{aligned}$$
The first inequality is trivial, we only need to prove the second one. In the following we will suppose that $H=H_+$ (so that $f=c$), otherwise we can apply the same argument by inverting the space. We denote by $b$ the leftmost point in the interval $[a,c]$ with minimal potential.
We extend the Markov chain on the interval $I$ to a Markov chain on the interval $I'=[a,c+1]$ in the following way. Let $V(c+1):=V(b)$, yielding $\omega_{c+1}=\bigl(1+e^{-(V(c)-V(b))}\bigr)^{-1}$. Again, with probability $1- \omega_{c+1}$, the Markov chain goes from $c+1$ to $c$, and with probability $\omega_{c+1}$, the Markov chain just stays in $c+1$.
Let us denote by ${\hat X}_t$ the continuous time version of the Markov chain on $I'$ (i.e., the transition probabilities become transition rates). The reason for considering continuous time is the following: we are going to use spectral gap estimates, and these are better suited for continuous time in this context (mainly due to the fact that the discrete-time random walk is periodic). We define the probability measure $\mu$ on $I'$ which is reversible (and therefore invariant) for ${\hat X}$ in the following way $$\mu(x)=\pi(x)\Bigl(\sum_{y\in I'} \pi(y)\Bigr)^{-1},$$ for all $x\in I'$, where $\pi$ is as in (\[pidef\]) with the potential defined above, satisfying $V(a-1) =0$ and $V(c+1)=V(b)$. Now, the goal is to bound the spectral gap $\lambda(I')$ from below. We can do this using a result of [@Miclo]: $$\label{micloprop}
\frac 1 {4B^{I'}} \leq \lambda(I') \leq \frac 2 {B^{I'}},$$ where $B^{I'}=\min_{i \in I'} ( B_-^{I'}(i) \wedge B_+^{I'}(i))$ and $$\begin{aligned}
B_+^{I'}(i) &= \max_{x>i}\left(\sum_{y=i+1}^x (\mu(y)(1-\omega_y))^{-1}\right)\mu[x,c+1],\quad i \in [a,c]\\
B_-^{I'}(i) &= \max_{x<i}\left(\sum_{y=x}^{i-1} (\mu(y)\omega_y)^{-1}\right)\mu[a,x], \quad i\in [a+1, c+1]\end{aligned}$$ and $ B_+^{I'}(c+1)= B_-^{I'}(a)= 0$. Obviously, we have $B^{I'} \leq B_-^{I'}(c+1)$. Moreover, since implies that $\omega_x\pi(x)=e^{-V(x)}$ for any $x\in I'$, we can write $$\begin{aligned}
B_-^{I'}(c+1) &= \max_{x\leq c} \Bigl(\sum_{y=x}^{c} \frac{1}{\omega_y\pi(y)}\Bigr)
\Bigl(\sum_{y=a}^x \pi(y)\Bigr) \\
&= \max_{x\leq c} \Bigl(\sum_{y=x}^{c} e^{V(y)}\Bigr)
\Bigl(\sum_{y=a}^x (e^{-V(y)}+e^{-V(y-1)})\Bigr) \\
& \leq 2 \max_{x\leq c}\Bigl(\sum_{y=x}^{c} e^{V(y)}\Bigr)\Bigl(\sum_{y=a}^x e^{-V(y)}\Bigr) \\
&\leq 2(c-a)^2 e^H.\end{aligned}$$ This yields $$\lambda(I') \geq \frac 1 {8(c-a)^2 e^H}.$$
Using Corollary 2.1.5 of [@Saloff], we obtain that for $x,y\in I'$ and $s>0$ $${\left\vert {P_{\omega}}^x[{\hat X}_s=y] - \mu(y)\right\vert} \leq \Bigl(\frac{\mu(y)}{\mu(x)}\Bigr)^{1/2} \exp(-\lambda(I')s).$$
We apply this formula for $y=c+1$. Note that, using (\[pidef\]), we obtain that $(\mu(c+1)/\mu(x))^{1/2} \leq \sqrt{2}e^{{\tilde M}/2}$ for any $x\in(a,c)$. So, for $s:=C_1(c-a)^2((c-a)+{\tilde M})e^H$, if $C_1>4$ is chosen large enough $${\left\vert {P_{\omega}}^x[{\hat X}_s=c+1] - \mu(c+1)\right\vert} \leq \sqrt{2}e^{-C_1(c-a)/8} < \frac {1}{8(c-a)},$$ and, since $\mu(c+1) \geq 1/2(c+1-a) \geq 1/(4(c-a))$, we obtain $$\min_{x\in I'} {P_{\omega}}^x[{\hat X}_s=c+1] \geq \frac 1 {8(c-a)}.$$
Let us divide $[0,t]$ into $N:=\left\lfloor t/s \right\rfloor$ subintervals. Using the above inequality and Markov’s property we obtain (${\hat T}$ stands for the hitting time with respect to ${\hat X}$) $$\begin{aligned}
{P_{\omega}}^x[{\hat T}_{c}>t] &\leq {P_{\omega}}^x[{\hat T}_{c+1}>t]\\
&\leq {P_{\omega}}^x[{\hat X}_{sk}\neq c+1, k=1,\ldots, N]\\
&\leq\Bigl(1-\frac 1 {8(c-a)}\Bigr)^N \\
&\leq \exp\Bigl(-\frac{N}{8(c-a)}\Bigr)\\
&\leq \exp\Bigl(-\frac t{8C_1(c-a)^3((c-a)+{\tilde M})e^H}\Bigr)\exp\Bigl(\frac 1 {8(c-a)}\Bigr).\end{aligned}$$
The estimates on the continuous time Markov chain transfer to discrete time. Indeed, there exists a family $({\mathbf e}_i)_{i\geq 1}$ of exponential random variables of parameter $1$, such that the $n$-th jump of the continuous time random walk occurs at $\sum_{i=1}^n {\mathbf e}_i$. These random variables are independent of the environment and the discrete-time random walk. Moreover, $P[{\mathbf e}_1+\cdots+ {\mathbf e}_n \geq n]$ $\geq 1/3$, for all $n$. So, for any $t$, $$\frac{1}{3} {\mathbf{P}}[ T_{c} \geq t] \leq {\mathbf{P}}[ T_{c} \geq t] {\mathbf{P}}[{\hat T}_{c}\geq T_{c}] = {\mathbf{P}}[ T_{c} \geq t, {\hat T}_{c}\geq T_{c} ]\leq {\mathbf{P}}[{\hat T}_{c} \geq t],$$ Hence, we have for all $v > 0$ $$\max_{x\in I} {P_{\omega}}^x\Bigl[\frac{{ T}_{c}}{8(1+v) C_1(c-a)^3((c-a)+{\tilde M})e^H}>u \Bigr]
\leq \Bigl(3 e^{1/8}e^{-v u}\Bigr)e^{- u},$$ for all $u\geq 0$. Hence for $u\geq 1$, choosing $v$ large enough in such a way that $3\exp(\frac{1}{8}-v)\leq 1$, we obtain the result with ${\gamma}_1=8C_1(1+v)$.
Next, we recall the following simple upper bound on hitting probabilities:
\[cost\_to\_climb\] There exists $\gamma_2$ such that for any $x,y$ and $h\in[x,y]$ we have $${P_{\omega}}^x[T_y<s] \leq \gamma_2(1+s)\frac{\pi(h)}{\pi(x)}.$$
We can adapt Lemma 3.4 of [@CP] (which used a uniform ellipticity condition). We remain in the continuous time setting and, considering the event that $y$ is visited before time $s$ and left again at least one time unit later (on which $\int_0^{s+1} {{\mathbf 1}{\{\hat X_u=y\}}}du \geq 1$), we have $$\int_0^{s+1} {P_{\omega}}^x[\hat X_u=y]du \geq {P_{\omega}}^x[\hat T_y<s]\cdot P[{\mathbf e}_1 \geq 1]$$ where ${\mathbf e}_1$ is an exponential random variable of parameter $1$. Hence $$\begin{aligned}
{P_{\omega}}^x[\hat T_y<s] & \leq {P_{\omega}}^x[\hat T_{h}<s] \\
& \leq e \int_0^{s+1} {P_{\omega}}^x[\hat X_u=h]du \\
& = e \int_0^{s+1} \frac{\pi(h)}{\pi(x)} {P_{\omega}}^{h}[\hat X_u=x]du \\
& \leq e (s+1) \frac{\pi(h)}{\pi(x)}.\end{aligned}$$
Again, one can easily transfer the estimates on the continuous time Markov chain to discrete time.
Let us now introduce $$\begin{aligned}
H_+^*&=\max_{x\in[a+1, c-1]} \Bigl(\max_{y \in [x,c-1]} V(y) -\min_{y\in [a+1,x)} V(y)\Bigr) ,\\
H_-^*&=\max_{x\in[a+1, c-1]} \Bigl(\max_{y \in [a+1,x]} V(y) -\min_{y\in (x,c-1]} V(y)\Bigr),\end{aligned}$$ and $$H^*=H_+^*\wedge H_-^*.$$
We obtain a lower bound on the confinement probability in the following proposition. Recall that $b$ is the leftmost point in the interval $[a,c]$ with minimal potential.
\[Lboundconf\] Suppose that $c-1$ has maximal potential on $[b,c-1]$ and $a$ has maximal potential on $[a,b]$. Then, there exists ${\gamma}_3>0$, such that for all $u \geq 1$ $$\min_{x\in I} {P_{\omega}}^x\Bigl[ {\gamma}_3 \ln(2(c-a)) \frac{T_{\{a,c\}}}{e^{H^*}}\geq u\Bigr] \geq \frac 1 {2(c-a)}e^{-u},$$ if $e^{H^*}\geq 16{\gamma}_2$.
Noticing that $$\min_{b<h<c-1} \frac{\pi(h)}{\pi(b)}\leq 2 e^{-H_+^*} \text{ and }\min_{a+1<h<b} \frac{\pi(h)}{\pi(b)}\leq 2 e^{-H_-^*},$$ we can apply Proposition \[cost\_to\_climb\] to obtain that $$\label{Lboundstep1}
\text{for all }s \geq 1,\qquad {P_{\omega}}^b[T_{\{a,c\}}<s]\leq 8{\gamma}_2 s e^{-H^*},$$ Hence for $s=e^{H^*}/(16 \gamma_2)\geq 1$, the right-hand side of the previous inequality equals $1/2$.
Now, using the exit probability formula (\[exit\_probs\]), we obtain that $$\label{Lboundstep2}
\min_{x\in I} {P_{\omega}}^x[T_b<T_{\{a,c\}}] \geq (c-a)^{-1}.$$ Denoting $N=\left\lceil t/s \right\rceil$, we obtain for $x\in I$, $$\begin{aligned}
{P_{\omega}}^x[T_{\{a,c\}}>t] & \geq (2(c-a))^{-(N+1)} \nonumber\\
& \geq \exp\Bigl( - \frac{ C_2 t \ln (2(c-a))}{e^{H^*}} \Bigr)(2(c-a))^{-1}.\end{aligned}$$ We used the following reasoning in the above calculation. Start from any $x\in (a,c)$, by (\[Lboundstep2\]) the particle hits $b$ before $\{a,c\}$ with probability at least $(c-a)^{-1}$. Then, during $s$ time units, $\{a,c\}$ will not be hit with probability at least $1/2$. After that, the particle is found in some $x'\in (a,c)$ and at least $s$ time units elapsed from the initial moment. So the cost of preventing the occurrence of $T_{\{a,c\}}$ during any time interval of length $s$ is at most $(2(c-a))^{-1}$. The result follows for ${\gamma}_3$ large enough.
Our main application of Proposition \[spectral\] and Proposition \[Lboundconf\], will be to control the exit times of valleys, more precisely we will be able to give upper bounds on the tail of $T_{\{K_i,K_{i+1}\}}$ and lower bounds on the tail of $T_{\{K_i-1,K_{i+1}+1\}}$ in terms of $H_i$.
Induced random walk {#s_induced_rw}
===================
Let us denote $(s_k(n))_{k\geq 0}$ the sequence defined by $$\begin{aligned}
s_0(n)&=0,\\
s_{i+1}(n) &=\min \{ j\geq s_i(n):\ X_j \in \{K_l(n),l\geq 0 \} \}.\end{aligned}$$
Then, we define $Y_i=X_{s_i}$, the embedded random walk with state space $\{K_l,l\geq 0\}$, enumerating the successive valleys we visit and $l_n(\nu)=\max \{i : s_i \leq T_{n^{\nu}}\}$ the numbers of steps made by the embedded random walk to reach $[n^{\nu}; \infty)$. For the reflected case, we will use the same notation, replacing $\{K_l,l\geq 0\}$ with $\{\widetilde{K}_l,l\geq 0\}$ defined in (\[tilddef\]).
Recall (\[indices\]) and let us denote $$\xi^{\nu}(i)= {\mathop{\mathrm{card}}}\{ j \in [0, l_n(\nu)]: Y_j=K_{i+1}, Y_{j+1}=K_{i} \} \text{ for } i=i_0+1, \ldots , i_1-1,$$ and in order to carry over the proofs to the reflected case $$\widetilde{\xi}^{\nu}(i)= {\mathop{\mathrm{card}}}\{ j \in [0, l_n(\nu)]: Y_j=\widetilde{K}_{i+1}, Y_{j+1}=\widetilde{K}_{i} \} \text{ for } i=i_0+1, \ldots , i_1-1.$$
Moreover, we introduce the real time elapsed, i.e. in the clock of $X_n$, during the first left-right crossing of the $i$-th valley $$T^{\text{next}}(i)=T_{K_{i+1}}\circ\theta(\text{next}(i))-\text{next}(i),$$ where $\theta$ denotes the time-shift for the random walk and $$\text{next}(i)=\inf\{n\geq 0: X_n=K_i, T_{K_{i+1}}\circ \theta(n)<T_{K_{i-1}}\circ \theta(n)\}.$$ In this way, each time the embedded random walk backtracks, $T^{\text{next}}(i)$ is the time the walk will need to make the necessary left-right crossing of the corresponding valley. Recall (\[Ndef\]). Conditionally on $(Y_i)_{i\geq 1}$ we have that (“dir” stands for “direct”, and “back” stands for “backtrack”) $$\label{decompos}
T_{n^{\nu}} = {{\mathcal T}_{\text{\itshape{init}}}}+{{\mathcal T}_{\text{\itshape{dir}}}}+{{\mathcal T}_{\text{\itshape{back}}}}+{{\mathcal T}_{\text{\itshape{left}}}}+{{\mathcal T}_{\text{\itshape{right}}}},$$ where $$\begin{aligned}
{{\mathcal T}_{\text{\itshape{init}}}}&=
\begin{cases}
T_{K_{i_0+1}}, &\text{ if $T_{K_{i_0+1}}<T_{K_{i_0}}$, }\\
T_{K_{i_0}}+T^{\text{next}}(i_0)\circ \theta(T_{K_{i_0}}), &\text{ else,}
\end{cases} \\
{{\mathcal T}_{\text{\itshape{left}}}}&=
\begin{cases}
{\mathop{\mathrm{card}}}\{i\leq T_{n^{\nu}}: X_i < K_1\}\qquad \qquad \qquad \qquad \ \ \ \ \text{ without reflection,}\\ \\
\sum_{j=0}^{l_n(\nu)} {{\mathbf 1}{\{Y_j=\widetilde{K}_{i_0+1},Y_{j+1}=\widetilde{K}_{i_0}\}}} &\\
\ \ \times \Bigl(T_{K_i}\circ \theta(s_j)-s_j+T^{\text{next}}(i) \circ (T_{K_i}\circ \theta(s_j))\Bigr)\text{ with reflection,}
\end{cases}\\
{{\mathcal T}_{\text{\itshape{right}}}}&=T_{n^{\nu}}\circ \theta(\text{next}^*(i_1))-\text{next}^*(i_1),\\
{{\mathcal T}_{\text{\itshape{dir}}}}&=\sum_{i=i_0+1}^{i_1-1} T^{\text{next}}(i)\circ \theta(T_{K_{i}}),\\
{{\mathcal T}_{\text{\itshape{back}}}}&=\begin{cases}
\sum_{i=1}^{i_1-1}\sum_{j=0}^{l_n(\nu)} {{\mathbf 1}{\{Y_j=K_{i+1},Y_{j+1}=K_i\}}} \\
\ \ \times \Bigl(T_{K_i}\circ \theta(s_j)-s_j+T^{\text{next}}(i) \circ (T_{K_i}\circ \theta(s_j))\Bigr) \text{ without reflection,}\\ \\
\sum_{i=i_0+1}^{i_1-1}\sum_{j=0}^{l_n(\nu)} {{\mathbf 1}{\{Y_j=K_{i+1},Y_{j+1}=K_i\}}} \\
\ \ \times \Bigl(T_{K_i}\circ \theta(s_j)-s_j+T^{\text{next}}(i) \circ (T_{K_i}\circ \theta(s_j))\Bigr) \text{ with reflection},\\
\end{cases}\end{aligned}$$ where $\text{next}^*(i_1)=\inf\{n\geq 0: X_n=K_{i_1}, T_{n^{\nu}}\circ \theta(n)<T_{K_{i_1-1}}\circ \theta(n)\}$. In the reflected case, replace $K_i$ with $\widetilde{K}_i$ in all the above definitions except for that of ${{\mathcal T}_{\text{\itshape{left}}}}$. This decomposition is illustrated on Figure \[f\_decompo\] for the non-reflected case.
![On the decomposition (\[decompos\]) of $T_{n^{\nu}}$[]{data-label="f_decompo"}](times){width="\textwidth"}
In the non-reflected case, we have the following equalities in law (for each $\omega$): $$\begin{aligned}
{{\mathcal T}_{\text{\itshape{init}}}}&=\overline{\tau}(0),\label{defTinit}\\
{{\mathcal T}_{\text{\itshape{right}}}}&=\overline{\tau}(n^{\nu}),\label{defTright}\\
{{\mathcal T}_{\text{\itshape{dir}}}}&=\sum_{i=i_0+1}^{i_1-1} \tau_+^{(0)}(i),\label{defTdirect}\\
{{\mathcal T}_{\text{\itshape{back}}}}&=\sum_{i=1}^{i_1-2} ( \tau_+^{(1)}(i)+\tau_-^{(1)}(i)+ \cdots + \tau_+^{(\xi^{\nu}(i))}(i)+\tau_-^{(\xi^{\nu}(i))}(i))
\label{defTbacktrack}\\ \nonumber & \qquad+ \sum\limits_{j=1}^{\xi^\nu(i_1 -1)} \tau_+^{(j)}(i_1-1)+ \tau_-^{\rm{last}, (j)}, \end{aligned}$$ where $\tau_+^{(j)}(i)$, $\tau_-^{(j)}(i)$ and $\tau_-^{\rm{last}, (j)}$ are independent sequences of i.i.d. random variables described as follows. First, $\tau_+^{(j)}(i)$ is a sequence of independent random variables with the same law as $T_{K_{i+1}}$ under ${P_{\omega}}^{K_i}[~\cdot\mid T_{K_{i+1}}<T_{K_{i-1}}]$. Then, $\tau_-^{(j)}(i)$ is a sequence of independent random variables with the same law as $T_{K_{i}}$ (under ${P_{\omega}}^{K_{i+1}}[~\cdot\mid T_{K_{i}}<T_{K_{i+2}}]$) and $\tau_-^{\rm{last},j}$ is a sequence of independent random variables with the same law as $T_{K_{i_1-1}}$ under ${P_{\omega}}^{K_{i_1}}[~\cdot\mid T_{K_{i_1-1}} < T_{n^{\nu}}]$. Clearly, the random variable $\overline{\tau}(0)$ (respectively, $\overline{\tau}(n^{\nu})$) has the same law as $T_{K_{i_0+1}}$ (respectively, $T_{n^{\nu}}$) under ${P_{\omega}}[~\cdot\mid T_{K_{i_0+1}}<T_{K_{i_0-1}}]$ (respectively, ${P_{\omega}}^{K_{i_1}}[~\cdot\mid T_{n^{\nu}}<T_{K_{i_1-1}}]$).
In the reflected case, we simply replace $K_i$ by $\widetilde{K}_i$, $\xi^{\nu}(i)$ by $\widetilde{\xi}^{\nu}_i$ and $\omega$ by $\tilde{\omega}$.
We want to give bounds on the number of backtracks between valleys before the walk reaches $\lfloor n^{\nu}\rfloor$. Denote $$\label{Bdef}
{{\mathfrak{B}}}(n):={\mathop{\mathrm{card}}}\{ i\geq 1 : s_{i+1}(n) \leq T_{n^{\nu}},\ Y_{i+1}<Y_i\}=\sum_{i=1}^{i_1-1} \xi^{\nu}(i).$$
By (\[exit\_probs\]), we obtain that for $i\leq i_1$, ${\mathbf{P}}$-a.s. for $n$ large enough, $$\begin{aligned}
{P_{\omega}}^{K_i}[T_{K_{i+1}}>T_{K_{i-1}}]&=\Bigl(\sum_{j=K_{i-1}}^{K_{i+1}-1} e^{V(j)}\Bigr)^{-1}\sum_{j=K_i}^{K_{i+1}-1} e^{V(j)}\label{crossingprob}\\
&\leq \max_{i\leq n} (K_{i}-K_{i-1}) \frac{(\ln n)^{2/\kappa}}{n^{2/(1\wedge\kappa)}}\nonumber\\
&\leq n^{-3/2}, \nonumber\end{aligned}$$ since $\max_{i\leq n} (K_{i+1}-K_{i})\leq (\ln n)^2$ on $A(n)$ and, due to Lemma \[depthvalley\], with the same argument as for (\[weightbottom\]), we have $V(K_{i-1})-V(x) \geq \frac{2}{1\wedge \kappa} \ln n -\frac{2}{\kappa}\ln \ln n$ for $x\in [K_i,K_{i+1}]$.
Using (\[exit\_probs\]) and (\[bound\_pot\]), we obtain a lower bound: for $\omega \in A(n) \cap F(n) \cap G_1(n)$ we have $$\label{crossingprob2}
{P_{\omega}}^{K_i}[T_{K_{i+1}}>T_{K_{i-1}}] \geq \frac 1 {K_{i+1}-K_{i-1}} \frac 1 {e^{V(K_{i-1})-V(K_{i+1})}} \geq n^{-(1+2\gamma_0)}.$$
During the first $3n$ steps of the embedded random walk there are two cases, either the walk has reached $n^{\nu}$ or there are at least $n$ steps back. But then if $n^{\nu}$ is reached in less than $3n$ steps, ${{\mathfrak{B}}}(n)$ is stochastically dominated by a $\text{Bin}(3n,n^{-3/2})$ by (\[crossingprob\]). Moreover, we get for $f(\cdot)$ such that $f(n)=O(n)$, ${\mathbf{P}}$-a.s. for $n$ large enough, $$P_{\omega}[{{\mathfrak{B}}}(n)\geq f(n)] \leq \binom{3n}{n} \Bigl(\frac1 {n^{3/2}}\Bigr)^{n}+P\Bigl[\text{Bin}(3n,n^{-3/2})\geq f(n)\Bigr],$$ and so using Stirling’s formula and Chebyshev’s exponential inequality, ${\mathbf{P}}$-a.s. for $n$ large enough, $$\begin{aligned}
P_{\omega}[{{\mathfrak{B}}}(n)\geq f(n)] &\leq \exp(-C_1 n) +C_2\exp(-f(n))
\label{boundback} \\
& \leq C_3\exp(-f(n)). \nonumber\end{aligned}$$
Quenched slowdown {#s_q_slowdown}
=================
In this section, we prove Theorem \[t\_q\_slow\]. Before going into technicalities, let us give an informal argument about why we obtain different answers in Theorem \[t\_q\_slow\].
![The three strategies for the slowdown[]{data-label="f_strategies"}](strategies){width="12cm"}
Suppose that $\frac{\kappa}{\kappa+1}<1-\frac{\nu}{\kappa}$, or equivalently, $\nu<\frac{\kappa}{\kappa+1}$. Consider the three strategies depicted on Figure \[f\_strategies\]:
- The particle goes to the biggest valley in the interval $[0,n^\nu]$, and stays there up to time $n$.
- The particle goes to the biggest valley in the interval $[0,n^{\frac{\kappa}{\kappa+1}}]$, stays there up to time $n-n^{\frac{\kappa}{\kappa+1}}$, and then goes back to the interval $[0,n^\nu]$.
- The particle goes to the biggest valley in the interval $[-n^{\frac{\kappa}{\kappa+1}},0]$ (so that typically it has to go roughly $n^{\frac{\kappa}{\kappa+1}}$ units to the left), and stays there up to time $n$.
By Lemmas \[depthvalley\] and \[depthvalley2\], the biggest valley in the interval $[0,n^\nu]$ has depth of approximately $\frac{\nu}{\kappa}\ln n$. Using Proposition \[Lboundconf\], we obtain that the probability of staying there up to time $n$ is roughly $\exp(-n^{1-\frac{\nu}{\kappa}})$. As for the strategy [**2**]{}, analogously we find that the biggest valley in the interval $[0,n^{\frac{\kappa}{\kappa+1}}]$ has depth around ${\frac{1}{\kappa+1}}\ln n$, and the probability of staying there is roughly $\exp(-n^{\frac{\kappa}{\kappa+1}})$. Then, the probability of backtracking is again around $\exp(-n^{\frac{\kappa}{\kappa+1}})$. The situation with the strategy [**3**]{} is the same as that with strategy [**2**]{} (for the strategy [**3**]{}, we first have to backtrack and then to stay in the valley, but the probabilities are roughly the same).
So, in the case $\nu<\frac{\kappa}{\kappa+1}$ the strategies [**2**]{} and [**3**]{} are better than the strategy [**1**]{}. The only situation when we cannot use neither [**2**]{} nor [**3**]{} is when the RWRE has reflection in the origin, and we are considering the hitting times.
Time spent in a valley {#s_time_crossing}
----------------------
We have
\[crossingtime\] There exists ${\gamma}_4>0$ such that for ${\mathbf{P}}$-almost all $\omega$, for all $n$ large enough we have for $i\leq 2n+1$ and $u\geq 1$, $$\begin{aligned}
{P_{\omega}}^{K_i}\bigl[T_{K_{i+1}}>u \bigl({\gamma}_4(\ln n)^{10}e^{H_{i-1}\vee H_i}\bigr)\mid T_{K_{i+1}}<T_{K_{i-1}}\bigr] &\leq e^{-u},\\
{P_{\omega}}^{K_i}\bigl[T_{K_{i-1}}>u \bigl({\gamma}_4(\ln n)^{10}e^{H_{i-1}\vee H_i}\bigr)\mid T_{K_{i-1}}<T_{K_{i+1}}\bigr] &\leq e^{-u}.\end{aligned}$$
We prove only the second part of the proposition, the first one uses the same arguments. First, we have $$\max_{x\in (K_{i-1},K_{i+1})} \Bigl(\max_{y \in [x,K_{i+1})} V(y) -\min_{y\in [K_{i-1},x)} V(y)\Bigr)=H_{i-1}\vee H_i.$$
Using (\[crossingprob2\]) (or (\[crossingprob\]) for the first part of the proposition), we obtain ${\mathbf{P}}$-a.s. for $n$ large enough, $$\begin{aligned}
& \lefteqn{{P_{\omega}}^{K_i}\bigl[T_{K_{i-1}}>u \bigl({\gamma}_4(\ln n)^{10}e^{H_{i-1}\vee H_i}\bigr)\mid
T_{K_{i+1}}>T_{K_{i-1}}\bigr]} \\
&\leq n^{1+2\gamma_0}{P_{\omega}}^{K_i}\bigl[T_{\{K_{i-1},K_{i+1}\}}>u \bigl({\gamma}_4(\ln n)^{10}e^{H_{i-1}\vee H_i}\bigr),T_{K_{i+1}}>T_{K_{i-1}}\bigr].
\end{aligned}$$ To estimate this last probability, we may consider the random walk reflected at $K_{i-1}$ and $K_{i+1}$. On $A(n)$ we have $K_{i+1}-K_{i-1} \leq 2 (\ln n)^2$ and on $G_1(n)\cap F(n)$ we have $\max_{y \in [K_{i-1},K_{i+1}]} V(y) -\min_{y\in [K_{i-1},K_{i+1})} V(y)\leq 2\gamma_0 \ln n$ by (\[bound\_pot\]). Hence for $n$ such that $\gamma_0\leq (\ln n)^2$ we can apply Proposition \[spectral\] with $a=K_{i-1}$, $c=K_{i+1}$, $\tilde{M}\leq 2 (\ln n)^2$ and $H=H_{i-1}\vee H_i$ to get $$\begin{aligned}
\lefteqn{ P_{\hat\omega}^{K_i}\left[T_{\{K_{i-1},K_{i+1}\}}>u \left({\gamma}_4(\ln n)^{10}e^{H_{i-1}\vee H_i}\right)\right]}\hphantom{***********}\\
&\leq \exp\bigl(-u{\gamma}_4 (\ln n)^2/(32{\gamma}_1)\bigr),\end{aligned}$$ where $\hat\omega$ denotes the environment with reflection at $K_{i-1}$ and $K_{i+1}$, so that $$\begin{aligned}
\lefteqn{{P_{\omega}}^{K_i}\bigl[T_{K_{i-1}}>u \bigl({\gamma}_4(\ln n)^{10}e^{H_{i-1}\vee H_i}\bigr)\mid
T_{K_{i+1}}>T_{K_{i-1}}\bigr]} \hphantom{*****}\\
&\leq \exp\bigl(-u{\gamma}_4 (\ln n)^2/(32{\gamma}_1)+(1+2\gamma_0) \ln n\bigr)\\
&\leq e^{-u},
\end{aligned}$$ for ${\gamma}_4>32{\gamma}_1((1+2\gamma_0)+1)$ and $n$ large enough.
Let $Z_i$ be a random variable with the same law as $T_{K_{i+1}}$ under ${P_{\omega}}^{K_i}[~\cdot\mid T_{K_{i+1}}<T_{K_{i-1}}]$. Then, for $i \in N(-n^{a},n^{b})$ and $H=\max_{i\in N(-n^{a},n^{b})} H_i$, we have that ${\mathbf{P}}$-a.s. for $n$ large enough $$\label{domstoch_1}
\frac{Z_i}{{\gamma}_4e^H(\ln n)^{10}}\prec 1 + \mathbf{e},$$ where $\mathbf{e}$ is an exponential random variable with parameter $1$. Since $\omega\in G_1(n^{a\vee b})$ ${\mathbf{P}}$-a.s. for $n$ large enough, there is a constant $\gamma > 0$ (depending only on $\kappa$) such that $$\label{domstoch}
\frac{Z_i}{{\gamma}_4n^{(a \vee b)/\kappa}(\ln n)^{{\gamma}}}\prec 1 + \mathbf{e}.$$ The same inequality is true when $K_{i-1}$ and $K_{i+1}$ are exchanged. We point out that the same stochastic domination holds in the reflected case, even for $T_{\tilde K_{i_0+2}}$ under ${P_{{{\tilde\omega}}}}^{\tilde K_{i_0+1}}[~\cdot\mid T_{\tilde K_{i_0+2}}<T_{\tilde K_{i_0}}]={P_{{{\tilde\omega}}}}^{\tilde K_{i_0+1}}[~\cdot~]$ in which case it is a direct consequence of Proposition \[spectral\].
Using the same kind of arguments as in the proof of Proposition \[crossingtime\] we obtain
\[initcrossingtime\] There exists a positive constant ${\gamma}_4$ (without restriction of generality, the same as in Proposition \[crossingtime\]) such that for ${\mathbf{P}}$-almost all $\omega$, we have for all $n$ large enough, with $i_0={\mathop{\mathrm{card}}}N_n(-n,0)$ and $u\geq 1$, $$\begin{aligned}
&{P_{\omega}}\bigl[T_{K_{i_0+1}(n)}>u \bigl({\gamma}_4(\ln n)^{10}e^{H_{i_0-1}\vee H_{i_0}}\bigr)\mid T_{K_{i_0+1}(n)}<T_{K_{i_0-1}(n)}\bigr] \leq e^{-u},\\
&{P_{{{\tilde\omega}}}}\bigl[T_{K_{i_0+1}(n)}>u \bigl({\gamma}_4(\ln n)^{10}e^{H_{i_0-1}\vee H_{i_0}}\bigr)\bigr] \leq e^{-u}.\end{aligned}$$
Similarly we obtain
\[rightcrossingtime\] There exists a positive constant ${\gamma}_4$ (without restriction of generality, the same as in Proposition \[crossingtime\]) such that for ${\mathbf{P}}$-almost all $\omega$, we have for all $n$ large enough with $i_1={\mathop{\mathrm{card}}}N_n(-n,n^{\nu})$ and $u\geq 1$, $${P_{\omega}}^{K_{i_1}}\bigl[T_{n^{\nu}}>u \bigl({\gamma}_4(\ln n)^{10}e^{H_{i_1-1}\vee H_{i_1}}\bigr)\mid T_{n^{\nu}}<T_{K_{i_1}(n)}\bigr] \leq e^{-u}.$$ and $${P_{\omega}}^{K_{i_1}}\bigl[T_{K_{i_1-1}}>u \bigl({\gamma}_4(\ln n)^{10}e^{H_{i_1-1}\vee H_{i_1}}\bigr)\mid T_{K_{i_1 -1}}<T_{n^\nu}\bigr] \leq e^{-u}.$$
This proposition implies that $$\label{lastdomstoch}
\frac{\tau_-^{\text{last}}}{{\gamma}_4n^{\nu/\kappa}(\ln n)^{{\gamma}}}\prec 1 + \mathbf{e}.$$
Time spent for backtracking {#s_time_backtracking}
---------------------------
Recalling the definitions (\[defTbacktrack\]) and (\[Bdef\]), we obtain, for the reflected case,
\[timebacktrack\] For $0<a<b<c<1$, we have ${\mathbf{P}}$-a.s. for $n$ large enough, $${P_{{{\tilde\omega}}}}\Bigl[\frac{{{\mathcal T}_{\text{\itshape{back}}}}}{{\gamma_4 n^{\nu/\kappa}(\ln n)^{{\gamma}}}} \geq n^c , {{\mathfrak{B}}}(n) \in[n^a,n^b)\Bigr]\leq \exp(-n^c/4),$$ where ${\gamma}$ is as in (\[domstoch\]).
On the event $\{{{\mathfrak{B}}}(n)\in [n^a,n^b)\}$, we have $\sum_{i\in N(0,n^{\nu})} \xi^{\nu}(i)= {{\mathfrak{B}}}(n)< n^b$, so we can use (\[domstoch\]) and (\[lastdomstoch\]) to get that ${\mathbf{P}}$-a.s. for $n$ large enough, $$\label{Gammaest}
\frac{{{\mathcal T}_{\text{\itshape{back}}}}}{\gamma_4 n^{\nu/\kappa} (\ln n)^{{\gamma}}} \prec 2n^b + \text{Gamma}(2n^b,1).$$ (note that ${{\mathcal T}_{\text{\itshape{back}}}}$ is the time spent in valleys from $0$ to $n^\nu$ because we have a reflection at 0). The factor $2$ arises from the fact that each backtracking creates one right-left crossing and one left-right crossing. We use the following bound on the tail of $\text{Gamma}(k,1)$: $$\label{tail_gamma} P[\text{Gamma}(k,1) \geq u]\leq e^{-u/2} E[\exp(\text{Gamma}(k,1)/2)]=e^{-u/2}2^k.$$
Hence we have ${\mathbf{P}}$-a.s. for $n$ large enough, $${P_{{{\tilde\omega}}}}\Bigl[\frac{{{\mathcal T}_{\text{\itshape{back}}}}}{{\gamma_4 n^{\nu/\kappa}(\ln n)^{{\gamma}}}} \geq n^c , {{\mathfrak{B}}}(n) \in [n^a,n^b)\Bigr]\leq P[\text{Gamma}(2n^b,1)\geq n^c-2n^b],$$ and since $(n^c-2n^b)/2-2n^b\ln 2\geq n^c/4$ for $n$ large enough, we conclude with (\[Gammaest\]).
In the same way, we get, still for the reflected case
\[timebacktrackleft\] For $0<a<b<c<1$, we have ${\mathbf{P}}$-a.s. for $n$ large enough, $${P_{{{\tilde\omega}}}}\Bigl[\frac{{{\mathcal T}_{\text{\itshape{left}}}}}{{\gamma_4 n^{\nu/\kappa}(\ln n)^{{\gamma}}}} \geq n^c , {{\mathfrak{B}}}(n) \in[n^a,n^b)\Bigr]\leq \exp(-n^c/4),$$ where ${\gamma}$ only depends on $\kappa$.
On the event $\{{{\mathfrak{B}}}(n) \in[n^a,n^b)\}$, ${{\mathcal T}_{\text{\itshape{left}}}}$ is lower than the time spent in the valleys of indexes $i_0$ and $i_0+1$ during backtrackings from $\widetilde{K}_{i_0+1}$ to $\widetilde{K}_{i_0}$. Since, there are at most $n^b$ backtracks for this valley and since (\[domstoch\]) is valid even for $T_{K_{i_0+2}}$ under ${P_{{{\tilde\omega}}}}^{K_{i_0+1}}[~\cdot~]$, we can use the same argument as in the proof of Proposition \[timebacktrack\].
Next, recalling the definition (\[defTbacktrack\]), we obtain
\[timebacktrack1\] For $0<a<b<1$ and $c\in(b\vee \nu,1)$, we have ${\mathbf{P}}$-a.s. for $n$ large enough, $${P_{\omega}}\Bigl[\frac{{{\mathcal T}_{\text{\itshape{back}}}}}{{n^{(b\vee \nu)/\kappa}(\ln n)^{{\gamma}}}} \geq n^c , {{\mathfrak{B}}}(n)\in [n^a,n^b) \Bigr]\leq \exp(-n^c/4),$$ where ${\gamma}$ only depends on $\kappa$.
On the event $\{{{\mathfrak{B}}}(n) \in[n^a,n^b)\}$, ${{\mathcal T}_{\text{\itshape{back}}}}$ consists of the time spent in the valleys indexed by $N_n(-n^b,n^{\nu})$, once this is noted we use the same argument as in the proof of Proposition \[timebacktrack\].
Time spent for the direct crossing {#s_direct_crossing}
----------------------------------
We can control ${{\mathcal T}_{\text{\itshape{dir}}}}$ with the following proposition. Recall (\[def\_b\_prime\]) and (\[Gdef\]).
\[timedirect\] For all $m\geq m_0(\kappa,\nu)$, we have for $n$ large enough $${P_{\omega}}\left[{{\mathcal T}_{\text{\itshape{dir}}}}\geq n \right]\leq C(m)\exp(-n^{1-(1+2/m)\frac{\nu}{\kappa}}).$$
Recall the definition (\[defTdirect\]) and let us take $\omega \in B'(n,\nu,m)\cap G_1(n)$. Let us introduce for $k=-1,\ldots,m$, $$\begin{aligned}
N(k)= &{\mathop{\mathrm{card}}}\{i \in N(-n^{\nu},n^{\nu}): \ H_i \geq \frac{\nu k}{\kappa m} \ln n + 2\ln \ln n \},\\
\sigma(k)=&{\mathop{\mathrm{card}}}\Bigl\{ i \leq T_{n^{\nu}}: X_i \in \left[K_j(n),K_{j+1}(n)\right)\text{ for some } j \nonumber\\
& \qquad\text{ with } H_j \in \Bigl[\ln n \frac{\nu k}{\kappa m}+2\ln \ln n,
\ln n \frac{\nu (k+1)}{\kappa m}+2\ln \ln n\Bigr]\Bigr\}.\label{deftauk}\end{aligned}$$
If ${{\mathcal T}_{\text{\itshape{dir}}}}\geq n$, then for some $k \in [-1,m]$ the particle spent an amount of time greater than $n/(4m)$ in the valleys of depth in $\bigl[\frac{\nu k}{\kappa m} \ln n + 2\ln \ln n, \frac{\nu (k+1)}{\kappa m} \ln n + 2\ln \ln n\bigr]$ because $\omega$ is in $G_1(n)$, so that $$\label{discretize}
{P_{\omega}}[{{\mathcal T}_{\text{\itshape{dir}}}}> n] \leq 4m \max_{k \in [-1,m]} {P_{\omega}}[\sigma(k) \geq n/(4m)].$$
Using Proposition \[crossingtime\], since $\omega \in B'(n,\nu,m)\cap G_1(n)$, we have $N(k) \leq n^{\nu(1- k/m)}$, and so $$\frac{\sigma(k)}{{\gamma}_4(\ln n)^{11}n^{\nu(k+1)/(\kappa m)}} \prec 2n^{\nu(1- k/m)} + \text{Gamma}(2n^{\nu(1- k/m)},1).$$ For $m>(1-\nu)^{-1}$ we have that $n^{\nu(1-k/m)}=o(n^{1-\nu(k+1)/m}(\ln n)^{-11})$, and for $n$ large enough (depending on $\nu$ and $m$), we use (\[tail\_gamma\]) to obtain $$\begin{aligned}
{P_{\omega}}[\sigma(k) \geq n/(4m)]& \leq P\Bigl[\text{Gamma}(2n^{\nu(1- k/m)},1) \geq \frac{n^{1-\nu(k+1)/(\kappa m)}}{(\ln n)^{12}} \Bigr]\\
& \leq 4^{n^{\nu(1-k/m)}}\exp\Bigl(-\frac{n^{1-\nu(k+1)/(\kappa m)}}{(\ln n)^{12}}\Bigr)\\
& \leq \exp\Bigl(-2n^{1-\nu(k+2)/(\kappa m)}+\ln 4 n^{\nu(1-k/m)}\Bigr).\end{aligned}$$
We need to check that $n^{1-(1+2/m)\nu/\kappa}\geq \ln 4 n^{\nu(1-k{\varepsilon})}$ for any $k$, if we take $m$ large enough, but this can be done by considering the cases $k=0$ and $k=m$. Hence we get Proposition \[timedirect\].
Upper bound for the probability of quenched slowdown for the hitting time {#s_upper_q_slowdown}
-------------------------------------------------------------------------
In this section we suppose that $\omega \in A(n)\cap G_1(n) \cap B'(n,\nu,m)$, which is satisfied ${\mathbf{P}}$-a.s. for $n$ large enough. First, we consider RWRE with reflection at the origin. Because of (\[decompos\]) $$\begin{aligned}
\label{decompoUB}
{P_{{{\tilde\omega}}}}\left[T_{n^{\nu}}>n\right] \leq &{P_{{{\tilde\omega}}}}\left[{{\mathcal T}_{\text{\itshape{dir}}}}\geq n/5 \right]+{P_{{{\tilde\omega}}}}\left[{{\mathcal T}_{\text{\itshape{back}}}}\geq n/5\right]+{P_{{{\tilde\omega}}}}\left[{{\mathcal T}_{\text{\itshape{init}}}}\geq n/5\right] \\ \nonumber
&+{P_{{{\tilde\omega}}}}\left[{{\mathcal T}_{\text{\itshape{right}}}}\geq n/5\right]+{P_{{{\tilde\omega}}}}\left[{{\mathcal T}_{\text{\itshape{left}}}}\geq n/5\right].\end{aligned}$$
Let ${\varepsilon}>0$ and recall (\[Bdef\]), then $$\begin{aligned}
{P_{{{\tilde\omega}}}}\left[{{\mathcal T}_{\text{\itshape{back}}}}\geq n/5 \right] \leq & {P_{{{\tilde\omega}}}}[{{\mathfrak{B}}}(n)> n^{1-(1+2/m)\nu/\kappa}]\\
&+{P_{{{\tilde\omega}}}}[{{\mathcal T}_{\text{\itshape{back}}}}\geq n/5,{{\mathfrak{B}}}(n)\leq n^{1-(1+2/m)\nu/\kappa}].\end{aligned}$$
Using (\[boundback\]), we can write $${P_{{{\tilde\omega}}}}[{{\mathfrak{B}}}(n)> n^{1-(1+2/m)\nu/\kappa}]\leq C_2 \exp(-n^{1-(1+2/m)\nu/\kappa}),$$ and for $n$ large enough by Proposition \[timebacktrack\], $$\begin{aligned}
\lefteqn{{P_{{{\tilde\omega}}}}[{{\mathcal T}_{\text{\itshape{back}}}}\geq n/5,{{\mathfrak{B}}}(n)\leq n^{1-(1+2/m)\nu/\kappa}]}\\
&\leq {P_{{{\tilde\omega}}}}\Bigl[\frac{{{\mathcal T}_{\text{\itshape{back}}}}}{n^{\nu/\kappa}(\ln n)^{{\gamma}}}\geq n^{1-(1+1/m)\nu/\kappa} ,{{\mathfrak{B}}}(n)\leq n^{1-(1+2/m)\nu/\kappa}\Bigr] \\
&\leq \exp(-n^{1-(1+1/m)\nu/\kappa}/4)\\
&\leq \exp(-n^{1-(1+2/m)\nu/\kappa}),\end{aligned}$$ so we obtain $$\label{backtrackUB}
{P_{{{\tilde\omega}}}}\left[{{\mathcal T}_{\text{\itshape{back}}}}\geq n/5\right] \leq \exp(-n^{1-(1+2/m)\nu/\kappa}).$$
By Proposition \[initcrossingtime\], recalling (\[defTinit\]), we have $$\label{initUB}
{P_{{{\tilde\omega}}}}\left[{{\mathcal T}_{\text{\itshape{init}}}}\geq n/5\right] \leq \exp(-n^{1-(1+2/m)\nu/\kappa}).$$
Recalling \[defTright\], using Proposition \[rightcrossingtime\] and the fact that $\omega \in G_1(n)$, we get $$\label{rightUB}
{P_{{{\tilde\omega}}}}\left[{{\mathcal T}_{\text{\itshape{right}}}}\geq n/5\right] \leq \exp(-n^{1-(1+2/m)\nu/\kappa}).$$
Finally, using (\[decompoUB\]), (\[backtrackUB\]), (\[initUB\]), (\[rightUB\]) and Proposition \[timedirect\], we get that for all ${\varepsilon}>0$ $${P_{{{\tilde\omega}}}}\left[T_{n^{\nu}}>n\right] \leq C_3\exp(-n^{1-(1+2/m) \nu/ \kappa}).$$ Hence, letting $m$ go to $\infty$ we obtain $$\label{UboundQST}
\liminf_{n\to\infty} \frac{\ln (- \ln {P_{{{\tilde\omega}}}}\left[T_{n^{\nu}}>n\right])}{\ln n} \geq 1- \frac{\nu}{\kappa},\qquad \text{${\mathbf{P}}$-a.s.}$$
Now, we consider RWRE without reflection. All estimates remain true except (\[backtrackUB\]) for ${{\mathcal T}_{\text{\itshape{back}}}}$. Concerning the estimates on ${{\mathcal T}_{\text{\itshape{left}}}}$ it is easy to see that since $\{{{\mathcal T}_{\text{\itshape{left}}}}>0\}$ implies that ${{\mathfrak{B}}}(n)\geq n/(\ln n)^2-1$, we have using (\[boundback\]) $$\label{leftUB2}
{P_{\omega}}[{{\mathcal T}_{\text{\itshape{left}}}}\geq n/5] \leq \exp(-n^{1-(1+2/m)\nu/\kappa}).$$
It remains to estimate ${P_{\omega}}[{{\mathcal T}_{\text{\itshape{back}}}}\geq n]$, hence we take $m$ and we note that $${P_{\omega}}[{{\mathcal T}_{\text{\itshape{back}}}}>n]\leq \sum_{k=0}^{m}{P_{\omega}}[{{\mathcal T}_{\text{\itshape{back}}}}>n ,{{\mathfrak{B}}}(n)\in [n^{k/m},n^{(k+1)/m})].$$
Using (\[boundback\]), we obtain that ${\mathbf{P}}$-a.s. for $n$ large enough, $${P_{\omega}}[{{\mathcal T}_{\text{\itshape{back}}}}>n , {{\mathfrak{B}}}(n)\in [n^{k/m},n^{(k+1)/m})]\leq C_3\exp(-n^{k/m}).$$
Using Proposition \[timebacktrack1\], we obtain that $${P_{\omega}}[{{\mathcal T}_{\text{\itshape{back}}}}>n , {{\mathfrak{B}}}(n)\in [n^{k/m},n^{(k+1)/m})]\leq C_4\exp(-C_5n^{1-(\nu\vee((k+1)/m))/\kappa}).$$ Hence, with these estimates on ${{\mathcal T}_{\text{\itshape{back}}}}$, (\[decompoUB\]), (\[initUB\]), (\[leftUB2\]), (\[rightUB\]) and Proposition \[timedirect\] we obtain that ${\mathbf{P}}$-a.s. for $n$ large enough, $$\liminf_{n\to\infty} \frac{\ln( - \ln {P_{\omega}}[T_{n^{\nu}}>n])}{\ln n} \geq \min_{k \in[-1,m+1]} \Bigl(\frac{k}{m} \vee \Bigl(1-\frac{\nu \vee ((k+1)/m)}{\kappa}\Bigr)\Bigr),$$ minimizing we obtain, $$\liminf_{n\to\infty} \frac{\ln( - \ln {P_{\omega}}[T_{n^{\nu}}>n])}{\ln n} \geq \Bigl(1- \frac{\nu}{\kappa}\Bigr) \wedge \frac{\kappa}{\kappa+1} -\frac{2}{(1\wedge\kappa) m}, \qquad \text{${\mathbf{P}}$-a.s.}$$
Taking the limit as $m$ goes to infinity yields the upper bound in (\[nonrefl\_hit\_q\_sl\]), i.e., $$\label{loboT}
\liminf_{n\to\infty} \frac{\ln(-\ln {P_{\omega}}[T_{n^{\nu}}>n])}{\ln n}
\geq \Bigl(1-\frac{\nu}{\kappa}\Bigr) \wedge \frac{\kappa}{\kappa+1},
\qquad \text{${\mathbf{P}}$-a.s.}$$
Upper bound for the probability of quenched slowdown for the walk {#s_inversion_q_sl}
-----------------------------------------------------------------
The argument of this section applies for both reflected and non-reflected RWREs, for the proof in the reflected case, just replace “${P_{\omega}}$” with “${P_{{{\tilde\omega}}}}$”. We assume that $\omega \in A(n) \cap G_1(n) \cap B'(n,\nu,m)$ which is satisfied ${\mathbf{P}}$-a.s. for $n$ large enough.
Set $m\in {\mathbb{Z}}^+$, we have using Markov’s property $$\begin{aligned}
\label{decompo1}
{P_{\omega}}[X_n<n^{\nu}]\leq& \sum_{k=0}^{m}{P_{\omega}}[T_{n^{\nu+(k-1)/m}}<n]\\
& \qquad\times\max_{i\leq n} {P_{\omega}}^{n^{\nu+(k-1)/m}}[X_i<n^{\nu},~T_{n^{\nu+k/m}}>n-i].
\nonumber\end{aligned}$$
First let us notice that $$\begin{aligned}
\lefteqn{
\max_{i\leq n} {P_{\omega}}^{n^{\nu+(k-1)/m}}[X_i\leq n^{\nu},~T_{n^{\nu+k/m}}>n-i]}\nonumber\\
& \leq\Bigl( \max_{i\leq n} {P_{\omega}}^{n^{\nu+(k-1)/m}}[X_i< n^{\nu}]\Bigr)\wedge{P_{\omega}}^{n^{\nu+(k-1)/m}}[T_{n^{\nu+k/m}}>n].
\label{decompo0}\end{aligned}$$
Using reversibility we have for any $x \in {\mathbb{Z}}$ (omitting integer parts for simplicity), $${P_{\omega}}^{n^{\nu+(k-1)/m}}[X_i=x] \leq \frac{\pi(x)}{\pi(n^{\nu+(k-1)/m})},$$ hence $$\max_{i\leq n} {P_{\omega}}^{n^{\nu+(k-1)/m}}[X_i< n^{\nu}]\leq 1\wedge \frac{\pi([-n,n^{\nu}])}{\pi(n^{\nu+(k-1)/m})}.$$
Recall (\[indices\]), then by (\[pidef\]) and the definition of $b_i$ we get $\pi(b_{i_1})\leq 2 e^{-V(b_{i_1})}$ and $$\pi(b_{i_1})\leq 2 e^{-V(b_{i_1})}\leq C_6 (\ln n)^{2/\kappa} n^{1/\kappa} e^{-V(K_{i_1+1}(n))},$$ since, due to (\[Gdef\]), the increase of potential in a valley is at most $\frac 1{\kappa}(\ln n + 2\ln \ln n)$. Hence, using (\[weightbottom\]) and the fact that the width of the valleys is at most $(\ln n)^2$, we get that $$\pi([-n,n^{\nu}])\leq C_7(\ln n)^{2+2/\kappa}n^{1/\kappa} e^{-V(K_{i_1+1}(n))}.$$ Furthermore, denoting by $i_2$ the index of the valley containing $n^{\nu+(k-1)/m}$, for $n$ large enough we have using (\[pidef\]) $$\pi(n^{\nu+(k-1)/m}) \geq \pi(K_{i_2-1}(n)),$$ since on the event $G(n)$ both $V(K_{i_2-1}(n))$ and $V(K_{i_2-1}(n)-1)$ are bigger than $V(n^{\nu+(k-1)/m})$ and $V(n^{\nu+(k-1)/m}-1)$.
On $A(n)$, we have ${\left\vert (i_2-1)-i_1\right\vert}\geq {\left\vert n^{\nu+(k-1){\varepsilon}}-n^{\nu}\right\vert}/(\ln n)^2-2$. Since $V(K_i)-V(K_{i+1}) \geq 1/(1\wedge \kappa) \ln n$ for $\omega \in G_1(n)$, we have for $k\geq 2$ $$\begin{aligned}
\label{decompo2}
\frac{\pi([-n,n^{\nu}])}{\pi(n^{\nu+(k-1)/m})} &\leq C_8 (\ln n)^{2+2/\kappa} n^{1/\kappa} \exp(-(V(K_{i_1+1})-V(K_{i_2-1}))) \\ \nonumber
&\leq C_9 (\ln n)^{2+2/\kappa} n^{1/\kappa} \exp\Bigl(-C_{10} \frac{ n^{\nu+(k-1)/m}-n^{\nu}}{(\ln n)^2}\Bigr).\end{aligned}$$
Moreover, using (\[nonrefl\_hit\_q\_sl\]) in the non-reflected case (or (\[UboundQST\]) in the reflected case), we have $${P_{\omega}}^{n^{\nu+(k-1)/m}}[T_{n^{\nu+k/m}}>n]\leq \exp(-n^{(1-(\nu+(k/m))/\kappa)\wedge (\kappa/(\kappa+1)) -1/m}).$$ Hence, using this last inequality and (\[decompo2\]), the inequality (\[decompo1\]) becomes $$\begin{aligned}
\lefteqn{{P_{\omega}}[X_n<n^{\nu}]}\\
&\leq \max_{k\in[-1, m+1]} \Bigl[ 1\wedge \Bigl[C_9m n^{1/\kappa} (\ln n)^{2+2/\kappa} \exp\Bigl(-C_{10} \frac{n^{\nu+(k-1)/m}-n^{\nu}}{(\ln n)^2}\Bigr)\Bigr]\\
&\qquad\qquad\qquad \wedge \exp(-n^{(1-(\nu+(k/m))/\kappa)\wedge (\kappa/(\kappa+1)) -1/m})\Bigr],\end{aligned}$$ so that ${\bf P}$-a.s., $$\begin{aligned}
\liminf_{n\to\infty} \frac{\ln (-\ln {P_{\omega}}[X_n<n^{\nu}])}{\ln n}
& \geq \min_{k\in[-1,m+1]}\Biggl[ \Bigl({\mathbf 1}\Bigl\{\frac{k-1}m\geq 0\Bigr\}\Bigl(\nu+\frac{k-1} m\Bigr)\Bigr)\\
& \qquad\qquad\vee \Bigl(\Bigl(1-\frac{\nu+ k/m}{\kappa}\Bigr) \wedge \frac{\kappa}{\kappa+1} -\frac 1m\Bigr)\Biggr].\end{aligned}$$
Minimizing over $k$, we obtain $$\liminf_{n\to\infty} \frac{\ln (-\ln {P_{\omega}}[X_n<n^{\nu}])}{\ln n} \geq \Bigl(1-\frac{\nu}{\kappa}\Bigr)\wedge \frac{\kappa}{\kappa+1}-\frac{1}{m}, \qquad \text{${\mathbf{P}}$-a.s.}$$ Letting $m$ goes to infinity, we obtain $$\label{upper_q_slowdown}
\liminf_{n\to\infty} \frac{\ln (-\ln {P_{\omega}}[X_n<n^{\nu}])}{\ln n} \geq \Bigl(1-\frac{\nu}{\kappa}\Bigr)\wedge \frac{\kappa}{\kappa+1},
\qquad \text{${\mathbf{P}}$-a.s.}$$
Lower bound for quenched slowdown {#s_lower_q_sl}
---------------------------------
In this section we assume $\omega \in A(n) \cap D(n) \cap F(n)$ which is satisfied ${\mathbf{P}}$-a.s. for $n$ large enough. First, we consider RWRE with reflection at the origin.
For all ${\varepsilon}>0$, note that for $n$ large enough there is a valley of depth at least $\frac{(1-{\varepsilon}) \nu}{\kappa} \ln n$ strictly before level $n^\nu$ and denote by $i_2$ the index of the first such valley. Hence $${P_{{{\tilde\omega}}}}[T_{n^{\nu}}>n] \geq {P_{{{\tilde\omega}}}}^{\tilde K_{i_2}}[T_{\tilde K_{i_2+1}+1}>n],$$ and using Proposition \[Lboundconf\] we obtain $${P_{{{\tilde\omega}}}}^{\tilde K_{i_2}}[T_{\tilde K_{i_2+1}+1}>n] \geq \exp(-n^{1-(1-{\varepsilon})\nu/\kappa+{\varepsilon}}).$$ Letting ${\varepsilon}$ go to $0$, yields $$\label{LboundQST}
\limsup_{n\to\infty} \frac{\ln(-\ln {P_{{{\tilde\omega}}}}[T_{n^{\nu}}>n])}{\ln n} \leq 1- \frac{\nu}{\kappa}.$$ This yields the lower bound for the exit time, so, recalling (\[UboundQST\]), we obtain (\[refl\_hit\_q\_sl\]).
Now let us deduce the results on the slowdown. Set $a\in[0,\kappa-\nu)$, for $n$ large enough there is a valley of depth $(\nu+(1-{\varepsilon})a)/\kappa \ln n$ strictly before $n^{\nu+a}$ whose index is denoted $i_3$. One possible strategy for the walk is to enter the $i_2$-th valley at $\tilde K_{i_2}+1\leq n^{\nu+a}$, stay there up to time $n-(n^{\nu+a}-n^{\nu})- (\ln n)^2$, then go to the left up to time $n$. The probability of this event can be bounded from below by $$\begin{aligned}
{P_{{{\tilde\omega}}}}[X_n< n^{\nu}] \geq& {P_{{{\tilde\omega}}}}\left[ T_{n^{\nu+a}}<n/2\right]\min_{j\leq n} {P_{{{\tilde\omega}}}}^{\widetilde K_{i_3}+1}\left[T_{\{\widetilde K_{i_3}-1,\widetilde K_{i_3+1}+1\}} >j\right]\\ &\times n^{-(3/{\varepsilon}_0)(n^{\nu+a}-n^{\nu}+(\ln n)^2)}.\end{aligned}$$ The first term is bigger than $1/2$ for $n$ large enough (one can see this by using e.g. (\[LboundQST\])). The second can be bounded by Proposition \[Lboundconf\] $$\min_{j\leq n} {P_{{{\tilde\omega}}}}^{\widetilde K_{i_3}+1}\left[T_{\{\widetilde K_{i_3}-1,\widetilde K_{i_3+1}+1\}} >j\right] \geq \exp(-n^{1-(\nu+(1-{\varepsilon})a)/\kappa+{\varepsilon}}),$$ for $n$ large enough. Then, the last term (going left) was dealt with using the fact that $\omega \in F(n)$.
This yields for any $a\geq 0$, $$\limsup_{n\to\infty} \frac{\ln (-\ln {P_{{{\tilde\omega}}}}[X_n< n^{\nu}])}{\ln n} \leq {{\mathbf 1}{\{a>0\}}}(\nu+a) \vee \Bigl(1-(1-{\varepsilon})\frac{\nu+a}{\kappa}+{\varepsilon}\Bigr),$$ and if we choose $a=0\vee(\kappa/(\kappa+1)-\nu)$, we obtain $$\limsup_{n\to\infty} \frac{\ln (-\ln {P_{{{\tilde\omega}}}}[X_n< n^{\nu}])}{\ln n} \leq \Bigl(1-\frac{\nu}{\kappa}\Bigr) \wedge \frac{\kappa}{\kappa+1} +\frac{2{\varepsilon}}{\kappa}+{\varepsilon},\qquad \text{${\mathbf{P}}$-a.s.}$$ Together with (\[upper\_q\_slowdown\]), this yields (\[refl\_loc\_q\_sl\]) by letting ${\varepsilon}$ go to $0$.
Now, we consider the case of RWRE without reflection. Using the same reasoning, we write $$\label{part_res_q_sl1}
\limsup_{n\to\infty} \frac{\ln(-\ln {P_{\omega}}[T_{n^{\nu}}>n])}{\ln n} \leq 1- \frac{\nu}{\kappa},\qquad \text{${\mathbf{P}}$-a.s.}$$
Now we can see that, if we denote by $i_4$ the index of a valley of depth at least $(1-{\varepsilon})/(\kappa+1)\ln n$ between $-n^{\kappa/(\kappa+1)}$ and 0, since we are on $D(n)$, we can go to this valley before reaching $n^{\nu}$ and then stay there for a time at least $n$. This yields, $${P_{\omega}}[T_{n^{\nu}}>n] \geq {P_{\omega}}[T_{-n^{\kappa/(\kappa+1)}}<T_{n^{\nu}}] {P_{\omega}}^{K_{i_4}}[T_{K_{i_4+1}+1}>n],$$ bounding the first term by the probability of going to the left on the $n^{\kappa/(\kappa+1)}$ first steps, we get using Proposition \[Lboundconf\] that for all $n$ large enough $${P_{\omega}}^0[T_{n^{\nu}}>n] \geq n^{-(3/{\varepsilon}_0)n^{\kappa/(\kappa+1)}} \exp(-n^{1-(1-2{\varepsilon})/(\kappa+1)}),$$ and hence $$\label{part_res_q_sl2}
\limsup_{n\to\infty} \frac{\ln(-\ln {P_{\omega}}^0[T_{n^{\nu}}>n])}{\ln n} \leq \frac{\kappa}{\kappa+1}+2\frac{{\varepsilon}}{\kappa+1},\qquad \text{${\mathbf{P}}$-a.s.}$$ Moreover, it is clear that $$\label{X<->T}
{P_{\omega}}[X_n<n^{\nu}] \geq {P_{\omega}}[T_{n^{\nu}} > n],$$ and letting ${\varepsilon}$ go to $0$ in (\[part\_res\_q\_sl2\]) and using (\[part\_res\_q\_sl1\]) and (\[loboT\]), we obtain (\[nonrefl\_hit\_q\_sl\]) and (\[nonrefl\_loc\_q\_sl\]). This finishes the proof of Theorem \[t\_q\_slow\].
Annealed slowdown {#s_a_sl}
=================
Lower bound for annealed slowdown {#s_lower_a_sl}
---------------------------------
Let us define the events $$A'(n,\nu,a)=\bigl\{\text{there exists }x \in [-n^{\nu},n^{\nu}]:\ \max_{y\in [x,n^{\nu}]} V(y)-V(x) \geq (1+a)\ln n\bigr\},$$ and $$A_+'(n,\nu,a)=\bigl\{\text{there exists }x \in [0,n^{\nu}]:\ \max_{y\in [x,n^{\nu}]} V(y)-V(x) \geq (1+a)\ln n\bigr\}.$$
\[A2\] We have for $a\in (-1,1)$, $$\lim_{n\to\infty}\frac{\ln {\mathbf{P}}[A'(n,\nu,a)]}{\ln n}= \lim_{n\to\infty}\frac{\ln {\mathbf{P}}[A_+'(n,\nu,a)]}{\ln n} = -(\kappa- \nu)-a\kappa.$$
From (\[fellerthm\]), it is straightforward to obtain that $$\begin{aligned}
{\mathbf{P}}[A_+'(n,\nu,a)]&\leq {\mathbf{P}}[A'(n,\nu,a)] \\
&\leq 2n^{\nu} {\mathbf{P}}\Bigl[ \max_{i \geq 0} V(i) \geq (1+a) \ln n\Bigr]\\ &=\Theta(n^{\nu-(1+a)\kappa}).\end{aligned}$$
In order to give the corresponding lower bound, let us define the event $$A_1(n,a)=\bigl\{\text{there exists }k \in[0,(\ln n)^2] \text{ such that } V(k) \geq (1+a)\ln n \bigr\},$$ we have $$\begin{aligned}
{\mathbf{P}}[A_1(n,a)]
\geq& {\mathbf{P}}\Bigl[ \max_{i \geq 0} V(i) \geq (1+a)\ln n\Bigr] - {\mathbf{P}}[V(\ln n)^2>- \ln n] \\
&-{\mathbf{P}}\Bigl[\max_{i \geq (\ln n)^2} V(i)-V((\ln n)^2) > (2+a) \ln n\Bigr]
\nonumber\\
=& \Theta( n^{-(1+a)\kappa}),\end{aligned}$$ where we used (\[fellerthm\]) and a reasoning similar to the proof of Lemma \[widthvalley\]. Now, we write $${\mathbf{P}}[A'(n,\nu,a)] \geq {\mathbf{P}}[A_+'(n,\nu,a)]\geq \frac {n^{\nu}} {\lfloor (\ln n)^2\rfloor } {\mathbf{P}}[A_1(n)] =\Theta\Bigl( \frac{n^{\nu-(1+a)\kappa}}{(\ln n)^2}\Bigr),$$ and Lemma \[A2\] follows.
For any ${\varepsilon}>0$, on the event $A_+'(n,\nu,{\varepsilon})$ there exists a valley $[K_i, K_{i+1}]$ with $V(K_{i+1})- V(b_i) \geq (1+{\varepsilon})\ln n$ contained in $[0, n^{\nu})$ and we denote by $i_5$ its index. Then we have by Proposition \[cost\_to\_climb\] $${P_{\omega}}[T_{n^{\nu}}>n] \geq P_{\omega}^{b_{i_5}}[T_{K_{i_5+1}+1}>n]\geq 1-\gamma_2(1+n)e^{-(1+{\varepsilon})\ln n}\geq \frac 12$$ for $n$ large enough. So $${\mathbb{P}}[T_{n^{\nu}}>n] \geq E[{{\mathbf 1}{\{A_+'(n,\nu,{\varepsilon})\}}}{P_{\omega}}[T_{n^{\nu}}>n]]\geq \frac 12 {\mathbf{P}}[A_+'(n,\nu,{\varepsilon})].$$ Hence we obtain by Lemma \[A2\] that for any ${\varepsilon}>0$ $$\liminf_{n\to\infty} \frac{\ln {\mathbb{P}}[T_{n^{\nu}}>n]}{\ln n} \geq -(\nu-\kappa) -\kappa {\varepsilon}.$$ Using (\[X<->T\]), we obtain the corresponding lower bound for ${\mathbb{P}}[X_n<n^{\nu}]$ as well. Replacing ${P_{\omega}}$ by ${P_{{{\tilde\omega}}}}$ and ${\mathbb{P}}$ by ${{\tilde{\mathbb{P}}}}$, exactly the same argument can be used to obtain the result in the reflected case.
Upper bound for annealed slowdown {#s_upper_a_sl}
---------------------------------
We prove the upper bound in the non-reflected case, the reflected case follows easily; indeed a simple coupling argument shows that $T_{n^{\nu}}$ in the environment ${{\tilde\omega}}$ is stochastically dominated by $T_{n^{\nu}}$ in the environment $\omega$. For $m \in {\mathbb{N}}$ such that $1/m\in (0,\nu)$, we have $${\mathbb{P}}[T_{n^{\nu}}>n] \leq {\mathbf{P}}[A'(n,\nu,-1/m)]+{\mathbf{E}}\big({{\mathbf 1}{\{A'(n,\nu,-1/m)^c\}}}{P_{\omega}}^0[T_{n^{\nu}}>n]\big).$$ The second term can be further bounded by $$\begin{aligned}
\lefteqn{{\mathbf{E}}\big({{\mathbf 1}{\{A'(n,\nu,-1/m)^c\}}}{P_{\omega}}^0[T_{n^{\nu}}>n]\big)}\\
& \leq {\mathbf{P}}[A(n)^c \cup B'(n,\nu,m)^c] \\
& \qquad +{\mathbf{E}}\big({{\mathbf 1}{\{A'(n,\nu,-1/m)^c \cap A(n)\cap B'(n,\nu,m)\}}}{P_{\omega}}^0[T_{n^{\nu}}>n]\big),\end{aligned}$$ where $B'(n,\nu,m)$ is defined in (\[def\_b\_prime\]).
Using Lemma \[A2\] we have that $1/n=o({\mathbf{P}}[A'(n,\nu,-1/m)])$, and thus Lemma \[widthvalley\] and Lemma \[depthvalley1\] imply that $${\mathbf{P}}[A(n)^c \cup B'(n,\nu,m)^c]=o({\mathbf{P}}[A'(n,\nu,-1/m)]).$$
We can turn (\[domstoch\_1\]) into the following, for $i \in N(-n^{{\varepsilon}},n^{\nu})$ we have $$\text{on $A'(n,\nu,-1/m)^c \cap A(n) \cap B'(n,\nu,m)$,} \qquad \frac{Z}{C_8n^{(1-1/m)}(\ln n)^{{\gamma}}}\prec 1 + \mathbf{e},$$ where $Z$ has the same law as $T_{K_{i+1}(n)}$ under ${P_{\omega}}^{K_i(n)}[~\cdot\mid T_{K_{i+1}(n)}<T_{K_{i-1}(n)}]$; ${\gamma}={\gamma}(\kappa)$ and $\mathbf{e}$ denotes an exponential random variable of parameter $1$. The same inequality is true when $K_{i-1}(n)$ and $K_{i+1}(n)$ are exchanged.
This stochastic domination is the key argument for Section \[s\_upper\_q\_slowdown\]. We can adapt the proof of Proposition \[timebacktrack\], so that on $A'(n,\nu,-1/m)^c \cap A(n) \cap B'(n,\nu,m)$ we obtain for all $u \geq 1$, $${P_{\omega}}\Bigl[\frac{{{\mathcal T}_{\text{\itshape{back}}}}}{{n^{1-1/m}(\ln n)^{{\gamma}}}} \geq \exp(n^{1/(2m)}) , {{\mathfrak{B}}}(n) \leq n^{1/(4m)} \Bigr]\leq e^{-n^{1/(2m)}/4},$$ and $${P_{\omega}}\Big[{{\mathcal T}_{\text{\itshape{right}}}}> \frac n5\Big]\leq C_1 \exp(-n^{1/(4m)}).$$
Moreover, (\[boundback\]) still holds, so that $${P_{\omega}}[{{\mathfrak{B}}}(n) \geq n^{1/(4m)} ]\leq C_2 \exp(-n^{1/(4m)}),$$ which yields $${P_{\omega}}\Big[{{\mathcal T}_{\text{\itshape{left}}}}> \frac n5\Big]\leq C_3 \exp(-n^{1/(4m)}).$$
Finally, recalling (\[defTinit\]) and using Proposition \[spectral\] on $A'(n,\nu,-1/m)^c \cap A(n)$, we obtain $${P_{\omega}}\Big[{{\mathcal T}_{\text{\itshape{init}}}}> \frac n5\Big]\leq C_4 \exp(-n^{1/(4m)}).$$
Since Proposition \[timedirect\] remains true and $A'(n,\nu,-1/m)^c \subset G(n)$, we get that for all $\omega \in A'(n,\nu,-1/m)^c \cap A(n) \cap B'(n,\nu,m)$ $${P_{\omega}}[T_{n^{\nu}}>n] \leq C_5\exp(-n^{1/(4m)}).$$ Loosely speaking it costs at least $\exp(-n^{1/(2m)})$ to backtrack $n^{1/m}$ times, hence, on $A'(n,\nu,-1/m)^c \cap A(n) \cap B'(n,\nu,m)$, we can only see valleys of size lower than $(1-1/m) \ln n$. To spend a time $n$ in those valleys would cost at least $\exp(-n^{1/(2m)})$. This finally implies that for all $m>0$, $$\limsup_{n\to \infty} \frac{\ln {\mathbf{E}}\Bigl[{{\mathbf 1}{\{A'(n,\nu,-1/m)^c,A(n)^c,B'(n,\nu,m)^c\}}}{P_{\omega}}^0[T_{n^{\nu}}>n]\Bigr]}{\ln n} = 0,$$ so that $$\label{UBAS}
\limsup_{n\to\infty} \frac{\ln {\mathbb{P}}[T_{n^{\nu}}>n]}{\ln n} \leq -(\kappa- \nu) +\frac{\kappa}m,$$ the result for the hitting time follows by letting $m$ go to infinity.
It is simple to extend this result to the position of the walk, indeed if $X_n<n^{\nu}$ then $T_{n^{(1+1/m)\nu}}>n$ or ${{\mathfrak{B}}}(n) \geq n^{1/(2m)}$ and hence using (\[boundback\]) , we get for all $m>0$ $${\mathbb{P}}[X_n<n^{\nu}]\leq {\mathbb{P}}[T_{n^{(1+1/m)\nu}}>n]+C_6e^{-n^{1/(2m)}},$$ and the result follows by using (\[UBAS\]) and letting $m$ go to infinity.
This concludes the proof of Theorem \[t\_a\_slow\].
Backtracking {#s_backtracking}
============
In this section we prove Theorem \[t\_backtrack\].
Quenched backtracking for the hitting time {#s_q_backtracking_t}
------------------------------------------
Set $\nu \in (0,1)$ and consider ${P_{\omega}}[T_{-n^{\nu}}<n]$. First, we get that $$\text{for all $\omega\in F(n)$,} \qquad {P_{\omega}}[T_{-n^{\nu}}<n]\geq n^{-(3/{\varepsilon}_0)n^{\nu}},$$ since the particle can go straight to the left during the first $n^{\nu}$ steps, hence $$\label{backtrack_limsup}
\limsup_{n\to \infty} \frac{\ln(-\ln {P_{\omega}}[T_{-n^{\nu}}<n])}{\ln n} \leq \nu.$$
Secondly, we remark that if $(-\infty,-n^{\nu}]$ has been hit before time $n$ then, at some time $i\leq n$ the particle is at $X_i\in [-n, -n^{\nu}]$ and hence for all $\omega$ $$\begin{aligned}
{P_{\omega}}[T_{-n^{\nu}}<n]&\leq \sum_{i=1}^n {P_{\omega}}[X_i \in [-n,-n^{\nu}] ] \nonumber\\
&\leq n \max_{i\leq n}{P_{\omega}}[X_i \in [-n,-n^{\nu}] ].
\label{back_pi}\end{aligned}$$
In order to estimate this quantity, we use arguments similar to those in Section \[s\_inversion\_q\_sl\], i.e., first we use the reversibility of the walk to write $$\max_{i\leq n}{P_{\omega}}[X_i \in [-n,-n^{\nu}]]\leq \frac{\pi([-n,-n^{\nu}])}{\pi (0)},$$ then, the right-hand side can be estimated in the same way as we obtained (\[decompo2\]), and so we get on $A(n)\cap G_1(n)$ that $$\frac{\pi([-n,-n^{\nu}])}{\pi (0)}\leq C_1 (\ln n)^{2+2/\kappa} n^{1/\kappa}\exp(-C_2 n^{\nu}/(\ln n)^2).$$
The previous inequality and (\[back\_pi\]) yield $$\text{for all $\omega\in A(n)\cap G(n)$,} \qquad {P_{\omega}}[T_{-n^{\nu}}<n]\leq C_3 n^{1+2/\kappa}\exp(-C_2 n^{\nu}/(\ln n)^2),$$ so that $$\liminf_{n\to \infty} \frac{\ln(-\ln {P_{\omega}}[T_{-n^{\nu}}<n])}{\ln n} \geq \nu.$$
Together with (\[backtrack\_limsup\]), this proves (\[backtrack\_T\]).
Quenched backtracking for the position of the random walk {#s_q_backtracking_rw}
---------------------------------------------------------
Let us denote $a_0=\frac{\kappa}{\kappa+1} \vee \nu$. We give a lower bound for ${P_{\omega}}[X_n<-n^{\nu}]$. For $n$ large enough, there exists ${\bf P}$-a.s. a valley of depth $(1-{\varepsilon})(a_0/\kappa)\ln n$ of index $i_2$, between $-n^{a_0}$ and $0$. Consider the event that the walker goes to this valley directly and stays there up to time $n-n^{a_0}$ and then goes to the left for the next $n^{a_0}+1$ steps. On this event we have $X_n< -n^{a_0}$, so we obtain $$\begin{aligned}
{P_{\omega}}[X_n<-n^{\nu}] &\geq n^{-(3/{\varepsilon}_0)2(n^{a_0}+1)} {P_{\omega}}^{K_{i_2+1}-1}[T_{\{K_{i_2}-1,K_{i_2+1}+1\}} \geq n]\\
& \geq n^{-(3/{\varepsilon}_0)2(n^{a_0}+1)}\exp(-n^{1-(1-2{\varepsilon})a_0/\kappa}),\end{aligned}$$ where we used Proposition \[Lboundconf\] and $\omega\in F(n)$. Hence we obtain $$\limsup_{n\to \infty} \frac{\ln(-\ln{P_{\omega}}[X_n<-n^{\nu}])}{\ln n} \leq a_0 +\frac{2{\varepsilon}a_0}{\kappa},$$ and letting ${\varepsilon}$ go to 0 we have $$\label{limsup_q_back_rw}
\limsup_{n\to \infty} \frac{\ln(-\ln{P_{\omega}}[X_n<-n^{\nu}])}{\ln n} \leq a_0.$$
Turning to the upper bound, we have for $m\in {\mathbb{N}}$, $$\label{back_step1}
{P_{\omega}}[X_n<-n^{\nu}]\leq \sum_{k=0}^{m}{P_{\omega}}[T_{n^{(k-1)/m}}<n] \max_{i\leq n} {P_{\omega}}^{n^{(k-1)/m}}[T_{n^{k/m}}>n-i,X_i< -n^{\nu}],$$ where once again $$\begin{aligned}
\lefteqn{
\max_{i\leq n} {P_{\omega}}^{n^{(k-1)/m}}[T_{n^{k/m}}>n-i,X_i< -n^{\nu}]}\\
& \leq \Bigl(\max_{i\leq n}
{P_{\omega}}^{n^{(k-1)/m}}[X_i< -n^{\nu}]\Bigr)\wedge{P_{\omega}}^{n^{(k-1)/m}}[T_{n^{k/m}}>n] .\end{aligned}$$
First, using (\[nonrefl\_hit\_q\_sl\]), for $n$ large enough $$\label{back_step2}
{P_{\omega}}^{n^{(k-1)/m}}[T_{n^{k/m}}>n ] \leq \exp(-n^{(1-(k/m)/\kappa)\wedge (\kappa/(\kappa+1))-1/m}).$$
Then, as in Section \[s\_inversion\_q\_sl\], the reversibility of the walk yields that $$\label{back_step3}
\max_{i\leq n}{P_{\omega}}^{n^{(k-1)/m}}[X_i \in [-n,-n^{\nu}]]\leq \frac{\pi([-n,-n^{\nu}])}{\pi (n^{(k-1)/m})},$$ the right-hand side can be estimated in the same way we obtained (\[decompo2\]) and we get on $A(n)\cap G(n)$ $$\label{back_step4}
\frac{\pi([-n,-n^{\nu}])}{\pi (n^{(k-1)/m})}\leq C_4 \exp(-C_5 (n^{(k-1)/m}+n^{\nu})/(\ln n)^2).$$
Putting together , , , and , we obtain $$\begin{aligned}
\lefteqn{\liminf_{n\to \infty} \frac{\ln(-\ln{P_{\omega}}[X_n<-n^{\nu}])}{\ln n}}\\
& \geq
\min_{k\in[0,m]} \Bigl(\Bigl(\Bigl(1-\frac{k}{m\kappa} \Bigr)\wedge \frac{\kappa}{\kappa+1}\Bigr)\vee \Bigl(\frac{k-1}m \vee \nu \Bigr) \Bigr)-\frac 1m,\end{aligned}$$ minimizing yields that $$\liminf_{n\to \infty} \frac{\ln(-\ln{P_{\omega}}[X_n<-n^{\nu}])}{\ln n} \geq a_0 -\frac 2{m},$$ letting $m$ go to infinity and recalling we obtain .
Annealed backtracking {#s_a_backtracking}
---------------------
Let $\theta_0={\mathbf{E}}\left[\ln\rho_0\right]<0$. Define $${{\mathcal R}}= \Bigl\{\omega: V(x)\leq \frac{\theta_0}{3}n^{\nu} \text{ for }x\in [0,n],
|V(x)+\theta_0x|\leq \frac{|\theta_0|}{3}n^\nu
\text{ for }x\in[-n^\nu,0)\Bigr\}.$$ Since $V$ is a sum of i.i.d. random variables having some finite exponential moments, we can use large deviations techniques to obtain $C_6$ such that $$\label{bound_PP_R}
{\mathbf{P}}[{{\mathcal R}}] \geq 1-2ne^{-C_6n^\nu}.$$
Then, on ${{\mathcal R}}$, using (\[exit\_probs\]), we obtain $$\begin{aligned}
{P_{\omega}}[T_{-n^\nu}<n] &\leq {P_{\omega}}[T_{-n^\nu}<T_n] \nonumber\\
&\leq C_7n \exp\Bigl(-\frac{2\theta_0}{3}n^\nu\Bigr). \label{Po_on_RR}\end{aligned}$$ Using (\[bound\_PP\_R\]) and (\[Po\_on\_RR\]), we obtain $$\label{b*}
{\mathbb{P}}[X_n<-n^\nu] \leq {\mathbb{P}}[T_{-n^\nu}<n]\leq e^{-C_8 n^\nu}.$$
On the other hand, we easily obtain that $$\label{b**}
{\mathbb{P}}[T_{-n^{\nu}}<n]\geq {\mathbb{P}}[X_n<-n^\nu] \geq \Bigl(\frac{{{\delta}}}2\Bigl)^{n^\nu} n^{-C_9},$$ where ${{\delta}}>0$ is such that ${\mathbf{P}}[1-\omega_0 \geq {{\delta}}]>1/2$. Indeed on the event of probability at least $(1/2)^{n^{\nu}}$ that $1-\omega_x\geq {{\delta}}$ for $x\in (-n^{\nu},0]$, the particle can go “directly” (to the left on each step) to $(-n^\nu)$, and then the cost of creating a valley of depth $2\ln n$ there is polynomial and then it costs nothing to stay there for a time $n$ by Proposition \[cost\_to\_climb\]. Now, (\[b\*\]) and (\[b\*\*\]) imply (\[a\_backtrack\]). This finishes the proof of Theorem \[t\_backtrack\].
Speedup {#s_speedup}
=======
In this section we prove Theorem \[t\_speedup\]. So, we have $\kappa<1$, $\nu\in (\kappa,1)$; let us denote $g(\alpha)=\nu+\frac{\alpha}{\kappa}-\alpha$, and let $\alpha_0=\kappa\frac{1-\nu}{1-\kappa}$. Clearly, $g(\alpha)$ is a linear function, $g(0)=\nu<1$, $g(\nu)=\frac{\nu}{\kappa}>1$, and $g(\alpha_0)=1$; note also that $\nu-\alpha_0=\frac{\nu-\kappa}{1-\kappa}$.
The discussion in this section is for the RWRE on ${\mathbb{Z}}$ (i.e., without reflection), the proof for the reflected case is quite analogous.
Lower bound for the quenched probability of speedup {#s_lower_q_speedup}
---------------------------------------------------
We are going to obtain a lower bound for ${P_{\omega}}[X_n> n^\nu]$.
By Lemma \[depthvalley1\] and Borel-Cantelli, for any fixed $m$, $\omega\in B'(n,\alpha_0,m)\cap A(n)\cap F(n)$ for all $n$ large enough, ${\mathbf{P}}$-a.s. (recall the definition of $A(n)$ and $B'(n,\alpha_0,m)$ from Section \[s\_estimates\_env\]). So, from now on we suppose that $\omega\in B'(n,\alpha_0,m)\cap A(n)$.
Let us denote $M=N_n(0,n^{\nu})$, define the index sets $$\begin{aligned}
{{\mathcal I}}_0 &= \{i\in M : H_{i-1}\vee H_i \leq \ln\ln n\},\\
{{\mathcal I}}_k &= \Bigl\{i\in M : (H_{i-1}\vee H_i) - \ln\ln n \in \Bigl[\frac{(k-1)\alpha_0}{m\kappa}\ln n,
\frac{k\alpha_0}{m\kappa}\ln n\Bigr)\Bigr\}\end{aligned}$$ for $k\in[1,m-1]$, and $$\begin{aligned}
{{\mathcal U}}&= \Bigl\{i\in M : H_{i-1}\vee H_i \geq \frac{(m-1)\alpha_0}{m\kappa}\ln n + \ln\ln n\Bigr\}.\end{aligned}$$ Note that on $B'(n,\alpha_0,m)$ $$\begin{aligned}
{\mathop{\mathrm{card}}}{{\mathcal U}}&\leq n^{\nu-\alpha_0 + \frac{\alpha_0}{m}}
= n^{\frac{\nu-\kappa}{1-\kappa} + \frac{\alpha_0}{m}}, \label{razmer_U}\\
{\mathop{\mathrm{card}}}{{\mathcal I}}_k &\leq n^{\nu-\frac{k\alpha_0}{m}}, \qquad \text{ for all } k=1,\ldots,m-1.\label{razmer_I_k}\end{aligned}$$
Recalling (\[indices\]) we define the quantities $\sigma_{i_0}=T_{K_{i_0+1}}$, $\sigma_{i_1}=T_{n^\nu}-T_{K_{i_1}}$, and $\sigma_j=T_{K_{j+1}}-T_{K_j}$ for $j=i_0+1,\ldots,i_1-1$. Then for ${\varepsilon}>0$, we can write $$\begin{aligned}
{P_{\omega}}[X_n> n^{(1-{\varepsilon})\nu}] \geq & {P_{\omega}}\Bigl[\sum_{k=0}^{m-1}\sum_{i\in{{\mathcal I}}_k}\sigma_i\leq \frac{n}{2}\Bigr]
{P_{\omega}}\Bigl[\sum_{i\in{{\mathcal U}}}\sigma_i\leq \frac{n}{2}\Bigr]\nonumber\\
& {}\times {P_{\omega}}^{n^\nu}\bigl[X_j> n^{(1-{\varepsilon})\nu} \text{ for all }
i\in[0,n-n^\nu]\bigr].\label{decompos_low_P_speed}\end{aligned}$$
Let us obtain lower bounds for the three terms in the right-hand side of (\[decompos\_low\_P\_speed\]). First, we write using (\[razmer\_I\_k\]) $$\begin{aligned}
{P_{\omega}}\Bigl[\sum_{k=0}^{m-1}\sum_{i\in{{\mathcal I}}_k}\sigma_i\leq \frac{n}{2}\Bigr] &\geq
\prod_{k=0}^{m-1} {P_{\omega}}\Bigl[\sum_{i\in{{\mathcal I}}_k}\sigma_i\leq \frac{n}{2m}\Bigr]\nonumber\\
&\geq \prod_{k=0}^{m-1} {P_{\omega}}\Bigl[\sigma_i\leq \frac{1}{2m}n^{1-(\nu-\frac{k\alpha_0}{m})}
\text{ for all }i\in{{\mathcal I}}_k\Bigr]. \label{coisa*}\end{aligned}$$ Now, consider any $\ell\in{{\mathcal I}}_k$ and write $$\begin{aligned}
\lefteqn{{P_{\omega}}\Bigl[\sigma_\ell\leq \frac{1}{2m}n^{1-(\nu-\frac{k\alpha_0}{m})} \Bigr]}\qquad\\
&\geq
{P_{\omega}}^{K_\ell}[T_{K_{\ell+1}}<T_{K_{\ell-1}}] \\ & \quad\times
{P_{\omega}}^{K_\ell}\Bigl[T_{\{K_{\ell-1},K_{\ell+1}\}}\leq \frac{1}{2m}n^{1-(\nu-\frac{k\alpha_0}{m})} \mid T_{K_{l+1}}<T_{K_{l-1}}\Bigr].\end{aligned}$$
By the formula (\[crossingprob\]), on $A(n)$ we have $${P_{\omega}}^{K_\ell}[T_{K_{\ell+1}}<T_{K_{\ell-1}}] \geq 1 - n^{-3/2},$$ and by Proposition \[crossingtime\], $$\begin{aligned}
{P_{\omega}}^{K_\ell}\Bigl[T_{\{K_{\ell-1},K_{\ell+1}\}}\leq \frac{1}{2m}n^{1-(\nu-\frac{k\alpha_0}{m})}\mid T_{K_{l+1}}<T_{K_{l-1}}\Bigr]
\\ \geq 1-\exp\Bigl(-\frac{C_1}{m(\ln n)^{{\gamma}}}n^{1-(\nu-\frac{k\alpha_0}{m})-\frac{k\alpha_0}{m\kappa}}\Bigr),\end{aligned}$$ so $$\label{coisa**}
{P_{\omega}}\Bigl[\sigma_\ell\leq \frac{1}{2m}n^{1-(\nu-\frac{k\alpha_0}{m})} \Bigr] \geq
(1 - n^{-3/2})
\Bigl(1-\exp\Bigl(-\frac{C_1}{m(\ln n)^{{\gamma}}}
n^{1-g(\frac{k\alpha_0}{m})}\Bigr)\Bigr).$$
Now, for $k\leq m-1$ we have $$1-g\Bigl(\frac{k\alpha_0}{m}\Bigr)\geq \frac{(1-\kappa)\alpha_0}{m\kappa},$$ so (\[coisa\*\]) and (\[coisa\*\*\]) imply that $$\begin{aligned}
{P_{\omega}}\Bigl[\sum_{k=0}^{m-1}\sum_{i\in{{\mathcal I}}_k}\sigma_i\leq \frac{n}{2}\Bigr] &\geq
\prod_{k=0}^{m-1} \Biggl[ (1 - n^{-3/2})\Bigr)
\Bigl(1-\exp\Bigl(-\frac{C_1}{m(\ln n)^{{\gamma}}}
n^{\frac{(1-\kappa)\alpha_0}{m\kappa}}\Bigr)\Bigr) \Biggr]^{n^\nu}\nonumber\\
&\to 1 \qquad \text{as }n\to\infty. \label{ocenka_term1}\end{aligned}$$
Now, we obtain a lower bound for the second term in the right-hand side of (\[decompos\_low\_P\_speed\]). On $G_1(n)$, we get an upper bound on $\rho_i$ for $i\in[-n,n]$ and hence we have $\omega_x \geq n^{-C_2}$, we obtain for any $\ell\in{{\mathcal U}}$ (imagine that, to cross the corresponding interval, the particle just goes to the right at each step) $${P_{\omega}}\Bigl[\sigma_\ell \leq \frac{1}{2}n^{1-(\nu-\alpha_0)-\frac{\alpha_0}{m}}\Bigr]
\geq n^{-C_2 (\ln n)^2},$$ so, $$\begin{aligned}
{P_{\omega}}\Bigl[\sum_{i\in{{\mathcal U}}}\sigma_i\leq \frac{n}{2}\Bigr] &\geq
{P_{\omega}}\Bigl[\sigma_\ell \leq \frac{1}{2}n^{1-(\nu-\alpha_0)-\frac{\alpha_0}{m}}
\text{ for all }\ell\in{{\mathcal U}}\Bigr] \nonumber\\
&\geq \bigl(n^{-C_2(\ln n)^2}\bigr)^{n^{\frac{\nu-\kappa}{1-\kappa}+\frac{\alpha_0}{m}}}\nonumber\\
& = \exp\Bigl(- C_2(\ln n)^3 n^{\frac{\nu-\kappa}{1-\kappa}+\frac{\alpha_0}{m}}\Bigr)
\label{ocenka_term2}\end{aligned}$$ (recall that $\nu-\alpha_0 = \frac{\nu-\kappa}{1-\kappa}$).
As for the third term in (\[decompos\_low\_P\_speed\]), using (\[exit\_probs\]) we easily obtain that, on $A(n)\cap G(n)$, $$\label{ocenka_term3}
{P_{\omega}}^{n^\nu}\bigl[X_j> n^{(1-{\varepsilon})\nu} \text{ for all } j\in[0,n-n^\nu]\bigr] \geq {P_{\omega}}^{n^\nu}[T_n<T_{n^{(1-{\varepsilon})\nu}}]>C_3>0.$$
Now, plugging (\[ocenka\_term1\]), (\[ocenka\_term2\]), and (\[ocenka\_term3\]) into (\[decompos\_low\_P\_speed\]) and sending $m$ to $\infty$, we obtain that $$\limsup_{n\to\infty} \frac{\ln(-\ln {P_{\omega}}[X_n> n^{(1-{\varepsilon})\nu}])}{\ln n} \leq \frac{\nu-\kappa}{1-\kappa}, \qquad \text{${\mathbf{P}}$-a.s.}$$ applying this for $\nu'=\nu/(1-{\varepsilon})$ and letting ${\varepsilon}$ go to 0, $$\label{lower_b_sp}
\limsup_{n\to\infty} \frac{\ln(-\ln {P_{\omega}}[X_n> n^{\nu'}])}{\ln n} \leq \frac{\nu'-\kappa}{1-\kappa}, \qquad \text{${\mathbf{P}}$-a.s.}$$ Since obviously ${P_{\omega}}[T_{n^\nu}< n]\geq{P_{\omega}}[X_n> n^\nu]$, (\[lower\_b\_sp\]) holds for ${P_{\omega}}[T_{n^\nu}< n]$ as well.
Upper bound for the quenched probability of speedup {#s_upper_q_speedup}
---------------------------------------------------
Fix ${\varepsilon}>0$ such that $\alpha_0+{\varepsilon}<\nu$. Define $$\begin{aligned}
{{\mathcal W}}&= \Bigl\{i\in N_n(0,n^\nu) : H_i\geq \frac{\alpha_0+{\varepsilon}}{\kappa}\ln n - 4\ln\ln n\Bigr\},\\
\Psi_n^{\varepsilon}&= \Bigl\{\omega: {\mathop{\mathrm{card}}}{{\mathcal W}}\geq \frac{1}{3} n^{\nu-\alpha_0-{\varepsilon}}\Bigr\}. \end{aligned}$$ By Lemma \[depthvalley2\], on each subinterval of length $n^{\alpha_0+{\varepsilon}}$ we find a valley of depth at least $\frac{\alpha_0+{\varepsilon}}{\kappa}\ln n - 4\ln\ln n$ with probability at least $1/2$. Since the interval $[0,n^\nu]$ contains $n^{\nu-\alpha_0-{\varepsilon}}$ such subintervals, we have $$\label{PP_Theta}
{\mathbf{P}}[\Psi_n^{\varepsilon}] \geq 1 - \exp(-C_4n^{\nu-\alpha_0-{\varepsilon}}),$$ in particular by Borel-Cantelli’s Lemma, ${\mathbf{P}}$-a.s. we have $\omega \in \Psi_n^{\varepsilon}$ for $n$ large enough.
For $i\in{{\mathcal W}}$, define ${{\tilde\sigma}}_i=T_{K_{i+1}+1}-T_{K_i+1}$, and let $$s_0 = \frac{1}{4{\gamma}_2(\ln n)^4} n^{\frac{\alpha_0+{\varepsilon}}{\kappa}}.$$ Then, by Proposition \[cost\_to\_climb\], for any $i\in{{\mathcal W}}$, $$\begin{aligned}
{P_{\omega}}[{{\tilde\sigma}}_i<s_0] &\leq 2{\gamma}_2 s_0 \exp\Bigl(-\frac{\alpha_0+{\varepsilon}}{\kappa}\ln n + 4\ln\ln n\Bigr)
\nonumber\\
&= 2{\gamma}_2 s_0n^{-\frac{\alpha_0+{\varepsilon}}{\kappa}}(\ln n)^4 \nonumber\\
&= \frac{1}{2}. \label{ocenka_climb}\end{aligned}$$
Define the family of random variables $\zeta_i={{\mathbf 1}{\{{{\tilde\sigma}}_i<s_0\}}}$, $i\in{{\mathcal W}}$. These random variables are independent with respect to ${P_{\omega}}$, and ${P_{\omega}}[\zeta_i=1]\leq 1/2$ by (\[ocenka\_climb\]). Suppose without restriction of generality that (recall that $g(\alpha_0)=1$) $$\frac{1}{3}s_0 \times \frac{1}{3}n^{\nu-\alpha_0-{\varepsilon}}
= \frac{1}{36{\gamma}_2 (\ln n)^4} n^{g(\alpha_0+{\varepsilon})} > n.$$ Then, since ${\mathop{\mathrm{card}}}{{\mathcal W}}\geq \frac{1}{3}n^{\nu-\alpha_0-{\varepsilon}}$ for $\omega \in \Psi_n^{\varepsilon}$, we see using large deviations techniques that for $n$ large enough $$\begin{aligned}
{P_{\omega}}[T_{n^\nu}<n] &\leq {P_{\omega}}\Bigl[\sum_{i\in{{\mathcal W}}}\zeta_i > \frac{2}{3} {\mathop{\mathrm{card}}}{{\mathcal W}}\Bigr]\nonumber\\
&\leq \exp\bigl(-C_5 n^{\frac{\nu-\kappa}{1-\kappa}-{\varepsilon}}\bigr) \label{oc_up_q_sp}\end{aligned}$$ (recall that $\nu-\alpha_0 = \frac{\nu-\kappa}{1-\kappa}$). Since ${\varepsilon}>0$ is arbitrary, we obtain $$\label{upper_b_sp}
\liminf_{n\to\infty} \frac{\ln(-\ln {P_{\omega}}[T_{n^\nu}<n])}{\ln n} \geq \frac{\nu-\kappa}{1-\kappa} \qquad \text{${\mathbf{P}}$-a.s.}$$ Together with (\[lower\_b\_sp\]), this shows (\[eq\_q\_speed\]).
Annealed speedup {#s_a_speedup}
----------------
As usual, the quenched lower bound obtained in Section \[s\_lower\_q\_speedup\] also yields the annealed one, i.e. (\[lower\_b\_sp\]) implies that $$\limsup_{n\to\infty} \frac{\ln(-\ln {\mathbb{P}}[X_n> n^\nu])}{\ln n} \leq \frac{\nu-\kappa}{1-\kappa},$$ Turning to the upper bound, we have by (\[PP\_Theta\]) and (\[oc\_up\_q\_sp\]) that $$\begin{aligned}
{\mathbb{P}}[T_{n^\nu}<n] &= \int {P_{\omega}}[T_{n^\nu}<n] \, d{\mathbf{P}}\\
&\leq \int_{\Psi_n^{\varepsilon}} {P_{\omega}}[T_{n^\nu}<n] \, d{\mathbf{P}}+ {\mathbf{P}}[(\Psi_n^{\varepsilon})^c]\\
&\leq \exp\bigl(-C_5 n^{\frac{\nu-\kappa}{1-\kappa}-{\varepsilon}}\bigr)
+ \exp\bigl(-C_4 n^{\frac{\nu-\kappa}{1-\kappa}-{\varepsilon}}\bigr),\end{aligned}$$ and this implies (\[eq\_a\_speed\]). This finishes the proof of Theorem \[t\_speedup\].
Acknowledgements {#acknowledgements .unnumbered}
================
A.F. would like to thank the ANR MEMEMO, the Accord France-Brésiland the ARCUS program.
S.P. is thankful to FAPESP (04/07276–2), CNPq (300328/2005–2 and 471925/2006–3), and N.G. and S. P. are thankful to CAPES/DAAD (Probral) for financial support.
We thank two anonymous referees whose extremely careful lecture of the first version lead to many improvements.
[19]{} Comets, F., Gantert, N. and Zeitouni, O. (2000). Quenched, annealed and functional large deviations for one-dimensional random walk in random environment. [*Probab. Theory Relat. Fields*]{} [**118**]{}, 65–114.
Comets, F. and Popov, S. (2003). Limit law for transition probabilities and moderate deviations for Sinai’s random walk in random environment. [*Probab. Theory Relat. Fields*]{} [**126**]{} (4), 571–609.
Comets, F. and Popov, S. (2004). A note on quenched moderate deviations for Sinai’s random walk in random environment. [*ESAIM: Probab. Statist.*]{} [**8**]{}, 56–65.
Enriquez, N., Sabot, C. and Zindy, O. (2007). Limit laws for transient random walks in random environment on ${\mathbb{Z}}$. arXiv:math/0703648.
Feller, W. (1971). [*An Introduction to Probability Theory and its Applications,*]{} Vol. II. (2nd ed.). Wiley, New York.
Gantert, N. and Shi, Z. (2002). Many visits to a single site for a transient random walk in random environment. [*Stochastic Process. Appl.*]{} [**99**]{}, 159–176.
Hu, Y. and Shi, Z. (2004). Moderate deviations for diffusions with Brownian potentials. [*Ann. Probab.*]{} [**32**]{} (4), 3191–3220.
Iglehart, D.L. (1972). Extreme values in the GI/G/$1$ queue. [*Ann. Math. Statist.*]{} [**43**]{}, 627–635.
Kesten, K., Kozlov, M.V. and Spitzer, F. (1975). A limit law for random walk in a random environment. [*Compositio Math*]{}. [**30**]{}, 145–168.
Miclo, L. (1999). An example of application of discrete Hardy’s inequalities. [*Markov Process. Relat. Fields*]{} [**5**]{}, 319–330.
Peterson, J. and Zeitouni, O. (2007) Quenched limits for transient, zero speed one-dimensional random walk in random environment. arXiv:0704.1778
Saloff-Coste, L. (1997). [*Lectures on finite Markov chains*]{}. Volume [**1665**]{}. Ecole d’Eté de Probabilités de Saint-Flour, P. Bernard, (ed.), Lectures Notes in Mathematics, Berlin: Springer.
Sinai, Ya.G. (1982). The limiting behavior of a one-dimensional random walk in a random medium. [*Theory Probab. Appl.*]{} [**27**]{}, 256–268.
Solomon, F. (1975). Random walks in a random environment. [*Ann. Probab.*]{} [**3**]{}, 1–31.
Spitzer, F. (1976). [*Principles of Random Walk*]{}. 2nd Ed., Springer-Verlag, New-York, 1976.
Zeitouni, O. (2004). [*Random Walks in Random Environment*]{}, XXXI summer school in probability, St Flour (2001), [*Lecture Notes in Math.*]{} [**1837**]{}, p.193–312. Springer, Berlin.
[^1]: apparently, this result is folklore, at least we were unable to find a precise reference in the literature. Anyhow, note that it is straightforward to obtain this result from Theorems \[t\_q\_slow\] and \[t\_speedup\]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
A phonon and charge transport simulation based on the Geant4 Monte Carlo toolkit is presented. The transport code is capable of propagating acoustic phonons, electrons and holes in cryogenic crystals. Anisotropic phonon propagation, oblique carrier propagation and phonon emission by accelerated carriers are all taken into account. The simulation successfully reproduces theoretical predictions and experimental observations such as phonon caustics, heat pulse propagation times and mean carrier drift velocities.
Implementation of the transport code using the Geant4 toolkit ensures availability to the wider scientific community.
author:
- 'D. Brandt'
- 'R. Agnese'
- 'P. Redl'
- 'K. Schneck'
- 'M. Asai'
- 'M. Kelsey'
- 'D. Faiez'
- 'E. Bagli'
- 'B. Cabrera'
- 'R. Partridge'
- 'T. Saab'
- 'B. Sadoulet'
title: Semiconductor phonon and charge transport Monte Carlo simulation using Geant4
---
Geant4, Phonon, Electron/Hole transport, Cryogenic detectors
Introduction {#sec:Introduction}
============
We present a Monte Carlo simulation of phonon and charge transport in semiconductor crystals using Geant4. Geant4 is a sophisticated C++ based Monte Carlo simulation toolkit maintained by an international collaboration and freely available under an open source license [@Geant-A; @Geant-B]. The toolkit was originally developed in support of High Energy Physics (HEP) experiments and provides a framework for simulating the passage of particles through complex geometries and materials. It aims to accurately simulate all particle-matter interactions and has become an important tool both for HEP particle accelerator based experiments , and experiments wishing to estimate backgrounds induced by comic rays and from radiogenic sources [@CDMS-D; @Brandt].
In its current incarnation, the Geant4 toolkit is entirely focused on free particles and does not take into account crystal physics or conduction/valence band interactions of the low-energy charge carriers and phonons relevant to condensed-matter physics. This paper documents our effort to build a cohesive Geant4 Condensed Matter Physics Monte Carlo simulation toolkit, G4CMP. The original purpose of this project was to accurately reproduce data from the Cryogenic Dark Matter Search (CDMS) [@CDMS-A; @CDMS-B; @CDMS-C], a dark-matter direct-detection experiment. The CDMS detectors are cylindrical Ge crystals approximately $75$ mm in diameter with a height of $25$ mm [@CDMS-E], cooled to approximately $60$ mK. Dark-matter particles may recoil from Ge nuclei and thus create phonons and electron-hole pairs within the crystal [@Lindhart]. Electron-hole pairs drift in a few V/cm field to the crystal faces where they are collected. Phonons are detected by superconducting Transition Edge Sensors (TES) [@TES]. We reproduce all of these processes in our simulation. The phonon and charge transport code described below models several physics processes relevant to phonon and charge collection at cryogenic temperatures. This includes anisotropic phonon transport and focusing, phonon isotope scattering, anharmonic downconversion, oblique charge carrier propagation with inter-valley scattering, and emission of Luke-Neganov phonons by accelerated carriers. The resulting G4CMP framework is sufficiently general that it should be useful to other experiments employing cryogenic phonon and/or ionization detectors.
Phonon Transport {#sec:PhononTransport}
================
Phonon transport was the first component of the G4CMP framework to be developed [@Brandt]. Since the phonon transport code described here is intended for temperatures $T<1 K$, scattering off thermally excited background phonons is ignored. Currently, only acoustic phonons are simulated, with optical phonons not supported.
Anisotropic transport and phonon focusing {#sec:Focusing}
-----------------------------------------
Phonons are quantized vibrations of the crystal lattice. The propagation of phonons is governed by the three-dimensional wave equation [@Wolfe]: $$\label{eq:3DWave}
\rho \omega ^2e_i=C_{ijml}k_jk_me_l,$$ where $\rho$ is the crystal mass density, $\omega$ is the angular phonon frequency, $\vec{e}$ is the polarization vector, $\vec{k}$ is a wave vector and $C_{ijml}$ is the elasticity tensor.
For any given wave vector, $\vec{k}$, Eq. \[eq:3DWave\] has three eigenvalues, $\omega$, and three eigenvectors, $\vec{e}$. These correspond to the three different polarization states: Longitudinal (L), Fast Transverse (FT) and Slow Transverse (ST). The actual direction and velocity of propagation of phonons is given by the group velocity vector $\vec{v}_g = d\omega/dk$. The group velocity can be calculated by interpreting $\omega$ in Eq. \[eq:3DWave\] as a function of $\vec{k}$:
$$\label{eq:GroupV}
\vec{v}_g=\nabla \omega (\vec{k}),$$
Due to the anisotropy in $C_{ijml}$, Eq. \[eq:GroupV\] yields a group velocity $\vec{v}_g$ that is not parallel to the phonon momentum $\hbar\vec{k}$. Instead, phonons are focused onto propagation directions that correspond to the highest density of eigenvectors $\vec{k}$. This focusing gives rise to caustics when observing the energy distribution resulting from a point-like phonon source that is isotropic in $\vec{k}$-space. The resulting caustics can be observed using micro-calorimeters [@Nothrop]. Figure \[fig:caustics\] shows that the caustics simulated by the Geant4 phonon transport code are in good agreement with experimental observations.
For the purposes of the G4CMP phonon transport code, the wave equation is not solved in real time. Instead, a look-up table is generated which maps $\vec{k}$ onto $\vec{v}_g$, and bilinear interpolation is used to generate a continuous mapping function. Phonon focusing and methods for solving the three-dimensional wave equations are treated in [@Wolfe].
![**Left:** outline of phonon caustics in Ge as predicted by Nothrop and Wolfe [@Nothrop]. **Right:** Phonon caustics as simulated using the Geant4 phonon transport code. This result is in good agreement with both the theoretical prediction and experimental observations reported by Nothrop and Wolfe [@Nothrop]. []{data-label="fig:caustics"}](caustics.png){width="50.00000%"}
Phonon processes {#sec:Processes}
----------------
In addition to the phonon equation of motion, Eq. \[eq:3DWave\], two processes are relevant to acoustic phonon transport in cryogenic crystals: isotope scattering and anharmonic downconversion [@Tamura1; @Tamura2; @Tamura3]. The scattering and downconversion rates for phonons in a cryogenic crystal are given by [@Brandt; @Tamura2]: $$\label{eq:anhRate}
\Gamma_{\mathrm{anh}} = A\nu^5,$$
$$\label{eq:ScatterRate}
\Gamma_{\mathrm{scatter}} = B\nu^4,$$
where $\Gamma_{\mathrm{anh}}$ is the number of anharmonic downconversion events per unit time, $\Gamma_{\mathrm{scatter}}$ is the number of scattering events per unit time, $\nu$ is the phonon frequency and $A$, $B$ are constants of proportionality related to the elasticity tensor. In Ge, $A = 6.43 \times 10^{-55} s^4$ and $B = 3.67 \times 10^{-41} s^3$ [@Brandt; @Tamura1].
Isotope scattering occurs when a phonon interacts with an isotopic substitution site in the lattice. It is effectively an elastic scattering during which the phonon momentum vector is randomized, and the phonon polarization state can change freely between the three states $L$, $ST$, $FT$. The partition between the three polarizations is determined by the relative density of allowed states for each polarization. This change between polarization states is often referred to as *mode mixing*.
Anharmonic downconversion causes a single phonon to decay into two phonons of reduced energy. This process conserves energy but not momentum, since momentum is exchanged with the crystal lattice. In theory all three polarization states can decay; however, the downconversion rate of $L$-phonons completely dominates the energy evolution of the phonon system, with downconversion events from other polarization states being negligible [@Tamura2].
Equations \[eq:anhRate\] and \[eq:ScatterRate\] indicate that both process rates strongly depend on phonon energy $\hbar \nu$. High-energy phonons ($\nu$ of order THz) start out in a diffusive regime with high isotope scattering and downconversion rates and mean free paths of order microns. Once a few downconversion events have occurred, phonon mean free paths increase to be of order $\sim0.1$ m (the size of a typical CDMS detector). This transition from a diffuse to a ballistic transport mode is commonly referred to as “quasi-diffuse” and controls the time evolution of phonon pulses. Simulation of phonon pulses with our Geant4 transport code as described in [@Brandt] shows good agreement with experiment. Anharmonic downconversion and isotope scattering are well understood and are discussed in great detail in the literature [@Tamura1; @Tamura2; @Wolfe; @Tamura3].
Charge Transport {#sec:ChargeTransport}
================
Conclusion {#sec:Conclusion}
==========
We have presented a new Monte Carlo transport code capable of propagating phonons on a cryogenic crystal lattice (Section \[sec:PhononTransport\]) as well as drifting electron-hole pairs, taking into account conduction band anisotropy (Section \[sec:ChargeTransport\]). The results produced by the transport code are in good agreement with experiment, reproducing phonon caustics and carrier drift velocity with acceptable accuracy. It was shown that this code reproduces heat-pulse propagation and dispersion in cryogenic Ge crystals with acceptable accuracy in [@Brandt]. The entire transport code is written in C++ for Geant4 and is easily adaptable for crystals other than Ge, provided that the Herring-Vogt contracted elasticity tensor and effective carrier masses are known. The Geant4 framework makes it possible to extend the code presented here. The phonon transport code is already freely available as part of the examples provided with Geant4 v9.6p02 and newer. We hope that the work presented here will establish Geant4 as a tool in condensed-matter physics and cryogenic calorimeter design as well as motivate others to add to the G4CMP framework. We would like to thank members of the CDMS Detector Monte Carlo Group for discussions that have made completing this work possible. Specifically we want to thank P.L. Brink, A. Anderson and C. Schlupf. Furthermore we would like to thank A. Phipps and K. Sundqvist for useful discussions and providing us with experimental data. S. Yellin and R. Bunkers’ on point feedback on earlier paper drafts have helped us improve the final paper tremendously. This work is supported in part by the National Science Foundation and by the United States Department of Energy. SLAC is operated under Contract No. DE-AC02-76SF00515 with the United States Department of Energy.
[99]{}
K. D. Irwin, S. W. Nam, B. Cabrera, B. Chugg and B. Young, [*Rev. Sci. Instrum.*]{} **66**, 5322, (1995).
J. Allison et al., [*IEEE Transactions on Nuclear Science*]{} **53**, 270 - 278, (2006).
S. Agostinelli et al., [*Nuclear Instruments and Methods A*]{} **506**, 250 - 303, (2003).
D.S. Akerib et al., [*Phys. Rev. D*]{} **72**, (2005).
D. Brandt et al., [*Journal of Low Temperature Physics*]{} **167**, 485 - 490, (2011).
Smith, P.F. and Lewin, J.D., [*Phys.Rept.*]{} **187**, 203, (1990).
Z. Ahmed et al., [*Science*]{} **327**, 1619, (2010).
Z. Ahmed et al., [*Phys. Rev. D*]{} **84**, 011102, (2011).
C. Collaboration et al., [*arXiv astro-ph*]{} **1304.4279**, (2013).
R. Agnese et al., [*arXiv astro-ph*]{} **1305.2405**, (2013).
P. Brink et al., [*NIM A*]{} **2**, 414, (2006).
G.A. Nothrop and J.P. Wolfe, [*Phys. Rev. Lett.*]{} **19**, 1424, (1979).
J.P. Wolfe, [*Imaging Phonons, Chapter 2*]{},42, Cambridge University Press, United Kingdom (1998).
S. Tamura, [*Journal of Low Temperature Physics*]{} **93**, 433, (1993).
S. Tamura, [*Phys. Rev. B.*]{} **48**, 13502, (1993).
S. Tamura,[*Phys. Rev. B.*]{} **31**, (1985).
S. W. Leman, [*Rev. Sci. Instrum.*]{} **83**, 091101, (2012).
Cabrera, B. et al., [*arXiv astro-ph*]{} **1004.1233**, (2010).
Broniatowski, A., [*Journal of Low Temperature Physics*]{}, **167**, (2012).
Juillard, A., [*Journal of Low Temperature Physics*]{}, **151**, (2008).
A. Phipps, K. M. Sundqvist, A. Lam, B. Sadoulet, [*Journal of Low Temperature Physics*]{}, **167**, (2012). V. Aubry-Fortuna, [*AIP Conference Series*]{}, **1185** 639, (2009).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that finding the lowest eigenvalue of a 3-local symmetric stochastic matrix is QMA-complete. We also show that finding the highest energy of a stoquastic Hamiltonian is QMA-complete and that adiabatic quantum computation using certain excited states of a stoquastic Hamiltonian is universal. We also show that adiabatic evolution in the ground state of a stochastic frustration free Hamiltonian is universal. Our results give a new QMA-complete problem arising in the classical setting of Markov chains, and new adiabatically universal Hamiltonians that arise in many physical systems.'
author:
- 'Stephen P. Jordan'
- David Gosset
- 'Peter J. Love'
bibliography:
- 'stoquastic.bib'
title: 'QMA-complete problems for stoquastic Hamiltonians and Markov matrices'
---
Introduction {#sec_intro}
============
Quantum complexity theory is the study of the capabilities and limitations of computational devices operating according to the principles of quantum mechanics [@QCT]. Because many of the classical constructs of computer science (e.g. circuits and clauses) are replaced by matrices, quantum complexity theory is sometimes referred to as [*matrix-valued*]{} complexity theory [@Bravyi2; @Bravyi_Terhal]. In addition to its intrinsic interest this subject has many connections to issues of practical relevance to physical science, such as the difficulty of computing properties of quantum systems using either quantum or classical devices [@SchuchVer2009; @LiuVer2007; @Bravyi1].
Perhaps the most basic classical complexity classes are P - the class of problems solved by a deterministic Turing machine in polynomial time, and NP - the class of problems whose verification lies in P. It is widely believed, but not proven, that NP is strictly larger than P [@Sipser].
Because quantum mechanics only predicts probabilities of events, the classical deterministic classes are not the most natural place to start if one seeks their quantum generalizations. The probabilistic generalization of P is BPP (Bounded-error Probabilistic Polynomial-time) - those problems solvable by a probabilistic Turing machine in polynomial time with bounded error [@BPP]. The quantum generalization of this class is BQP (Bounded-error Quantum Polynomial-time) - the class of problems solvable in polynomial time with bounded error on a quantum computer [@QCT].
The classical probabilistic generalization of NP is the class MA [@MAAM]. This generalizes NP to problems whose verification is in BPP. MA stands for Merlin-Arthur. Merlin, who is computationally unbounded but untrustworthy, provides a proof that Arthur can verify using his BPP machine. The class MA possesses a quantum generalization to QMA (Quantum Merlin Arthur) [@WatrousQMA; @Kitaev_book; @QNP]. QMA may be intuitively understood as the class of decision problems that can be efficiently verified by a quantum computer.
Given a classical description of a decision problem $x$ of length $n$, the prover, Merlin, provides a witness state ${|\psi\rangle}$ to the verifier, Arthur. Arthur then peforms a $\mathrm{poly}(n)$-time quantum computation on the witness ${|\psi\rangle}$ and either accepts or rejects. A problem is contained in QMA if, for all YES instances, there exists a witness causing Arthur to accept with probability greater than $2/3$ and for NO instances, there does not exist any witness that causes Arthur to accept with probability greater than $1/3$. A problem $X$ is said to be QMA-complete if it is contained in QMA and every problem in QMA can be converted to an instance of $X$ in classical polynomial time.
Let us consider the following question: What is the ground state energy of a quantum system? This question lies at the core of many areas of physical science, including electronic structure theory and condensed matter physics. In quantum complexity theory this problem has been formalized (originally by Kitaev [@Kitaev_book], see also, for example, [@Kempe]) as the $k$-local Hamiltonian problem. For some systems, complexity-theoretic arguments suggest that efficient computation of the ground state energy is likely to remain beyond reach [@barahona; @Kitaev_book].
A Hamiltonian $H$, acting on $n$ qubits, is said to be $k$-local if it is of the form $$H=\sum_{s}H_{s},$$ where each $H_{s}$ acts on at most $k$-qubits. Thus, for example, 1-local Hamiltonians consist only of external fields acting on individual qubits, and 2-local Hamiltonians consist of 1-local terms and pairwise couplings between qubits. Physically realistic Hamiltonians are usually $k$-local with small $k$, often $2$ or $1$, and each local term has bounded norm. Note that this notion of locality has nothing to do with spatial locality; a 2-local Hamiltonian may have long-range couplings but they must be pairwise.\
\
**Problem:** $k$-local Hamiltonian\
**Input:** We are given a classical description of a $k$-local Hamiltonian $H$ on $n$ qubits $H=\sum_{j=1}^{r}H_{j}$ with $r=\mathrm{poly}(n)$. Each $H_{j}$ acts on at most $k$ qubits and has $O(1)$ operator norm. In addition we are given two constants $a$ and $b$ such that $0\leq a\leq b$, and $b-a=\epsilon>1/\mathrm{poly}(n)$.\
**Output:** If $H$ has an eigenvalue $\leq a$ answer YES. If all eigenvalues of $H$ are $>b$ answer NO.\
**Promise:** the Hamiltonian is such that it will produce either YES or NO.\
\
Perhaps a more obvious formulation of this problem is to ask for an approximate ground state energy to within $\pm\epsilon$ of the correct answer. However, if one can decide the answer to $k$-local Hamiltonian in polynomial time, then one can solve the approximation version in polynomial time by a binary search. Thus, the approximation problem is of equivalent difficulty to the “decision” version, to within a polynomial factor.
The problem $k$-local Hamiltonian is QMA-complete for $k\geq2$ [@Kempe]. The $k$-local Hamiltonian problem is specified by the matrix elements of the local terms of $H$. YES instances possess the ground state as a witness. The verification circuit is the phase estimation algorithm - a suitably formalized version of the notion of energy measurement [@Kitaev_book]. If the lowest eigenvalue of $H$ is less than $a$ (a YES instance) then Arthur will accept the ground state as a witness. However, if the lowest eigenvalue of $H$ is greater than $b$ (a NO instance) then Merlin cannot supply any eigenstate or superposition of eigenstates that will result in a measurement of energy less than $b$.
It is considered unlikely that $\mathrm{QMA} \subseteq \mathrm{BQP}$ and therefore it is probably impossible to construct a general quantum (or classical) algorithm that finds ground state energies in polynomial time. However, many Hamiltonians studied in practice have additional restrictions beyond $k$-locality. In particular, many physical systems are *stoquastic*, meaning that all of their off-diagonal matrix elements are nonpositive in the standard basis. This includes the ferromagnetic Heisenberg model, the quantum transverse Ising model, and most Hamiltonians achievable with Josephson-junction flux qubits [@Bravyi1]. In [@Bravyi1] it was shown that for any fixed $k$, stoquastic $k$-local Hamiltonian is contained in the complexity class AM. Thus, unless $\mathrm{QMA}\subseteq\mathrm{AM}$ (which is believed to be unlikely), stoquastic $k$-local Hamiltonian is not QMA-complete[^1]. It was also shown in [@Bravyi1] that, for any fixed $k$, adiabatic quantum computation in the ground state of a $k$-local stoquastic Hamiltonian can be simulated in $\mathrm{BPP}_{\mathrm{path}}$. Thus, unless $\mathrm{BQP}\subseteq\mathrm{BPP}_{\mathrm{path}}$ (which is also believed to be unlikely), such quantum computation is not universal. The work of [@Bravyi2] also defines a random stoquastic local Hamiltonian problem which is complete for the class AM.
These results were tightened further for stoquastic frustration free (SFF) Hamiltonians in [@Bravyi_Terhal]. A local Hamiltonian is frustration free if it can be written as a sum of terms $$H=\sum_{s}^{m}H_{s},$$ such that
1. [Each local operator $H_{s}$ is positive semidefinite]{}
2. [The ground state $|\psi\rangle$ of H satisfies $H_{s}|\psi\rangle=0$ for each $s\in\{1,\dots,m\}$]{}
The work of [@Bravyi_Terhal] showed that an adiabatic evolution along a path composed entirely of SFF Hamiltonians may be simulated by a sequence of classical random walks - that is, the adiabatic evolution may be simulated in the complexity class BPP.
These results were extended to the quantum $k$-satisfiability problem in [@Bravyi2; @Bravyi_Terhal]. The quantum $k$-satisfiability problem was defined in [@Bravyi4] and we reproduce the definition here:\
\
**Problem:** Quantum $k$-SAT\
**Input:** A set of $k$-local projectors $\{\Pi_{q}\}$ for $q\in\{1,\dots,m\}$ where $m=\mathrm{poly}(n)$ and a parameter $\tilde{\epsilon}>1/\mathrm{poly}(n)$\
**Output:** If there is a state ${|\phi\rangle}$ such that $\Pi_{1}{|\phi\rangle}=0$ for each $q\in\{1,\dots,M\}$, then this is a YES instance. If every state ${|\phi\rangle}$ satisfies $$\sum_{q=1}^{M}{\langle \phi|}\Pi_{q}{|\phi\rangle}\geq\tilde{\epsilon}$$ then it is a NO instance.\
**Promise:** The instance is either YES or NO.\
\
In [@Bravyi2] the stoquastic restriction of quantum $k$-SAT was shown to be contained in MA for any constant $k$, and MA-complete for $k=6$ – the first nontrivial example of an MA-complete problem. In [@Bravyi_Terhal] these results were extended to a simplified form of stoquastic quantum $k$-SAT in which projectors $\Pi_{a}$ all have matrix elements taken from the set $\{0,1/2,1\}$, and the stoquastic constraints which appear as terms in the Hamiltonian are of the form $H_{a}={\mathds{1}}-\Pi_{a}$.
The main intuition behind these results is that, by the Perron-Frobenius theorem, the ground state of a stoquastic Hamiltonian consists entirely of real positive amplitudes (given the appropriate choice of global phase). Thus the ground state is proportional to a classical probability distribution. For this reason, ground state properties are amenable to classical random walk algorithms and certain problems such as stoquastic $k$-local Hamiltonian fall into classical probabilistic complexity classes such as AM. Diffusion Quantum Monte Carlo calculations for stoquastic Hamiltonians do not suffer from the sign problem because the negativity of the nonzero off-diagonal matrix elements guarantees that the transition probabilities in the associated random walk are all positive.
In this paper we first demonstrate that stoquastic Hamiltonians may be constructed which allow universal adiabatic quantum computation in a subspace. Then we show that the $3$-local Hamiltonian problem is QMA-complete when restricted to stochastic Hamiltonians. These are Hamiltonians in which all matrix elements are real and nonnegative, and the sum of matrix elements in any row or column is one. Hence determining the lowest eigenstate of a symmetric stochastic matrix is QMA-hard. If $H$ is a stochastic Hamiltonian, then $-H$ is stoquastic. Thus, our result also shows that determination of the highest lying eigenstate of a stoquastic matrix is QMA-hard, sharpening the intuition that it is the positivity of the ground state which causes its local Hamiltonian problem to fall in a classical class. We then show that universal adiabatic quantum computation is possible in the ground state of a stochastic frustration free Hamiltonian. Defining the computational problem stochastic $k$-SAT in analogy to the definition of stoquastic $k$-SAT given in [@Bravyi_Terhal], we show that this problem is QMA$_{1}$-complete for $k=6$. (QMA$_1$ is a slight variant of QMA such that in YES instances, Arthur can be made to accept with probability one[@Bravyi4].)
QMA-completeness and adiabatic universality of stoquastic Hamiltonians {#QMA_section}
======================================================================
We start with the result of [@Biamonte], which shows that for a Hamiltonian of the form $$\label{xzform}
H_{XZ} = \sum_{i}d_{i}X_{i} + \sum_{i}h_{i}Z_{i} +
\sum_{i,j}K_{ij}X_{i}X_{j} + \sum_{i,j}J_{ij}Z_{i}Z_{j},$$ the 2-local Hamiltonian problem is QMA-complete if the coefficients $d_{i}$, $h_{i}$, $K_{ij}$, $J_{ij}$ are allowed to have both signs. Furthermore, time-dependent Hamilonians that take the form $H_{XZ}$ at all times can perform universal adiabatic quantum computation [@Biamonte].
Starting with a Hamiltonian of the form $H_{XZ}$ on $n$ qubits we can eliminate the negative matrix elements in each term using a technique from [@Wocjan_walks]. Essentially, the idea is that instead of representing the group $Z_{2}$ by $\{1,-1\}$ we use its regular representation: $$\left\{ \left[\begin{array}{cc}
1 & 0\\
0 & 1\end{array}\right],\left[\begin{array}{cc}
0 & 1\\
1 & 0\end{array}\right]\right\} .$$ $H_{XZ}$ can be rewritten as $$H_{XZ}=-\sum_{k}\alpha_{k}T_{k}\label{alpha}$$ Where each coefficient $\alpha_{k}$ is positive and for each $k$, $T_{k}$ is one of $$\pm X,\pm Z,\pm X_{i}X_{j},\pm Z_{i}Z_{j}\label{possibilities}$$ with identity acting on the remaining qubits. For any $k$, $T_{k}$ is a $2^{n}\times2^{n}$ matrix in which each entry is either +1,-1, or 0. From $T_{k}$ we construct a $2^{n+1}\times2^{n+1}$ matrix $\widetilde{T}_{k}$ by making the following replacements $$1\to\left[\begin{array}{cc}
1 & 0\\
0 & 1\end{array}\right],\ -1 \to \left[\begin{array}{cc}
0 & 1\\
1 & 0\end{array}\right],\ 0 \to \left[\begin{array}{cc}
0 & 0\\
0 & 0\end{array}\right].\label{replacement}$$ We can interpret $\widetilde{T}_{k}$ as acting on $n+1$ qubits. The $2\times2$ matrices of (\[replacement\]) act on the ancilla qubit that has been added. Each $T_{k}$ is 2-local or 1-local, thus each corresponding $\widetilde{T}_{k}$ is 3-local or 2-local. Furthermore, each $\widetilde{T}_{k}$ is a permutation matrix. Let $$\widetilde{H}_{XZ}=-\sum_{k}\alpha_{k}\widetilde{T}_{k}.$$ This is a linear combination of permutation matrices with negative coefficients. By construction, $\widetilde{H}_{XZ}$ is therefore a 3-local stoquastic Hamiltonian. We can rewrite $\widetilde{H}_{XZ}$ as $$\widetilde{H}_{XZ}=H_{XZ}\otimes {| - \rangle \langle - |} - \bar{H}_{XZ} \otimes
{| + \rangle \langle + |} \label{nicestoq}$$ where $$\bar{H}_{XZ}=\sum_{k}\alpha_{k}|T_{k}|,$$ $|T_{k}|$ is the entry-wise absolute value of $T_k$, and $$\begin{aligned}
{|+\rangle} & = & \frac{1}{\sqrt{2}}\left({|0\rangle}+{|1\rangle}\right)\\
{|-\rangle} & = & \frac{1}{\sqrt{2}}\left({|0\rangle}-{|1\rangle}\right).\end{aligned}$$ The projectors ${\left|-\right>}{\left<-\right|}$ and ${\left|+\right>}{\left<+\right|}$ act on the ancilla qubit.
Equation \[nicestoq\] makes the relationship between the spectra of $H_{XZ}$ and $\widetilde{H}_{XZ}$ clear. Let ${|\psi_{0}\rangle},{|\psi_{1}\rangle},\ldots,{|\psi_{N-1}\rangle}$ denote the eigenstates of $H_{XZ}$ with corresponding eigenvalues $\lambda_{0}\leq\lambda_{1}\leq\ldots\leq\lambda_{N-1}$, and let ${|\bar{\psi}_{0}\rangle},{|\bar{\psi}_{1}\rangle},\ldots,{|\bar{\psi}_{N-1}\rangle}$ denote the eigenstates of $\bar{H}_{XZ}$ with corresponding eigenvalues $\bar{\lambda}_{0}\leq\bar{\lambda}_{1}\leq\ldots\leq\bar{\lambda}_{N-1}$. ($H_{XZ}$ acts on $n$ qubits, so $N=2^{n}$.) $\widetilde{H}_{XZ}$, which acts on a $2N$-dimensional Hilbert space, has two $N$-dimensional invariant subspaces. The first is spanned by ${\left|\psi_{j}\right>}{\left|-\right>}$ with eigenvalues $\lambda_{j}$. The second is spanned by eigenvectors ${\left|\bar{\psi}_{j}\right>}{\left|+\right>}$ with eigenvalues $-\bar{\lambda}_{j}$.
We can perform universal adiabatic quantum computation in such an eigenstate of a stoquastic Hamiltonian. To prove this, we make use of the universal adiabatic Hamiltonian $H_{XZ}(t)$ from [@Biamonte], which at all $t$ takes the form shown in equation \[xzform\]. One can use the construction described above to obtain a stoquastic Hamiltonian $\widetilde{H}_{XZ}(t)$ corresponding to each instantaneous Hamiltonian $H_{XZ}(t)$. In this way we obtain a time varying Hamiltonian $\widetilde{H}_{XZ}(t)$ whose spectrum in the ${|-\rangle}$ subspace exactly matches the spectrum of $H_{XZ}(t)$, the only difference being the addition of an ancilla qubit in the ${|-\rangle}$ state. Because $\widetilde{H}_{XZ}(t)$ has no coupling between the ${|-\rangle}$ subspace and the ${|+\rangle}$ subspace, the adiabatic theorem may be applied within the ${|-\rangle}$ subspace. The relevant eigenvalue gap is thus the same as that of $H_{XZ}(t)$, and so is the runtime.
In standard adiabatic quantum computation, the qubits are in the ground state of the instantaneous Hamiltonian. Thus, any disturbance to the state costs energy. This is thought to offer some protection against thermal noise [@Childs]. When performing universal adiabatic quantum computation with $\widetilde{H}_{XZ}(t)$, the qubits are not in the ground state. Thus, it is possible for the system to thermally relax out of the computational state. However, this can only occur by disturbing the ancilla qubit out of the state ${|-\rangle}$. By protecting the ancilla qubit, one can to a large degree protect the entire computation. Note that an energy penalty against the ancilla qubit leaving the state ${|-\rangle}$ would be non-stoquastic. This is why the above construction fails to prove QMA-completeness and universal adiabatic quantum computation using the *ground state* of a stoquastic Hamiltonian, as we expect it must, based on the complexity-theoretic results of [@Bravyi1; @Bravyi2; @Liu; @Bravyi_Terhal].
QMA-Complete problems for Markov Matrices {#Markov_section}
=========================================
The second main result of our paper provides an example of a QMA-complete classical problem: finding the lowest eigenvalue of a symmetric Markov matrix. A matrix with all nonnegative entries, such that the entries in any given column sum to one is called a stochastic or Markov matrix. These matrices are named after Markov chains, which are stochastic processes such that given the present state, the future state is independent of the past states. Suppose a system has $d$ possible states. Then, its probability distribution at time $t$ is described by the $d$-dimensional vector $x_{t}$ whose entries are nonnegative and sum to one. If the system is evolving according to a Markov process then its dynamics are completely specified by the equation $x_{t+1}=Mx_{t}$ where $M$ is a $d\times d$ stochastic matrix. Note that, like quantum Hamiltonians, Markov matrices often have tensor product structure. For example, suppose we have two independent simultaneous Markov chains governed by $x_{t+1}=Mx_{t}$ and $y_{t+1}=Ny_{t}$. Then their joint probability distribution $z$ is governed by $z_{t+1}=(M\otimes N)z_{t}$.
Markov processes for which the Markov matrix is symmetric correspond to random walks on undirected weighted graphs. (Self-loops are allowed and correspond to diagonal matrix elements.) These matrices are doubly-stochastic: the sum of the entries in any row or column is one. By the Perron-Frobenius theorem, the highest eigenvalue of a symmetric stochastic matrix is one, and the corresponding eigenvector is the uniform distribution. The eigenvalue with next largest magnitude controls the rate of convergence of the process to its fixed point. A symmetric stochastic matrix is Hermitian and therefore one can also think of these matrices as Hamiltonians.
To prove that finding the lowest eigenvalue of a 3-local symmetric stochastic matrix is QMA-complete, we again use a reduction from the QMA-complete $H_{XZ}$ Hamiltonian of [@Biamonte]. We must take the opposite sign convention from equation \[alpha\]: $$H_{XZ}=\sum_{k}\alpha_{k}S_{k},$$ where the coefficients $\alpha_{k}$ are the same as before (all positive) and $S_{k}=-T_{k}$. Now define: $$\hat{H}_{XZ}=\frac{1}{N} \sum_k \alpha_k \widetilde{S}_k
\label{eq:normalize}$$ where $$N = \sum_k \alpha_k$$ and $\widetilde{S}_{k}$ is the permutation matrix obtained by applying the replacement rules (\[replacement\]) to $S_{k}$. By construction, $\hat{H}_{XZ}$ is a 3-local, symmetric, doubly stochastic matrix. We can rewrite $\hat{H}_{XZ}$ as $$\hat{H}_{XZ}=\frac{1}{N} \left( H_{XZ}\otimes {|-\rangle} {\langle -|} +
\bar{H}_{XZ} \otimes {|+\rangle} {\langle +|} \right),$$ where $\bar{H}_{XZ} = \sum_k \alpha_k |S_k|$. Thus, to determine an eigenvalue of $H_{XZ}$ to within $\pm\epsilon$ we must find the corresponding eigenvalue of $\hat{H}_{XZ}$ to within $\pm\epsilon/N$. Because $H_{XZ}$ is a two-local Hamiltonian on $n$ qubits with coupling strengths of order unity, $N$ is at most $O(n^{2})$. Thus the problem of determining the eigenvalue of $\hat{H}_{XZ}$ corresponding to the ground state of $H_{XZ}$ to polynomial precision is QMA-hard.
To obtain a cleaner QMA-hard problem we would like to construct a stochastic matrix whose *lowest* eigenvalue is QMA-hard to find. To do this, let $$H_{p}=(1-p)\sigma^{+}_{n+1}+p\hat{H}_{XZ}.$$ Here $\sigma^{+}_{n+1} = {|+\rangle}{\langle +|} = \frac{1}{2}({\mathds{1}}+X_{n+1})$ acts on the ancilla qubit, thereby giving it an energy penalty of size $(1-p)$ against leaving the state ${|-\rangle}$. For $0\leq p\leq1$, $H_{p}$ is a stochastic Hamiltonian. For $p<1/3$ the energy penalty is large enough that the highest eigenvalue in the ${\left|-\right>}$ subspace lies below the lowest eigenvalue in the ${\left|+\right>}$ subspace. In this case the lower half of the spectrum of $H_{p}$ is the spectrum of $H_{XZ}$ scaled by $p/N$, and the upper half of the spectrum of $H_{p}$ is the spectrum of $\bar{H}_{XZ}$ scaled by $p/N$ and shifted up by $1-p$.
Thus, we can obtain the ground energy of $H_{XZ}$ to polynomial precision by computing the lowest eigenvalue of $H_{p}$ to a higher but still polynomial precision. This reduction proves that finding the lowest eigenvalue of $H_{p}$ to polynomial precision is QMA-hard. Using the quantum algorithm for phase estimation, one easily shows that the problem of estimating the lowest eigenvalue of $H_{p}$ is contained in QMA (see [@Kitaev_book]). Thus this problem is QMA-complete.
Frustration Free Adiabatic Computation
======================================
It was stated in [@Bravyi_Terhal] that universal adiabatic quantum computation can be performed in the ground state of a 5-local frustration-free Hamiltonian. Let $U = U_L \ldots U_2 U_1$ be a quantum circuit acting on $n$ qubits with $L = \mathrm{poly}(n)$ gates. Let $$\label{psij}
{|\psi_j\rangle} = U_j \ldots U_1 {|0\rangle}^{\otimes n}$$ be a state of $n$ qubits corresponding to the $j{^\mathrm{th}}$ state of the time evolution of a quantum circuit specified by gates $U_j$ and $$\label{unary}
{|c_t\rangle} = {|1^{t+1} 0^{L-t}\rangle}.$$ be a state of $L+1$ clock qubits. Bravyi and Terhal construct a parametrized 5-local Hamiltonian $H(s)$ such that the ground state ${|\psi(s)\rangle}$ satisfies $$\begin{aligned}
{|\psi(1)\rangle} & = & \frac{1}{\sqrt{L+1}} \sum_{j=0}^L
{|\psi_j\rangle} {|c_j\rangle} \\
{|\psi(0)\rangle} & = & \frac{1}{\sqrt{L+1}} \sum_{j=0}^L {|0^n\rangle} {|c_j\rangle}.\end{aligned}$$ We can think of first register in ${|\psi(s)\rangle}$ as consisting of “work” qubits on which the computation happens and the second register in ${|\psi(s)\rangle}$ as being a clock containing a time written in unary.
For $s \in [0,1]$, the minimal eigenvalue gap between the ground state and first excited state of $H(s)$ is $O(1/L^2)$. By the adiabatic theorem [^2], $1/\mathrm{poly}(L)$ eigenvalue gap ensures that given ${|\psi_0\rangle}$, one obtains ${|\psi_1\rangle}$ by applying $H(s)$ and varying $s$ from zero to one over $\mathrm{poly}(L)$ time. By measuring the clock register of ${|\psi(1)\rangle}$, one obtains the result ${|1^{L+1}\rangle}$ with probability $1/(L+1)$. If this result is obtained, one finds the output of the circuit $U$ by measuring the first register of qubits in the computational basis. By repeating this process with $O(L)$ copies of ${|\psi(1)\rangle}$ one succeeds with high probability. Alternatively, one can pad the underlying circuit with $L$ identity gates, in which case each trial succeeds with probability $1/2$.
The construction from [@Bravyi_Terhal] invokes the fact that the spectrum of $H(1)$ is independent of the form of the gates $U_j$. By choosing a gate set which is composed of elements of simply connected unitary groups such as $SU(2)$ and $SU(4)$ one may construct a continuous path connecting each gate to the identity, and use a single parameter $s$ to transform all gates from the identity to the final circuit at once. The Hamiltonian at $s=0$ corresponds to the identity circuit, and its ground state is the uniform superposition of the clock states tensored with the initial data on the work qubits. In this ground state, the qubits of the clock register are entangled. It is standard to design adiabatic computations such that the initial Hamiltonian has a product state as its ground state, because such states should be easily produced by cooling or single-qubit measurements. In this section we construct a modified version of the construction from [@Bravyi_Terhal] that satisfies this condition and is still frustration free.
Let $c(j)$ indicate the $j{^\mathrm{th}}$ clock qubit and let $w(j)$ indicate the $j{^\mathrm{th}}$ work qubit. Let $$\begin{aligned}
H^{\mathrm{init}}_j & = & {|1\rangle}{\langle 1|}_{w(j)} \otimes {|10\rangle}
{\langle 10|}_{c(1),c(2)} \\
H^{\mathrm{clock}}_j & = & {|01\rangle}{\langle 01|}_{c(j-1),c(j)} \\\end{aligned}$$ For $j\in\{1,...,L-1\}$ define $$\begin{aligned}
H^{\mathrm{prop}}_j(s) & = & s
{|100\rangle}{\langle 100|}_{c(j),c(j+1),c(j+2)} + \\
& & (1-s) {|110\rangle}{\langle 110|}_{c(j),c(j+1),c(j+2)} - \\
& & \sqrt{s (1-s)} \left(
U_j \otimes {|110\rangle}{\langle 100|}_{c(j),c(j+1),c(j+2)} + \right. \\
& & \left. U_j^\dag \otimes {|100\rangle}{\langle 110|}_{c(j),c(j+1),c(j+2)}
\right). \\\end{aligned}$$ and let $$\begin{aligned}
H^{\mathrm{prop}}_L(s) & = & s
{|10\rangle}{\langle 10|}_{c(L),c(L+1)} + \\
& & (1-s) {|11\rangle}{\langle 11|}_{c(L),c(L+1)} - \\
& & \sqrt{s (1-s)} \left(
U_L \otimes {|11\rangle}{\langle 10|}_{c(L),c(L+1))} + \right. \\
& & \left. U_L^\dag \otimes {|10\rangle}{\langle 11|}_{c(L),c(L+1)}
\right). \\\end{aligned}$$ It can be directly verified that each $H^{\mathrm{clock}}_j$, $H^{\mathrm{init}}_j$, and $H^{\mathrm{prop}}_j(s)$ is a projector. Here, for convenience, we define $s$ so that it varies from zero to one half rather than from zero to one as is done in [@Bravyi_Terhal]. Our frustration-free Hamiltonian is the following sum of projectors: $$\begin{aligned}
H^{\mathrm{clock}} & = & {|0\rangle}{\langle 0|}_{c(0)} + \sum_{j=1}^L
H^{\mathrm{clock}}_j \\
H^{\mathrm{init}} & = & \sum_{j=1}^n H^{\mathrm{init}}_j \\
H^{\mathrm{prop}}(s) & = & \sum_{j=1}^L H^{\mathrm{prop}}_j(s) \\
H^{\mathrm{FF}}(s) & = & H^{\mathrm{clock}} + H^{\mathrm{init}} +
H^{\mathrm{prop}}(s)\end{aligned}$$
If $U_1 \ldots U_L$ are chosen from a universal set of two-qubit gates then $H^{\mathrm{FF}}(s)$ is an efficient 5-local frustration-free adiabatic quantum computer. To see how this Hamiltonian achieves universal adiabatic computation, we examine the various terms one by one. The ground state of $H^{\mathrm{FF}}(s)$ is the simultaneous zero eigenspace of $H^{\mathrm{clock}}$, $H^{\mathrm{init}}$, and $H^{\mathrm{prop}}$. $H^{\mathrm{clock}}$ commutes with $H^{\mathrm{prop}}(s) + H^{\mathrm{init}}$ and provides an energy penalty of at least unit size if the clock register is not in one of the unary states ${|c_t\rangle} = {|1^{t+1} 0^{L-t}\rangle}$. Thus the low lying spectrum of $H^{\mathrm{FF}}(s)$ is strictly contained in the ground space of $H^{\mathrm{clock}}$.
For any bit string $x \in \{0,1\}^n$ and integer $j \in
\{1,2,\ldots,L\}$, let $${|\chi_x^j\rangle} = \left( U_j U_{j-1} \ldots U_1 {|x\rangle} \right) \otimes
{|c_j\rangle},$$ where ${|c_j\rangle}$ is as defined in equation \[unary\]. (We also define ${|\chi_x^0\rangle} = {|x\rangle} \otimes {|c_0\rangle}$.) There are $2^n(L+1)$ such states and they form an orthonormal basis for the ground space of $H^{\mathrm{clock}}$. In this basis, $H^{\mathrm{prop}}(s) +
H^{\mathrm{init}}$ takes the block-diagonal form $$H^{\mathrm{prop}}(s) + H^{\mathrm{init}} = \bigoplus_{x \in \{0,1\}^n} M_x$$ where $$M_x = \left[ \begin{array}{cccccc}
s+|x| & -b & & & & \\
-b & 1 & -b & & & \\
& -b & 1 & -b & & \\
& & {\ddots}& {\ddots}& {\ddots}& \\
& & & -b & 1 & -b \\
& & & & -b & 1-s
\end{array} \right]$$ is an $L+1$ by $L+1$ matrix and $b = \sqrt{s(1-s)}$. Here $|x|$ denotes the Hamming weight of the bit string $x$. The appearance of $|x|$ is the sole manifestation of $H^{\mathrm{init}}$. The rest of the matrix elements all come from the “hopping” action of $H^{\mathrm{prop}}(s)$.
$M_{00\ldots0}$ has the unique ground state $$\label{newhist}
N \sum_{j=0}^L r^j {|\psi_j\rangle} {|c_j\rangle}$$ where ${|\psi_j\rangle}$ and ${|c_j\rangle}$ are as defined in equations \[psij\] and \[unary\], $r = \sqrt{\frac{s}{1-s}}$, and $N$ is a normalization factor. This constitutes the ground state of $H^{\mathrm{FF}}(s)$. The first excited state of $M_{00\ldots0}$ has energy $1-2 \sqrt{s (1-s)} \cos \left( \frac{\pi}{L+1}
\right)$. Because of the direct sum structure of $H^{\mathrm{FF}}(s)$, we can apply the adiabatic theorem directly to $M_{00\ldots0}$. The runtime of the adiabatic algorithm is thus determined by the gap between the ground and first excited states of $M_{00\ldots0}$. This takes its minumum at $s=1/2$, where it is equal to $1-\cos \left(
\frac{\pi}{(L+1)} \right) = O(1/L^2)$. For questions of fault tolerance it is also useful to know the eigenvalue gap between the ground and first excited states of the full Hamiltonian $H^{\mathrm{FF}}(s)$. The first excited energy of $H^{\mathrm{FF}}(s)$ is equal to the ground energy of $M_{10\ldots0}$, which is $1-2
\sqrt{s(1-s)} \cos \left( \frac{\pi}{2(L+1)} \right)$. Thus the minumum eigenvalue gap of $H^{\mathrm{FF}}(s)$ occurs at $s=1/2$ and is equal to $1-\cos \left( \frac{\pi}{2 (L+1)} \right) = O(1/L^2)$.
By equation \[newhist\], the ground state of $H^{\mathrm{FF}}(0)$ is ${|000\ldots\rangle} \otimes {|1000\ldots\rangle}$, and the ground state of $H^{\mathrm{FF}}(1/2)$ is the same state $\frac{1}{\sqrt{L+1}}
\sum_{j=0}^L {|\psi_j\rangle} {|c_j\rangle}$ produced by the scheme of [@Bravyi_Terhal].
Stochastic Frustration Free Computation
=======================================
In [@Bravyi_Terhal], Bravyi and Terhal showed that adiabatic quantum computation in the ground state of a stoquastic frustration free Hamiltonian can be efficiently simulated by a classical computer. In this section we show that in contrast, one can perform universal adiabatic quantum computation in the ground state of a *stochastic* frustration free (StochFF) Hamiltonian $H^{\mathrm{StochFF}}(s)$. (Alternatively, we can view this as computation in the highest energy state of the stoquastic Hamiltonian $-H^{StochFF}(s)$.)
It has been shown that the two-qubit CNOT gate, together with any one-qubit rotation whose square is not basis preserving, are sufficient to perform universal quantum computation [@Shi]. All matrix elements in these gates are real numbers. If we choose $U_1,\ldots,U_L$ from this gate set then $H^{\mathrm{FF}}(s)$ is a 5-local real frustration free Hamiltonian. Examining the construction of section \[Markov\_section\] one sees that it can be applied to any Hamiltonian with real matrix elements, and it increases the locality by one. This construction also preserves frustration-freeness, as we will show in the next paragraph. We can thus use this construction on $H^{FF} (s)$ to obtain a 6-local stochastic frustration free Hamiltonian whose ground state is universal for adiabatic quantum computation
To show that the mapping of section \[Markov\_section\] preserves frustration-freeness, consider applying this mapping to a frustration free local Hamiltonian $H=\sum_{j=1}^{m}H_j$, where $H_j=\sum_{k}\alpha_{k}^{j} S_k^j$ (where each $S_k^j$ is, up to an overall sign, a tensor product of Pauli operators and each $\alpha_k^j$ is positive). We obtain the Hamiltonian $$\begin{aligned}
H_p&=&p\hat{H}+\left(1-p\right)\left(\frac{{\mathds{1}}+X_{n+1}}{2}\right)\nonumber \\
&=&\sum_j \frac{N_j}{N} \left[p\hat{H_j}+\left(1-p\right)\left(\frac{{\mathds{1}}+X_{n+1}}{2}\right)\right] \label{fr_free}\end{aligned}$$ where $N_j=\sum_k {\alpha_k^j}$ and $N=\sum_j N_j$. When $p<\frac{1}{3}$, $H_p$ is stochastic and has a zero energy ground state with an eigenvalue gap which is $\frac{p}{N}$ times the gap of $H$. Furthermore we see from (and the fact that each $H_j$ is positive semidefinite) that $H_p$ is a sum of positive semidefinite operators. Hence the Hamiltonian $H_p$ is frustration free.
Generalizations
===============
The constructions of sections \[Markov\_section\] and \[QMA\_section\] replace Hamiltonians with real matrix elements of both signs by computationally equivalent Hamiltonians with real positive matrix elements. In this section we show that this technique can be generalized to directly replace Hamiltonians with complex matrix elements by computationally equivalent Hamiltonians with only real positive matrix elements. However, in the process we necessarily introduce a two-fold degeneracy of the ground state.
Let $H$ be an arbitrary $k$-local Hamiltonian. We may expand $H$ as $$\label{pdecomp}
H = \sum_{j} \alpha_j O_j$$ where each $O_j$ is a tensor product of $k$ or fewer Pauli matrices and each $\alpha_j$ is positive. Each entry in each $O_j$ is $\pm 1$ or $\pm i$. We can replace the group $\{1,i,-1,-i\}$ with its left-regular representation $$\label{freplace}
\begin{split}i & \mapsto F\\
-1 & \mapsto F^{2}\\
-i & \mapsto F^{3}\\
1 & \mapsto F^{4}\end{split}$$ where $$F=\begin{pmatrix}0 & 1 & 0 & 0\\
0 & 0 & 1 & 0\\
0 & 0 & 0 & 1\\
1 & 0 & 0 & 0\end{pmatrix} .$$ The eigenvectors of $F$ are ${|v_0\rangle},{|v_1\rangle},{|v_2\rangle},{|v_3\rangle}$ where $${|v_j\rangle} = \frac{1}{2}\sum_{l = 0}^3 i^{lj} {|l\rangle}.$$ The corresponding eigenvalues are $$\label{eigenvalues}
F {|v_j\rangle} = i^j {|v_j\rangle}.$$
Let $S_j$ and $A_j$ be the real and imaginary parts of $\alpha_j
O_j$. That is, $S_j$ and $A_j$ are the unique real symmetric and anti-symmetric matrices such that $$\alpha_j O_j = S + iA$$ Further let $S_j^+ = (|S_j|+S_j)/2)$ and $S_j^- = (|S_j|-S_j)/2$ and similarly for $A_j^{\pm}$, where $|\cdot|$ denotes the entrywise absolute value. Applying the replacement (\[freplace\]) to $H$ and dividing by $N = \sum_j \alpha_j$ yields the stochastic Hamiltonian $\tilde{H}$ with the decomposition $$\begin{aligned}
\label{tildeH}
\tilde{H} & = \frac{1}{N} & \left( H^{(0)} \otimes {|v_0\rangle} {\langle v_0|} +
H^{(1)} \otimes {|v_1\rangle} {\langle v_1|} + \right. \nonumber \\
& & \left. H^{(2)} \otimes {|v_2\rangle} {\langle v_2|} +
H^{(3)} \otimes {|v_3\rangle} {\langle v_3|} \right)\end{aligned}$$ where $$\begin{aligned}
H^{(0)} & = & \sum_j S_j^+ + S_j^- + A_j^+ + A_j^- \\
H^{(1)} & = & \sum_j S_j^+ - S_j^- + iA_j^+ - iA_j^- \\
H^{(2)} & =& \sum_j S_j^+ + S_j^- - A_j^+ - A_j^-\\
H^{(3)} & = & \sum_j S_j^+ - S_j^- - iA_j^+ + iA_j^-. \end{aligned}$$
$H^{(1)} = H$, thus the spectrum of $\tilde{H}$ in the ${|v_1\rangle}$ subspace matches that of $H$ up to a normalization factor of $N$ and a pair of extra ancilla qubits. If we write each projector ${|v_j\rangle}{\langle v_j|}$ in terms of the Pauli basis we obtain $$\begin{aligned}
{|v_0\rangle}{\langle v_0|} & = & \Pi_+^X \Pi_+^X \\
{|v_1\rangle}{\langle v_1|} & = & \Pi_-^X \Pi_+^Y \\
{|v_2\rangle}{\langle v_2|} & = & \Pi_+^X \Pi_-^X \\
{|v_3\rangle}{\langle v_3|} & = & \Pi_-^X \Pi_-^Y \\\end{aligned}$$ where $\Pi_\pm^a$ is the projector onto the eigenvalue $\pm 1$ eigenstate of the Pauli matrix $a$. Thus an $X$ penalty on the first ancilla qubit will separate the ${|v_1\rangle},{|v_3\rangle}$ subspace from the ${|v_0\rangle},{|v_2\rangle}$ subspace. So, taking, $0<p < \frac{1}{3}$, the stochastic Hamiltonian $$H_p^\prime = \left(1-p\right)\left(\frac{{\mathds{1}}+X_{n+1}}{2}\right) + p\tilde{H}$$ has ground space spanned by ${|\psi^{(1)}\rangle} {|v_1\rangle}$ and ${|\psi^{(3)}\rangle} {|v_3\rangle}$, where ${|\psi^{(1)}\rangle}$ is the ground state of $H^{(1)}$ and ${|\psi^{(3)}\rangle}$ is the ground state of $H^{(3)}$. $H^{(1)} = H$, thus ${|\psi^{(1)}\rangle}$ is the ground state of $H$. $H^{(3)} = H^*$, thus ${|\psi^{(3)}\rangle}$ is the complex conjugate of the ground state of $H$.
A simple argument shows that the doubling in the spectrum of $H_p^\prime$ is a necessary property for any construction which maps an arbitrary Hamiltonian onto a real Hamiltonian $H_R$, where $H_R$ is equal to $H$ within a fixed 1D subspace of the ancillas. Suppose that we have such a map which sends an arbitrary Hamiltonian $H$ which acts on a Hilbert space $\mathcal{H}_1$ to a real Hamiltonian $H_R$ on a larger Hilbert space $\mathcal{H}_1\otimes\mathcal{H}_2$ with the property that $$\label{HR}
H_R=H\otimes |\phi\rangle \langle \phi|+H^{\text{other}}\otimes({\mathds{1}}-|\phi\rangle\langle \phi|).$$ where the state $|\phi\rangle\in\mathcal{H}_2$ does not depend on the particular Hamiltonian $H$ but the operator $H^{\text{other}}$ may depend on $H$. Then for any eigenvector $|\psi\rangle$ of $H$ with energy $E$ we have: $$H_R|\psi\rangle|\phi\rangle=E|\psi\rangle|\phi\rangle$$ Since $H_R$ is real, complex conjugating this equation gives: $$\label{ccHR}
H_R|\psi^\star \rangle|\phi^\star \rangle=E|\psi^\star \rangle|\phi^\star \rangle$$ To show that doubling exists in the spectrum it is sufficient to show that $\langle \phi|\phi^\star\rangle=0$. To prove this, first use equation to obtain $$\left({\mathds{1}}\otimes|\phi\rangle\langle \phi|\right)H_R |\psi^\star \rangle |\phi^\star \rangle=E|\psi^\star \rangle |\phi\rangle \langle \phi|\phi^\star\rangle$$ Then use equation to obtain $$\left({\mathds{1}}\otimes|\phi\rangle\langle \phi|\right)H_R |\psi^\star \rangle |\phi^\star \rangle=\left(H|\psi^\star \rangle \right) |\phi\rangle \langle \phi|\phi^\star\rangle.$$ Equating these expressions gives $$H|\psi^\star\rangle \langle \phi|\phi^\star\rangle=E|\psi^\star \rangle \langle \phi|\phi^\star\rangle.$$ This must hold for all Hamiltonians $H$ and eigenstates $|\psi\rangle$ and therefore it must be the case that $\langle
\phi|\phi^\star\rangle=0$. So we have shown that the doubling in the spectrum of $H_p^\prime$ is a necessary feature of the type of maps we consider. For constructing universal adiabatic quantum computers the degeneracy induced by this construction may be problematic. However, for proving complexity-theoretic completeness results it is often irrelevant, as we see in the next section.
Stochastic $k$-SAT
==================
The methods of the previous section can be used to show, roughly speaking, that deciding whether or not a Hamiltonian which is a sum of positive semidefinite stochastic operators is frustration free is as difficult as the general problem of deciding whether a Hamiltonian is frustration free. In this section we formalize this by defining a problem called stochastic $k$-SAT, which we show to be $\mathrm{QMA}_1$-complete for $k=6$.
We first recall the definition of stoquastic $k$-SAT which is given in [@Bravyi_Terhal].\
\
**Problem:** Stoquastic $k$-SAT\
**Input:** Input: A set of $k$-local Hermitian operators $\{H_{j}\}$ for $j\in\{1,\dots,m\}$ where $m=\mathrm{poly}(n)$ and a parameter $\epsilon>1/\mathrm{poly}(n)$
1. [Each $H_{j}$ is positive semidefinite]{}
2. [Each $H_{j}$ has norm which is bounded by a polynomial in $n$.]{}
3. [Every $H_{j}$ is stoquastic.]{}
**Output:** If $H=\sum_j H_j$ has a zero energy ground state, then this is a YES instance. Otherwise if every eigenstate of $H$ has energy $>\epsilon$ then it is a NO instance.\
**Promise:** Either the ground state of H has energy 0, or else it has energy $>\epsilon$.\
\
The stoquastic $k$-SAT problem is therefore the problem of deciding if a given stoquastic Hamiltonian that is a sum of positive definite operators is frustration free, given that either this is the case or else its ground energy exceeds $\epsilon$ [@Bravyi_Terhal]. Note that this definition of stoquastic $k$-SAT looks somewhat different from the definition of quantum $k$-SAT which was given in section \[sec\_intro\], which was stated entirely in terms of projectors. Given an instance of stoquastic $k$-SAT, we can define operators $\Pi_j$ which project onto the zero eigenspaces of the $H_j$. When the Hamiltonians $H_j$ are stoquastic, these projectors are guaranteed to have nonnegative matrix elements in the computational basis [@Bravyi_Terhal]. So given an instance of stoquastic $k$-SAT with Hermitian positive semidefinite operators $H_j$, it is possible to construct another instance of stoquastic $k$-SAT with operators $\tilde{H}_j=\{1-\Pi_j\}$ that are all projectors.
We now define a problem called stochastic $k$-SAT, which is identical to stoquastic $k$-SAT except that condition 3 is replaced by\
$3'$. Every $H_j$ is a stochastic matrix.\
\
We note that there does not appear to be an equivalence between this definition of stochastic $k$-SAT and the corresponding definition where all the $H_j$ are (in addition) required to be projectors.
Given these two definitions and the foregoing map from an arbitrary Hamiltonian to a stochastic Hamiltonian, we now show how to reduce any instance of quantum 4-SAT to an instance of stochastic 6-SAT. Starting with an instance of quantum 4-SAT specified by a set of projectors $\{\Pi_j\}$ (for $j\in{1,...,m}$) we use the map of the previous section (with $p=\frac{1}{3}$ for concreteness) on each projector to obtain a set of 6-local positive semidefinite stochastic Hamiltonians $\{H_j\}$ where $$H_j= \frac{2}{3}\left(\frac{{\mathds{1}}+X_{n+1}}{2}\right) + \frac{1}{3}\tilde{\Pi}_j$$ (Note that $\tilde{\Pi}_j$ refers to the operator obtained by applying the mapping from equation .) If the 4-SAT instance is satisfiable, then the stochastic 6-SAT instance will also be satisfiable. Define $N_{max}$ to be the maximum value of $N$ obtained for one of the terms $\Pi_j$ when using the mapping of equation . If the 4-SAT instance is not satisfiable then for any state $|\phi\rangle$ there is some projector $\Pi_k$ such that $\langle \phi|\Pi_k |\phi \rangle\geq \frac{\epsilon}{m}$. If we take the parameter $\tilde{\epsilon}$ of the stochastic 6-SAT instance to be related to the parameter $\epsilon$ of the quantum 4-SAT instance by $\tilde{\epsilon}=\frac{\epsilon}{3m N_{max}}$ then the stochastic 6-SAT instance will also be unsatisfiable. Therefore stochastic 6-SAT is $\mathrm{QMA}_1$ hard. Stochastic 6-SAT is contained in $\mathrm{QMA}_1$ since every instance of stochastic 6-SAT can be mapped to an instance of quantum 6-SAT by taking projectors $\Pi_j$ which project onto everything but the zero eigenspaces of the $H_j$.
So $\mathrm{QMA}_{1}$ completeness of stochastic 6-SAT follows from the results of Bravyi [@Bravyi4] on quantum $k$-SAT. This is in contrast to stoquastic $k$-SAT, which is contained in MA for every constant $k$ [@Bravyi_Terhal].
QMA-completeness for excited states {#excited}
===================================
The local Hamiltonian problem refers specifically to ground state energies. Similarly, we have formulated a computational problem based on the highest energy of a given Hamiltonian. It is natural to ask about the complexity of estimating the $c{^\mathrm{th}}$ excited state. We can formulate this as follows. Let $H$ be a $k$-local Hamiltonian on the Hilbert space $\mathcal{H}$ of $n$ qubits. Let $\lambda_1 \leq
\lambda_2 \leq \ldots \leq \lambda_{2^n}$ denote the eigenvalues of $H$, with corresponding eigenvectors ${|\psi_1\rangle}, {|\psi_2\rangle},
\ldots, {|\psi_{2^n}\rangle}$. The $(k,c,\epsilon)$-energy problem is as follows.\
\
**Problem:** $(k,c,\epsilon)$-energy\
**Input:** We are given a classical description of $H$, an integer $c \geq 1$, and a pair of parameters $a,b$ such that $b-a = \epsilon >
1/\mathrm{poly}(n)$.\
**Output:** If $\lambda_c \leq a$ answer YES. If $\lambda_c \geq b$ output NO.\
**Promise:** $H$ is such that the answer is YES or NO.\
In this section we will show that the $(k,c,\epsilon)$-energy problem is QMA-complete for any $c = O(1)$. Showing QMA-hardness is the easier of the two proofs. This can be achieved as follows. Let $H^{(a)}$ and $H^{(b)}$ be a pair of $k$-local Hamiltonians on $n$ qubits, with spectra $\lambda_1^{(a)},\ldots,\lambda_{2^n}^{(a)}$, ${|\psi_1^{(a)}\rangle},\ldots,{|\psi_{2^n}^{(a)}\rangle}$ and $\lambda_1^{(b)},\ldots,\lambda_{2^n}^{(b)}$, ${|\psi_1^{(b)}\rangle},\ldots,{|\psi_{2^n}^{(b)}\rangle}$, respectively. Then $$H^{(ab)} = H^{(a)} \otimes {|0\rangle} {\langle 0|} + H^{(b)} \otimes {|1\rangle} {\langle 1|}$$ is a $(k+1)$-local Hamiltonian on $n+1$ qubits. Its complete set of eigenvalues is $\lambda_1^{(a)},\ldots,\lambda_{2^n}^{(b)};\lambda_1^{(b)},\ldots,\lambda_{2^n}^{(b)}$, with corresponding eigenvectors ${|\psi_1^{(a)}\rangle}{|0\rangle}, \ldots,
{|\psi_{2^n}^{(a)}\rangle}{|0\rangle}; {|\psi_1^{(b)}\rangle}{|1\rangle}, \ldots,
{|\psi_{2^n}^{(b)}\rangle}{|1\rangle}$. To prove QMA-hardness of a low-lying excited state let $H_0$ to be a Hamiltonian such that determining whether the ground energy is close to zero is QMA-hard. Given an integer $c$, let $d = \lceil
\log_2 c \rceil$, $P_k = \frac{1}{2} (Z_k + {\mathds{1}})$, and $$H_c = \sum_{k=0}^d 2^k P_k + \sum_{k=d+1}^n
2^{d+1} P_k - \left( c-\frac{1}{2} \right) {\mathds{1}}$$ $H_c$ has exactly $c$ states with negative energy, and its lowest nonnegative eigenvalue is $\frac{1}{2}$. Thus determining the $c{^\mathrm{th}}$ excited energy of $H_c \otimes {|0\rangle} {\langle 0|} + H_0 \otimes
{|1\rangle} {\langle 1|}$ is QMA-hard. In particular, it is interesting to note that by choosing $c=2$ we construct a Hamiltonian whose eigenvalue gap between the ground state and first excited state is QMA-hard to compute.
Next we show containment in QMA. The naive protocol would be for Merlin to provide Arthur with the state ${|\psi_1\rangle} {|\psi_2\rangle} \ldots {|\psi_c\rangle}$ and for Arthur to use phase estimation to check that the $c$ registers each contain a state of energy at most $a$. The problem is that for “no” instances there are many ways for Merlin to cheat. For example if $\lambda_1 \leq a$ but $\lambda_c \geq b$, the answer is “no” but Merlin can provide the state ${|\psi_1\rangle}^{\otimes c}$ as a supposed witness. To prevent this, Arthur needs to somehow check that he has been given a set of $c$ orthogonal states that each have energy at most $a$. Thus we propose the following protocol.
Arthur demands that Merlin give him the state $$\label{witness}
{|W\rangle} = \frac{1}{\sqrt{c!}} \sum_{\pi \in S_c} \mathrm{sign}(\pi)
{|\psi_{\pi(1)}\rangle} {|\psi_{\pi(2)}\rangle} \ldots {|\psi_{\pi(c)}\rangle}.$$ Arthur performs the projective measurement to see that the state given to him by Merlin lies in the antisymmetric subspace of $\mathcal{H}^{\otimes c}$. If this fails he rejects. He then throws away all but the first register and performs phase estimation of $H$ to precision better than $\epsilon$. If the state has energy above $b$ he rejects. Otherwise he accepts.
It is clear that for YES instances, Arthur will accept the state ${|W\rangle}$ with high probability. (The only source of error is imprecision in phase estimation.) We will next prove that for NO instances the acceptance probability is at most $1-\frac{1}{c}$. Using standard methods[@Kitaev_book; @MW; @NWZ] we can amplify this protocol to obtain polynomially small acceptance probability for NO instances.
\[symm\] For any state ${|\phi\rangle}$ in the antisymmetric subspace of $\mathcal{H}^{\otimes c}$ and any state ${|\alpha_1\rangle} \in
\mathcal{H}$, ${\langle \phi|} \left( {|\alpha_1\rangle} {\langle \alpha_1|} \otimes
{\mathds{1}}\right) {|\phi\rangle} \leq \frac{1}{c}$ where ${\mathds{1}}$ is the identity operator on $\mathcal{H}^{\otimes (c-1)}$.
Extend ${|\alpha_1\rangle}$ to an orthonormal basis ${|\alpha_1\rangle},{|\alpha_2\rangle},\ldots,{|\alpha_{2^n}\rangle}$ for $\mathcal{H}$. Let $F$ be the set of functions $f:\{1,2,\ldots,c\} \to
\{1,2,\ldots,2^n\}$ such that $f(1) < f(2) < \ldots < f(c)$. Thus $|F|
= \binom{2^n}{c}$. For any $f \in F$ we have the corresponding Slater determinant state. $${|D_f\rangle} = \frac{1}{\sqrt{c!}} \sum_{\pi \in S_c} \mathrm{sign}(\pi)
{|\alpha_{f(\pi(1))}\rangle} {|\alpha_{f(\pi(2))}\rangle} \ldots
{|\alpha_{f(\pi(c))}\rangle}$$ It a standard result that these $\binom{2^n}{c}$ states form a complete orthonormal basis for the antisymmetric subspace of $\mathcal{H}^{\otimes c}$. For any $f,g \in F$ we have $$\label{deltas}
{\langle D_f|} \left( {|\alpha_1\rangle} {\langle \alpha_1|} \otimes {\mathds{1}}\right)
{|D_g\rangle} = \frac{ \delta_{f,g} \delta_{f(1),1}}{c}$$ where each $\delta$ denotes a generalized Kronecker-$\delta$. Because ${|\phi\rangle}$ is antisymmetric, it can be decomposed in the Slater determinant basis: $${|\phi\rangle} = \sum_{f \in F} \phi_f {|D_f\rangle}$$ and $p_f = \phi^*_f \phi_f$ is a corresponding probability distribution on $F$. Thus $${\langle \phi|} \left( {|\alpha_1\rangle} {\langle \alpha_1|} \otimes {\mathds{1}}\right)
{|\phi\rangle} =$$ $$\sum_{f,g \in F} \phi^*_f {\langle D_f|} \left( {|\alpha_1\rangle}
{\langle \alpha_1|} \otimes {\mathds{1}}\right) {|D_g\rangle} \phi_g.$$ By equation \[deltas\] this is $$= \frac{1}{c} \sum_{f \in F} p_f \delta_{f(1),1} \leq \frac{1}{c}.$$
The quantity $$p_j(\phi) = {\langle \phi|} \left( {|\psi_j\rangle}{\langle \psi_j|} \otimes {\mathds{1}}\right)
{|\phi\rangle}$$ is the probability of obtaining ${|\psi_j\rangle}$ if we measure the first register of a state ${|\phi\rangle}$ in the eigenbasis of $H$. By lemma \[symm\] $$\sum_{j=1}^{c-1} p_j(\phi) \leq 1 - \frac{1}{c}.$$ Thus, with probability at least $\frac{1}{c}$, such a measurement would yield ${|\psi_j\rangle}$ with $j > c-1$. Thus if $\lambda_c > b$ then with probability at least $\frac{1}{c}$ a measurement of the observable $H$ would yield energy at least $b$. The phase estimation algorithm can in $\mathrm{poly}(1/\epsilon)$ time perform such an energy measurement with exponentially small chance of making an error as large as $\epsilon$. Thus the protocol is sound, which completes the proof that the $(c,k,\epsilon)$-energy problem is QMA-complete for constant $c$ and $k$ and polynomially small $\epsilon$.
Discussion and conclusions
==========================
The results presented in this paper have several applications. Although calculating the ground state energy of stoquastic Hamiltonians appears easier than calculating the ground energy of generic Hamiltonians, our results suggest that calculating other eigenstates of stoquastic Hamiltonians remains hard. Because the wavefunctions of these states have amplitudes which are both positive and negative, the hardness of determining their energy supports the intuition that it is the positivity of the amplitudes which makes the ground state problem for stoquastic Hamiltonians easier. An extreme distinction between stochastic and stoquastic Hamiltonians arises when the Hamiltonians are also frustration free. Although adiabatic evolution with stoquastic frustration free Hamiltonians is simulable in BPP [@Bravyi_Terhal], we have shown that adiabatic evolution in the ground state of a stochastic frustration free Hamiltonian is universal.
Secondly, these results may be relevant for the physical implementation of quantum computers. The first proof of universality of adiabatic quantum computation used 5-local interactions [@Aharonov_adiabatic]. Since then, the Hamiltonians have been brought into incrementally more physically feasible form by various techniques while retaining universality [@Kempe; @Kempe_Regev; @Oliveira]. The universal Hamiltonian $H_{XZ}$ of [@Biamonte] is one outcome of this chain of reductions. Here we add one more step to this chain, obtaining universal stochastic and stoquastic Hamiltonians which resemble those arising in some systems of superconducting qubits [@Bravyi1]. The constructions given here are at least three local, and so would require the use of perturbative gadgets to implement in terms of physical two-local interactions.
Finally, our results are of interest from a purely complexity-theoretic point of view. Stochastic matrices arise outside the context of quantum mechanics, in Markov chains. Our reduction shows that finding the lowest eigenvalue of a certain class of exponentially large but efficiently describable doubly-stochastic matrices is QMA-complete. (These stochastic matrices correspond to Markov Chains in which the “update rule” is a probabilistic selection over some set of updates which are local in the tensor product sense.) In general, the problem of finding eigenvalues of stochastic matrices is of interest because the eigenvalue of second largest magnitude determines the mixing time of the corresponding Markov chain. There exist Markov chains in which the eigenvalue of second largest magnitude is negative, and is the lowest lying eigenvalue. We hope that the demonstration of a QMA-complete problem arising in a classical setting will help shed further light on the class QMA itself.
We thank Sergey Bravyi for a helpful discussion. SJ thanks MIT’s Center for Theoretical Physics, RIKEN’s Digital Materials Laboratory, and Caltech’s Institute for Quantum Information, Franco Nori, Sahel Ashhab, ARO/DTO’s QuaCGR program, the DOE, the Sherman Fairchild Foundation, and the NSF under grant number PHY-0803371. DG thanks Eddie Farhi, the W. M. Keck Foundation Center for Extreme Quantum Information Theory, and the Natural Sciences and Engineering Research Council of Canada. PJL thanks John Preskill and the Institute of Quantum Information at Caltech for hosting an extended visit during which part of this work was completed. This research was supported in part by the National Science Foundation under Grant No. PHY05-51164.
[^1]: Like MA, the class AM is a probabilistic generalization of NP, see [@Arora].
[^2]: Many versions of the adiabatic therem have been proven. For one example see appendix F of [@mythesis].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report Atacama Large Millimeter/submillimeter Array (ALMA) 1.3 mm continuum upper limits for 5 planetary-mass companions DH Tau B, CT Cha B, GSC 6214-210 B, 1RXS 1609 B, and GQ Lup B. Our survey, together with other ALMA studies, have yielded null results for disks around young planet-mass companions and placed stringent dust mass upper limits, typically less than 0.1 $M_{\oplus}$, when assuming dust continuum is optically thin. Such low-mass gas/dust content can lead to a disk lifetime estimate (from accretion rates) much shorter than the age of the system. To alleviate this timescale discrepancy, we suggest that disks around wide companions might be very compact and optically thick, in order to sustain a few Myr of accretion yet have very weak (sub)millimeter flux so as to still be elusive to ALMA. Our order-of-magnitude estimate shows that compact optically-thick disks might be smaller than 1000 $R_{\rm Jup}$ and only emit $\sim$$\mu$Jy of flux in the (sub)millimeter, but their average temperature can be higher than that of circumstellar disks. The high disk temperature could impede satellite formation, but it also suggests that mid- to far-infrared might be more favorable than radio wavelengths to characterize disk properties. Finally, the compact disk size might imply that dynamical encounters between the companion and the star, or any other scatterers in the system, play a role in the formation of planetary-mass companions.'
author:
- 'Ya-Lin Wu$^1$, Laird M. Close$^1$, Josh A. Eisner$^1$, and Patrick D. Sheehan$^2$'
title: '**A E V L R F Y P C**'
---
**1. I** {#i .unnumbered}
========
Direct imaging of circumplanetary disks is crucial to study how planets and their satellites actually form. Theoretical modeling has suggested that circumplanetary disks can emit significant infrared fluxes, and magnetospheric accretion onto the surface of protoplanets can create strong line emission and UV/optical continuum excess (e.g., @E15 [@Z15; @Z16]). Recently, high-contrast imaging campaigns have been targeting transition disks to search for these accretion signatures inside the central holes or gaps. While a few red sources have been identified as protoplanet candidates, their nature remains elusive and some may actually be disk features or data reduction artifacts (e.g., @Biller14 [@S15a; @F17]). So far the only confirmed accreting planet is LkCa 15 b, which emits at H$\alpha$ [@S15b], but its accretion disk has yet to be directly resolved [@I14].
One can also search for disks around wide-orbit planetary-mass companions, which are $\sim$5–40 $M_{\rm Jup}$ gas giants orbiting at tens to hundreds of astronomical units. Wide companions, unlike the copious transiting planets, are intrinsically rare (e.g., @NC10), so their formation scenario may not resemble that of planets (i.e., core accretion) but more like brown dwarfs or low-mass stars (e.g., @Brandt14). Nonetheless, their wide separations offer a direct view into the physical mechanisms that can regulate planet formation in circumplanetary disks. Analytic and numerical analyses have shown that circumplanetary disks are truncated at $\sim$1/3 of the planet’s Hill radius (e.g., @QT98 [@AB09]). [@SB13] further demonstrated that gas giants formed via disk fragmentation at 100 au can harbor such a truncated disk. Simulations have predicted that circumplanetary disks are luminous at radio wavelengths and may be easily detectable by the Atacama Large Millimeter/submillimeter Array (ALMA) (e.g., @SB13 [@Z16]).
Recently, there have been several attempts to search for accretion disks around wide companions with radio interferometers such as the NOrthern Extended Millimeter Array (NOEMA) and ALMA (@B15 [@M17; @W17; @Wolff17; @R17]). None of these studies has successfully detected a wide companion disk. In Table \[tb0\] and Figure \[fig0\], we show our ALMA 1.3 mm survey for 5 systems, GQ Lup, CT Cha, DH Tau, GSC 6214-210, and 1RXS 1609. Data have been self-calibrated and CLEANed with natural weighting. Details on the observational setup and data reduction can be found in [@W17] and [@WS17]. As with other studies, no companions are detected in our survey. We note that a disk has been detected around the proposed planet-mass companion FW Tau C [@K14; @K15; @C15], but a recent dynamical mass measurement has shown that FW Tau C is a $\sim$0.1 $M_\sun$ star [@WS17].
[@lccccccccrc@]{} GQ Lup & 2015 Nov 01 & 41 & 85–14969 & $\sim$11 & J1534$-$3526 & J1337$-$1257 & J1427$-$4206 & $0.054\times0.031$ & 68.7 & $\sim$40\
DH Tau & 2016 Sep 14 & 36 & 15–3247 & $\sim$13 & J0433$+$2905 & J0510$+$1800 & J0510$+$1800 & $0.286\times0.153$ & $-$18.8 & $\sim$43\
GSC 6214-210 & 2016 Sep 16 & 36 & 15–3143 & $\sim$12 & J1634$-$2058 & J1517$-$2422 & J1517$-$2422 & $0.202\times0.172$ & $-$2.3 & $\sim$30\
1RXS 1609 & 2016 Sep 16 & 36 & 15–3143 & $\sim$12 & J1634$-$2058 & J1517$-$2422 & J1517$-$2422 & $0.202\times0.176$ & $-$1.5 & $\sim$30\
CT Cha & 2016 Sep 27 & 36 & 15–3247 & $\sim$14 & J1058$-$8003 & J1107$-$4449 & J1107$-$4449 & $0.259\times0.134$ & 16.5 & $\sim$50
[@lccccc@]{} Mass ($M_{\rm Jup}$) & $\sim$10–40 & $\sim$15 & $\sim$15 & $\sim$13 & $\sim$20\
SpT & L1 $\pm$ 1 & M9.25 $\pm$ 0.25 & M9.5 $\pm$1 & L2 $\pm$ 1 & $\sim$M8\
log($L/L_\sun$) & $-2.47\pm0.28$ & $-2.71\pm0.12$ & $-3.1\pm0.1$ & $-3.36\pm0.09$ & $-2.68\pm0.21$\
$T_{\rm eff}$ (K) & $2400\pm100$ & $2400\pm100$ & $2200\pm100$ & $2000\pm100$ & $2600\pm100$\
Age (Myr) & 2–5 & $\sim$2 & $\sim$10 & $\sim$10 & $\sim$2\
$D$ (pc) & $\sim$150 & $\sim$140 & $\sim$150 & $\sim$150 & $\sim$180\
$\rho$ () & 0.72 & 2.35 & 2.17 & 2.21 & 2.68\
\
Accretion and & H$\alpha$, Pa-$\beta$ & H$\alpha$, Pa-$\beta$ & H$\alpha$, Pa-$\beta$, Br-$\gamma$ & high $A_V$ & Pa-$\beta$\
disk markers & red $K'-L'$ & red $K'-L'$ & red $K'-L'$ & red $K'-L'$ & high $A_V$\
& && 24 excess & 24 excess &\
\
$\dot{M}$ ($M_{\rm Jup}$ yr$^{-1}$) & $5.3\times10^{-7}$ & $4.2\times10^{-9}$ & $1.3\times10^{-8}$ &$\cdots$&$\cdots$\
3$\sigma$ Flux Limit ($\mu$Jy) &150 (880 ) & 130 (1.3 mm) & 220 (880 ) & 90 (1.3 mm)& 150 (1.3 mm)\
&120 (1.3 mm) && 90 (1.3 mm) & &\
3$\sigma$ Dust Limit ($M_\oplus$) & $<$0.04 & $<$0.07 & $<$0.05 & $<$0.06 & $<$0.14\
References &1, 2, 3, 4 & 3, 4, 5, 6 & 3, 4, 7, 8, 9 & 4, 7, 10 & 11, 12
{width="0.33\linewidth"} {width="0.33\linewidth"}\
{width="0.33\linewidth"} {width="0.33\linewidth"} {width="0.33\linewidth"}
{width="\linewidth"}
**2. O D T P** {#o-d-t-p .unnumbered}
==============
Assuming optically thin dust, one can estimate the dust mass from [@H83], $$\label{eq1}
M_{\rm dust} = \frac{F_{\nu}~D^2}{\kappa_{\nu}~B_{\nu}(T)},$$ where $F_{\nu}$ is the observed flux, $D$ is the distance to the source, $\kappa_{\nu}$ is the dust opacity, and $B_{\nu}$ is the Planck function. In Table \[tb1\], we list the physical properties, disk evidence, 3$\sigma$ radio flux limits (from our observations and the literature), and the corresponding dust mass limits for 5 planetary-mass companions in our survey. We notice that the $\sim$5 $M_{\rm Jup}$ planet 2M1207 b has a 880 flux limit of $\sim$80 $\mu$Jy recently measured by [@R17]. Hence, current ALMA observations have reached 3$\sigma$ sensitivities of 80 to 220 $\mu$Jy at 880 and 90 to 150 $\mu$Jy at 1.3 mm for wide companions. These flux upper limits translate to a dust mass of 0.002 to 0.14 $M_\oplus$ (0.2 to 11.4 $M_{\rm moon}$) assuming a characteristic disk temperature of 20 K. Since many wide companions have multiple features suggestive of accretion disks, it is surprising and puzzling that radio observations have placed such a strong constraint on the amount of dust. The small amount of dust in turn implies a very short disk lifetime due to accretion. For instance, GQ Lup B has an accretion rate of $5\times10^{-7}~M_{\rm Jup}$ yr$^{-1}$ [@Z14], yet an optically thin disk has $<$0.04 $M_\oplus$ of dust [@M17]. Hence, GQ Lup B’s accretion disk could potentially disappear in $\sim$20 kyr, much shorter than the 2–5 Myr age of the system. If the actual gas-to-dust ratio is lower than the canonical value of 100, as hinted by recent ALMA surveys on T Tauri disks (e.g., @A16 [@E16]), then GQ Lup B’s disk would be depleted even faster.
This dramatic timescale difference seems to suggest that we are observing planet-mass companions at a very special time close to the very end of accretion, which is [*a priori*]{} unlikely. We note that at an age of $\sim$10 Myr, GSC 6214-210 B still has lines of evidence arguing for active accretion (see Table \[tb1\]). This implies that disks around planet-mass companions can indeed survive for a long period, presumably longer than the average lifetime of protoplanetary disks ($\sim$5 Myr; @H01).
**3. C O D** {#c-o-d .unnumbered}
============
Low radio flux does not necessarily mean low dust mass. A more natural explanation is that their disks may have the mass needed to sustain accretion and satellite formation for a few to even $>$10 Myr ($>$0.1 $M_{\rm Jup}$ or even $>$1 $M_{\rm Jup}$), but they appear faint in (sub)millimeter because the disks are compact. This is also hinted by infrared observations as some companions have near- or mid-infrared excesses likely from hot inner disks (Table \[tb1\]), implying that the lack of (sub)millimeter detections requires small disk radii.
Compact dust continuum emission is optically thick, rather than optically thin as usually assumed. We note that similar thoughts have been discussed and applied to derive an upper limit on disk size [@B15; @M17; @R17; @Wolff17]. Here, we further study the brightness and implications of compact optically-thick disks.
Assuming disk heating is dominated by the irradiation from the companion, we can roughly estimate the brightness of an optically-thick disk following [@P81]. The temperature profile is $$\label{eq2}
T(r) = T_\star~(r/R_\star)^{-3/4}~\big(1-\sqrt{R_\star/r}\big)^{1/4},$$ where $T_\star$ and $R_\star$ are the effective temperature and radius of the central object, respectively. For $r \gg R_\star$, $T(r)\propto r^{-3/4}$, which is similar to the profile seen in some simulated circumplanetary disks (e.g., @AB09). We can calculate the disk flux $F_\nu$ by integrating the Planck function $B_\nu$ over solid angle: $$\label{eq3}
F_\nu = \int B_\nu(T(r))\,d\Omega = \frac{2\pi}{D^2} \int^{R_{\rm max}}_{R_{\rm min}} B_\nu(T(r))\,r\,dr,$$ where $R_{\rm min}$ and $R_{\rm max}$ are the disk inner and outer radii, respectively. At long wavelengths, $B_\nu~\propto~T$, so $F_\nu$ approximately scales as $R_{\rm max}$.
We adopt $R_\star=2.5~R_{\rm Jup}$, which is typical for young substellar objects (e.g., @Baraffe15), $R_{\rm min} = 2~R_\star$, $R_{\rm max}$ = 10 to $10^4~R_{\rm Jup}$ ($\sim$0.005 to 5 au), and $D$ = 150 pc as most nearby star-forming regions are approximately at that distance. Finally, we choose $T_\star$ = 2500 K and 1000 K as most companions have spectral types from late M to mid L (Table \[tb1\]).
Figure \[fig1\] shows the disk radius versus the 880 (ALMA band 7) and 1.3 mm (ALMA band 6) disk fluxes given by Equation \[eq3\]. We also label the 3$\sigma$ flux upper limits from ALMA observations. We can see that compact optically-thick disks are indeed very faint at radio wavelengths — current observational constraints suggest that they are fainter than $\sim$100 $\mu$Jy. As a result, they cannot be as large as 1/3 of the Hill radius ($\sim$5 to 30 au in radius), or we would have easily detected them with ALMA. Instead, Figure \[fig1\] suggests that they are probably smaller than 1000 $R_{\rm Jup}$, or $\sim$0.5 au. Such small optically-thick disks can still contain all the gas and dust needed for a few Myr of steady accretion onto the companion, hence solving the lifetime problem created by the ALMA non-detections.
{width="\linewidth"}
**4. I** {#i-1 .unnumbered}
========
If these planet-mass companions do indeed have compact optically-thick disks, then our calculations have a few implications:
1\. ALMA may not be the ideal instrument to detect and characterize circumplanetary disks. As Figure \[fig1\] shows, disk flux roughly scales with disk size, so detecting a disk that is 10 times smaller in size would require $\sim$100 times longer integration to reach the same signal-to-noise ratio. Therefore, an unrealistically long integration is required for ALMA to reach an rms of 1 $\mu$Jy in order to detect disks smaller than 100 $R_{\rm Jup}$.
2\. Mid-infrared observations, which probe the peak of the spectral energy distribution (SED) of compact disks, may be more favorable for constraining disk sizes in order to compare with theories. The average temperature of compact optically-thick disks can be higher than that of T Tauri disks ($\sim$25 K; e.g., @A13). The area-weighted disk temperatures, $T_{\rm disk} = \int^{R_{\rm max}}_{R_{\rm min}} T(r)\,2\pi r\,dr / (\pi R_{\rm max}^2-\pi R_{\rm min}^2)$, for disk radii of 10–1000 $R_{\rm Jup}$ ($\sim$0.005–0.5 au) range from $\sim$40 K to $\sim$880 K for a 2500 K companion, and $\sim$20 K to $\sim$350 K for a 1000 K companion. As a result, compact disks are bright in the mid- to far-infrared. Figure \[fig2\] shows the SEDs derived from Equation \[eq3\] for 10, 100, and 1000 $R_{\rm Jup}$ compact disks. It is evident that the peak of SED is size-dependent, ranging from $\sim$10 to $\sim$100 . We also plot the 3$\sigma$ point source detection limits for 1000 s integration time for the Mid-Infrared Instrument (MIRI) on the [*James Webb Space Telescope (JWST)*]{}. MIRI’s superior sensitivity at 10 to 25 will be able to constrain disk sizes for most young planetary-mass companions.
3\. The origin of planetary-mass companions needs further exploration. Non-detections of wide companion disks are in stark contrast to recent studies which reveal that brown dwarf disks can have 0.1–1 $M_\oplus$ of dust and tens of au in radius (e.g., @Ricci14 [@V16; @H17]). Moreover, the disk around the 12 $M_{\rm Jup}$ free-floating planet OTS 44 has recently been imaged with ALMA and is found to have 0.1–0.6 $M_\oplus$ of dust [@B17]. Hence, even though planet-mass companions share similar masses to field brown dwarfs and free-floating planets, they probably have a different formation pathway that involves dynamical encounters with other massive bodies to truncate their disks and also scatter them to outer orbits (e.g., @B03). However, searches for inner massive bodies that can serve as the scatterers seem to disagree with the scattering scenario but favor the in situ formation via disk fragmentation or prestellar core collapse [@B16].
4\. Moons could be hard to form in compact optically-thick disks in the first few Myr because of the high disk temperature. In the inner disk, the temperature can be even higher than the dust sublimation temperature. As a result, satellite formation might have to wait until the companion cools off, or only happen in the outer disk where the temperature is lower (e.g., @Z16 [@S17]).
We thank the referee for helpful comments. Y.-L.W. is supported by the NSF AAG award and the TRIF fellowship. This paper makes use of the following ALMA data: ADS/JAO.ALMA\#2015.1.00773.S. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada), NSC and ASIAA (Taiwan), and KASI (Republic of Korea), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO, and NAOJ. The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.
Allard, F., Homeier, D., & Freytag, B. 2011, in ASP Conf. Ser. 448, XVI Cambridge Workshop on Cool Stars, Stellar Systems, and the Sun, ed. C. Johns-Krull, M. K. Browning, & A. A. West (San Francisco, CA: ASP), 91 Allers, K. N., & Liu, M. C. 2013, , 772, 79 Andrews, S. M., Rosenfeld, K. A., Kraus, A. L., & Wilner, D. J. 2013, , 771, 129 Ansdell, M., Williams, J. P., van der Marel, N., et al. 2016, , 828, 46 Ayliffe, B. A., & Bate, M. R. 2009, , 397, 657 Bailey, V., Hinz, P. M., Currie, T., et al. 2013, , 767, 31 Baraffe, I., Homeier, D., Allard, F., & Chabrier, G. 2015, , 577, A42 Barman, T. S., Macintosh, B., Konopacky, Q. M., & Marois, C. 2011, , 735, L39 Bate, M. R., Bonnell, I. A., & Bromm, V. 2003, , 339, 577 Bayo, A., Joergens, V., Liu, Y., et al. 2017, , 841, L11 Beckwith, S. V. W., Sargent, A. I., Chini, R. S., & Guesten, R. 1990, , 99, 924 Biller, B. A., Males, J. R., Rodigas, T. J., et al. 2014, , 792, L22 Bonnefoy, M., Chauvin, G., Lagrange, A.-M., et al. 2014, , 562, A127 Bowler, B. P., Andrews, S. M., Kraus, A. L., et al. 2015, , 805, L17 Bowler, B. P., Liu, M. C., Kraus, A. L., & Mann, A. W. 2014, , 784, 65 Bowler, B. P., Liu, M. C., Kraus, A. L., Mann, A. W., & Ireland, M. J. 2011, , 743, 148 Bryan, M. L., Bowler, B. P., Knutson, H. A., et al. 2016, , 827, 100 Brandt, T. D., McElwain, M. W., Turner, E. L., et al. 2014, , 794, 159 Caceres, C., Hardy, A., Schreiber, M. R., et al. 2015, , 806, L22 Eisner, J. A. 2015, , 803, L4 Eisner, J. A., Bally, J. M., Ginsburg, A., & Sheehan, P. D. 2016, , 826, 16 Follette, K. B., Rameau, J., Dong, R., et al. 2017, , 153, 264 Haisch, K. E., Lada, E. A., & Lada, C. J. 2001, , 553, L153 Hendler, N., Mulders, G. D., Pascucci, I., et al. 2017, , 841, 116 Hildebrand, R. H. 1983, QJRAS, 24, 267 Isella, A., Chandler, C. J., Carpenter, J. M., Pérez, L. M., & Ricci, L. 2014, , 788, 129 Kraus, A. L., Andrews, S. M., Bowler, B. P., et al. 2015, , 798, L23 Kraus, A. L., Ireland, M. J., Cieza, L. A., et al. 2014, , 781, 20 MacGregor, M. A., Wilner, D. J., Czekala, I., et al. 2017, , 835, 17 Males, J. R., Close, L. M., Morzinski, K., et al. 2014, , 786, 32 Manara, C. F., Prusti, T., Voirin, J., & Zari, E. 2017, arXiv:1707.03179 Nielsen, E. L., & Close, L. M. 2010, , 717, 878 Pringle, J. E. 1981, , 19, 137 Quillen, A. C., & Trilling, D. E. 1998, , 508, 707 Ricci, L., Cazzoletti, P., Czekala, I., et al. 2017, , 154, 24 Ricci, L., Testi, L., Natta, A., et al. 2014, , 791, 20 Sallum, S., Eisner, J. A., Close, L. M., et al. 2015, , 801, 85 Sallum, S., Follette, K. B., Eisner, J. A., et al. 2015, , 527, 342 Shabram, M., & Boley, A. C. 2013, , 767, 63 Szulágyi, L. 2017, , 842, 103 van der Plas, G., Ménard, F., Ward-Duong, K., et al. 2016, , 819, 102 Wolff, S. G., Ménard, F., Caceres. C., et al. 2017, , 154, 26 Wu, Y.-L., Close, L. M., Males, J. R., et al. 2015a, , 801, 4 Wu, Y.-L., Close, L. M., Males, J. R., et al. 2015b, , 807, L13 Wu, Y.-L., Sheehan, P. D., Males, J. R., et al. 2017, , 836, 223 Wu, Y.-L., & Sheehan, P. D. 2017, , 846, L26 Zhou, Y., Herczeg, G. J., Kraus, A. L., Metchev, S., & Cruz, K. L. 2014, , 783, L17 Zhu, Z. 2015, , 799, 16 Zhu, Z., Ju, W., & Stone, J. M. 2016, , 832, 193
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We determine the stability and instability of a sufficiently small and periodic traveling wave to long wavelength perturbations, for a nonlinear dispersive equation which extends a Camassa-Holm equation to include all the dispersion of water waves and the Whitham equation to include nonlinearities of medium amplitude waves. In the absence of the effects of surface tension, the result qualitatively agrees with the Benjamin-Feir instability of a Stokes wave. In the presence of the effects of surface tension, it qualitatively agrees with those from formal asymptotic expansions of the physical problem and it improves upon that for the Whitham equation, correctly predicting the limit of strong surface tension. We discuss the modulational stability and instability in the Camassa-Holm equation and related models.'
address: 'Department of Mathematics, University of Illinois at Urbana-Champaign, Urbana, IL 61801 USA'
author:
- Vera Mikyoung Hur
- 'Ashish K. Pandey'
bibliography:
- 'stability.bib'
title: |
Modulational instability in\
the full-dispersion Camassa-Holm equation
---
Introduction {#sec:intro}
============
In the 1960s, Whitham (see [@Whitham], for instance) proposed $$\label{E:whitham0}
\eta_t+c_{\rm ww}(\sqrt{\beta}|\partial_x|)\eta_x+(3\sqrt{(1+\alpha\eta)}-3)\eta_x=0,$$ to argue for wave breaking in shallow water. That is, the solution remains bounded but its slope becomes unbounded in finite time. Here $t\in\mathbb{R}$ is proportional to elapsed time, and $x\in\mathbb{R}$ is the spatial variable in the primary direction of wave propagation; $\eta=\eta(x,t)$ is the fluid surface displacement from the undisturbed depth, $$\alpha=\frac{\text{a typical amplitude}}{\text{the undisturbed fluid depth}}\quad\text{and}\quad
\beta=\frac{(\text{the undisturbed fluid depth})^2}{(\text{a typical wavelength})^2}.$$ Moreover, $c_{\rm ww}(|\partial_x|)$ is a Fourier multiplier operator, defined as $$\label{def:c1}
\widehat{c_{\rm ww}(|\partial_x|)f}({\kappa})=\sqrt{\frac{\tanh{\kappa}}{{\kappa}}}\widehat{f}({\kappa}).$$ Note that ${c_{\rm ww}}({\kappa})$ means the phase speed in the linear theory of water waves. For small amplitude waves satisfying $\alpha\ll1$, we may expand the nonlinearity of up to terms of order $\alpha$ to arrive at $$\label{E:whitham1}
\eta_t+c_{\rm ww}(\sqrt{\beta}|\partial_x|)\eta_x+\frac32\alpha\eta\eta_x=0.$$ For relatively shallow water or, equivalently, relatively long waves satisfying $\beta\ll1$, we may expand the right side of up to terms of order $\beta$ to find $$c_{\rm ww}(\sqrt{\beta}k)=1-\frac16\beta k^2+O(\beta^2).$$ Therefore, for small amplitude and long waves satisfying $\alpha=O(\beta)$ and $\beta\ll1$, we arrive at the famous Korteweg-de Vries equation $$\label{E:KdV}
\eta_t+\eta_x+\frac16\beta\eta_{xxx}+\frac32\alpha\eta\eta_x=0.$$ As a matter of fact, for well-prepared initial data, the solutions of the Whitham equation and the Korteweg-de Vries equation differ from those of the water wave problem merely by higher order terms over the relevant time scale; see [@Lannes], for instance, for details. But and offer improvements over for short waves. Whitham conjectured wave breaking for and . One of the authors [@Hur-breaking] recently proved this. In stark contrast, no solutions of break.
Moreover, Johnson and one of the authors [@HJ2] showed that a sufficiently small and $2\pi/{\kappa}$ periodic traveling wave of the Whitham equation be spectrally unstable to long wavelength perturbations, provided that ${\kappa}>1.145\dots$. In other words, (or ) and predict the Benjamin-Feir instability of a Stokes wave; see [@BF; @Benjamin1967; @Whitham1967] and [@BM1995], for instance. In contrast, periodic traveling waves of the Korteweg-de Vries equation are all modulationally stable. By the way, under the assumption that $\eta_t+\eta_x$ is small, we may modify to arrive at the Benjamin-Bona-Mahony equation $$\label{E:BBM}
\eta_t+\eta_x-\frac16\beta\eta_{xxt}+\frac32\alpha\eta\eta_x=0.$$ It agrees with for long waves but is preferable for short waves. Note that the phase speed for is bounded for all frequencies. The authors [@HP1] showed that a sufficiently small and $2\pi/{\kappa}$ periodic traveling wave of be modulationally unstable if ${\kappa}>\sqrt{3}$. Hence the Benjamin-Bona-Mahony equation seems to predict the Benjamin-Feir instability of a Stokes wave. But the instability mechanism is different from that in the Whitham equation or the water wave problem; see [@HP1] for details.
Furthermore, in the presence of the effects of surface tension, Johnson and one of the authors [@HJ3] determined the modulational stability and instability of a sufficiently small and periodic traveling wave of and $$\label{def:cT1}
\widehat{c_{\rm ww}(|\partial_x|;T)f}({\kappa})=\sqrt{(1+T{\kappa}^2)\frac{\tanh{\kappa}}{{\kappa}}}\widehat{f}({\kappa}),$$ where $T$ is the coefficient of surface tension. The result agrees by and large with those in [@Kawahara; @DR], for instance, from formal asymptotic expansions of the physical problem. But it fails to predict the limit of “strong surface tension." Perhaps, this is not surprising because neglects higher order nonlinearities of the water wave problem. It is interesting to find an equation, which predicts the modulational stability and instability of a gravity capillary wave. This is the subject of investigation here.
By the way, the authors [@HP2] recently extended the Whitham equation to include bidirectional propagation, and they showed that the “full-dispersion shallow water equations" correctly predict capillary effects on the Benjamin-Feir instability. But the modulation calculation is very lengthy and tedious. Here we seek higher order nonlinearities suitable for unidirectional propagation. Such a model is likely to be Hamiltonian and is potentially useful for other purposes.
As a matter of fact, for medium amplitude and long waves satisfying $\alpha=O(\sqrt{\beta})$ and $\beta\ll1$, the Camassa-Holm equations for the fluid surface displacement $$\label{E:CH1}
\eta_t+\eta_x+\beta(a\eta_{xxx}+b\eta_{xxt})+\frac32\alpha\eta\eta_x-\frac38\alpha^2\eta^2\eta_x+\frac{3}{16}\alpha^3\eta^3\eta_x=-\alpha\beta(c\eta\eta_{xxx}+d\eta_x\eta_{xx})\hspace*{-20pt}$$ and for the average horizontal velocity $$\label{E:CHu1}
u_t+u_x+\beta(au_{xxx}+bu_{xxt})+\frac32\alpha uu_x=-\alpha\beta(cuu_{xxx}+du_xu_{xx}),$$ where $$0{\leqslant}a{\leqslant}\frac16, \qquad b=a-\frac16,\qquad c=\frac32a+\frac16,\quad\text{and}\quad d=\frac92a+\frac{5}{24},$$ extend the Korteweg-de Vries equation to include higher order nonlinearities, and they approximate the physical problem; see [@Lannes], for instance, for details. In the case of $a=1/12$, reads $$\eta_t+\eta_x+\frac{1}{12}\beta(\eta_{xxx}-\eta_{xxt})+\frac32\alpha\eta\eta_x-\frac38\alpha^2\eta^2\eta_x+\frac{3}{16}\alpha^3\eta^3\eta_x=-\frac{7}{24}\alpha\beta(\eta\eta_{xxx}+2\eta_x\eta_{xx}),$$ which is particularly interesting because it predicts wave breaking; see [@Lannes] and references therein. Note that $$\frac{3\alpha\eta}{1+\sqrt{1+\alpha\eta}}=\frac32\alpha\eta-\frac38\alpha^2\eta^2+\frac{3}{16}\alpha^3\eta^3+O(\alpha^4).$$ Lannes [@Lannes] combined the dispersion relation of water waves and a Camassa-Holm equation, to propose the [*full-dispersion Camassa-Holm*]{} (FDCH) equation for the fluid surface displacement $$\label{E:FDCH1}
\eta_t+c_{\rm ww}(\sqrt{\beta}|\partial_x|)\eta_x+\frac{3\alpha\eta}{1+\sqrt{1+\alpha\eta}}\eta_x
=-\alpha\beta\Big(\frac{5}{12}\eta\eta_{xxx}+\frac{23}{24}\eta_x\eta_{xx}\Big),$$ where ${c_{\rm ww}}(|\partial_x|)$ is in , or in the presence of the effects of surface tension. For relatively long waves satisfying $\beta\ll1$, and agree with , where $a=1/6$, up to terms of order $\beta$. But, including all the dispersion of water waves, and may offer an improvement over for short waves. For small amplitude waves satisfying $\alpha\ll1$, agrees with up to terms of order $\alpha$. But, including higher order nonlinearities, may offer an improvement over for medium amplitude waves. For the average horizontal velocity, we may combine , or in the presence of the effects of surface tension, and to introduce $$\label{E:FDCHu1}
u_t+c_{\rm ww}(\sqrt{\beta}|\partial_x|)u_x+\frac32\alpha uu_x
=-\alpha\beta\Big(\frac{5}{12}uu_{xxx}+\frac{23}{24}u_xu_{xx}\Big).$$
We follow along the same line as the arguments in [@HJ2; @HJ3; @HP1] (see also [@BHJ]) and investigate the modulational stability and instability in the FDCH equation. A main difference lies in that the nonlinearities of involve higher order derivatives and, hence, a periodic traveling wave is not a priori smooth. We examine the mapping properties of various operators to construct a smooth solution.
In the absence of the effects of surface tension, we show that a sufficiently small and $2\pi/{\kappa}$ periodic traveling wave of and is spectrally unstable to long wavelength perturbations, provided that $${\kappa}>1.420\dots,$$ and stable to square integrable perturbations otherwise. The result qualitatively agrees with the Benjamin-Feir instability of a Stokes wave (see [@BF; @Benjamin1967; @Whitham1967], for instance) and that for the Whitham equation (see [@HJ2]). The critical wave number compares reasonably well with $1.363\dots$ in the Benjmain-Feir instability. Including the effects of surface tension, in the ${\kappa}$ and ${\kappa}\sqrt{T}$ plane, we determine the regions of modulational stability and instability for a sufficiently small and periodic traveling wave of and ; see Figure \[f:ST\] for details. The result qualitatively agrees with those in [@Kawahara; @DR], for instance, from formal asymptotic expansions of the physical problem, and it improves upon that in [@HJ3] for the Whitham equation. In particular, the limit of ${\kappa}(T)\sqrt{T}\to1.283\dots$ as $T\to \infty$, where ${\kappa}(T)$ is a critical wave number, whereas the limit is unbounded for the Whitham equation (see [@HJ3]).
Moreover, we show that a sufficiently small and $2\pi/{\kappa}$ periodic traveling wave of is modulationally unstable if ${\kappa}>6$. To the best of the authors’ knowledge, this is new. The Camassa-Holm equation seems to predict the Benjamin-Feir instability of a Stokes wave. But the instability mechanism is different from that in and , or the water wave problem. One may use the Evans function and other ODE methods to determine the modulational stability and instability for all amplitudes. This is an interesting direction of future research. The result herein indicates that the stability and instability depend on the carrier wave.
In the absence of the effects of surface tension, we show that a sufficiently small and $2\pi/{\kappa}$ periodic traveling wave of and is modulationally unstable if ${\kappa}$ is greater than a critical value, similarly to the Benjamin-Feir instability. But, in the presence of the effects of surface tension, the modulational stability and instability in and qualitatively agree with that in the Whitham equation (see [@HJ3]). In particular, it fails to predict the limit of strong surface tension. Therefore, we learn that the higher power nonlinearities of improve the result, not the higher derivative nonlinearities.
It is interesting to explain breaking, peaking, and other phenomena of water waves in and (or ).
### Notation {#notation .unnumbered}
The notation to be used is mostly standard, but worth briefly reviewing. Let $\mathbb{T}$ denote the unit circle in $\mathbb{C}$. We identify functions over $\mathbb{T}$ with $2\pi$ periodic functions over $\mathbb{R}$ via $f(e^{iz})=F(z)$ and, for simplicity of notation, we write $f(z)$ rather than $f(e^{iz})$. For $p$ in the range $[1,\infty]$, let $L^p(\mathbb{T})$ consist of real or complex valued, Lebesgue measurable, and $2\pi$ periodic functions over $\mathbb{R}$ such that $$\|f\|_{L^p(\mathbb{T})}:=\Big(\frac{1}{2\pi}\int^\pi_{-\pi}|f(z)|^p~dz\Big)^{1/p}<\infty
\qquad\text{if\quad$p<\infty$},$$ and $\|f\|_{L^\infty(\mathbb{T})}:=\text{ess\,sup}_{-\pi<z{\leqslant}\pi}|f(z)|<\infty$ if $p=\infty$. Let $H^1(\mathbb{T})$ consist of $L^2(\mathbb{T})$ functions whose derivatives are in $L^2(\mathbb{T})$. Let $H^\infty(\mathbb{T})=\bigcap_{k=0}^\infty H^k(\mathbb{T})$.
For $f \in L^1(\mathbb{T})$, the Fourier series of $f$ is defined by $$\sum_{n\in \mathbb{Z}}\widehat{f}(n)e^{inz}, \qquad\text{where}\quad
\widehat{f}(n)=\frac{1}{2\pi}\int^\pi_{-\pi}f(z)e^{-inz}~dz.$$ If $f\in L^2(\mathbb{T})$ then its Fourier series converges to $f$ pointwise almost everywhere. We define the $L^2(\mathbb{T})$ inner product as $$\label{def:prod1}
\langle f_1,f_2\rangle_{L^2(\mathbb{T})}=\frac{1}{2\pi}\int^\pi_{-\pi} f_1(z)f_2^*(z)~dz
=\sum_{n\in\mathbb{Z}}\widehat{f_1}(n)\widehat{f_2}^*(n).$$
Sufficiently small and periodic traveling waves {#sec:periodic}
===============================================
We determine periodic traveling waves of the FDCH equation, after normalization of parameters, $$\label{E:FDCH}
\eta_t+c_{\rm ww}(|\partial_x|;T)\eta_x+\frac{3\eta}{1+\sqrt{1+\eta}}\eta_x
=-\Big(\frac{5}{12}\eta\eta_{xxx}+\frac{23}{24}\eta_x\eta_{xx}\Big),$$ where ${c_{\rm ww}}(|\partial_x|;T)$ is in , and we calculate their small amplitude expansion.
Properties of ${c_{\rm ww}}(\cdot\,;T)$ {#properties-of-c_rm-wwcdott .unnumbered}
---------------------------------------
For any $T{\geqslant}0$, ${c_{\rm ww}}(\cdot\,;T)$ is even and real analytic, and ${c_{\rm ww}}(0;T)=1$. Note that ${c_{\rm ww}}(|\partial_x|;0):H^s(\mathbb{R})\to H^{s+1/2}(\mathbb{R})$ for any $s\in\mathbb{R}$, and for $T>0$, ${c_{\rm ww}}(|\partial_x|;T):H^{s+1/2}(\mathbb{R})\to H^{s}(\mathbb{R})$; see [@HJ2; @HJ3], for instance, for details.
Note that ${c_{\rm ww}}(\cdot;0)$ decreases to zero monotonically away from the origin. For $T{\geqslant}1/3$, ${c_{\rm ww}}(\cdot\,;T)$ increases monotonically and unboundedly away from the origin. For $0<T<1/3$, on the other hand, ${c_{\rm ww}}'(0;T)=0$, ${c_{\rm ww}}''(0;T)<0$ and ${c_{\rm ww}}({\kappa};T)\to\infty$ as ${\kappa}\to\infty$. Hence ${c_{\rm ww}}(\cdot\,;T)$ possesses a unique minimum over the interval $(0,\infty)$; see Figure \[fig:cT\].
(a) ![Schematic plots of ${c_{\rm ww}}(\cdot\,;T)$ for (a) $T=0$, (b) $T{\geqslant}1/3$, and (c) $0<T<1/3$.[]{data-label="fig:cT"}](ck.pdf "fig:")(b) ![Schematic plots of ${c_{\rm ww}}(\cdot\,;T)$ for (a) $T=0$, (b) $T{\geqslant}1/3$, and (c) $0<T<1/3$.[]{data-label="fig:cT"}](ck1.pdf "fig:")(c) ![Schematic plots of ${c_{\rm ww}}(\cdot\,;T)$ for (a) $T=0$, (b) $T{\geqslant}1/3$, and (c) $0<T<1/3$.[]{data-label="fig:cT"}](ck2.pdf "fig:")
By a traveling wave of and , we mean a solution of the form $\eta(x,t)=\eta(x-ct)$ for some $c>0$, the wave speed, where $\eta$ satisfies by quadrature $$({c_{\rm ww}}(|\partial_x|;T)-c-3)\eta+2(1+\eta)^{3/2}-2+\frac{5}{12}\eta \eta_{xx}+\frac{13}{48} \eta_x^2 = (1-c)^2 b$$ for some $b\in\mathbb{R}$. We seek a periodic traveling wave of and . That is, $\eta$ is a $2\pi$ periodic function of $z:={\kappa}x$ for some ${\kappa}>0$, the wave number, and it satisfies $$\label{E:pFDCH}
({c_{\rm ww}}({\kappa}|\partial_z|;T) -c -3)\eta +2(1+\eta)^{3/2}-2+\frac{5}{12}{\kappa}^2\eta\eta_{zz}+\frac{13}{48}{\kappa}^2\eta_z^2 =(1-c)^2b.$$ Note that $$\label{E:c-Hk}
{c_{\rm ww}}({\kappa}|\partial_z|;0): H^s(\mathbb{T})\to H^{s+1/2}(\mathbb{T}),\quad\text{and}\quad
{c_{\rm ww}}({\kappa}|\partial_z|;T): H^{s+1/2}(\mathbb{T})\to H^{s}(\mathbb{T})$$ for $T>0$, for any ${\kappa}>0$ and $s\in\mathbb{R}$. Note that $$\label{def:ck}
{c_{\rm ww}}({\kappa}|\partial_z|;T)e^{inz}={c_{\rm ww}}(n{\kappa};T)e^{inz}\quad \text{for $n\in\mathbb{Z}$}.$$
Note that remains invariant under $$\label{E:invariance}
z\mapsto z+z_0\quad\text{and}\quad z\mapsto -z$$ for any $z_0\in \mathbb{R}$. Hence we may assume that $\eta$ is even. But does not possess scaling invariance. Hence we may not a priori assume that ${\kappa}=1$. Rather, the (in)stability result herein depends on the carrier wave number. Moreover, does not possess Galilean invariance. Hence we may not a priori assume that $b=0$. Rather, we exploit the variation of in the $b$ variable in the instability proof. To compare, the Whitham equation for periodic traveling waves possesses Galilean invariance; see [@HJ3], for instance.
We follow along the same line as the arguments in [@HJ2; @HJ3; @HP1], for instance, to construct periodic traveling waves of and . A main difference lies in the lack of a priori smoothness of solutions of . We examine the mapping properties of various operators to construct smooth solutions.
For any $T{\geqslant}0$ and an integer $k{\geqslant}0$, let $$F:H^{k+2}(\mathbb{T})\times\mathbb{R}_+\times \mathbb{R}\times\mathbb{R}_+\to H^k(\mathbb{T})$$ denote $$\begin{gathered}
\label{def:F}
F(\eta,c;b,{\kappa},T)=({c_{\rm ww}}({\kappa}|\partial_z|;T) -c -3)\eta +2(1+\eta)^{3/2}-2\\
\quad+\frac{5}{12} {\kappa}^2 \eta \eta_{zz}+\frac{13}{48} {\kappa}^2 \eta_z^2 -(1-c)^2 b.\end{gathered}$$ It is well defined by and a Sobolev inequality. We seek a solution $\eta\in H^{k+2}(\mathbb{T})$, $c>0$, and $b\in \mathbb{R}$ of $$\label{E:F=0}
F(\eta,c;b, {\kappa},T)=0.$$ Since $k$ is arbitrary, $\eta\in H^\infty(\mathbb{T})$. Note that $F$ is invariant under . Hence we may assume that $\eta$ is even.
For any $T{\geqslant}0$, and $c>0$, $b\in\mathbb{R}$, ${\kappa}>0$, note that $$\begin{gathered}
F_\eta(\eta,c; b, {\kappa},T)\zeta=\Big({c_{\rm ww}}({\kappa}|\partial_z|;T)-c-3+3(1+\eta)^{1/2}\\
+{\kappa}^2\Big(\frac{5}{12}(\eta_{zz}+\eta \partial_z^2)+\frac{13}{24}\eta_z \partial_z\Big)\Big)\zeta:
H^{k+2}(\mathbb{T})\to H^k(\mathbb{T})\end{gathered}$$ is continuous by and a Sobolev inequality. Here a subscript means Fréchet differentiation. Moreover, for any $T{\geqslant}0$, and $\eta\in H^{k+2}(\mathbb{T})$, ${\kappa}>0$, $b\in\mathbb{R}$, note that $F_c(\eta;{\kappa},c,b)=-\eta+2(1-c)b:\mathbb{R}\to H^k(\mathbb{T})$ is continuous. Since $F_b(\eta;{\kappa},c,b)=-(1-c)^2$ and $$F_{\kappa}(\eta;{\kappa},c,b):={c_{\rm ww}}'({\kappa}|\partial_z|;T) \eta + \frac{5}{6} {\kappa}\eta\eta_{zz}+\frac{13}{24} {\kappa}\eta_z^2$$ are continuous likewise, $F$ depends continuously differentiably on its arguments. Furthermore, since the Fréchet derivatives of $F$ with respect to $\eta$, and $c$, $b$ of all orders ${\geqslant}3$ are zero everywhere by brutal force, and since ${c_{\rm ww}}$ is a real analytic function, $F$ is a real analytic operator.
Bifurcation condition {#bifurcation-condition .unnumbered}
---------------------
For any $T{\geqslant}0$, ${\kappa}>0$, for any $c>0$, $b\in \mathbb{R}$ and $|b|$ sufficiently small, note that $$\label{E:eta0'}
\eta_0(c;b,{\kappa},T)=b(1-c)+O(b^2)$$ makes a constant solution of - and, hence, . It follows from the implicit function theorem that if non-constant solutions of - and, hence, bifurcate from $\eta=\eta_0$ for some $c=c_0$ then, necessarily, $$L_0:=F_\eta(\eta_0,c_0;b,{\kappa},T):H^{k+2}(\mathbb{T}) \to H^k(\mathbb{T}),$$ where $$\label{def:L0}
L_0={c_{\rm ww}}({\kappa}|\partial_z|;T) -c_0 -3 +3(1+\eta_0)^{1/2}+\frac{5}{12}{\kappa}^2\eta_0\partial_z^2,$$ is not an isomorphism. Here $\eta_0$ depends on $c_0$. But we suppress it for simplicity of notation. A straightforward calculation reveals that $L_0e^{inz}=0$, $n\in\mathbb{Z}$, if and only if $$\label{E:bifurcation}
c_0={c_{\rm ww}}(n{\kappa};T)-3+3(1+\eta_0)^{1/2}-\frac{5}{12}{\kappa}^2 n^2\eta_0.$$ For $b=0$ and, hence, $\eta_0=0$ by , it simplifies to $c_0={c_{\rm ww}}(n{\kappa};T)$. Without loss of generality, we restrict the attention to $n=1$. For $|b|$ sufficiently small, and become $$\begin{aligned}
c_0(b,{\kappa},T)=&{c_{\rm ww}}({\kappa};T)+b\Big(\frac32-\frac{5}{12}{\kappa}^2\Big)(1-c_{\rm ww}({\kappa};T))+O(b^2)\label{E:c0}
\intertext{and}
\eta_0(b,{\kappa},T)=&b(1-c_{\rm ww}({\kappa};T))+O(b^2).\label{E:eta0}\end{aligned}$$ For $T=0$, since ${c_{\rm ww}}({\kappa};0)>{c_{\rm ww}}(n{\kappa};0)$ for $n=2,3,\dots$ everywhere in $\mathbb{R}$ (see Figure \[fig:cT\]a), it is straightforward to verify that for any ${\kappa}>0$, $b\in\mathbb{R}$ and $|b|$ sufficiently small, the kernel of $L_0:H^{k+2}(\mathbb{T})\to H^k(\mathbb{T})$ is two dimensional and spanned by $e^{\pm iz}$. Moreover, the co-kernel of $L_0$ is two dimensional. Therefore, $L_0$ is a Fredholm operator of index zero.
Similarly, for $T{\geqslant}1/3$, since ${c_{\rm ww}}({\kappa};T)<{c_{\rm ww}}(n{\kappa};T)$ for $n=2,3,\dots$ everywhere in $\mathbb{R}$ (see Figure \[fig:cT\]b), for any ${\kappa}>0$, $b\in\mathbb{R}$ and $|b|$ sufficiently small, $L_0:H^{k+2}(\mathbb{T})\to H^k(\mathbb{T})$ is a Fredholm operator of index zero, whose kernel is two dimensional and spanned by $e^{\pm iz}$.
For $0<T<1/3$, on the other hand, for any integer $n{\geqslant}2$, it is possible to find some ${\kappa}$ such that ${c_{\rm ww}}({\kappa};T)={c_{\rm ww}}(n{\kappa};T)$ (see Figure \[fig:cT\]c). If $$\label{A:resonance}
{c_{\rm ww}}({\kappa};T)\neq{c_{\rm ww}}(n{\kappa};T)\qquad\text{for any $n=2,3,\dots$}$$ then $L_0:H^{k+2}(\mathbb{T})\to H^k(\mathbb{T})$ is likewise a Fredholm operator of index zero, whose kernel is two dimensional and spanned by $e^{\pm iz}$. But if ${c_{\rm ww}}({\kappa};T)={c_{\rm ww}}(n{\kappa};T)$ for some integer $n{\geqslant}2$, resulting in the resonance of the fundamental mode and the $n$-th harmonic, then the kernel is four dimensional. One may follow along the same line as the argument in [@Jones], for instance, to construct a periodic traveling wave. But we do not pursue this here.
Lyapunov-Schmidt procedure {#lyapunov-schmidt-procedure .unnumbered}
--------------------------
For any $T{\geqslant}0$, ${\kappa}>0$ satisfying , $b\in\mathbb{R}$ and $|b|$ sufficiently small, we employ a Lyapunov-Schmidt procedure to construct non-constant solutions of - and, hence, bifurcating from $\eta=\eta_0$ and $c=c_0$, where $\eta_0$ and $c_0$ are in and . Throughout the proof, $T$, ${\kappa}$, and $b$ are fixed and suppressed for simplicity of notation.
Recall that $F(\eta_0,c_0)=0$, where $F$ is in , and $L_0e^{\pm iz}=0$, where $L_0$ is in . We write that $$\label{def:eta2}
\eta(z)=\eta_0+\frac12(ae^{iz}+a^*e^{-iz})+\eta_r(z)\quad\text{and}\quad c=c_0+c_r,$$ and we require that $a\in \mathbb{C}$, $\eta_r\in H^{k+2}(\mathbb{T})$ be even and $$\label{E:v-condition}
{\langle}\eta_r, e^{\pm iz}{\rangle}_{L^2(\mathbb{T})}=0,$$ and $c_r\in \mathbb{R}$. Substituting into -, we use $F(\eta_0,c_0)=0$, $L_0e^{\pm iz}=0$, and we make an explicit calculation to arrive at $$\begin{aligned}
L_0\eta_r=&(3(1+\eta_0)^{1/2}+c_r)\Big(\frac12(ae^{iz}+a^*e^{-iz})+\eta_r\Big)\notag \\
&-2\Big(1+\eta_0+\frac12(ae^{iz}+a^*e^{-iz})+\eta_r\Big)^{3/2} \notag\\
&-\frac{13}{48}{\kappa}^2 \Big(\frac{i}{2}(ae^{iz}-a^*e^{-iz})+\eta_r'\Big)^2 \notag\\
&-\frac{5}{12}{\kappa}^2\Big(\frac12(ae^{iz}+a^*e^{-iz})+\eta_r\Big)\Big(-\frac12(ae^{iz}+a^*e^{-iz})+\eta_r''\Big)\notag\\
=:&g(\eta_r;a,a^*,c_r). \label{def:g}\end{aligned}$$ Here and elsewhere, the prime means ordinary differentiation. Note that $$g:H^{k+2}(\mathbb{T})\times{\mathbb{C}}\times {\mathbb{C}}\times {\mathbb{R}}\to H^k(\mathbb{T}).$$ Recall that $F$ is a real analytic operator. Hence $g$ depends analytically on its arguments. Clearly, $g(0;0,0,c_r)=0$ for all $c_r\in\mathbb{R}$.
Let $\Pi:L^2(\mathbb{T})\to \ker L_0$ denote the spectral projection, defined as $$\Pi f(z)=\widehat{f}(1)e^{iz}+\widehat{f}(-1)e^{-iz}.$$ Since $\Pi\eta_r=0$ by , we may rewrite as $$\label{E:LS1}
L_0\eta_r=(1-\Pi)g(\eta_r;a,a^*,c_r)\quad \text{and} \quad 0=\Pi g(\eta_r;a,a^*,c_r).$$ Moreover, for any $T{\geqslant}0$, and ${\kappa}>0$ satisfying , note that $L_0$ is invertible on $(1-\Pi)H^k(\mathbb{T})$. Specifically, $${L_0}^{-1}f(z)=\sum_{n\neq \pm 1}\frac{\widehat{f}(n)}{{c_{\rm ww}}({\kappa}n;T)-{c_{\rm ww}}({\kappa};T)+\frac{5}{12}{\kappa}^2 \eta_0(1- n^2)}e^{inz}.$$ Hence we may rewrite as $$\label{E:LS2}
\eta_r=L_0^{-1}(I-\Pi)g(\eta_r;a,a^*,c_r)\quad \text{and}\quad 0=\Pi g(\eta_r;a,a^*,c_r).$$ Note that $L_0^{-1}: (1-\Pi)H^k(\mathbb{T}) \to H^k(\mathbb{T})$ is bounded. We claim that $$L_0^{-1}: (1-\Pi)H^k(\mathbb{T}) \to H^{k+2}(\mathbb{T})$$ is bounded. As a matter of fact, $$\left|\frac{n^2\widehat{f}(n)}{{c_{\rm ww}}({\kappa}n;T)-{c_{\rm ww}}({\kappa};T)+\frac{5}{12}{\kappa}^2 \eta_0(1- n^2)}\right|{\leqslant}C|\widehat{f}(n)|$$ for some constant $C>0$ for $n\in\mathbb{Z}$ and $|n|$ sufficiently large. Therefore, for any $a, a^*\in\mathbb{C}$ and $c_r\in\mathbb{R}$, $$L_0^{-1}(1-\Pi)g:H^{k+2}(\mathbb{T})\to H^{k+2}(\mathbb{T})$$ is bounded. Note that it depends analytically on its argument. Since $g(0;0,0,c_r)=0$ for any $c_r\in\mathbb{R}$, it follows from the implicit function theorem that a unique solution $$\eta_2=\eta_r(a,a^*,c_r)$$ exists to the former equation of near $\eta_r=0$ for $a\in\mathbb{C}$ and $|a|$ sufficiently small for any $c_r\in\mathbb{R}$. Note that $\eta_2$ depends analytically on its arguments and it satisfies for $|a|$ sufficiently small for any $c_r\in\mathbb{R}$. The uniqueness implies $$\label{Vdef}
\eta_2(0,0,c_r)=0 \qquad\text{for any $c_r\in \mathbb{R}$}.$$ Moreover, since - and, hence, are invariant under for any $z_0\in\mathbb{R}$, it follows that $$\begin{aligned}
\eta_2(a,a^*,c_r)(z+z_0)=\eta_2(ae^{iz_0},a^*e^{-iz_0},c_r)\quad\text{and}\quad
\eta_2(a,a^*,c_r)(-z)=\eta_2(a,a^*,c_r)(z) \label{Vprop}\end{aligned}$$ for any $z_0\in\mathbb{R}$ for any $a\in\mathbb{C}$ and $|a|$ sufficiently small, and $c_r\in\mathbb{R}$.
To proceed, we rewrite the latter equation in as $$\Pi g(\eta_2(a,a^*,c_r);a,a^*,c_r)=0$$ for $a\in\mathbb{C}$ and $|a|$ sufficiently small for $c_r\in\mathbb{R}$. This is solvable, provided that $$\label{def:pi}
\pi_{\pm}(a,a^*,c_r):=\left{\langle}g(\eta_2(a,a^*,c_r);a,a^*,c_r),ae^{iz}\pm a^*e^{-iz}\right{\rangle}_{L^2(\mathbb{T})}=0.$$ We use , where $z_0=-2\arg (a)$, and to show that $$\pi_- (a^*,a,c_r)=\pi_- (a,a^*,c_r)=-\pi_- (a^*,a,c_r).$$ Hence $\pi_- (a,a^*,c_r)=0$ holds for any $a\in\mathbb{C}$ and $|a|$ sufficiently small for any $c_r\in\mathbb{R}$. Moreover, we use , where $z_0=-\arg (a)$, and to show that $$\pi_+ (a,a^*,c_r)=\pi_+ (|a|,|a|,c_r).$$ Hence it suffices to solve $\pi_+ (a,a,c_r)=0$ for any $a, c_r\in \mathbb{R}$ and $|a|$ sufficiently small.
Substituting into , where $\eta_r=\eta_2(a,a,c_r)$, we make an explicit calculation to arrive at $$\pi_+ (a,a,c_r)=a^2(\pi c_r+\pi_r(a,c_r)),$$ where $$\begin{aligned}
\pi_r(a,c_r)=&-2a^{-1} {\langle}(1+\eta_0+a\cos z+\eta_2(a,a,c_r)(z))^{3/2},\cos z{\rangle}\\
&-\frac{5}{12}{\kappa}^2({\langle}\eta_2''(a,a,c_r)(z) - \eta_2(a,a,c_r)(z),\cos^2 z{\rangle}-a^{-1}{\langle}\eta_2\eta_2''(a,a,c_r)(z), \cos z{\rangle})\\
& -\frac{13}{48}{\kappa}^2 a^{-1}( {\langle}\eta_2'(a,a,c_r)(z)^2, \cos z{\rangle}-{\langle}\eta_2'(a,a,c_r)(z),\sin 2z{\rangle}),\end{aligned}$$ and ${\langle}\cdot\,,\cdot{\rangle}$ means the $L^2(\mathbb{T})$ inner product. We merely pause to remark that $\pi_r$ is well defined. As a matter of fact, $a^{-1}\eta_2$ is not singular for $a\in\mathbb{R}$ and $|a|$ sufficiently small by . Clearly, $\pi_r$ and, hence, $\pi_{\pm}$ depend analytically on its arguments. Since $\pi_r(0,0)=\partial \pi_r/\partial c_r(0,0)=0$ by , it follows from the implicit function theorem that a unique solution $$c_r=c_1(a)$$ exists to $\pi_+(a,a,c_r)=0$ and, hence, the latter equation of near $c_r=0$ for $a\in\mathbb{R}$ and $|a|$ sufficiently small. Clearly, $c_1$ depends analytically on $a$.
To recapitulate, $$\eta_r=\eta_2(a,a,c_1(a))\quad\text{and}\quad c_r=c_1(a)$$ uniquely solve for $a\in\mathbb{R}$ and $|a|$ sufficiently small, and by virtue of , $$\label{E:soln}
\eta(a)(z)=\eta_0+a\cos z+\eta_2(a,a,c_1(a))(z)\quad\text{and}\quad c(a)=c_0+c_1(a)$$ uniquely solve - and, hence, for $a\in\mathbb{R}$ and $|a|$ sufficiently small. Note that $\eta$ is $2\pi$ periodic and even in $z$. Moreover, $\eta \in H^\infty(\mathbb{T})$.
For $a,b\in\mathbb{R}$ and $|a|,|b|$ sufficiently small, we write that $$\begin{aligned}
\eta(a;b,{\kappa},T)(z):=&\eta_0(b,{\kappa},T)+ a \cos z+a^2\eta_2(z)+a^3\eta_3(z)+\cdots\label{E:eta(a,b)}
\intertext{and}
c(a;b,{\kappa},T):=&c_0(b,{\kappa},T)+ac_1+a^2c_2+\cdots,\label{E:c(a,b)}\end{aligned}$$ where $\eta_2,\eta_3, \dots$ are $2\pi$ periodic, even, and smooth functions of $z$, and $c_1, c_2, \dots \in \mathbb{R}$.
We claim that $c_1=0$. As a matter of fact, note that and, hence, - remain invariant under $z\mapsto z+\pi$ by . Since $\partial \eta/\partial a(0)(z)=\cos z$, however, $\eta(z)\neq\eta(z+\pi)$ must hold. Thus $\partial c/\partial a(0)=0$. This proves the claim. If ${\langle}\eta_{j-1},\eta_j{\rangle}_{L^2(\mathbb{T})}=0$ for any integer $j{\geqslant}1$, in addition, then $c_{2j-1}=0$ for any integer $j{\geqslant}1$. Hence $c$ is even in $a$.
Substituting and into , we may calculate the small amplitude expansion. The proof is very similar to that in [@HP1], for instance. Hence we omit the details.
Below we summarize the conclusion.
\[lem:existence\] For any $T{\geqslant}0$, ${\kappa}>0$ satisfying , $b\in\mathbb{R}$ and $|b|$ sufficiently small, a one parameter family of solutions of exists, denoted $\eta(a;b,{\kappa},T)$ and $c(a;b,{\kappa},T)$, for $a\in\mathbb{R}$ and $|a|$ sufficiently small; $\eta\in H^\infty(\mathbb{T})$ and it is even in $z$; $\eta$ and $c$ depend analytically on $a$, and $b$, ${\kappa}$. Moreover, $$\begin{aligned}
\eta(a;b,{\kappa},T)(z)=&b(1-c_{\rm ww}({\kappa};T))+a\cos z+a^2(h_0+h_2\cos 2z)+O(a(a+b)^2),\hspace*{-20pt}\label{E:n(k,a,b)} \\
c(a;b,{\kappa},T)=&c_{\rm ww}({\kappa};T)+b\Big(\frac32-\frac{5}{12}{\kappa}^2\Big)(1-c_{\rm ww}({\kappa};T))+a^2 c_2+O(a(a+b)^2) \hspace*{-20pt}\label{E:c(k,a,b)}\end{aligned}$$ as $a,b \to 0$, where $$\begin{gathered}
h_0=\Big(\frac38-\frac{7}{96}{\kappa}^2\Big)\frac{1}{c_{\rm ww}({\kappa};T)-1},\qquad
h_2 =\Big(\frac38-\frac{11}{32}{\kappa}^2\Big)\frac{1}{c_{\rm ww}({\kappa};T)-c_{\rm ww}(2{\kappa};T)}, \label{def:A0A2}
\intertext{and}
c_2=\Big(\frac32-\frac{5}{12}{\kappa}^2\Big)h_0+\Big(\frac34-\frac12{\kappa}^2\Big)h_2-\frac{3}{32}. \label{def:c2}\end{gathered}$$
Modulational instability index {#sec:MI}
==============================
For $T{\geqslant}0$, ${\kappa}>0$ satisfying , $a, b\in\mathbb{R}$ and $|a|, |b|$ sufficiently small, let $\eta=\eta(a;b,{\kappa},T)$ and $c=c(a;b,{\kappa},T)$, denote a sufficiently small and $2\pi/{\kappa}$ periodic traveling wave of and , whose existence follows from the previous section. We address its modulational stability and instability.
Linearizing about $\eta$ in the coordinate frame moving at the speed $c$, we arrive at $$\zeta_t+{\kappa}\partial_z\Big({c_{\rm ww}}({\kappa}|\partial_z|;T) -c-3+3(1+\eta)^{1/2}+ {\kappa}^2\Big(\frac{5}{12}(\eta \partial_z^2+\eta_{zz})+\frac{13}{24} \eta_z \partial_z\Big)\Big)\zeta=0,$$ where ${c_{\rm ww}}({\kappa}|\partial_z|;T)$ is in . Seeking a solution of the form $\zeta(z,t)=e^{\lambda {\kappa}t}\zeta(z)$, $\lambda\in\mathbb{C}$, we arrive at $$\begin{aligned}
\label{E:eigen}
\lambda \zeta&=\partial_z\Big(-{c_{\rm ww}}({\kappa}|\partial_z|;T)+c+3-3(1+\eta)^{1/2}- {\kappa}^2\Big(\frac{5}{12}(\eta \partial_z^2+\eta_{zz})+\frac{13}{24} \eta_z \partial_z\Big)\Big)\zeta \\
&=:\mathcal{L}(a;b,{\kappa},T)\zeta. \notag\end{aligned}$$ We say that $\eta$ is [*spectrally unstable*]{} to square integrable perturbation if the $L^2(\mathbb{R})$ spectrum of ${\mathcal{L}}$ intersects the open right-half plane of $\mathbb{C}$, and it is spectrally stable otherwise. Note that $\eta$ is $2\pi$ periodic in $z$, but $\zeta$ needs not. Note that the spectrum of ${\mathcal{L}}$ is symmetric with respect to the reflections in the real and imaginary axes. Hence $\eta$ is spectrally unstable if and only if the spectrum of ${\mathcal{L}}$ is [*not*]{} contained in the imaginary axis.
It is well known (see [@BHJ], for instance, and references therein) that the $L^2(\mathbb{R})$ spectrum of ${\mathcal{L}}$ contains no eigenvalues. Rather, it consists of the essential spectrum. Moreover, a nontrivial solution of does not belong to $L^p(\mathbb{R})$ for any $p\in [1,\infty)$. Rather, if $\zeta\in L^\infty(\mathbb{R})$ solves then, necessarily, $$\zeta(z)=e^{i\xi z}\phi(z),\qquad\text{where}\quad \phi(z+2\pi)=\phi(z),$$ for some $\xi$ in the range $(-1/2,1/2]$. It follows from Floquet theory (see [@BHJ], for instance, and references therein) that $\lambda$ belongs to the $L^2(\mathbb{R})$ spectrum of $\mathcal{L}$ if and only if $$\label{def:Lx}
\lambda\phi=e^{-i\xi z}\mathcal{L}(a;b,{\kappa},T)e^{i\xi z}\phi=:\mathcal{L}(\xi)(a;b,{\kappa},T)\phi$$ for some $\xi\in(-1/2,1/2]$ and $\phi\in L^2(\mathbb{T})$. Thus $$\text{spec}_{L^2(\mathbb{R})}(\mathcal{L}(a;b,{\kappa},T))=
\bigcup_{\xi\in(-1/2,1/2]}\text{spec}_{L^2(\mathbb{T})}(\mathcal{L}(\xi)(a;b,{\kappa},T)).$$ Note that for any $\xi\in(-1/2,1/2]$, the $L^2(\mathbb{T})$ spectrum of $\mathcal{L}(\xi)$ comprises of eigenvalues of finite multiplicities. Thus the essential spectrum of ${\mathcal{L}}$ may be characterized as a one parameter family of point spectra of ${\mathcal{L}}(\xi)$ for $\xi\in(-1/2,1/2]$. Note that $$\text{spec}_{L^2(\mathbb{T})}({\mathcal{L}}(\xi))=(\text{spec}_{L^2(\mathbb{T})}({\mathcal{L}}(-\xi)))^*.$$ Hence it suffices to take $\xi\in [0,1/2]$.
Note that $\xi=0$ corresponds to the same period perturbations as $\eta$. Moreover, $\xi>0$ and small corresponds to long wavelength perturbations, whose effects are to slowly vary the period and other wave characteristics. They furnish the spectral information of ${\mathcal{L}}$ in the vicinity of the origin in $\mathbb{C}$; see [@BHJ], for instance, for details. Therefore, we say that $\eta$ is [*modulationally unstable*]{} if the $L^2(\mathbb{T})$ spectra of ${\mathcal{L}}(\xi)$ are not contained in the imaginary axis near the origin for $\xi>0$ and small, and it is modulationally stable otherwise.
For an arbitrary $\xi$, one must in general study by means of numerical computation. But, for $\xi>0$ and small for $\lambda$ in the vicinity of the origin in $\mathbb{C}$, we may take a spectral perturbation approach in [@HJ2; @HJ3; @HP1], for instance, to address it analytically.
### Notation {#notation-1 .unnumbered}
Throughout the section, $T{\geqslant}0$ and ${\kappa}>0$ satisfying are suppressed for simplicity of notation, unless specified otherwise. We assume that $b=0$. For nonzero $b$, one may explore in like manner. But the calculation becomes lengthy and tedious. Hence we do not discuss the details. We use $$\label{def:Lxia}
{\mathcal{L}}(\xi,a)={\mathcal{L}}(\xi)(a;0,{\kappa},T).$$
Spectra of ${\mathcal{L}}(\xi,0)$ and ${\mathcal{L}}(0,a)$ {#spectra-of-mathcallxi0-and-mathcall0a .unnumbered}
----------------------------------------------------------
For $a=0$ — namely, the rest state — a straightforward calculation reveals that $$\label{eq:a=0FDCH}
{\mathcal{L}}(\xi,0)e^{inz}=i\omega(n+\xi)e^{inz}
\quad \text{for $n\in\mathbb{Z}$ and $\xi\in[0,1/2]$},$$ where $$\label{def:w-FDCH}
\omega(n+\xi)=(\xi+n)(c_{\rm ww}({\kappa};T)-{c_{\rm ww}}({\kappa}(n+\xi);T)).$$ For $\xi=0$, $$\omega(1)=\omega(-1)=\omega(0)=0,$$ and $\omega(n)\neq 0$ otherwise. Hence zero is an $L^2(\mathbb{T})$ eigenvalue of $\mathcal{L}(0,0)$ with multiplicity three. Moreover, $$\label{eq:eigen0-FDCH}
\cos z, \qquad \sin z,\quad\text{and}\quad 1$$ are the associated eigenfunctions, real valued and orthogonal to each other. For $\xi>0$ sufficiently small, $$i\omega(\pm1+\xi)\quad\text{and}\quad i\omega(\xi)$$ are the $L^2(\mathbb{T})$ eigenvalues of ${\mathcal{L}}(\xi,0)$ in the vicinity of the origin in $\mathbb{C}$, and are the associated eigenfunctions.
For $a\in\mathbb{R}$ and $|a|$ sufficiently small for $\xi=0$, zero is an $L^2(\mathbb{T})$ eigenvalue of $\mathcal{L}(0,a)$ with algebraic multiplicity three and geometric multiplicity two, and $$\label{def:p123}
\begin{aligned}
\phi_1(z) &:=\Big(\eta_a-\frac{c_a}{c_b}\eta_b\Big)(a;0,{\kappa},T)(z)=\cos z+ap_1+2a h_2\cos 2z+O(a^2),\\
\phi_2(z) &:=-\frac{1}{a}\eta_z(a;0,{\kappa},T)(z)= \sin z+2a h_2\sin 2z+O(a^2),\\
\phi_3(z) &:=\frac{1}{1-{c_{\rm ww}}({\kappa};T)}\eta_b({\kappa},a,0;T)(z)= 1+O(a^2)
\end{aligned}$$ are the associated eigenfunctions, where $$\label{def:p1}
p_1 = 2h_0-\frac{24c_2}{18-5 {\kappa}^2} = \frac{1}{18-5 {\kappa}^2}\Big( \frac94 -\frac{3}{16}\frac{(3-2 {\kappa}^2)(12-11 {\kappa}^2)}{c_{\rm ww}({\kappa};T)-c_{\rm ww}(2{\kappa};T)}\Big)$$ and $h_2$ is defined in . The proof is nearly identical to that in [@HJ2], for instance. Hence we omit the details. For $a=0$, note that becomes .
Spectral perturbation calculation {#spectral-perturbation-calculation .unnumbered}
---------------------------------
Recall that for $\xi>0$ and sufficiently small for $a=0$, the $L^2(\mathbb{T})$ spectrum of ${\mathcal{L}}(\xi,0)$ contains three purely imaginary eigenvalues $i\omega(\pm1+\xi)$ and $i\omega(\xi)$ in the vicinity of the origin in $\mathbb{C}$, and spans the associated eigenspace, which does not depend on $\xi$. For $\xi=0$ for $a\in\mathbb{R}$ and $|a|$ sufficiently small, the spectrum of ${\mathcal{L}}(0,a)$ contains three eigenvalues at the origin, and spans the associated eigenspace, which depends analytically on $a$.
For $\xi>0$, $a\in\mathbb{R}$ and $\xi, |a|$ sufficiently small, it follows from perturbation theory (see [@K], for instance, for details) that the $L^2(\mathbb{T})$ spectrum of ${\mathcal{L}}(\xi,a)$ contains three eigenvalues in the vicinity of the origin in $\mathbb{C}$, and spans the associated eigenspace. Let $$\label{def:BI}
\mathbf{L}(\xi,a)=\left( \frac{{\langle}{\mathcal{L}}(\xi,a)\phi_j, \phi_k{\rangle}}{{\langle}\phi_j, \phi_j{\rangle}}\right)_{j,k=1,2,3}
\quad\text{and}\quad
\mathbf{I}(a)=\left( \frac{{\langle}\phi_j, \phi_k{\rangle}}{{\langle}\phi_j, \phi_j{\rangle}}\right)_{j,k=1,2,3},$$ where $\phi_1,\phi_2,\phi_3$ are in . Throughout the subsection, $\langle\cdot\,,\cdot\rangle$ means the $L^2(\mathbb{T})$ inner product. Note that $\mathbf{L}$ represents the action of $\mathcal{L}$ on the eigenspace, spanned by $\phi_1,\phi_2,\phi_3$, and $\mathbf{I}$ is the projection of the identity onto the eigenspace. It follows from perturbation theory (see [@K], for instance for details) that for $\xi>0$, $a\in\mathbb{R}$ and $\xi, |a|$ sufficiently small, the eigenvalues of ${\mathcal{L}}(\xi,a)$ agree in location and multiplicity with the roots of $\det(\mathbf{L}-\lambda\mathbf{I})$ up to terms of order $a$.
For $a\in\mathbb{R}$ and $|a|$ sufficiently small, a Baker-Campbell-Hausdorff expansion reveals that $${\mathcal{L}}(\xi,a)={\mathcal{L}}(0,a)+i\xi [{\mathcal{L}}(0,a),z]-\frac12\xi^2[[{\mathcal{L}}(0,a),z],z]+O(\xi^3)$$ as $\xi\to0$, where $[\cdot\,,\cdot]$ means the commutator. We merely pause to remark that $[{\mathcal{L}},z]$ and $[[{\mathcal{L}},z],z]$ are well defined in the periodic setting even though $z$ is not. We use , and , to write $$\begin{aligned}
{\mathcal{L}}(\xi,a)=&\mathcal{M}
-a\partial_z\Big(\frac32\cos z+{\kappa}^2\Big(\frac{5}{12}\cos z(\partial_z^2-1)-\frac{13}{24}\sin z\partial_z\Big)\Big) \label{def:L-FDCH}\\
&-i\xi a\Big(\frac32\cos z + {\kappa}^2\Big(\frac{5}{12}(2\partial_z\cos z\partial_z+\cos z(\partial_z^2-1))
-\frac{13}{24} (\sin z\partial_z+\partial_z \sin z)\Big)\Big)\notag \\&+O(\xi^3+\xi^2 a+a^2)\notag\end{aligned}$$ as $\xi, a\to 0$, where $$\mathcal{M}={\mathcal{L}}(0,0)+i\xi [{\mathcal{L}}(0,0),z]-\frac12\xi^2[[{\mathcal{L}}(0,0),z],z]$$ agrees with ${\mathcal{L}}(\xi,0)$ up to terms of order $\xi^2$ as $\xi\to 0$. We may then resort to , , and we make an explicit calculation to find that $$\begin{aligned}
{\mathcal{L}}(\xi,0) e^{ inz}=&in(c_{\rm ww}({\kappa};T)-c_{\rm ww}(n{\kappa};T))e^{ inz}\\
&+i\xi(c_{\rm ww}({\kappa};T)-c_{\rm ww}(n{\kappa};T)-{\kappa}c'_{\rm ww}(n{\kappa};T))e^{ inz} \\
& -\frac12\xi^2(2{\kappa}c'_{\rm ww}(n{\kappa};T)+{\kappa}^2 c''_{\rm ww}(n{\kappa};T))e^{ inz}+O(\xi^3)\end{aligned}$$ as $\xi\to 0$. Therefore, for $T{\geqslant}0$ but $T\neq1/3$, for ${\kappa}>0$ satisfying , ${\mathcal{M}}1=i\xi(c_{\rm ww}({\kappa};T)-1)$, $$\begin{aligned}
{\mathcal{M}}\left\{ \begin{matrix} \cos z\\ \sin z\end{matrix}\right\}=&
-i\xi {\kappa}{c_{\rm ww}}'({\kappa};T)\left\{ \begin{matrix} \cos z\\ \sin z\end{matrix}\right\}
\pm \frac12\xi^2(2{\kappa}{c_{\rm ww}}'({\kappa};T)+{\kappa}^2{c_{\rm ww}}''({\kappa};T))\left\{ \begin{matrix} \sin z\\ \cos z\end{matrix}\right\},\end{aligned}$$ and $$\begin{aligned}
{\mathcal{M}}\left\{ \begin{matrix} \cos 2z\\ \sin 2z\end{matrix}\right\}=&
\mp 2({c_{\rm ww}}({\kappa};T)-{c_{\rm ww}}(2{\kappa};T))\left\{ \begin{matrix} \sin 2z\\ \cos 2z\end{matrix}\right\}\\
&+i\xi({c_{\rm ww}}({\kappa};T)-{c_{\rm ww}}(2{\kappa};T)-2{\kappa}c'_{\rm ww}(2{\kappa};T))\left\{ \begin{matrix} \cos 2z\\ \sin 2z\end{matrix}\right\} \\
&\pm \frac12\xi^2(2{\kappa}c'_{\rm ww}(2{\kappa};T)+{\kappa}^2 c''_{\rm ww}(2{\kappa};T))\left\{ \begin{matrix} \sin 2z\\ \cos 2z\end{matrix}\right\}.\end{aligned}$$ For $T=1/3$, one must calculate higher order terms in the expansion of ${\mathcal{L}}(\xi,0)$. We do not pursue this here.
We use , and the above formula for $\mathcal{M}$, and we make a lengthy but explicit calculation to find that $$\begin{aligned}
{\mathcal{L}}\phi_1=&-i\xi {\kappa}{c_{\rm ww}}'({\kappa};T)\cos z \\
&-i\xi a\Big(\frac34-\frac{7}{48}{\kappa}^2-p_1({c_{\rm ww}}({\kappa};T)-1)\Big)\\
&+i\xi a\Big(-\frac34+\frac{33}{16}{\kappa}^2+2h_2({c_{\rm ww}}({\kappa};T)-{c_{\rm ww}}(2{\kappa};T)-2{\kappa}c'_{\rm ww}(2{\kappa};T))\Big)\cos 2z\notag \\
&+\frac12\xi^2(2{\kappa}{c_{\rm ww}}'({\kappa};T)+{\kappa}^2{c_{\rm ww}}''({\kappa};T))\sin z+O(\xi^3+\xi^2a+a^2), \\
{\mathcal{L}}\phi_2 =&-i\xi {\kappa}{c_{\rm ww}}'({\kappa};T)\sin z \\
&+i\xi a\Big(-\frac34+\frac{33}{16}{\kappa}^2+2h_2({c_{\rm ww}}({\kappa};T)-{c_{\rm ww}}(2{\kappa};T)-2{\kappa}c'_{\rm ww}(2{\kappa};T))\Big)\sin 2z\notag\\
&-\frac12\xi^2(2{\kappa}{c_{\rm ww}}'({\kappa};T)+{\kappa}^2 {c_{\rm ww}}''({\kappa};T))\cos z+O(\xi^3+\xi^2a+a^2), \\
{\mathcal{L}}\phi_3=&a\Big(3-\frac{5}{6}{\kappa}^2\Big)\sin z+i\xi ({c_{\rm ww}}({\kappa};T)-1)-i\xi a\Big(\frac32-\frac{23}{24}{\kappa}^2\Big)\cos z+O(\xi^3+\xi^2a+a^2) \end{aligned}$$ as $\xi, a\to 0$, where $p_1$ is in and $h_2$ is in .
To proceed, we take the $L^2(\mathbb{T})$ inner product of the above and , and we make a lengthy but explicit calculation to find that $$\begin{aligned}
\langle {\mathcal{L}}\phi_1,\phi_1\rangle &=\langle{\mathcal{L}}\phi_2,\phi_2 \rangle
= -\frac12i\xi {\kappa}{c_{\rm ww}}'({\kappa};T)+O(\xi^3+\xi^2a+a^2),\\
\langle {\mathcal{L}}\phi_1,\phi_2\rangle &=-\langle {\mathcal{L}}\phi_2,\phi_1\rangle= \frac14\xi^2(2{\kappa}{c_{\rm ww}}'({\kappa};T)+{\kappa}^2 {c_{\rm ww}}''({\kappa};T))+O(\xi^3+\xi^2a+a^2),\\
\langle {\mathcal{L}}\phi_1,\phi_3\rangle &=
i\xi a\Big(-\frac34+\frac{36}{48}{\kappa}^2+p_1 ({c_{\rm ww}}({\kappa};T)-1)\Big)+O(\xi^3+\xi^2a+a^2),
\intertext{and}
\langle {\mathcal{L}}\phi_2,\phi_3\rangle &=0+O(\xi^3+\xi^2a+a^2),\\
\langle {\mathcal{L}}\phi_3,\phi_1\rangle &=i\xi a\Big(-\frac34+\frac{23}{48}{\kappa}^2+p_1 ({c_{\rm ww}}({\kappa};T)-1) \Big)+O(\xi^3+\xi^2a+a^2),\\
\langle {\mathcal{L}}\phi_3,\phi_2\rangle &=a\Big(\frac34-\frac{5}{24}{\kappa}^2\Big)+O(\xi^3+\xi^2a+a^2),\\
\langle {\mathcal{L}}\phi_3,\phi_3\rangle &=i\xi( {c_{\rm ww}}({\kappa};T)-1)+O(\xi^3+\xi^2a+a^2)\end{aligned}$$ as $\xi,a\to 0$, where $p_1$ is in . Moreover, we take the $L^2(\mathbb{T})$ inner products of , and we make an explicit calculation to find that $$\begin{aligned}
\langle \phi_1,\phi_1\rangle &= \langle \phi_2,\phi_2 \rangle=\frac{1}{2}+O(\xi^3+\xi^2a+a^2),\\
\langle \phi_1,\phi_2\rangle &, \langle \phi_2,\phi_3\rangle=0+O(\xi^3+\xi^2a+a^2),\\
\langle \phi_1,\phi_3\rangle &=ap_1+O(\xi^3+\xi^2a+a^2), \\
\langle \phi_3,\phi_3\rangle &=1+O(\xi^3+\xi^2a+a^2)\end{aligned}$$ as $\xi,a\to 0$, where $p_1$ is in . Together, becomes $$\begin{aligned}
\label{eq:B-FDCH}
\mathbf{L}(\xi,a) &=a\Big(\frac34-\frac{5}{24}{\kappa}^2\Big)\begin{pmatrix} 0&0&0\\0&0&0\\0&1&0\end{pmatrix}\\
&+i\xi \begin{pmatrix} -{\kappa}{c_{\rm ww}}'({\kappa};T)&0&0\\ 0&-{\kappa}{c_{\rm ww}}'({\kappa};T)&0\\ 0&0&{c_{\rm ww}}({\kappa};T)-1\end{pmatrix} \notag\\
&+i\xi a\begin{pmatrix} 0&0&{\displaystyle -\tfrac34+\tfrac{7}{24} {\kappa}^2+2p_1 ({c_{\rm ww}}({\kappa};T)-1)}\\ 0&0&0\\
{\displaystyle -\tfrac34+\tfrac{23}{48}{\kappa}^2+p_1 ({c_{\rm ww}}({\kappa};T)-1)}&0&0\end{pmatrix}\notag\\
&+\xi^2({\kappa}{c_{\rm ww}}'({\kappa};T)+\tfrac12{\kappa}^2{c_{\rm ww}}''({\kappa};T))\begin{pmatrix}0&1&0\\ -1&0&0\\ 0&0&0 \end{pmatrix}
+O(\xi^3+\xi^2a+a^2),\notag\end{aligned}$$ and $$\mathbf{I}(a)=\mathbf{I}+ap_1 \begin{pmatrix} 0&0&2\\0&0&0\\1&0&0\end{pmatrix}+O(a^2)\label{eq:I-FDCH}$$ as $\xi,a\to 0$, where $p_1$ is in and $\mathbf{I}$ is the $3\times3$ identity matrix.
Modulational instability index {#modulational-instability-index .unnumbered}
------------------------------
For any $T{\geqslant}0$ but $T\neq1/3$, ${\kappa}>0$ satisfying , for $\xi>0$, $a\in\mathbb{R}$ and $\xi, |a|$ sufficiently small, we turn the attention to the roots of $$\begin{gathered}
\det(\mathbf{L}-\lambda\mathbf{I})(\xi,a;{\kappa},T) \\
=:p_3(\xi,a;{\kappa},T)\lambda^3+ip_2(\xi,a;{\kappa},T)\lambda^2+p_1(\xi,a;{\kappa},T)\lambda+ip_0(\xi,a;{\kappa},T),\end{gathered}$$ where $\mathbf{L}$ and $\mathbf{I}$ are in and . Details are found in [@HJ2], for instance. Hence we merely hit the main points.
Let $$q(-i\xi\lambda)(\xi,a;{\kappa},T)=(i\xi^3(q_3\lambda^3-q_2\lambda^2-q_1\lambda+q_0)(\xi,a;{\kappa},T),$$ where $p_j=\xi^{3-j}q_j$, $j=0,1,2,3$. Note that $q_0,q_1,\dots,q_3$ are real valued and depend analytically on $\xi,a$ and ${\kappa}$ for any $\xi>0$ and $|a|$ sufficiently small for any ${\kappa}>0$ satisfying . Moreover, they are odd in $\xi$ and even in $a$. For any $T{\geqslant}0$ but $T\neq1/3$, ${\kappa}>0$ satisfying , $a\in\mathbb{R}$ and $|a|$ sufficiently small, a periodic traveling wave $\eta(a;0,{\kappa},T)$ and $c(a;0,{\kappa},T)$ of and is modulationally unstable, provided that $q$ possesses a pair of complex roots or, equivalently, $$\Delta_0(\xi,a;{\kappa},T):=(18q_3q_2q_1q_0+q_2^2q_1^2+4q_2^3q_0+4q_3q_1^3-27q_3^2q_0^2)(\xi,a;{\kappa},T)<0$$ for $\xi>0$ and small, and it is modullationally stable if $\Delta_0>0$. Note that $\Delta_0$ is even in $\xi$ and $a$. Hence we write that $$\Delta_0(\xi,a;{\kappa},T)=:\Delta_0(\xi,0;{\kappa},T)+a^2\Delta({\kappa};T)+O(a^2(\xi^2+a^2))$$ as $a\to 0$ for $\xi>0$ and small. We then use and , and we make a Mathematica calculation to show that $$\Delta_0(\xi,0;{\kappa},T)={\kappa}^2 i_1^2({\kappa};T)\Big(\xi i_2^2+\frac14\xi^3{\kappa}^2 i_1^2\Big)^2({\kappa};T)>0$$ as $\xi\to0$. Therefore, if $\Delta<0$ then $\Delta_0<0$ for $\xi>0$ and sufficiently small, depending on $a\in\mathbb{R}$ and $|a|$ sufficiently small, implying modulational instability, whereas if $\Delta>0$ then $\Delta_0>0$ for $\xi>0$, $a\in\mathbb{R}$ and $\xi, |a|$ sufficiently small, implying modulational stability. We use and , and we make a Mathematica calculation to find $\Delta$ explicitly.
Below we summarize the conclusion.
\[thm:FDCH\] For any $T {\geqslant}0$ but $T\neq 1/3$, for any ${\kappa}>0$ satisfying , a sufficiently small and $2\pi/{\kappa}$ periodic traveling wave of and is modulationally unstable, provided that $$\label{def:index}
\Delta({\kappa};T):=\frac{i_1i_2 }{i_3}i_4({\kappa};T)<0,$$ where
\[def:i1234\] $$\begin{aligned}
i_1({\kappa};T)=&({\kappa}c_{\rm ww}({\kappa};T))'', \\
i_2({\kappa};T)=&({\kappa}c_{\rm ww}({\kappa};T))'-1,\\
i_3({\kappa};T)=&c_{\rm ww}({\kappa};T)-c_{\rm ww}(2{\kappa};T), \\
i_4({\kappa};T)=&\Big(3i_2-i_2i_3+6 i_3
-\frac{1}{12}{\kappa}^2(57i_2+34 i_3)+\frac{1}{108}{\kappa}^4(198 i_2+35i_3)\Big)({\kappa},T), \hspace*{-30pt}\label{def:i4}\end{aligned}$$
and ${c_{\rm ww}}({\kappa};T)$ is in . It is modulationally stable if $\Delta({\kappa};T)>0$.
Theorem \[thm:FDCH\] elucidates four resonance mechanisms which contribute to the sign change in $\Delta$ and, ultimately, the change in the modulational stability and instability in and . Note that $$\text{${c_{\rm ww}}({\kappa};T)=$the phase speed\quad and\quad $({\kappa}{c_{\rm ww}}({\kappa};T))'=$the group speed}$$ in the linear theory. Specifically,
- $i_1({\kappa};T)=0$ at some ${\kappa}$; the group speed achieves an extremum at the wave number ${\kappa}$;
- $i_2({\kappa};T)=0$ at some ${\kappa}$; the group speed at the wave number ${\kappa}$ coincides with the phase speed in the limit as ${\kappa}\to0$, resulting in the “resonance of short and long waves;"
- $i_3({\kappa};T)=0$ at some ${\kappa}$; the phase speeds of the fundamental mode and the second harmonic coincide at the wave number ${\kappa}$, resulting in the “second harmonic resonance;"
- $i_4({\kappa};T)=0$ at some ${\kappa}$.
Resonances (R1), (R2), (R3) are determined from the dispersion relation. For instance, $i_1, i_2, i_3$ appear in an index formula for the Whitham equation; see [@HJ3] for details. Resonance (R4), on the other hand, results from the resonance of the dispersion and nonlinear effects, and it depends on the nonlinearity of the equation.
Results {#sec:results}
=======
For $\mathbf{T=0}$ {#for-mathbft0 .unnumbered}
------------------
Since $({\kappa}{c_{\rm ww}}({\kappa};0))'<1$ for any ${\kappa}>0$ and decreases monotonically over the interval $(0,\infty)$ by brutal force, $i_1({\kappa};0)<0$ and $i_2({\kappa};0)<0$ for any ${\kappa}>0$. Since ${c_{\rm ww}}({\kappa};0)>0$ for any ${\kappa}>0$ and decreases monotonically over the interval $(0,\infty)$ (see Figure \[fig:cT\]a), $i_3({\kappa};0)>0$ for any ${\kappa}>0$.
We use and make an explicit calculation to show that $$\lim_{{\kappa}\to 0+}\frac{i_4({\kappa})}{{\kappa}^2}=\frac92\quad\text{and}\quad\lim_{{\kappa}\to\infty}i_4({\kappa})=- \infty.$$ Hence $\Delta({\kappa};0)>0$ for ${\kappa}>0$ sufficiently small, implying the modulational stability, and it is negative for ${\kappa}>0$ sufficiently large, implying the modulational instability. The intermediate value theorem asserts a root of $i_4$. A numerical evaluation of reveals a unique root ${\kappa}_c$, say, of $i_4$ over the interval $(0,\infty)$ such that $i_4({\kappa})>0$ if $0<{\kappa}<{\kappa}_c$ and it is negative if ${\kappa}_c<{\kappa}<\infty$. Upon close inspection (see Figure \[fig:i4\]), moreover, ${\kappa}_c=1.420\dots$.
![The graph of $i_4({\kappa})$ for ${\kappa}\in(0,1.5)$.[]{data-label="fig:i4"}](i4 "fig:")
Therefore, a sufficiently small and $2\pi/{\kappa}$ periodic traveling wave of and is modulationally unstable if ${\kappa}>{\kappa}_c$, where ${\kappa}_c=1.420\dots$ is a unique root of $i_4$ in over the interval $(0,\infty)$. It is modulationally stable if $0<{\kappa}<{\kappa}_c$. The result qualitatively agrees with the Benjamin-Feir instability of a Stokes wave. The critical wave number compares reasonably well with that in [@Whitham1967] and [@BM1995]. The critical wave number for the Whitham equation is $1.146\dots$ (see [@HJ2]).
For $\mathbf{T>0}$ {#for-mathbft0-1 .unnumbered}
------------------
For $T>1/3$, since ${c_{\rm ww}}({\kappa};T)$ and $({\kappa}{c_{\rm ww}}({\kappa};T))'$ increase monotonically over the interval $(0,\infty)$ and since $({\kappa}{c_{\rm ww}}({\kappa};T))'$ does not possess an extremum (see Figure \[fig:cT\]b), $i_1$, $i_2$, $i_3$ do not vanish over the interval $(0,\infty)$. A numerical evaluation reveals that $i_4$ changes its sign once over the interval $(0,\infty)$. Together, a sufficiently small and periodic traveling wave of and is modulationally unstable, provided that the wave number is greater than a critical value, and modulationally stable otherwise, similarly to the gravity wave setting. Moreover, a numerical evaluation reveals that the critical wave number ${\kappa}_c(T)$, say, satisfies $$\lim_{T\to \infty}{\kappa}_c(T)\sqrt{T}\approx 1.283.$$
For $0<T<1/3$, on the other hand, $({\kappa}{c_{\rm ww}}({\kappa};T))'$ achieves a unique minimum over the interval $(0,\infty)$. Moreover, $i_2$ and $i_3$ each takes one transverse root over the interval $(0,\infty)$ (see Figure \[fig:cT\]c). Hence, $i_1$ through $i_4$ each contributes to the change in the modulational stability and instability.
Figure \[f:ST\] illustrates in the ${\kappa}$ and ${\kappa}\sqrt{T}$ plane the regions of modulational stability and instability for a sufficiently small and periodic traveling wave of and . Along Curve 1, $i_1=0$ and the group speed achieves an extremum at the wave number ${\kappa}$. Curve 2 is associated with $i_2=0$, along which the group speed coincides with the phase speed in the limit as ${\kappa}\to0$. In the deep water limit, as ${\kappa}\to\infty$ while ${\kappa}\sqrt{T}$ is fixed, it is asymptotic to ${\kappa}=\frac94{\kappa}^2T-\frac34$. Curve 3 is associated with $i_3=0$, along which the phase speeds of the fundamental mode and the second harmonic coincide. In the deep water limit, it is asymptotic to $k^2T=\frac12$. Moreover, along Curve 4, $i_4$ vanishes because of the resonance of the dispersion and nonlinear effects. The “lower" branch of Curve 4 passes through ${\kappa}=1.420\dots$, the critical wave number for $T=0$. The “upper" branch passes through ${\kappa}\sqrt{T}=1.283\dots$, the limit of strong surface tension.
\[f:ST\] The result qualitatively agrees with those in [@Kawahara] and [@DR], for instance, from formal asymptotic expansions for the physical problem, and it improves upon that in [@HJ3] for the Whitham equation, for which ${\kappa}_c(T)\sqrt{T}\to \infty$ as $T\to \infty$.
Discussion {#sec:discussion}
==========
The Camassa-Holm equation {#the-camassa-holm-equation .unnumbered}
-------------------------
We may write , in the case of $a=1/12$, after normalization of parameters, as $$\label{E:CH}
\eta_t +c_\text{\Tiny CH}(|\partial_x|)\Big(2(1+\eta)^{3/2}-2\eta+\frac{7}{24}\eta \eta_{xx}+\frac{7}{48}\eta_x^2\Big)_x = 0,$$ where $$\label{def:c-CH}
\widehat{c_\text{\Tiny CH}(|\partial_x|)f}({\kappa})=\frac{12-{\kappa}^2}{12+{\kappa}^2}\widehat{f}({\kappa}).$$ Note that $c_\text{\Tiny CH}({\kappa})$ approximates .
For any ${\kappa}>0$, we may repeat the argument in Section \[sec:periodic\] to determine sufficiently small and $2\pi/{\kappa}$ periodic traveling waves. Specifically, a two parameter family of periodic traveling waves of and exists, denoted $$\eta(a,b;{\kappa})(z)\qquad\text{where $z={\kappa}(x-c(a,b;{\kappa})t)$},$$ for $a,b\in\mathbb{R}$ and $|a|,|b|$ sufficiently small; $\eta$ is $2\pi$ periodic, even, and smooth in $z$. Moreover, $$\begin{aligned}
\eta(a,b;{\kappa})(z)=&b(1-c_\text{\Tiny CH}({\kappa}))+a \cos z + a^2 (h_0 + h_2 \cos 2z)+O(a(a^2+b^2)),\\
c(a,b;{\kappa})=&c_\text{\Tiny CH}({\kappa})+b\Big(\frac32 - \frac{7}{24} {\kappa}^2 \Big)c_\text{\Tiny CH}({\kappa})(1-c_\text{\Tiny CH}({\kappa}))+a^2 c_2+O(a(a^2+b^2)) \end{aligned}$$ as $a,b \to 0$, where $$\begin{gathered}
h_0 =\frac{36-7 {\kappa}^2}{96(c_\text{\Tiny CH}({\kappa})-1)}, \qquad h_2 = \frac{(12-7 {\kappa}^2)c_\text{\Tiny CH}(2{\kappa})}{32(c_\text{\Tiny CH}({\kappa})-c_\text{\Tiny CH}(2{\kappa}))},\\
c_2=c_\text{\Tiny CH}({\kappa})\Big( \frac{36-7 {\kappa}^2}{24} h_0+\frac{12-7 {\kappa}^2}{16}h_2-\frac{3}{32}\Big).\end{gathered}$$ We then proceed as in Section \[sec:MI\], to determine a modulational instability index $$\Delta_\text{\Tiny{CH}}({\kappa}):=\frac{i_1({\kappa})i_2({\kappa})}{i_3({\kappa})}i_4({\kappa}),$$ where $$\begin{aligned}
i_1({\kappa})=&({\kappa}c_\text{\Tiny{CH}}({\kappa}))'',\\
i_2({\kappa})=&({\kappa}c_\text{\Tiny{CH}}({\kappa}))'-1, \\
i_3({\kappa})=&c_\text{\Tiny{CH}}({\kappa})-c_\text{\Tiny{CH}}(2{\kappa}),\\
i_4({\kappa})=&1296(c_\text{\Tiny{CH}}(2{\kappa}) i_2({\kappa})+2i_3({\kappa}))-432 i_2({\kappa})i_3({\kappa}) \\
&-1512 {\kappa}^2 c_\text{\Tiny{CH}}(2{\kappa}) i_2({\kappa})+49{\kappa}^4(9c_\text{\Tiny{CH}}(2{\kappa}) i_2({\kappa})-2i_3({\kappa})).\end{aligned}$$ We omit the details.
A straightforward calculation shows that $\frac{i_2i_4}{i_3}({\kappa})< 0$ for any ${\kappa}> 0$ while $i_1({\kappa})$ changes its sign from negative to positive across ${\kappa}= 6$. Therefore, a sufficiently small and $2\pi/{\kappa}$ periodic traveling wave of and is modulationally unstable if ${\kappa}>6$. For other values of $a$ in , the result is qualitatively the same. Thus, the Camassa-Holm equation seems to predict the Benjamin-Feir instability of a Stokes wave. But Resonance (R1) following Theorem \[thm:FDCH\] results in the instability in and , whereas it does not take place in the water wave problem.
The velocity equation {#the-velocity-equation .unnumbered}
---------------------
The FDCH equation for the average horizontal velocity, after normalization of parameters, $$\label{E:FDCHu}
u_t+c_{\rm ww}(|\partial_x|)u_x+\frac32 uu_x=-\Big(\frac{5}{12}uu_{xxx}+\frac{23}{24}u_xu_{xx}\Big),$$ where ${c_{\rm ww}}(|\partial_x|)$ is in , differs from by higher power nonlinearities. We may repeat the arguments in Section \[sec:periodic\] and Section \[sec:MI\] to derive the modulational instability index, where is replaced by $$\begin{aligned}
\label{def:iFDCHu}
i_u({\kappa};T)=&i_2({\kappa};T)+2i_3({\kappa};T) \\
&+\frac{1}{36}{\kappa}^2 (57i_2+34 i_3)({\kappa};T)+\frac{1}{324}{\kappa}^4(198 i_2+35i_3)({\kappa};T). \notag\end{aligned}$$ We omit the details. The index formula for and agrees with that for the Whitham equation (see [@HJ2; @HJ3] for details) except the terms in explicitly depending on ${\kappa}^2$ and ${\kappa}^4$, which higher derivative nonlinearities of seem to contribute.
For $T=0$, it is straightforward to show that a sufficiently small and $2\pi/{\kappa}$ periodic traveling wave of and is modulationally unstable if ${\kappa}>0.637\dots$. Thus the FDCH equation for the average horizontal velocity predicts the Benjamin-Feir instability of a Stokes wave. For $T>0$, the modulational instability result for and qualitatively agrees with that in [@HJ3] for the Whitham equation; see Figure \[f:STu\]. In particular, ${\kappa}_c(T)\sqrt{T}\to \infty$ as $T\to\infty$, where ${\kappa}_c(T)$ is a critical wave number, depending on $T$, whereas the limit is finite in the FDCH equation for the fluid surface displacement and the water wave problem. That means, the higher power nonlinearities of improve the modulational instability result in the presence of the effects of surface tension, not the higher derivative nonlinearities.
\[f:STu\]
Acknowledgements {#acknowledgements .unnumbered}
----------------
VMH is supported by the National Science Foundation’s Faculty Early Career Development (CAREER) Award DMS-1352597, a Simons Fellowship in Mathematics, and the University of Illinois at Urbana-Champaign under the Arnold O. Beckman Research Award RB16227. She is grateful to the Department of Mathematics at Brown University for its generous hospitality.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We propose a numerical pipeline for shape optimization in naval engineering involving two different non-intrusive reduced order method (ROM) techniques. Such methods are proper orthogonal decomposition with interpolation (PODI) and dynamic mode decomposition (DMD). The ROM proposed will be enhanced by active subspaces (AS) as a pre-processing tool that reduce the parameter space dimension and suggest better sampling of the input space.
We will focus on geometrical parameters describing the perturbation of a reference bulbous bow through the free form deformation (FFD) technique. The ROM are based on a finite volume method (FV) to simulate the multi-phase incompressible flow around the deformed hulls.
In previous works we studied the reduction of the parameter space in naval engineering through AS [@tezzele2018dimension; @demo2018isope] focusing on different parts of the hull. PODI and DMD have been employed for the study of fast and reliable shape optimization cycles on a bulbous bow in [@demo2018shape].
The novelty of this work is the simultaneous reduction of both the input parameter space and the output fields of interest. In particular AS will be trained computing the total drag resistance of a hull advancing in calm water and its gradients with respect to the input parameters. DMD will improve the performance of each simulation of the campaign using only few snapshots of the solution fields in order to predict the regime state of the system. Finally PODI will interpolate the coefficients of the POD decomposition of the output fields for a fast approximation of all the fields at new untried parameters given by the optimization algorithm. This will result in a non-intrusive data-driven numerical optimization pipeline completely independent with respect to the full order solver used and it can be easily incorporated into existing numerical pipelines, from the reference CAD to the optimal shape.
author:
- 'Marco Tezzele[^1]'
- 'Nicola Demo[^2]'
- 'Gianluigi Rozza[^3]'
title: Shape optimization through proper orthogonal decomposition with interpolation and dynamic mode decomposition enhanced by active subspaces
---
Introduction {#sec:intro}
============
In a shape optimization problem, we aim to find the shape — among all the admissible geometries — that minimizes a certain objective function. In this work we propose a novel approach to optimize the total resistance of a ship hull advancing in calm water by deforming the original hull, a common problem in the naval engineering field.
First we define the total resistance as the sum of the viscous and lift forces acting on the hull. Formally, our optimization problem can be expressed as $$\min_{\forall {\ensuremath{\boldsymbol{\mu}}}\in D} f(\Omega, {\ensuremath{\boldsymbol{\mu}}}) =
\min_{\forall {\ensuremath{\boldsymbol{\mu}}}\in D} \oint p \cos{\theta}\:\mathrm{d} \Omega_{{\ensuremath{\boldsymbol{\mu}}}}
+ \oint \tau_x\:\mathrm{d} \Omega_{{\ensuremath{\boldsymbol{\mu}}}} ,$$ where the $D \subset \mathbb{R}^P$ is the parametric domain, $P$ the number of parameters, $\Omega \in
\mathbb{R}^3$ is the reference hull domain, and $\Omega_{{\ensuremath{\boldsymbol{\mu}}}} =
\mathcal{M}(\Omega, {\ensuremath{\boldsymbol{\mu}}})$ is the defomed hull. The morphing map $\mathcal{M}(\cdot, {\ensuremath{\boldsymbol{\mu}}}): \mathbb{R}^3 \to \mathbb{R}^3$ we use in this work is the free form deformation (FFD) and will be properly defined in Section \[sec:ffd\]. Examples of other deformation techniques are radial basis functions (RBF) interpolation [@buhmann2003radial; @morris2008cfd; @manzoni2012model], and inverse distance weighting (IDW) interpolation [@shepard1968; @forti2014efficient; @BallarinDAmarioPerottoRozza2018]. The unknowns $p$ and $\tau_x$ denote respectively the pressure and the $x$ component of the wall shear stress over the hull surface, while $\theta$ is the angle between the flow direction and the surface. The evaluation of the objective function requires a numerical simulation of the flow around the ship, which has a high computational cost. The purpose of this work is beyond an analysis of the adopted full order model for the fluid dynamics, we just provide a brief summary in order to facilitate the understanding of the pipeline. We resolve the Reynolds-averaged Navier Stokes (RANS) equations with the $k$–$\omega$ SST turbulence model using a finite volume approach, a typical benchmark in industrial hydrodynamics analysis. Such model deals very well with turbulent fluid, but at high computational cost. Moreover, due to the complexity of the optimization problem, we typically need many evaluations of the objective function to converge to the optimal shape.
For this work we choose to simulate the flow around the DTMB 5415 hull due to the existence of a vast amount of literature and benchmark tests. In Figure \[fig:hull\] the undeformed hull domain.
![Complete hull domain of the DTMB 5415.[]{data-label="fig:hull"}](dtmb5415_hull.png){width=".7\textwidth"}
In order to reduce the computational cost, we introduce in the optimization framework two reduced order modeling (ROM) techniques. These techniqes are able to represent complex systems in a low dimensional space, reducing the number of degrees of freedom used in the full order model discretization and providing an efficient and reliable approximation of the solution. The ROM methods initially collect a database of high-fidelity solutions — the solutions computed using the full-order model — during the most computationally expensive phase, also called [*offline*]{} phase. Then, the solutions are combined to build the reduced space we query during the [*online*]{} phase to obtain the new solution. In this work, we adopt the dynamic mode decomposition (DMD) and proper orthogonal decomposition with interpolation (PODI), two emerging data-driven techniques. PODI is used to approximate, given the high-fidelity solutions computed for some defomed hulls, the solution for any new parametric point in the domain $D$. DMD algorithm instead provides a simplification of the dynamics of complex system: we use it in order to accelerate the single high-fidelity simulations we need for PODI method, by storing few system outputs and exploiting them to approximate the flow dynamics. For more details about equation-free ROM methods, we suggest [@tezzele2018ecmi], while for a complete overview — including intrusive approaches — we cite [@salmoiraghi2016advances; @rozza2018advances; @morhandbook2019].
Moreover, additionally to these methods, we use the active subspace (AS) property as preprocessing tool in order to be able to reduce the dimension of the parameter space and obtain a better accuracy in ROM solution approximation.
In this contribution, we focus on all the components of the computational pipeline: in Section \[sec:ffd\] we provide a brief overview of the FFD method, the Section \[sec:dmd\] illustrates the DMD algorithm, in Section \[sec:active\] the AS property is explained, while Section \[sec:podi\] describes the idea behind PODI technique. Finally, Sections \[sec:results\] and \[sec:the\_end\] provide respectively the numerical results collected during this work and the final conclusions.
The free form deformation technique {#sec:ffd}
===================================
Free form deformationi (FFD) is a widespread deformation technique. Proposed in [@sederbergparry1986], FFD was initially employed in computer graphics, getting more popular both in academia and industry in the last decades. In this section, we provide an overview of the method: for more details about FFD, among all the works in literature, we recommend [@rozza2013free; @demo2018shape; @garotta2018quiet].
![Graphical representation of the FFD morphing map $\mathcal{M}$ as composition of the maps $\psi$, $T$, and $\psi^{-1}$. The displacements of the control points $\boldsymbol{P}$ define the morphing of the domain.[]{data-label="fig:ffd_scheme"}](ffd_scheme_marine_as.png){width=".60\textwidth"}
The idea of FFD is very intuitive: the domain is deformed by manipulating a lattice of points surrounding the object to morph. The displacements of these control points are the input parameters ${\ensuremath{\boldsymbol{\mu}}}_{\text{GEO}}$. To achieve this result $1$) the physical domain $\Omega$ is mapped to the reference domain $\widehat{\Omega}$ using the function $\psi$, and a lattice of control points $\boldsymbol{P}$ is constructed around the object to deform, then $2$) through the map $T$ the reference domain is morphed using B-splines or Bernstein polynomials tensor product and finally $3$) the deformed domain is remapped to the physical one by using $\psi^{-1}$. In Figure \[fig:ffd\_scheme\] is shown a sketch of the free form deformation map as a composition of the three functions presented above. Formally, we can define the deformation map $\mathcal{M}$ as $$\mathcal{M} ({\ensuremath{\boldsymbol{x}}}, {\ensuremath{\boldsymbol{\mu}}}_{\text{GEO}}) := (\boldsymbol{\psi}^{-1} \circ T
\circ \boldsymbol{\psi}) ({\ensuremath{\boldsymbol{x}}}, {\ensuremath{\boldsymbol{\mu}}}_{\text{GEO}})
\quad \forall {\ensuremath{\boldsymbol{x}}}\in \Omega.$$ This technique allows to manipulate complex geometries and also computational grids, since it is able to preserve derivatives continuity and perform global deformation using only few parameters. Figure \[fig:ffd\_lattice\] shows the position of the lattice of control points around a bulbous bow, which is the part of the hull we want to parametrize and morph. Regarding the implementation, the results in this contribution are obtained using PyGeM [@pygem], an open source Python package implementing several deformation techniques.
![Example of bulbous bow deformation using the FFD method. The red dots are the FFD control points, already manipulated.[]{data-label="fig:ffd_lattice"}](ffd_lattive_as.png){width=".60\textwidth"}
Dynamic mode decomposition as accelerator of the single simulations {#sec:dmd}
===================================================================
Dynamic mode decomposition is a data-driven modal decomposition technique for analysing the dynamics of nonlinear systems [@schmid2010dynamic; @schmid2011applications]. A comprehensive overview on DMD and its major variants is in [@kutz2016dynamic]. Other nonintrusive approaches with randomized DMD can be found in [@bistrian2017randomized; @bistrian2018efficiency], while naval engineering applications are in [@demo2018shape; @demo2018isope].
Here we present a brief overview of the method and how we integrate it in the computational pipeline we propose. Let us consider $m$ snapshots representing the state of the system for a given time interval: $\{\boldsymbol{x}_i\}_{i=1}^{m} \in \mathbb{R}^n$. We seek a linear operator $\mathbf{A}$ to approximate the nonlinear dynamics of the state variable $\boldsymbol{x}$, that is $\boldsymbol{x}_{k+1}
= \mathbf{A} \boldsymbol{x}_k$. In order to find the DMD decomposition we only need to approximate the eigenpairs of the operator $\mathbf{A}$, without explicitly compute it. We proceed by dividing the snapshots in two matrices $\mathbf{X}$ and $\mathbf{Y}$ as in the following: $$\label{eq:matarranged}
\mathbf{X} =
\begin{bmatrix}
x_1^1 & x_2^1 & \cdots & x_{m-1}^1 \\
x_1^2 & x_2^2 & \cdots & x_{m-1}^2 \\
\vdots & \vdots & \ddots & \vdots \\
x_1^n & x_2^n & \cdots & x_{m-1}^n
\end{bmatrix},\quad\quad
\mathbf{Y} =
\begin{bmatrix}
x_2^1 & x_3^1 & \cdots & x_m^1 \\
x_2^2 & x_3^2 & \cdots & x_m^2 \\
\vdots & \vdots & \ddots & \vdots \\
x_2^n & x_3^n & \cdots & x_m^n
\end{bmatrix}.$$ With this representation we seek $\mathbf{A}$ such that $\mathbf{Y}
\approx \mathbf{A} \mathbf{X}$. Using the Moore-Penrose pseudo-inverse operator, denoted by $^\dagger$, we express the best-fit matrix as $\mathbf{A} = \mathbf{Y} \mathbf{X}^\dagger$. We can compute the POD modes of the matrix $\mathbf{X}$ and project the data onto the subspace defined by them. We use the truncated singular value decomposition obtaining $\mathbf{X} \approx \mathbf{U}_r
\boldsymbol{\Sigma}_r \mathbf{V}^*_r$, where the unitary matrix $\mathbf{U}_r$ contains the first $r$ modes. With these modes we can compute the reduced operator $\mathbf{\tilde{A}} \in
\mathbb{C}^{r\times r}$ as $\mathbf{\tilde{A}} = \mathbf{U}_r^* \mathbf{A} \mathbf{U}_r =
\mathbf{U}_r^* \mathbf{Y} \mathbf{X}^\dagger \mathbf{U}_r =
\mathbf{U}_r^* \mathbf{Y} \mathbf{V}_r \boldsymbol{\Sigma}_r^{-1} \mathbf{U}_r^* \mathbf{U}_r =
\mathbf{U}_r^* \mathbf{Y} \mathbf{V}_r \boldsymbol{\Sigma}_r^{-1}$, without the explicit computation of the full operator $\mathbf{A}$. The reduced operator describe the evolution of the low-rank approximated state $\boldsymbol{\tilde{x}}_k \in \mathbb{R}^r$ as $\boldsymbol{\tilde{x}}_{k+1} = \mathbf{\tilde{A}}
\boldsymbol{\tilde{x}}_k$. We can then recover the high-dimensional state $\boldsymbol{x}_k$ using the POD modes already computed: $\boldsymbol{x}_k = \mathbf{U}_r \boldsymbol{\tilde{x}}_k$.
Using the eigendecomposition of the matrix $\mathbf{\tilde{A}}$, that is $\mathbf{\tilde{A}} \mathbf{W} = \mathbf{W} \boldsymbol{\Lambda}$, we are able to compute the eigenpairs of the full operator $\mathbf{A}$. In particular the eigenvalues in $\boldsymbol{\Lambda}$ correspond to the nonzero eigenvalues of $\mathbf{A}$, while the eigenvectors $\boldsymbol{\Phi}$ of $\mathbf{A}$ can be computed in two ways: by projecting the low-rank approximation $\mathbf{W}$ on the high-dimensional space $\boldsymbol{\Phi} = \mathbf{U}_r \mathbf{W}$, or by computing them exactly with $\boldsymbol{\Phi} =
\mathbf{Y}\mathbf{V}_r \boldsymbol{\Sigma}_r^{-1} \mathbf{W}$.
![Example of total drag reconstructed using the DMD algorithm. On the left the reconstructed field, while on the right the absolute error with respect to the high-fidelity solution.[]{data-label="fig:dmd"}](dmd.png){width=".7\textwidth"}
The actual implementation of the DMD algorithm we used and many different variants from multiresolution DMD [@kutz2016multiresolution], to DMD with control [@proctor2016dynamic], and higher order DMD [@le2017higher], can be found in the open source Python package PyDMD [@demo2018pydmd].
In this work the DMD is used to accelerate the computation of the total drag resistance for a given deformed hull. It uses only few snapshots of the high-fidelity simulation, equispaced in time, to predict the evolution of the target output. In particular we are interested in the value of the total drag at regime. Figure \[fig:dmd\] reports the forces field approximated using DMD and the absolute error with respect to the high-fidelity solution for a particular geometrical configuration.
How to reduce the parameter space dimension with active subspaces {#sec:active}
=================================================================
The active subspaces (AS) property has been formalized by Constantine in [@constantine2015active; @constantine2014active]. It is a property of a scalar function $f : \mathbb{R}^N \to \mathbb{R}$ and a probability density function $\rho : \mathbb{R}^N \to \mathbb{R}^+$, where $N$ is the number of the input parameters. Taking linear combinations of the original parameters we can approximate $f$ using these new parameters, thus reducing the parameter space dimension. The output of interest $f({\ensuremath{\boldsymbol{\mu}}}_{\text{GEO}})$, in our case the total drag of the hull advancing in calm water, depends on the geometrical parameters introduced in Section \[sec:ffd\], while $\rho$ describe the uncertainty in the model inputs, i.e. how we sample the parameter space. For sake of clarity we will drop the pedix and from now on $f({\ensuremath{\boldsymbol{\mu}}}) := f({\ensuremath{\boldsymbol{\mu}}}_{\text{GEO}})$. The general idea is to rotate the input domain, after a proper rescale, in order to unveil a low dimensional parametrization of $f$, which means to find proper directions in the input space where $f$ varies the most on average. We do so by checking the gradients of the output of interest with respect to the parameters.
To proper exploit the AS property we introduce some hypotheses: $f$ has to be continuous and differentiable with square-integrable partial derivatives in the support of $\rho$. Then we introduce the uncentered covariance matrix $\mathbf{C}$ of the gradients of the target function, which is the matrix constructed with the average products of partial derivatives of the map $f$ as follows
$$\label{eq:uncentered}
\mathbf{C} = \mathbb{E}\, [\nabla_{{\ensuremath{\boldsymbol{\mu}}}} f \, \nabla_{{\ensuremath{\boldsymbol{\mu}}}} f
^T] =\int (\nabla_{{\ensuremath{\boldsymbol{\mu}}}} f) ( \nabla_{{\ensuremath{\boldsymbol{\mu}}}} f )^T
\rho \, d {\ensuremath{\boldsymbol{\mu}}},$$
where with $\mathbb{E}$ we identify the expected value, and $\nabla_{{\ensuremath{\boldsymbol{\mu}}}} f =
\nabla f({\ensuremath{\boldsymbol{\mu}}}) = \left [ \frac{\partial f}{\partial \mu_1}, \dots,
\frac{\partial f}{\partial \mu_p} \right ]^T$ is the column vector of partial derivatives of $f$. Since $\mathbf{C}$ is symmetric we can express it with its real eigenvalue decomposition $\mathbf{C} =
\mathbf{W} \mathbf{\Lambda} \mathbf{W}^T$, where $\mathbf{W}$ is the eigenvectors matrix, and $\mathbf{\Lambda}$ the diagonal matrix with the eigenvalues in descending order. It can be proven that the eigenvalues express the amount of variance of the gradient along the corresponding eigenvector direction. This means that taking the first $M$ most energetic eigenvalues and the corresponding eigenvectors, we can approximate the target function with a reduce number of input parameters. So the eigenpairs of $\mathbf{C}$ define the active subspaces of the pair $(f, \rho)$. We proceed by partitioning $\mathbf{W}$ and $\mathbf{\Lambda}$ as follows $$\mathbf{\Lambda} = \begin{bmatrix} \mathbf{\Lambda}_1 & \\
&
\mathbf{\Lambda}_2\end{bmatrix},
\qquad
\mathbf{W} = \left [ \mathbf{W}_1 \quad \mathbf{W}_2 \right ],$$ where the pedix $1$ means the first $M$ eigenvalues and eigenvectors respectively. Now we can use $\mathbf{W}_1$ to project the original parameters to the active subspace, that is the span of the first $M$ eigenvectors. This means to align the input parameter space to $\mathbf{W}_1$ and retain only the directions where $f$ varies the most on average. We call active variable ${\ensuremath{\boldsymbol{\mu}}}_M$ the range of $\mathbf{W}_1^T$, that is ${\ensuremath{\boldsymbol{\mu}}}_M = \mathbf{W}_1^T{\ensuremath{\boldsymbol{\mu}}}\in
\mathbb{R}^M$. We can thus introduce a lower-dimension approximation $g : \mathbb{R}^M \rightarrow \mathbb{R}$ of the quantity of interest $f$, which is a function of ${\ensuremath{\boldsymbol{\mu}}}_M$ as follows $$f ({\ensuremath{\boldsymbol{\mu}}}) \approx g (\mathbf{W}_1^T {\ensuremath{\boldsymbol{\mu}}}) = g({\ensuremath{\boldsymbol{\mu}}}_M).$$
Active subspaces have been proven useful in naval applications in [@tezzele2018dimension; @tezzele2018model; @demo2018isope], but also coupled with POD-Galerkin model order reduction [@tezzele2018combined]. A gradient-free algorithm for the discovery of active subspaces has been proposed in [@coleman2019gradient], while an AS variant using average gradients in [@lee2019modified].
We are going to find the active subspace for the total drag resistance of the deformed hulls obtained by the FFD method and the application of the DMD algorithm. Then we are going to exploit this active subspace to perform a better sampling of the parameter space and thus enhancing the construction of the reduced order model.
Proper orthogonal decomposition with interpolation {#sec:podi}
==================================================
Reduced order modeling (ROM) is a popular technique to reduce the computational cost of numerical simulations. Among all the available methods to achieve this reduction, we focus in this contribution to the reduced basis method using the proper orthogonal decomposition (POD) algorithm for the basis identification. This method allows to reduce the number of degrees of freedom of a parametric system by collecting the [*snapshots*]{} — the full order system outputs — for several different configurations and combining them in an efficient way for a real-time approximation of new solutions (for any new configuration). In the POD reduction framework, we can discern two main techniques: POD-Galerkin, which requires all the details of the full order system to generate a consistent low-dimensional representation of the physical problem, and POD with interpolation (PODI), which instead requires only the snapshots. Due to these requirements, the PODI method is particularly suited for industrial problem, since it is able to been coupled to all the numerical solvers, even commercial ones. In this contribution, we adopt PODI method. For more information about POD-Galerkin, we suggest [@stabile2018finite; @StabileHijaziLorenziMolaRozza2017; @KaratzasBallarinRozza2019; @georgaka2018parametric], while for other examples of PODI applications we recommend [@garotta2018quiet; @demo2018ezyrb; @salmoiraghi2018].
To calculate the POD modes we use the singular value decomposition (SVD) applied to the snapshots matrix $\mathbf{X}$ such that $\mathbf{X} = \mathbf{U}
\mathbf{\Sigma} \mathbf{V}^*$. The columns of the unitary matrix $\mathbf{U}$ are the POD modes end the corresponding singular values, the elements in the diagonal matrix $\mathbf{\Sigma}$ in decresing order, indicate the energy associated to each mode. Hence it is possible to select the first modes — the most energetic — to span the reduced space and project onto it the high-fidelity snapshots. In matricial form, we have: $$\mathbf{X}^{\text{POD}} = \mathbf{U}_N^T \mathbf{X} ,$$ where $\mathbf{U}_N$ is the matrix containing the first $N$ modes, and $\mathbf{X}^{\text{POD}}$ is the matrix whose columns $\mathbf{x}^{\text{POD}}_i$ are the reduced snapshots. We note that $\mathbf{x}^{\text{POD}}_i \in V^N$and $\mathbf{x}_i \in V^\mathcal{N}$ where $\mathcal{N}$ refers to the number of degrees of freedom of the full-order system. Finally, due to the reduced dimension, we are able to interpolate the reduced snapshots in order to approximate the solution manifold. The new interpolated reduced snapshots are then mapped back to the high-dimensional space for a real-time evaluation of the solution. To perform the non-intrusive model order reduction, we use the open source package EZyRB [@demo2018ezyrb].
Numerical results {#sec:results}
=================
Here we are going to present the results of the complete numerical pipeline applied to the DTMB 5415 hull. Moreover we demonstrate the improvements obtained using the proposed pipeline, called POD+AS, with respect to the POD approach on the full parameter space.
After generating $N_{\text{POD}} = 100$ deformed hulls, we perform the high-fidelity simulations accelerated via the DMD algorithm. We construct the snapshots matrix and compute the POD modes and the corresponding eigenvalues for the construction of the reduced output space. We compare this approach with the one proposed in this work that exploits a preprocessing step with the finding of the active subpace for the total drag resistance. With the $N_{\text{POD}}$ input/output couples, we individuate an eigenvector $\mathbf{W}_1$ (compare Section \[sec:active\]) describing an active subspace of dimension 1, and we sample the full space only along the active direction described by this vector. After this second sampling we collect a new set of high-fidelity simulations formed by $N_{\text{POD+AS}} = 80$ snapshots. For this new snapshots matrix we compute again the POD modes and eigenvalues, and we compare the two approaches looking at the POD singular values decay. A faster decay means a better approximation of the output fields for a fixed number of modes. In Figure \[fig:svd\] the blue line shows the singular values $\sigma_i$ divided by the first and greatest singular value $\sigma_{\text{max}}$ for the sampling of the full parameter space; with the dashed red line the POD singular values decay for the POD+AS approach. A faster decay is observed, especially for the first few modes. This translates in an enhanced reduced order model, which exhibits a better approximation of the solutions manifold, with respect to the classical approach.
![POD singular values decay as a function of the number of modes. The blue line corresponds to the original sampling, while the red dotted line, called POD+AS approach, corresponds to the sampling along the active direction.[]{data-label="fig:svd"}](svd_decay.png){width=".8\textwidth"}
Since we are relying on multidimensional interpolation to reconstruct the solutions at untried parameters, having a new reduced parameter space improves the creation of such interpolator. In the POD+AS approach we have to interpolate a univariate function in $N$-dimension, where $N$ is the number of POD modes we retain. In the POD approach on the full parameter space instead, we have the same number of modes to fit but a multivariate function depending on 5 input parameters, resulting in a difficult interpolation.
Conclusions and perspectives {#sec:the_end}
============================
In this work we presented a nonintrusive numerical pipeline for shape optimization of the bulbous bow of a benchmark hull. It comprises automatic geometrical parametrization and morphing through FFD, estimation of the total drag resistance via DMD using only few snapshots of the time-dependent high fidelity simulations, the reduction of the parameter space exploiting the AS property, and the construction of a surrogate model with PODI for the real-time evaluation of the many-query problem solved by an optimization algorithm. We proved that the reduction of the parameter space can further enhance the reduced order model creation. Moreover all this parts of the pipeline can be used and integrated separately into an existing computational workflow resulting in a great interest for industrial applications.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was partially performed in the context of the project SOPHYA - “Seakeeping Of Planing Hull YAchts” supported by Regione FVG, POR-FESR 2014-2020, Piano Operativo Regionale Fondo Europeo per lo Sviluppo Regionale, and partially supported by European Union Funding for Research and Innovation — Horizon 2020 Program — in the framework of European Research Council Executive Agency: H2020 ERC CoG 2015 AROMA-CFD project 681447 “Advanced Reduced Order Methods with Applications in Computational Fluid Dynamics” P.I. Gianluigi Rozza.
[10]{}
: <https://github.com/mathLab/PyGeM>.
F. Ballarin, A. D’Amario, S. Perotto, and G. Rozza. . , 117(8):860–884, 2018.
D. Bistrian and I. Navon. Efficiency of randomised dynamic mode decomposition for reduced order modelling. , 32(2-3):88–103, 2018.
D. A. Bistrian and I. M. Navon. Randomized dynamic mode decomposition for nonintrusive reduced order modelling. , 112(1):3–25, 2017.
M. D. Buhmann. , volume 12. Cambridge university press, 2003.
K. D. Coleman, A. Lewis, R. C. Smith, B. Williams, M. Morris, and B. Khuwaileh. Gradient-free construction of active subspaces for dimension reduction in complex models with applications to neutronics. , 7(1):117–142, 2019.
P. G. Constantine. , volume 2. SIAM, 2015.
P. G. Constantine, E. Dow, and Q. Wang. Active subspace methods in theory and practice: applications to kriging surfaces. , 36(4):A1500–A1524, 2014.
N. Demo, M. Tezzele, G. Gustin, G. Lavini, and G. Rozza. Shape optimization by means of proper orthogonal decomposition and dynamic mode decomposition. In [*Technology and Science for the Ships of the Future: Proceedings of NAV 2018: 19th International Conference on Ship & Maritime Research*]{}, pages 212–219. IOS Press, 2018.
N. Demo, M. Tezzele, A. Mola, and G. Rozza. An efficient shape parametrisation by free-form deformation enhanced by active subspace for hull hydrodynamic ship design problems in open source environment. In [*The 28th International Ocean and Polar Engineering Conference*]{}, 2018.
N. Demo, M. Tezzele, and G. Rozza. . , 3(24):661, 2018.
N. Demo, M. Tezzele, and G. Rozza. . , 3(22):530, 2018.
D. Forti and G. Rozza. Efficient geometrical parametrisation techniques of interfaces for reduced-order modelling: application to fluid–structure interaction coupling problems. , 28(3-4):158–169, 2014.
F. Garotta, N. Demo, M. Tezzele, M. Carraturo, A. Reali, and G. Rozza. . , 2018.
S. Georgaka, G. Stabile, G. Rozza, and M. J. Bluck. Parametric pod-galerkin model order reduction for unsteady-state heat transfer problems. , 2018.
E. N. Karatzas, F. Ballarin, and G. Rozza. Projection-based reduced order models for a cut finite element method in parametrized domains. , 2019.
J. N. Kutz, S. L. Brunton, B. W. Brunton, and J. L. Proctor. , volume 149. SIAM, 2016.
J. N. Kutz, X. Fu, and S. L. Brunton. Multiresolution dynamic mode decomposition. , 15(2):713–735, 2016.
S. Le Clainche and J. M. Vega. Higher order dynamic mode decomposition. , 16(2):882–925, 2017.
M. R. Lee. Modified active subspaces using the average of gradients. , 7(1):53–66, 2019.
A. Manzoni, A. Quarteroni, and G. Rozza. Model reduction techniques for fast blood flow simulation in parametrized geometries. , 28(6-7):604–625, 2012.
A. Morris, C. Allen, and T. Rendall. -based optimization of aerofoils using radial basis functions for domain element parameterization and mesh deformation. , 58(8):827–860, 2008.
J. L. Proctor, S. L. Brunton, and J. N. Kutz. Dynamic mode decomposition with control. , 15(1):142–161, 2016.
G. Rozza, M. W. Hess, G. Stabile, M. Tezzele, and F. Ballarin. Preliminaries and warming-up: Basic ideas and tools. In P. Benner, S. Grivet-Talocia, A. Quarteroni, G. Rozza, W. H. A. Schilders, and L. M. Silveira, editors, [*Handbook on Model Order Reduction*]{}, volume 1, chapter 1. De Gruyter, 2019.
G. Rozza, A. Koshakji, and A. Quarteroni. . , 4(0):1–26, 2013.
G. Rozza, M. H. Malik, N. Demo, M. Tezzele, M. Girfoglio, G. Stabile, and A. Mola. . Glasgow, UK, 2018. ECCOMAS Proceedings.
F. Salmoiraghi, F. Ballarin, G. Corsi, A. Mola, M. Tezzele, and G. Rozza. Advances in geometrical parametrization and reduced order models and methods for computational fluid dynamics problems in applied sciences and engineering: Overview and perspectives. , 1:1013–1031, 2016.
F. Salmoiraghi, A. Scardigli, H. Telib, and G. Rozza. Free-form deformation, mesh morphing and reduced-order methods: enablers for efficient aerodynamic shape optimisation. , 0(0):1–15, 2018.
P. J. Schmid. Dynamic mode decomposition of numerical and experimental data. , 656:5–28, 2010.
P. J. Schmid, L. Li, M. P. Juniper, and O. Pust. Applications of the dynamic mode decomposition. , 25(1-4):249–259, 2011.
T. Sederberg and S. Parry. Free-[F]{}orm [D]{}eformation of solid geometric models. In [*Proceedings of SIGGRAPH - Special Interest Group on GRAPHics and Interactive Techniques*]{}, pages 151–159. SIGGRAPH, 1986.
D. Shepard. A two-dimensional interpolation function for irregularly-spaced data. In [*Proceedings-1968 ACM National Conference*]{}, pages 517–524. ACM, 1968.
G. Stabile, S. N. Hijazi, S. Lorenzi, A. Mola, and G. Rozza. Pod-galerkin reduced order methods for cfd using finite volume discretisation: Vortex shedding around a circular cylinder. , 8(1):210–236, dec 2017.
G. Stabile and G. Rozza. Finite volume pod-galerkin stabilised reduced order methods for the parametrised incompressible navier–stokes equations. , 173:273–284, 2018.
M. Tezzele, F. Ballarin, and G. Rozza. . In D. Boffi, L. F. Pavarino, G. Rozza, S. Scacchi, and C. Vergara, editors, [*Mathematical and Numerical Modeling of the Cardiovascular System and Applications*]{}, pages 185–207. Springer International Publishing, 2018.
M. Tezzele, N. Demo, M. Gadalla, A. Mola, and G. Rozza. Model order reduction by means of active subspaces and dynamic mode decomposition for parametric hull shape design hydrodynamics. In [*Technology and Science for the Ships of the Future: Proceedings of NAV 2018: 19th International Conference on Ship & Maritime Research*]{}, pages 569–576. IOS Press, 2018.
M. Tezzele, N. Demo, A. Mola, and G. Rozza. . , 2018.
M. Tezzele, F. Salmoiraghi, A. Mola, and G. Rozza. Dimension reduction in heterogeneous parametric spaces with application to naval engineering shape design problems. , 5(1):25, Sep 2018.
[^1]: [email protected]
[^2]: [email protected]
[^3]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Qichuan Geng$^1$Hong ZhangXiaojuan Qi$^2$Ruigang Yang$^{3}$Zhong Zhou$^1$Gao Huang$^{4}$\
$^1$Beihang University $^2$University of Oxford $^3$University of Kentucky$^4$ Tsinghua University\
[{zhaokefirst,zz}@buaa.edu.com]{}\
bibliography:
- 'citation.bib'
title: Gated Path Selection Network for Semantic Segmentation
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We introduce a notion of full field algebra which is essentially an algebraic formulation of the notion of genus-zero full conformal field theory. For any vertex operator algebras $V^{L}$ and $V^{R}$, $V^{L}\otimes V^{R}$ is naturally a full field algebra and we introduce a notion of full field algebra over $V^{L}\otimes V^{R}$. We study the structure of full field algebras over $V^{L}\otimes
V^{R}$ using modules and intertwining operators for $V^{L}$ and $V^{R}$. For a simple vertex operator algebra $V$ satisfying certain natural finiteness and reductive conditions needed for the Verlinde conjecture to hold, we construct a bilinear form on the space of intertwining operators for $V$ and prove the nondegeneracy and other basic properties of this form. The proof of the nondegenracy of the bilinear form depends not only on the theory of intertwining operator algebras but also on the modular invariance for intertwining operator algebras through the use of the results obtained in the proof of the Verlinde conjecture by the first author. Using this nondegenerate bilinear form, we construct a full field algebra over $V\otimes V$ and an invariant bilinear form on this algebra.
author:
- 'Yi-Zhi Huang and Liang Kong'
title: ' [**Full field algebras**]{} '
---
Introduction
============
In the present paper, we solve the problem of constructing a genus-zero full conformal field theory (a conformal field theory on genus-zero Riemann surfaces containing both chiral and antichiral parts) from representations of a simple vertex operator algebra $V$ satisfying the following conditions: (i) $V_{(n)}=0$ for $n<0$, $V_{(0)}=\mathbb{C}\mathbf{1}$, and $W_{(0)}=0$ for any irreducible $V$-module $W$ which is not equivalent to $V$. (ii) Every $\mathbb{N}$-gradable weak $V$-module is completely reducible. (iii) $V$ is $C_{2}$-cofinite. Note that the last two conditions are equivalent to a single condition that every weak $V$-module is completely reducible (see [@L] and [@ABD]).
Conformal field theories in its original form, as formulated by Belavin, Polyakov and Zamolodchikov [@BPZ] and by Kontsevich and Segal [@S1] [@S2] [@S3], have both chiral and antichiral parts. The fundamental work [@MS1] [@MS2] of Moore and Seiberg is also based on the existence of such full conformal field theories with both chiral and antichiral parts. In mathematics, however, it is mostly chiral conformal field theories that are constructed and studied. To use conformal field theory to solve mathematical problems and to understand mathematical results such as mirror symmetry, we need full conformal field theories, not just chiral or antichiral ones.
In the case of conformal field theories associated to tori, Tsukada first constructed and studied these theories mathematically in his Ph. D. thesis under the direction of I. Frenkel (see [@Ts]). Assuming the existence of the structure of a modular tensor category on the category of modules for a vertex operator algebra, the existence of conformal blocks with monodromies compatible with the modular tensor category and all the necessary convergence properties, Felder, Fröhlich, Fuchs and Schweigert [@FFFS] and Fuchs, Runkel, Schweigert and Fjelstad [@FRS1] [@FRS2] [@FRS3] [@FFRS] studied open-closed conformal field theories (in particular full (closed) conformal field theories) using the theory of tensor categories and three-dimensional topological field theories. They constructed correlation functions as states in some three-dimensional topological field theories and they showed the existence of consistent operator product expansion coefficients for bulk operators. However, since these works were based on the fundamental assumptions mentioned above, an explicit construction of the corresponding full conformal field theories, even in the genus-zero case, is still needed.
In [@KO], Kapustin and Orlov studied full conformal field theories associated to tori. They introduced a notion of vertex algebra which is more general than the original notion of vertex algebra [@B] or vertex operator algebra [@FLM2] by allowing both chiral and antichiral parts. In [@R1] and [@R2], Rosellen studied these algebras in details. However, the construction of the full conformal field theories associated to affine Lie algebras (the WZNZ models) and to the Virasoro algebra (the minimal models) was still an open problem. More generally, we would like to construct full conformal field theories from the representations of a vertex operator algebra satisfying reasonable conditions. Also, since the braid group representations obtained from the representations of these vertex operator algebras are not one dimensional in general, it seems that the corresponding full conformal field theories in general do not satisfy the axioms for the algebras introduced and studied in [@KO], [@R1] and [@R2].
In [@H5], [@H6] and [@H8], the first author constructed genus-zero chiral theories, genus-one chiral theories and modular tensor categories from the representations of simple vertex operator algebras satisfying the three conditions above. Since modular tensor categories give modular functors (see [@T] and [@BK]), these results also give modular functors. One of the remaining problems is to construct a full conformal field theory from a chiral theory and an antichiral theory obtained from the chiral theory. In the present paper, we solve this problem in the genus-zero case by constructing a full conformal field theory corresponding to what physicists call a diagonal theory (see, for example, [@MS3]) from the representations of a simple vertex operator algebra satisfying the conditions above. The same full conformal field theory is also constructed by the second author using the theory of tensor categories in [@K]. The genus-one case and the higher-genus case can be obtained from the construction of the genus-zero theories in this paper and the properties of genus-one and higher-genus chiral theories. These will be discussed in future publications.
Technically, we construct a genus-zero full conformal field theory as follows: We first introduce a notion of full field algebra and several variants, which are essentially algebraic formulations of the notion of genus-zero full conformal field theory (for a precise discussion of the equivalence of this notion of full field algebra and its variants with geometric formulations of genus-zero conformal field theories in terms of operads, see [@K]). For a simple vertex operator algebra satisfying the three conditions above, by the results in [@H5], we have an intertwining operator algebra, which is equivalent to a genus-zero chiral conformal field theory (see [@H3] and [@H3.5]). The genus-zero chiral conformal field theory also gives a genus-zero antichiral conformal field theory. We construct a nondegenerate bilinear form on the space of intertwining operators and use this bilinear form to put the genus-zero chiral and antichiral conformal field theories together. We show that the resulting mathematical object is a full field algebra satisfying additional properties and thus gives a genus-zero full conformal field theory. One interesting aspect of our construction is that our construction (actually the proof of the nondegeneracy of the bilinear form on the space of the intertwining operators) needs the theorem proved in [@H7] (see also [@H9]) stating that the Verlinde conjecture holds for such a vertex operator algebra. This theorem in [@H7], and thus also our construction of genus-zero full conformal field theories, depend not only on genus-zero chiral theories constructed in [@H5], but also on genus-one chiral theories constructed in [@H6].
This paper is organized as follows: In Section 1, we introduce the notion of full field algebra and several variants and discuss their basic properties. In Section 2, we discuss basic relations between intertwining operator algebras and full field algebras. This is a section preparing for our construction in Section 3. Our construction of full field algebras is given in Section 3. We also construct invariant bilinear forms on these full field algebras in the same section.
#### Acknowledgment
The first author is partially supported by NSF grant DMS-0401302.
Definitions and basic properties
================================
Let $\mathbb{F}_{n}({\mathbb{C}}) =\{ (z_{1}, \ldots, z_{n})\in {\mathbb{C}}^n \;|\; z_i\neq z_j
\mbox{ if } i\neq j \}$. For an ${\mathbb{R}}$-graded vector space $F=\coprod_{r\in {\mathbb{R}}}F_{(r)}$, we let $\overline{F}=\prod_{r\in {\mathbb{R}}}F_{(r)}$ be the algebraic completion of $F$. For $r\in {\mathbb{R}}$, let $P_{r}$ be the projection from $F$ or $\overline{F}$ to $F_{(r)}$. A series $\sum f_{n}$ in $\overline{F}$ is said to be [*absolutely convergent*]{} if for any $f'\in F'$, $\sum|\langle f', f_{n}\rangle|$ is convergent. The sums $\sum|\langle f', f_{n}\rangle|$ for $f'\in
F'$ define a linear functional on $F'$. We call this linear functional the [*sum*]{} of the series and denote it by the same notation $\sum f_{n}$. If the homogeneous subspaces of $F$ are all finite-dimensional, then $\overline{F}=(F')^{*}$ and, in this case, the sum of an absolutely convergent series is always in $\overline{F}$. When the sum is in $\overline{F}$, we say that the series is [*absolutely convergent in $\overline{F}$*]{}.
\[ffa\]
A [*full field algebra*]{} is an ${\mathbb{R}}$-graded vector space $F=\coprod_{r\in{\mathbb{R}}}F_{(r)}$ (graded by [*total conformal weight*]{} or simply [*total weight*]{}), equipped with [*correlation function maps*]{} $$\begin{array}{rrcl}
m_{n}: & F^{\otimes n}\times
\mathbb{F}_{n}({\mathbb{C}})&\to& \overline{F}\\
&(u_{1}\otimes \cdots \otimes u_{n}, (z_{1}, \dots, z_{n}))&\mapsto&
m_{n}(u_{1}, \dots, u_{n}; z_{1}, \bar{z}_{1}, \dots, z_{n}, \bar{z}_n),
\end{array}$$ for $n\in {\mathbb{Z}}_+$ and a distinguished element $\mathbf{1}$ called [*vacuum*]{} satisfying the following axioms:
1. For $n\in {\mathbb{Z}}_+$, $m_n(u_{1}, \dots, u_{n}; z_{1}, \bar{z}_{1}, \dots, z_{n}, \bar{z}_n)$ is linear in $u_{1}, \dots, u_{n}$ and smooth in the real and imaginary parts of $z_{1}, \dots, z_{n}$.
2. For $u\in F$, $m_{1}(u; 0, 0)=u$.
3. For $n\in {\mathbb{Z}}_{+}$, $u_{1}, \dots, u_{n}\in F$, $$\begin{aligned}
&m_{n+1}(u_{1}, \dots, u_{n}, {\mathbf{1}}; z_{1}, \bar{z}_{1}, \dots, z_{n}, \bar{z}_{n},
z_{n+1}, \bar{z}_{n+1})&{\nonumber \\}&=m_n(u_{1}, \dots, u_{n}; z_{1}, \bar{z}_{1}, \dots, z_{n}, \bar{z}_{n}).&\end{aligned}$$
4. The [*convergence property*]{}: For $k, l_{1}, \ldots, l_{k} \in {\mathbb{Z}}_+$ and $u_{1}^{(1)},\ldots, u_{l_{1}}^{(1)},\ldots, u_{1}^{(k)}$, $\dots,
u_{l_{k}}^{(k)} \in F$, the series $$\begin{aligned}
\label{ffa-conv-axiom}
\lefteqn{\sum_{r_{1}, \ldots, r_k} m_{k}(P_{r_{1}} m_{l_{1}}(u_{1}^{(1)},
\ldots, u_{l_{1}}^{(1)}; z_{1}^{(1)}, \bar{z}_{1}^{(1)}, \dots,
z_{l_{1}}^{(1)}, \bar{z}_{l_{1}}^{(1)} ), \dots,} {\nonumber \\}&&
P_{r_k} m_{l_{k}} (u_{1}^{(k)}, \ldots, u_{l_{k}}^{(k)},
z_{1}^{(k)}, \bar{z}_{1}^{(k)}, \ldots, z_{l_{k}}^{(k)},
\bar{z}_{l_{k}}^{(k)}); z^{(0)}_{1}, \bar{z}_{1}^{(0)}, \ldots,
z^{(0)}_{k}, \bar{z}^{(0)}_k){\nonumber \\}&&\end{aligned}$$ converges absolutely to $$\begin{aligned}
\label{sum}
\lefteqn{m_{l_{1}+ \cdots + l_k}(u_{1}^{(1)}, \dots, u_{l_k}^{(k)};
z_{1}^{(1)}+z^{(0)}_{1}, \bar{z}_{1}^{(1)}+ \bar{z}^{(0)}_{1}, \dots,
z_{l_{1}}^{(1)}+z^{(0)}_{1}, } {\nonumber \\}&&\quad
\bar{z}_{l_{1}}^{(1)}+\bar{z}^{(0)}_{1},
\dots,
z_{1}^{(k)}+z^{(0)}_k,
\bar{z}_{1}^{(k)}+ \bar{z}^{(0)}_k, \dots,
z_{l_k}^{k}+z^{(0)}_k, \bar{z}_{l_k}^{k}+\bar{z}^{(0)}_k). {\nonumber \\}&&\end{aligned}$$ when $|z_p^{(i)}| + |z_q^{(j)}|< |z^{(0)}_i
-z^{(0)}_j|$ for $i,j=1, \ldots, k$, $i\ne j$ and for $p=1,
\dots, l_i$ and $q=1, \dots, l_j$.
5. The [*permutation property*]{}: For any $n\in {\mathbb{Z}}_{+}$ and any $\sigma\in S_{n}$, we have $$\begin{aligned}
\lefteqn{m_{n}(u_{1}, \dots, u_{n}; z_{1}, \bar{z}_{1},
\dots, z_{n}, \bar{z}_{n})} {\nonumber \\}&& = m_{n}(u_{\sigma(1)}, \dots, u_{\sigma(n)};
z_{\sigma(1)}, \bar{z}_{\sigma(1)}, \dots,
z_{\sigma(n)}, \bar{z}_{\sigma(n)}) \label{ffa-perm-axiom}\end{aligned}$$ for $u_{1}, \dots, u_{n}\in F$ and $(z_{1}, \dots, z_{n})\in
\mathbb{F}_{n}({\mathbb{C}})$.
6. Let $\mathbf{d}$ be the grading operator, that is, the operator defined by $\mathbf{d}f=rf$ for $f\in F_{(r)}$. Then for $n\in {\mathbb{Z}}_{+}$, $a\in {\mathbb{R}}$, $u_{1},
\dots, u_{n}\in F$, $$\begin{aligned}
\lefteqn{e^{a\mathbf{d}}m_{n}(u_{1}, \dots, u_{n}; z_{1}, \bar{z}_{1},
\dots, z_{n}, \bar{z}_{n})}{\nonumber \\}&&=m_{n}(e^{a\mathbf{d}}u_{1}, \dots, e^{a\mathbf{d}}u_{n};
e^{a}z_{1}, e^{a}\bar{z}_{1},
\dots, e^{a}z_{n}, e^{a}\bar{z}_{n}).\end{aligned}$$
We denote the full field algebra defined above by $(F, m, \mathbf{1})$ or simply by $F$. In the definition above, we use the notations $$m_{n}(u_{1}, \dots, u_{n}; z_{1},
\bar{z}_{1} \dots, z_{n}, \bar{z}_n)$$ instead of $$m_{n}(u_{1}, \dots,
u_{n}; z_{1},\dots, z_{n})$$ to emphasis that these are in general not holomorphic in $z_{1},\dots, z_{n}$. For $u'\in F'$, $u_{1}, \dots, u_{n}\in F$, $$\langle u', m_{n}(u_{1}, \dots, u_{n}; z_{1},
\bar{z}_{1} \dots, z_{n}, \bar{z}_n)\rangle$$ as a function of $z_{1}, \dots, z_{n}$ is called a [*correlation function*]{}. [*Homomorphisms*]{} and [*isomorphisms*]{} for full field algebras are defined in the obvious way.
\[weak-c-p\] [Note that in the convergence property, we require that the multisum is absolutely convergent. This is stronger than the following convergence property: For $k, l \in {\mathbb{Z}}_+$ and $u_{1},\ldots, u_{k-1}, v_{1}, \dots, v_{l}\in F$, the series $$\begin{aligned}
\lefteqn{\sum_{r} m_{k}(u_{1}, \dots, u_{k-1},
P_{r} m_{l} (v_{1}, \dots, v_{l};
z_{1}^{(k)}, \bar{z}_{1}^{(k)}, \dots, z_{l}^{(k)},
\bar{z}_{l}^{(k)}); } {\nonumber \\}&& \quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad z_{1}^{(0)}, \bar{z}_{1}^{(0)}, \ldots,
z_{k}^{(0)}, \bar{z}_k^{(0)}){\nonumber \\}&&\end{aligned}$$ converges absolutely to $$\begin{aligned}
\lefteqn{m_{k+ l}(u_{1}, \dots, u_{k-1}, v_{1}, \dots, v_{l};
z^{(0)}_{1}, \bar{z}^{(0)}_{1}, \dots,
z^{(0)}_{k-1}, } {\nonumber \\}&&\quad
\bar{z}^{(0)}_{k-1},
\dots,
z_{1}^{(k)}+z^{(0)}_k,
\bar{z}_{1}^{(k)}+ \bar{z}^{(0)}_k, \dots,
z_{l}^{k}+z^{(0)}_k, \bar{z}_{l}^{k}+\bar{z}^{(0)}_k). {\nonumber \\}&&\end{aligned}$$ when $z_{i}^{(0)}\ne z_{j}^{(0)}$ for $i,j=1, \ldots, k$ and $|z_p^{(k)}|< |z^{(0)}_k
-z^{(0)}_i|$ for $i=1, \ldots, k-1$, and for $p=1, \dots, l$. However, for the purpose of constructing genus-zero conformal field theory satisfying geometric axioms, this version of the convergence property is actually enough.]{}
Let $(F, m, \mathbf{1})$ be a full field algebra and let $$\begin{array}{rrcl}
\mathbb{Y}:& F^{\otimes 2} \times {\mathbb{C}}^{\times} &\to&
\overline{F}\\
&(u\otimes v, z, \bar{z})&\mapsto& \mathbb{Y}(u; z, \bar{z})v
\end{array}$$ be given by $$\mathbb{Y}(u; z, \bar{z})v = m_{2}(u \otimes v; z,\bar{z}, 0, 0)$$ for $u, v\in F$. The map $\mathbb{Y}$ is called the [*full vertex operator map*]{} and for $u\in F$, $\mathbb{Y}(u; z, \bar{z})$ is called the [ *full vertex operator*]{} associated to $u$. We have the following immediate consequences of the definition:
1. The [*identity property*]{}: $\mathbb{Y}(\mathbf{1}; z, \bar{z})=I_{F}$.
2. The [*creation property*]{}: ${{\displaystyle}\lim_{z\to 0}\mathbb{Y}(u; z, \bar{z})
\mathbf{1}=u}$.
3. For $f\in F$, $$\label{d-bracket}
\left[ \mathbf{d}, \mathbb{Y}(u; z, \bar{z})\right]
= \left( z\frac{\partial}{\partial z} + \bar{z}\frac{\partial}{\partial
\bar{z}} \right) \mathbb{Y}(u; z, \bar{z})
+\mathbb{Y}(\mathbf{d} u; z,\bar{z})$$
4. The total weight of the vacuum ${\mathbf{1}}$ is $0$, that is, $\mathbf{d}{\mathbf{1}}=0$.
[[*Proof.*]{}]{}For $u\in F$, $$\begin{aligned}
\mathbb{Y}(\mathbf{1}; z, \bar{z})u&=&m_{2}({\mathbf{1}}, u; z, \bar{z},
0, 0){\nonumber \\}&=&m_{2}(u, {\mathbf{1}}; 0, 0, z, \bar{z}){\nonumber \\}&=&m_{1}(u; 0, 0){\nonumber \\}&=&u.\end{aligned}$$
For $u\in F$, $$\begin{aligned}
\lim_{z\to 0}\mathbb{Y}(u; z, \bar{z})
\mathbf{1}&=&\lim_{z\to 0}m_{2}(u, {\mathbf{1}}; z, \bar{z},
0, 0){\nonumber \\}&=&\lim_{z\to 0}m_{1}(u; z, \bar{z}){\nonumber \\}&=&m_{1}(u; 0, 0){\nonumber \\}&=&u.\end{aligned}$$
For $u, v\in F$ and $a\in {\mathbb{R}}$, $$\begin{aligned}
\label{d-bracket-0}
e^{a\mathbf{d}}\mathbb{Y}(u; z, \bar{z})v
&=&\mathbb{Y}(e^{a\mathbf{d}}u; e^{a}z, e^{a}\bar{z})e^{a\mathbf{d}}v.\end{aligned}$$ Taking derivatives of both sides of (\[d-bracket-0\]) with respect to $a$, letting $a=0$ and noticing that $v$ is arbitrary, we obtain (\[d-bracket\]).
From the identity property, $$\left(z\frac{\partial}{\partial z}
+\bar{z}\frac{\partial}{\partial \bar{z}}\right)
{\mathcal{Y}}({\mathbf{1}}; z, \bar{z})=0.$$ Then by (\[d-bracket\]), we obtain $$\label{d-bracket-1}
\left[\mathbf{d}, \mathbb{Y}({\mathbf{1}}; z, \bar{z})\right]
=\mathbb{Y}(\mathbf{d} {\mathbf{1}}; z,\bar{z}).$$ Since $\mathbb{Y}({\mathbf{1}}; z, \bar{z})=I_{F}$, (\[d-bracket-1\]) gives $$\label{d-bracket-2}
\mathbb{Y}(\mathbf{d} {\mathbf{1}}; z,\bar{z})=0.$$ Applying (\[d-bracket-2\]) to ${\mathbf{1}}$, taking the limit $z\to 0$ on both sides of the resulting formula and using the creation property, we obtain $\mathbf{d} {\mathbf{1}}=0$. So the total weight of ${\mathbf{1}}$ is $0$.
Now we discuss two important properties of full field algebras which follow also immediately from the definition.
For $u_{1}, u_{2}, u_{3}\in F$, $$\label{asso-1}
\mathbb{Y}(u_{1}; z_{1}, \bar{z}_{1})\mathbb{Y}(u_{2}; z_{2}, \bar{z}_{2})u_{3} =
\mathbb{Y}(\mathbb{Y}(u_{1}; z_{1}-z_{2}, \bar{z}_{1}-\bar{z}_{2})u_{2}; z_{2},\bar{z}_{2})u_{3}$$ when $|z_{1}|>|z_{2}|>|z_{1}-z_{2}|>0$.
[[*Proof.*]{}]{}The convergence property says, in particular, that the series $$\label{prod-series}
\mathbb{Y}(u_{1}; z_{1}, \bar{z}_{1})\mathbb{Y}(u_{2}; z_{2},
\bar{z}_{2})u_{3}=
\sum_{n\in {\mathbb{R}}} \mathbb{Y}(u_{1}; z_{1}, \bar{z}_{1}) P_n
\mathbb{Y}(u_{2}; z_{2}, \bar{z}_{2})u_{3},$$ (a [*product*]{} of full vertex operators) converges absolutely in $\overline{F}$ for $u_{1}, u_{2}, u_{3} \in F$ when $|z_{1}|>|z_{2}|>0$. The convergence property also says, in particular, that the series $$\begin{aligned}
\label{it-series}
&&\mathbb{Y}(\mathbb{Y}(u_{1}; z_{1}-z_{2},\bar{z}_{1}-\bar{z}_{2})u_{2},
z_{2},\bar{z}_{2})u_{3} {\nonumber \\}&&\hspace{2cm}= \sum_{n\in {\mathbb{R}}}\mathbb{Y}(P_n
\mathbb{Y}(u_{1}; z_{1}-z_{2},\bar{z}_{1}-\bar{z}_{2})u_{2}, z_{2},\bar{z}_{2})u_{3}\end{aligned}$$ (an [*iterate*]{} of full vertex operators) converges absolutely for $u_{1}, u_{2}, u_{3}\in F$ when $|z_{2}|>|z_{1}-z_{2}|>0$. Moreover, the convergence property also says that both (\[prod-series\]) and (\[it-series\]) converge absolutely to $$m_{3}(u_{1}, u_{2}, u_{3}; z_{1}, \bar{z}_{1}, z_{2}, \bar{z}_{2}, 0, 0).$$ This proves the associativity.
\[commu-general\] For $u_{1}, u_{2}, u_{3}\in F$, $$\begin{aligned}
\mathbb{Y}(u_{1}; z_{1}, \bar{z}_{1})\mathbb{Y}(u_{2}; z_{2},
\bar{z}_{2})u_{3}, \label{prod-series-1}\\
\mathbb{Y}(u_{2}; z_{2}, \bar{z}_{2})\mathbb{Y}(u_{1}; z_{1},
\bar{z}_{1})u_{3}, \label{prod-series-2}\end{aligned}$$ are the expansions of $$m_{3}(u_{1}, u_{2}, u_{3}; z_{1}, \bar{z}_{1}, z_{2}, \bar{z}_{2}, 0, 0).$$ in the sets given by $|z_{1}|>|z_{2}|>0$ and $|z_{2}|>|z_{1}|>0$, respectively.
[[*Proof.*]{}]{}By the convergence property, we know that (\[prod-series-1\]) and (\[prod-series-2\]) converge absolutely to $$m_{3}(u_{1}, u_{2}, u_{3}; z_{1}, \bar{z}_{1}, z_{2}, \bar{z}_{2}, 0, 0)$$ and $$m_{3}(u_{2}, u_{1}, u_{3}; z_{2}, \bar{z}_{2}, z_{1}, \bar{z}_{1}, 0, 0),$$ respectively, when $|z_{1}|>|z_{2}|>0$ and $|z_{2}|>|z_{1}|>0$, respectively. By the permutation property, $$m_{3}(u_{1}, u_{2}, u_{3}; z_{1}, \bar{z}_{1}, z_{2}, \bar{z}_{2}, 0, 0)
=m_{3}(u_{2}, u_{1}, u_{3}; z_{2}, \bar{z}_{2}, z_{1}, \bar{z}_{1}, 0, 0).$$ Thus (\[prod-series-1\]) and (\[prod-series-2\]) converge absolutely to $$m_{3}(u_{1}, u_{2}, u_{3}; z_{1}, \bar{z}_{1}, z_{2}, \bar{z}_{2}, 0, 0)$$ when $|z_{1}|>|z_{2}|>0$ and $|z_{2}|>|z_{1}|>0$, respectively. So they are the expansions of $$m_{3}(u_{1}, u_{2}, u_{3}; z_{1}, \bar{z}_{1}, z_{2}, \bar{z}_{2}, 0, 0).$$ in the sets given by $|z_{1}|>|z_{2}|>0$ and $|z_{2}|>|z_{1}|>0$, respectively.
Before proving more properties, we would like to discuss first the problem of constructing full field algebras. It is clear that vertex operator algebras have structures of full field algebras. Let $(V^L, Y^L, {\mathbf{1}}^L, \omega^L)$ and $(V^R, Y^R, {\mathbf{1}}^R, \omega^R)$ be two vertex operator algebras. Consider the graded vector space $V^L\otimes V^R$ equipped with the correlation function maps, the vacuum and the operator $\mathbf{d}$ given as follows: For $n\in {\mathbb{Z}}_{+}$, $u^{L}_{1}, \dots, u^{L}_{n}\in V^{L}$ and $u^{R}_{1}, \dots, u^{R}_{n}\in V^{R}$, $m_{n}(u^{L}_{1}\otimes u^{R}_{1},
\dots, u^{L}_{n}\otimes u^{R}_{n}; z_{1}, \bar{z}_{1},
\dots, z_{n}, \bar{z}_{n})$ are given by the analytic extensions of $$(Y^{L}(u^{L}_{1}, z_{1})\otimes Y^{R}(u^{R}_{1}, \bar{z}_{1}))
\cdots (Y^{L}(u^{L}_{n}, z_{n})\otimes Y^{R}(u^{R}_{n}, \bar{z}_{n}))
\mathbf{1}.$$ Then we take the vacuum $\mathbf{1}=\mathbf{1}^{L}\otimes \mathbf{1}^{R}$ and the operators $\mathbf{d}=
L^L(0)\otimes I_{V^R} + I_{V^L}\otimes L^R(0)$. In particular, the full vertex operators are given by $$\mathbb{Y}(u^L\otimes u^R, z) v^L\otimes v^R
= Y^L(u^L, z)v^L \otimes Y^R(u^R, \bar{z})v^R.$$ for $u^L, v^L\in V^L, u^R, v^R\in V^R$ and $z\in {\mathbb{C}}^{\times}$.
We have:
The vector space $V^L\otimes V^R$ equipped with the correlation function maps and the vacuum $\mathbf{1}$ given above is a full field algebra.
[[*Proof.*]{}]{}The proof is a straightforward and easy verification.
Note that there is also a vertex operator algebra structure on $V^{L}\otimes V^{R}$. For simplicity, we shall use the notation $V^{L}\otimes V^{R}$ to denote both the vertex operator algebra and the full field algebra structure. It should be easy to see which structure we will be using in the remaining part of this paper.
The full field algebra $V^L\otimes V^R$ in general does not give a genus-one theory, that is, suitable $q$-traces, even in the case that they are convergent, of the full vertex operators in general are not modular invariant. For chiral theories, we know from [@H6] that if we consider the intertwining operator algebras constructed from irreducible modules for suitable vertex operator algebras, we do have modular invariance. So it is then natural to look for full field algebras from suitable extensions of $V^L\otimes V^R$ by $V^L\otimes V^R$-modules.
Note that $V^L\otimes V^R$ has an ${\mathbb{Z}}\times {\mathbb{Z}}$-grading with grading operators being $L^L(0)\otimes I_{V^R}$ and $I_{V^L}\otimes L^R(0)$. If a full field algebra is an extension of $V^L\otimes V^R$ by $V^L\otimes V^R$-modules, it has an ${\mathbb{R}}\times {\mathbb{R}}$-grading.
For any ${\mathbb{R}}\times {\mathbb{R}}$-graded vector space $F=\coprod_{(m,n)\in {\mathbb{R}}\times {\mathbb{R}}} F_{(m,n)}$, we have a [*left grading operator*]{} $\mathbf{d}^L$ and a [*right grading operator*]{} $\mathbf{d}^R$ defined by $\mathbf{d}^L u=m u, \mathbf{d}^R u =nu$ for $u\in E_{(m,n)}$, where $m$ ($n$) is called the left (right) weight of $u$ and is denoted by ${\mbox{\rm \rm wt}}^L u$ (${\mbox{\rm \rm wt}}^R u$). For $m, n\in {\mathbb{R}}$, let $P_{m,n}$ be the projection from $F \to F_{(m,n)}$. We still use $F'$ and $\overline{F}$ to denote the graded dual and the algebraic completion of $F$, but note that they are with respect to the ${\mathbb{R}}\times {\mathbb{R}}$-grading, not any ${\mathbb{R}}$-grading induced from the ${\mathbb{R}}\times {\mathbb{R}}$-grading.
\[RR-grading-ffa\]
An ${\mathbb{R}}\times {\mathbb{R}}$-graded full field algebra is a full field algebra $(F, m, \mathbf{1})$ equipped with an ${\mathbb{R}}\times {\mathbb{R}}$-grading on $F$ (graded by [*left conformal weight*]{} or [*left weight*]{} and [*right conformal weight*]{} or [*right weight*]{} and thus equipped with left and right grading operators $\mathbf{d}^{L}$ and $\mathbf{d}^{R}$) and operators $D^{L}$ and $D^{R}$ satisfying the following conditions:
1. The [*grading compatibility*]{}: $\mathbf{d}=\mathbf{d}^{L}+\mathbf{d}^{R}$.
2. The [*single-valuedness property*]{}: $e^{2\pi i (\mathbf{d}^L-\mathbf{d}^R)} = I_F$.
3. The [*convergence property*]{}: For $k, l_{1}, \ldots,
l_{k} \in {\mathbb{Z}}_+$ and $u_{1}^{(1)},\ldots, u_{l_{1}}^{(1)},\ldots, u_{1}^{(k)}$, $\dots,
u_{l_{k}}^{(k)} \in F$, the series $$\begin{aligned}
\label{series-case-2-0}
\lefteqn{\sum_{p_1,q_1, \dots, p_k, q_k}
m_k(P_{p_1,q_1}m_{l_1}(u_1^{(1)},\dots, u_{l_1}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)},
\dots, z_{l_1}^{(1)}, \bar{z}_{l_1}^{(1)}), \dots,} {\nonumber \\}&&\quad
P_{p_k,q_k}m_{l_k}(u_1^{(k)},\dots, u_{l_k}^{(k)}; z_1^{(k)},
\bar{z}_1^{(k)},
\dots, z_{l_k}^{(k)}, \bar{z}_{l_k}^{(k)});
z_1^{(0)}, \bar{z}_1^{(0)}, \dots, z_k^{(0)}, \bar{z}_k^{(0)} ){\nonumber \\}&&\end{aligned}$$ converges absolutely to (\[sum\]) when $|z_p^{(i)}| + |z_q^{(j)}|< |z^{(0)}_i
-z^{(0)}_j|$ for $i,j=1, \ldots, k$, $i\ne j$ and for $p=1,
\dots, l_i$ and $q=1, \dots, l_j$.
4. [*The $\mathbf{d}^L$- and $\mathbf{d}^R$-bracket properties*]{}: For $u\in F$, $$\begin{aligned}
\left[ \mathbf{d}^L, \mathbb{Y}(u; z,\bar{z})\right]
&=& z\frac{\partial}{\partial z} \mathbb{Y}(u; z, \bar{z})
+\mathbb{Y}(\mathbf{d}^L u; z,\bar{z}) \label{d-l} \\
\left[ \mathbf{d}^R, \mathbb{Y}(u; z, \bar{z})\right]
&=& \bar{z}\frac{\partial}{\partial \bar{z}} \mathbb{Y}(u; z,\bar{z})
+ \mathbb{Y}(\mathbf{d}^R u; z, \bar{z}). \label{d-r}\end{aligned}$$
5. The [*$D^{L}$- and $D^{R}$-derivative property*]{}: For $u\in F$, $$\begin{aligned}
&&\left[ D^L, \mathbb{Y}(u; z,\bar{z})\right] =
\mathbb{Y}(D^{L}u; z,\bar{z}) =
\frac{\partial}{\partial z}\mathbb{Y}(u; z, \bar{z}),\\
&&\left[ D^R, \mathbb{Y}(u; z, \bar{z})\right] =
\mathbb{Y}(D^{R}u; z, \bar{z}) =
\frac{\partial}{\partial \bar{z}}\mathbb{Y}(u; z, \bar{z}). \label{D-L-R-b-d}\end{aligned}$$
We denote the ${\mathbb{R}}\times {\mathbb{R}}$-graded full field algebra defined above by $(F, m, \mathbf{1}, D^{L}, D^{R})$ or simply by $F$. But note that there is a refined grading on $F$ now.
[Note that for ${\mathbb{R}}\times {\mathbb{R}}$-graded full field algebra, there is also a weaker convergence property similar to the one in Remark \[weak-c-p\].]{}
[The single-valuedness property actually says that $F$ is graded by a subgroup $\{(m,n)\in {\mathbb{R}}\times {\mathbb{R}}\; |\; m-n \in {\mathbb{Z}}\}$ of ${\mathbb{R}}\times {\mathbb{R}}$. This single-valuedness indeed corresponds to a certain single-valuedness condition in the geometric axioms for full conformal field theories. ]{}
We have the following immediate consequences of the definition above.
1. The pair (left weight, right weight) for ${\mathbf{1}}$ is $(0, 0)$, that is, $\mathbf{d}^{L}{\mathbf{1}}=\mathbf{d}^{R}{\mathbf{1}}=0$.
2. The pairs (left weight, right weight) for $D^{L}$ and $D^{R}$ are $(1, 0)$ and $(0, 1)$, respectively, that is, $$\begin{aligned}
[\mathbf{d}^{L}, D^{L}]&=&D^{L},\\
{[\mathbf{d}^{R}, D^{L}]}&=&0,\\
{[\mathbf{d}^{L}, D^{R}]}&=&0,\\
{[\mathbf{d}^{R}, D^{R}]}&=&D^{R}.\end{aligned}$$
3. $D^{L}{\mathbf{1}}=D^{R}{\mathbf{1}}=0$.
[[*Proof.*]{}]{}From the identity property, $$z\frac{\partial}{\partial z}
\mathbb{Y}({\mathbf{1}}; z, \bar{z})=0.$$ Then by (\[d-l\]), we obtain $$\label{d-l-1}
\left[\mathbf{d}^{L}, \mathbb{Y}({\mathbf{1}}; z, \bar{z})\right]
=\mathbb{Y}(\mathbf{d}^{L} {\mathbf{1}}; z,\bar{z}).$$ Since $\mathbb{Y}({\mathbf{1}}; z, \bar{z})=I_{F}$, we obtain $$\label{d-l-2}
\mathbb{Y}(\mathbf{d}^{L} {\mathbf{1}}; z,\bar{z})=0$$ from (\[d-l-1\]). Applying (\[d-l-2\]) to ${\mathbf{1}}$, taking the limit $z\to 0$ on both sides of (\[d-l-2\]) and using the creation property, we obtain $\mathbf{d}^{L} {\mathbf{1}}=0$. So the left weight of ${\mathbf{1}}$ is $0$. Similarly, we can prove that the right weight of ${\mathbf{1}}$ is $0$.
Applying both sides of (\[d-l\]) to ${\mathbf{1}}$, taking the limit $z\to 0$ and then using the creation property and the fact $\mathbf{d}^{L} {\mathbf{1}}=0$ we have just proved, we obtain $$\label{limit-z-der}
\lim_{z\to 0}z\frac{\partial}{\partial z}\mathbb{Y}(u; z, \bar{z}){\mathbf{1}}=0$$ for $u\in F$. Applying $\frac{\partial}{\partial z}$ to both sides of (\[d-l\]) and using the $D^{L}$-derivative property, we obtain $$\begin{aligned}
\label{d-l-3}
\lefteqn{\left[\mathbf{d}^{L}, \mathbb{Y}(D^{L}u; z, \bar{z})\right]}{\nonumber \\}&&= \frac{\partial}{\partial z}z\frac{\partial}{\partial z}
\mathbb{Y}(u; z, \bar{z})
+\mathbb{Y}(D^{L}\mathbf{d}^{L} u; z,\bar{z}){\nonumber \\}&&=z\frac{\partial}{\partial z}
\frac{\partial}{\partial z}\mathbb{Y}(u; z, \bar{z})
+\frac{\partial}{\partial z}\mathbb{Y}(u; z, \bar{z})
+\mathbb{Y}(D^{L}\mathbf{d}^{L} u; z,\bar{z}){\nonumber \\}&&=z\frac{\partial}{\partial z}
\mathbb{Y}(D^{L}u; z, \bar{z})
+\mathbb{Y}(D^{L}u; z, \bar{z})
+\mathbb{Y}(D^{L}\mathbf{d}^{L} u; z,\bar{z}).\end{aligned}$$ Applying (\[d-l-3\]) to ${\mathbf{1}}$, taking the limit $z\to 0$ on both sides of (\[d-l-3\]) and using the creation property, (\[limit-z-der\]) and $\mathbf{d}^{L}{\mathbf{1}}=0$, we obtain $$\mathbf{d}^{L}D^{L}u=D^{L}u+D^{L}\mathbf{d}^{L} u,$$ proving that the left weight of $D^{L}$ is $1$. Similarly, we can prove that the right weight of $D^{L}$ is $0$, the left weight of $D^{R}$ is $0$ and the right weight of $D^{R}$ is $1$.
Using the $D^{L}$- and $D^{R}$-derivative properties and the creation property, we see immediately that $D^{L}{\mathbf{1}}=D^{R}{\mathbf{1}}=0$.
For an ${\mathbb{R}}\times {\mathbb{R}}$-graded full field algebra $(F, m, \mathbf{1}, D^{L}, D^{R})$, we now introduce a formal vertex operator map. We shall use the convention that for any $z\in {\mathbb{C}}^{\times}$, $\log z=\log |z|+\sqrt{-1}\arg z$ where $0\le \arg z <2\pi$. For $u\in F$, we use ${\mbox{\rm wt}^{L}\;}u$ and ${\mbox{\rm wt}^{R}\;}u$ to denote the left and right weights, respectively, of $u$. Let $u, v\in F$ and $w'\in F'$ be homogeneous elements. We have $$\begin{aligned}
\label{L-L-0-equ-1}
\langle w', [\mathbf{d}^L, \mathbb{Y}(u; z,\bar{z})]v\rangle &=& \langle
(\mathbf{d}^L)' w', \mathbb{Y}(u; z, \bar{z})v \rangle-
\langle w', \mathbb{Y}(u; z, \bar{z}) \mathbf{d}^Lv\rangle {\nonumber \\}&=& ({\mbox{\rm wt}^{L}\;}w'- {\mbox{\rm wt}^{L}\;}v)\langle w', \mathbb{Y}(u; z, \bar{z})v\rangle \end{aligned}$$ where $(\mathbf{d}^L)'$ is the adjoint of $\mathbf{d}^L$. On the other hand, $$\begin{aligned}
\label{L-L-0-equ-2}
\lefteqn{\langle w', \mathbb{Y}(\mathbf{d}^L u; z, \bar{z})v +
z\frac{\partial}{\partial z} \mathbb{Y}(u; z, \bar{z})v \rangle} {\nonumber \\}&&=\left( {\mbox{\rm wt}^{L}\;}u + z\frac{\partial}{\partial z}\right)
\langle w', \mathbb{Y}(u; z, \bar{z})v\rangle .\end{aligned}$$ Let $f(z, \bar{z})=\langle w', \mathbb{Y}(u; z, \bar{z})v\rangle$. Then by (\[d-l\]), (\[L-L-0-equ-1\]) and (\[L-L-0-equ-2\]), we have $$\label{L-L-0-equ}
z\frac{\partial}{\partial z} f(z, \bar{z}) = ({\mbox{\rm wt}^{L}\;}w'- {\mbox{\rm wt}^{L}\;}u
-{\mbox{\rm wt}^{L}\;}v) f(z, \bar{z}).$$ Similarly, using (\[d-r\]), we have $$\label{L-R-0-equ}
\bar{z}\frac{\partial}{\partial \bar{z}} f(z, \bar{z}) =
({\mbox{\rm wt}^{R}\;}w'- {\mbox{\rm wt}^{R}\;}u -{\mbox{\rm wt}^{R}\;}v) f(z, \bar{z}).$$ The general solution of the system (\[L-L-0-equ\]) and (\[L-R-0-equ\]) is $$\label{solution-1}
C z^{{\mbox{\rm {\scriptsize wt}}^{L}\,}w'- {\mbox{\rm {\scriptsize wt}}^{L}\,}u -{\mbox{\rm {\scriptsize wt}}^{L}\,}v}\bar{z}^{{\mbox{\rm {\scriptsize wt}}^{R}\,}w'- {\mbox{\rm {\scriptsize wt}}^{R}\,}u -{\mbox{\rm {\scriptsize wt}}^{R}\,}v}$$ where $C\in {\mathbb{C}}$. Note that $f(z, \bar{z})$ is a single-valued function and that by the single-valuedness of the full field algebra $F$, $$({\mbox{\rm wt}^{L}\;}w'- {\mbox{\rm wt}^{L}\;}u -{\mbox{\rm wt}^{L}\;}v)- ({\mbox{\rm wt}^{R}\;}w'- {\mbox{\rm wt}^{R}\;}u -{\mbox{\rm wt}^{R}\;}v) \in {\mathbb{Z}}.$$ This means that if we choose any branches of $z^{{\mbox{\rm {\scriptsize wt}}^{L}\,}w'- {\mbox{\rm {\scriptsize wt}}^{L}\,}u -{\mbox{\rm {\scriptsize wt}}^{L}\,}v}$ and $\bar{z}^{{\mbox{\rm {\scriptsize wt}}^{R}\,}w'- {\mbox{\rm {\scriptsize wt}}^{R}\,}u -{\mbox{\rm {\scriptsize wt}}^{R}\,}v}$, then there must be a unique constant $C$ such that $f(z, \bar{z})$ is equal to (\[solution-1\]). We choose the branches of $z^{{\mbox{\rm {\scriptsize wt}}^{L}\,}w'- {\mbox{\rm {\scriptsize wt}}^{L}\,}u -{\mbox{\rm {\scriptsize wt}}^{L}\,}v}$ and $\bar{z}^{{\mbox{\rm {\scriptsize wt}}^{R}\,}w'- {\mbox{\rm {\scriptsize wt}}^{R}\,}u -{\mbox{\rm {\scriptsize wt}}^{R}\,}v}$ to be $e^{({\mbox{\rm {\scriptsize wt}}^{L}\,}w'- {\mbox{\rm {\scriptsize wt}}^{L}\,}u -{\mbox{\rm {\scriptsize wt}}^{L}\,}v)\log z}$ and $e^{({\mbox{\rm {\scriptsize wt}}^{R}\,}w'- {\mbox{\rm {\scriptsize wt}}^{R}\,}u -{\mbox{\rm {\scriptsize wt}}^{R}\,}v)\;\overline{\log z}}$, respectively. So there is a unique $C\in {\mathbb{C}}$ such that $$\label{solution-1.5}
f(z, \bar{z})=C e^{({\mbox{\rm {\scriptsize wt}}^{L}\,}w'- {\mbox{\rm {\scriptsize wt}}^{L}\,}u -{\mbox{\rm {\scriptsize wt}}^{L}\,}v)\log z}
e^{({\mbox{\rm {\scriptsize wt}}^{R}\,}w'- {\mbox{\rm {\scriptsize wt}}^{R}\,}u -{\mbox{\rm {\scriptsize wt}}^{R}\,}v)\;\overline{\log z}}.$$ Hence $P_{p,q} \mathbb{Y}(u; z,\bar{z}) P_{m, n}$, $m, n, p, q \in {\mathbb{R}}$ can be written as $$u_{m, n}^{p, q} e^{(p-{\mbox{\rm {\scriptsize wt}}^{L}\,}u -m)\log z} e^{(q-{\mbox{\rm {\scriptsize wt}}^{R}\,}u-n)\;\overline{\log
z}},$$ where $u_{m ,n}^{p, q}$ are linear maps from $F_{(m, n)}\rightarrow F_{(p, q)}$ for $m, n, p, q \in {\mathbb{R}}$. Thus we have the following expansion: $$\label{expansion}
\mathbb{Y}(u; z,\bar{z}) =\sum_{r, s\in {\mathbb{R}}} \mathbb{Y}_{l, r}(u)
e^{(-l-1)\log z} e^{(-r-1)\;\overline{\log
z}}$$ where $\mathbb{Y}_{l, r}(u)\in {\mbox{\rm End}\ }F$ with ${\mbox{\rm \rm wt}}^L \mathbb{Y}_{l, r}(u) = {\mbox{\rm \rm wt}}^L u -l-1$ and ${\mbox{\rm \rm wt}}^R \mathbb{Y}_{l, r}(u) = {\mbox{\rm \rm wt}}^R u -r-1$. Moreover, the expansion above is unique. Let $x$ and $\bar{x}$ be independent and commuting formal variables. We define the [*formal full vertex operator*]{} $\mathbb{Y}_f$ associated to $u\in F$ by $$\label{formal-op-def}
\mathbb{Y}_f(u; x,\bar{x}) =
\sum_{l, r\in {\mathbb{R}}}\mathbb{Y}_{l, r}(u) x^{-l-1}\bar{x}^{-r-1}.$$ These formal full vertex operators give a [*formal full vertex operator map*]{} $$\mathbb{Y}_{f}: F\otimes F\mapsto F\{x, \bar{x}\}.$$ For nonzero complex numbers $z$ and $\zeta$, we can substitute $e^{r\log z}$ and $e^{s\; \overline{\log \zeta}}$ for $x^{r}$ and $\bar{x}^{s}$, respectively, in $\mathbb{Y}_f(u; x,\bar{x})$ to obtain a map $\mathbb{Y}_{\rm an}(u; z, \zeta): F\to \overline{F}$ called the [*analytic full vertex operator map*]{}.
The following propositions are clear:
For $u\in F$ and $z, \zeta \in {\mathbb{C}}^{\times}$, we have $$\label{d-conj1}
\mathbb{Y}_{\rm an}(u; z, \zeta)
= z^{\mathbf{d}^L} \zeta^{\mathbf{d}^R} \mathbb{Y}(
z^{-\mathbf{d}^L} \zeta^{-\mathbf{d}^R}u; 1, 1)
z^{-\mathbf{d}^L} \zeta^{-\mathbf{d}^R}.$$ For formal full vertex operators, we have $$\label{d-conj2}
\mathbb{Y}_{f}(u; x, \bar{x})
= x^{\mathbf{d}^L} \bar{x}^{\mathbf{d}^R}
\mathbb{Y}(x^{-\mathbf{d}^L} \bar{x}^{-\mathbf{d}^R}u; 1, 1)
x^{-\mathbf{d}^L} \bar{x}^{-\mathbf{d}^R}.$$
For $u\in F$, $$\begin{aligned}
\mathbb{Y}_f({\mathbf{1}}; x, \bar{x}) u &=& u, \label{form-ident} \\
\lim_{x\rightarrow 0, \bar{x}\rightarrow 0} \mathbb{Y}_f(u; x, \bar{x})
{\mathbf{1}}&=& u \label{form-creat},\end{aligned}$$ where $\lim_{x\rightarrow 0, \bar{x}\rightarrow 0}$ means taking the constant term of a power series in $x$ and $\bar{x}$. In particular, $\mathbb{Y}_{l, r}(u){\mathbf{1}}=0$ for all $l, r\in
\mathbb{R}$ and $\mathbb{Y}_{-1,-1}(u){\mathbf{1}}= u$.
For $u\in F$, we have $$\begin{aligned}
\left[ D^L, \mathbb{Y}_f(w; x, \bar{x})\right] &=&
\mathbb{Y}_f(D^L w; x, \bar{x})
= \frac{\partial}{\partial x} \mathbb{Y}_f (w; x, \bar{x}),
\label{D-l-formal}\\
\left[ D^R, \mathbb{Y}_f (w; x, \bar{x})\right] &=&
\mathbb{Y}_f (D^R w; x, \bar{x})
= \frac{\partial}{\partial \bar{x}} \mathbb{Y}_f(w, x, \bar{x}).
\label{D-r-formal}\end{aligned}$$ In particular, we have $D^L{\mathbf{1}}= D^R{\mathbf{1}}=0$ and for $l, r\in {\mathbb{R}}$, $$\begin{aligned}
\left[ D^L, \mathbb{Y}_{l, r}(u)\right] &=& \mathbb{Y}_{l, r}
(D^{L}u) = -l \mathbb{Y}_{l-1, r}(u),
\label{D-l-comp}\\
\left[ D^R, \mathbb{Y}_{l, r}(u)\right] &=& \mathbb{Y}_{l, r}
(D^{R}u) = -r \mathbb{Y}_{l, r-1}(u),
\label{D-r-comp}\end{aligned}$$
We need the following strong version of the creation property:
For $u\in F$, $$\label{strong-creat}
\mathbb{Y}_f(u; x, \bar{x}) {\mathbf{1}}= e^{xD^L + \bar{x}D^R} u.$$
[[*Proof.*]{}]{}Using (\[D-l-formal\]) and (\[D-r-formal\]), we have $$\begin{aligned}
\label{conj-D-l-r}
\mathbb{Y}_f(e^{x_0D^L + \bar{x}_0D^R}u; x, \bar{x})
&=&\mathbb{Y}_f(u; x+x_0, \bar{x}+\bar{x}_0). \end{aligned}$$ Now let both sides of (\[conj-D-l-r\]) act on the vacuum ${\mathbf{1}}$. Since $\mathbb{Y}_f(u; x+x_0, \bar{x}+\bar{x}_0){\mathbf{1}}$ involves only nonnegative integer powers of $x+x_0$ and $\bar{x}+\bar{x}_0$, we can take the limit $x\to 0, \bar{x}\to 0$. Then replacing $x_0$ and $\bar{x}_0$ by $x$ and $\bar{x}$, we obtain (\[strong-creat\]).
For any $u,v\in F$ and $z\in {\mathbb{C}}^{\times}$, we have $$\label{skew-symm}
\mathbb{Y}(u; z, \bar{z})v = e^{zD^L + \bar{z}D^R}
\mathbb{Y}(v; -z, \overline{-z})u$$ and $$\label{skew-formal}
\mathbb{Y}_f (u; x, \bar{x})v = e^{xD^L + \bar{x}D^R}
\mathbb{Y}_f(v; e^{\pi i}x, e^{-\pi i}\bar{x})u.$$
[[*Proof.*]{}]{} From the convergence property, it is clear that, for any $u,v\in F$, $$\mathbb{Y}(\mathbb{Y}(u; z_{1}-z_{2}, \bar{z}_{1}-\bar{z}_{2})v; z_{2},\bar{z}_{2}) {\mathbf{1}}$$ converges absolutely to $m_{3}(u,v,{\mathbf{1}};z_{1},\bar{z}_{1}, z_{2}, \bar{z}_{2},0,0)$ when $|z_{2}|>|z_{1}-z_{2}|>0$, and $$\mathbb{Y}(\mathbb{Y}(v; z_{2}-z_{1}, \bar{z}_{2}-\bar{z}_{1})u; z_{1},\bar{z}_{1}) {\mathbf{1}}$$ converges absolutely when $|z_{1}|>|z_{1}-z_{2}|>0$ to $m_{3}(v,u,{\mathbf{1}};z_{2},\bar{z}_{2}, z_{1}, \bar{z}_{1}, 0,0)$ which is equal to $m_{3}(u,v,{\mathbf{1}};z_{1},\bar{z}_{1}, z_{2}, \bar{z}_{2}, 0,0)$ by the permutation property. Hence, using (\[strong-creat\]), we obtain $$\label{skew-proof-1}
e^{z_{2}D^L + \bar{z}_{2}D^R} \mathbb{Y}(u; z_{1}-z_{2},
\bar{z}_{1}-\bar{z}_{2})v
= e^{z_{1}D^L + \bar{z}_{1}D^R} \mathbb{Y}(v; z_{2}-z_{1},
\bar{z}_{2}-\bar{z}_{1})u$$ when $|z_{2}|>|z_{1}-z_{2}|>0$ and $|z_{1}|>|z_{1}-z_{2}|>0$. We change the variables from $z_{1}, z_{2}$ to $z=z_{1}-z_{2}$ and $z_{2}$. Then (\[skew-proof-1\]) gives $$\label{skew-proof-1.5}
e^{z_{2}D^L + \bar{z}_{2}D^R} \mathbb{Y}(u; z, \bar{z})v
= e^{(z_{2}+z)D^L + (\bar{z}_{2}+\bar{z})D^R}
\mathbb{Y}(v; -z, \overline{-z})u$$ when $|z_{2}|>|z|>0$ and $|z_{2}+z|>|z|>0$.
Notice that for fixed $z\ne 0$ and $w'\in F'$, $$\label{skew-proof-2}
\langle w', e^{z_{2}D^L + \bar{z}_{2}D^R} \mathbb{Y}(u; z,
\bar{z})v\rangle$$ involves only positive integral powers of $z_{2}, \bar{z}_{2}$ and thus is a power series in $z_{2}$ and $\bar{z}_{2}$ absolutely convergent when $|z_{2}|>|z|>0$. From complex analysis, we know that a power series in two variables $z_{2}$ and $\zeta_{2}$ convergent at $z_{2}=z_{2}^{0}$ and $\zeta_{2}=\zeta_{2}^{0}$ must be convergent absolutely when $|z_{2}|<|z_{2}^{0}|$ and $|\zeta_{2}|<|\zeta_{2}^{0}|$. In particular, when $\zeta_{2}^{0}=
\bar{z}_{2}^{0}$, such a power series must be absolutely convergent when $|z_{2}|<|z_{2}^{0}|$ and $\zeta_{2}=\bar{z}_{2}$. In our case, since for any fixed $z$, (\[skew-proof-2\]) is absolutely convergent when $|z_{2}|>|z|>0$, we conclude that (\[skew-proof-2\]) converges absolutely for all $z_{2}$. Since $w'$ is arbitrary, we see that the left-hand side of (\[skew-proof-1.5\]) is absolutely convergent in $\overline{F}$ for all $z_{2}$. Since $z$ is also arbitrary, by the convergence property again, we see that the left-hand side of (\[skew-proof-1.5\]) is absolutely convergent in $\overline{F}$ for all $z$ and $z_{2}$ such that $z\ne 0$. Similarly, the right hand side of (\[skew-proof-1.5\]) also converges absolutely in $\overline{F}$ for all $z$ and $z_{2}$ such that $z\ne 0$.
If $e^{-z_{2}D^L - \bar{z}_{2}D^R}$ gives a linear operator on $\overline{F}$, then we can just multiply both sides of (\[skew-proof-1.5\]) by $e^{-z_{2}D^L - \bar{z}_{2}D^R}$ to obtain (\[skew-symm\]). In the case that the total weights of $F$ is lower-truncated, $e^{-z_{2}D^L - \bar{z}_{2}D^R}$ is indeed a linear operator on $\overline{F}$. In the most general case, this might not be true. But we can still obtain (\[skew-symm\]) as follows: Consider the formal series $$\begin{aligned}
\label{skew-proof-2.5}
\lefteqn{e^{x_{1}D^L + \bar{x}_{1}D^R}
e^{x_{2}D^L + \bar{x}_{2}D^R} \mathbb{Y}(u; x, \bar{x})v}{\nonumber \\}&&=e^{(x_{1}+x_{2})D^L + (\bar{x}_{1}+\bar{x}_{2})D^R}
\mathbb{Y}(u; x, \bar{x})v\end{aligned}$$ where $x$, $\bar{x}$, $x_{1}$, $\bar{x}_{1}$, $x_{2}$ and $\bar{x}_{2}$ are commuting formal variables. Since $\mathbb{Y}(u; z, \bar{z})v$ is absolutely convergent in $\overline{F}$ when $z\ne 0$, we can substitute $z$, $\bar{z}$, $-z_{2}$, $-\bar{z}_{2}$, $z_{2}$ and $\bar{z}_{2}$ for $x$, $\bar{x}$, $x_{1}$, $\bar{x}_{1}$, $x_{2}$ and $\bar{x}_{2}$ on the right-hand side of (\[skew-proof-2.5\]), respectively, and the resulting series is absolutely convergent in $\overline{F}$. So we can do the same substitution on the left-hand side of (\[skew-proof-2.5\]) and the resulting series is absolutely convergent in $\overline{F}$. Similarly, consider the formal series $$\begin{aligned}
\label{skew-proof-3}
\lefteqn{e^{x_{1}D^L + \bar{x}_{1}D^R}
e^{(x_{2}+x)D^L + (\bar{x}_{2}+\bar{x})D^R}
\mathbb{Y}(v; e^{\pi i}x, e^{-\pi i}\bar{x})u}{\nonumber \\}&&=e^{(x_{1}+x_{2}+x)D^L + (\bar{x}_{1}+\bar{x}_{2}+\bar{x})D^R}
\mathbb{Y}(v; e^{\pi i}x, e^{-\pi i}\bar{x})u.\end{aligned}$$ Since $e^{zD^L + \bar{z}D^R}
\mathbb{Y}(v; -z, \overline{-z})u$ is absolutely convergent in $\overline{F}$ when $z\ne 0$, we can substitute $z$, $\bar{z}$, $-z_{2}$, $-\bar{z}_{2}$, $z_{2}$ and $\bar{z}_{2}$ for $x$, $\bar{x}$, $x_{1}$, $\bar{x}_{1}$, $x_{2}$ and $\bar{x}_{2}$ on the right-hand side and thus also on the left-hand side of (\[skew-proof-3\]) and the resulting series is absolutely convergent in $\overline{F}$. The convergence of these series and (\[skew-proof-1.5\]) with suitably chosen $z_{2}$ gives (\[skew-symm\])
Now (\[skew-formal\]) follows immediately: On the one hand, by (\[d-conj2\]), we have $$\begin{aligned}
\label{skew-formal-1}
x^{\mathbf{d}^{L}} \bar{x}^{\mathbf{d}^{L}}
\mathbb{Y}(x^{-\mathbf{d}^{L}} \bar{x}^{-\mathbf{d}^{L}}u; 1, 1)
x^{-\mathbf{d}^{L}} \bar{x}^{-\mathbf{d}^{L}}v =
\mathbb{Y}_f(u; x, \bar{x}) v.\end{aligned}$$ On the other hand, we have $$\begin{aligned}
\label{skew-formal-2}
&x^{\mathbf{d}^{L}} \bar{x}^{\mathbf{d}^{L}}e^{D^L+D^R}
\mathbb{Y}(x^{-\mathbf{d}^{L}} \bar{x}^{-\mathbf{d}^{L}}v;
-1, -1) x^{-\mathbf{d}^{L}} \bar{x}^{-\mathbf{d}^{L}}u &{\nonumber \\}&=
e^{xD^L+\bar{x} D^R}\mathbb{Y}_f(v; e^{\pi i}x, e^{-\pi i}\bar{x})
u.&\end{aligned}$$ Using skew symmetry (\[skew-symm\]), (\[skew-formal-1\]) and (\[skew-formal-2\]), we obtain (\[skew-formal\]).
\[g-r-RR-g-ffa\]
An ${\mathbb{R}}\times {\mathbb{R}}$-graded full field algebra $(F, m, {\mathbf{1}},
D^{L}, D^{R})$ is called [*grading restricted*]{} if it satisfies the following grading-restriction conditions:
1. There exists $M\in {\mathbb{R}}$ such that $F_{(m,n)}=0$ if $n<M$ or $m<M$.
2. $\dim F_{(m,n)}<\infty$ for $m, n\in {\mathbb{R}}$.
We say that $F$ is [*lower truncated*]{} if $F$ satisfies the first grading restriction condition.
In this case, for $u\in F$ and $k\in {\mathbb{R}}$, we have $\sum_{l+r=k} \mathbb{Y}_{l, r}(u) \in {\mbox{\rm End}\ }F$ with total weight ${\mbox{\rm wt}\ }u -k-2$. We denote $\sum_{l+r=k} \mathbb{Y}_{l, r}(u) $ by $\mathbb{Y}_{k-1}(u)$. Then we have the expansion $$\label{Y-x-x}
\mathbb{Y}_f(u; x, x) =
\sum_{k\in {\mathbb{R}}} \mathbb{Y}_k(u) x^{-k-1},$$ where ${\mbox{\rm wt}\ }\mathbb{Y}_k(u) = {\mbox{\rm wt}\ }u - k -1$. For given $u, v\in F$, we have $\mathbb{Y}_k(u) w=0$ for sufficiently large $k$.
Let $(V^{L}, Y^{L}, {\mathbf{1}}^{L}, \omega^{L})$ and $(V^{R}, Y^{R}, {\mathbf{1}}^{R}, \omega^{R})$ be vertex operator algebras. Let $\rho$ be an injective homomorphism from the full field algebra $V^{L}\otimes V^{R}$ to $F$. Then we have ${\mathbf{1}}=\rho({\mathbf{1}}^{L}\otimes {\mathbf{1}}^{R})$, $\mathbf{d}^{L}\circ \rho=\rho\circ (L^{L}(0)\otimes I_{V^{R}})$, $\mathbf{d}^{R}\circ \rho=\rho\circ (I_{V^{L}}\otimes L^{R}(0)))$, $D^{L}\circ \rho=\rho\circ (L^{L}(-1)\otimes I_{V^{R}})$ and $D^{R}\circ \rho=\rho\circ (I_{V^{L}}\otimes L^{R}(-1))$. Moreover, $F$ has a [*left conformal element*]{} $\rho(\omega^{L}\otimes {\mathbf{1}}^{R})$ and an [*right conformal element*]{} $\rho({\mathbf{1}}^{L}\otimes \omega^{R})$. We have the following operators on $F$: $$\begin{aligned}
L^{L}(0)&=&{\mbox{\rm Res}}_{x}{\mbox{\rm Res}}_{\bar{x}}\bar{x}^{-1}\mathbb{Y}_{f}(
\rho(\omega^{L}\otimes
\mathbf{1}^{R}); x, \bar{x}),\\
L^{R}(0)&=&{\mbox{\rm Res}}_{x}{\mbox{\rm Res}}_{\bar{x}}x^{-1}\mathbb{Y}_{f}(
\rho(\mathbf{1}^{L}\otimes
\omega^{L}); x, \bar{x}),\\
L^{L}(-1)&=&{\mbox{\rm Res}}_{x}{\mbox{\rm Res}}_{\bar{x}}x\bar{x}^{-1}\mathbb{Y}_{f}(
\rho(\omega^{L}\otimes
\mathbf{1}^{R}); x, \bar{x}),\\
L^{R}(-1)&=&{\mbox{\rm Res}}_{x}{\mbox{\rm Res}}_{\bar{x}}x^{-1}\bar{x}\mathbb{Y}_{f}(
\rho(\mathbf{1}^{L}\otimes
\omega^{L}); x, \bar{x}).\end{aligned}$$ Since these operators are operators on $F$, it should be easy to distinguish them from those operators with the same noptation but acting on $V^{L}$ or $V^{R}$.
[Let $(V^{L}, Y^{L}, {\mathbf{1}}^{L}, \omega^{L})$ and $(V^{R}, Y^{R}, {\mathbf{1}}^{R}, \omega^{R})$ be vertex operator algebras. A [*full field algebra over $V^{L}\otimes V^{R}$*]{} is a grading-restricted ${\mathbb{R}}\times {\mathbb{R}}$-graded full field algebra $(F, m, {\mathbf{1}},
D^{L}, D^{R})$ equipped with an injective homomorphism $\rho$ from the full field algebra $V^{L}\otimes V^{R}$ to $F$ such that $\mathbf{d}^{L}=L^{L}(0)$, $\mathbf{d}^{R}=L^{R}(0)$, $D^{L}=L^{L}(-1)$ and $D^{R}=L^{R}(-1)$.]{}
We shall denote the full field algebra over $V^{L}\otimes V^{R}$ defined above by $(F, m, \rho)$ or simply by $F$.
The following result allows us to construct full field algebras using the representation theory of vertex operator algebras:
\[mod-int-op\] Let $(V^{L}, Y^{L}, {\mathbf{1}}^{L}, \omega^{L})$ and $(V^{R}, Y^{R}, {\mathbf{1}}^{R}, \omega^{R})$ be vertex operator algebras. Let $(F, m, \rho)$ be a full field algebra over $V^{L}\otimes V^{R}$. Then $F$ is a module for $V^L\otimes V^R$ viewed as a vertex operator algebra. Moreover, $\mathbb{Y}_f(\cdot, x,x)$, is an intertwining operator of type ${{{F}\choose {FF}}}$.
[[*Proof.*]{}]{}Let $\mathbb{Y}^{L, R}$ be the vertex operator map for the full field algebra $\rho(V^{L}\otimes V^{R})$. Then we have $$\label{Y-split}
\mathbb{Y}^{L, R}(\rho(u^L\otimes u^R); z,\bar{z})\rho(v^L\otimes v^R)
=\rho(Y^L(u^L, z)v^L \otimes Y^R(u^R, \bar{z})v^R)$$ for $u^L, v^L \in V^L$, $u^R, v^R\in V^R$ and $z\in {\mathbb{C}}^{\times}$.
Now we show that a splitting formula similar to (\[Y-split\]) holds for vertex operators of the form $\mathbb{Y}(\rho(u^{L}\otimes u^{R}); z, \bar{z}): F\to \overline{F}$. By the associativity of $\mathbb{Y}$, we have $$\begin{aligned}
\label{asso-K-K-alg-1}
\lefteqn{\langle w', \mathbb{Y}(\rho(u^L\otimes u^R); z_{1},\bar{z}_{1})
\mathbb{Y}(\rho(v^L\otimes v^R); z_{2},\bar{z}_{2})w\rangle} {\nonumber \\}&& = \langle w', \mathbb{Y}(\mathbb{Y}^{L, R}
(\rho(u^L\otimes u^R); z_{1}-z_{2}, \bar{z}_{1}-\bar{z}_{2})
\rho(v^L\otimes v^R); z_{2},\bar{z}_{2})w\rangle{\nonumber \\}&&\end{aligned}$$ when $|z_{1}|>|z_{2}|>|z_{1}-z_{2}|>0$ for $u^L, v^{L}\in V^L$, $u^R, v^{R}\in V^R$, $w\in F$ and $w'\in F'$. Take $v^{L}={\mathbf{1}}^{L}$ and $u^R, v^{R}={\mathbf{1}}^R$. Then we have $$\begin{aligned}
\label{asso-K-K-alg-2}
\lefteqn{\langle w', \mathbb{Y}(\rho(u^L\otimes {\mathbf{1}}^R); z_{1},\bar{z}_{1})
w\rangle} {\nonumber \\}&&= \langle w', \mathbb{Y}(\rho(u^L\otimes {\mathbf{1}}^R); z_{1},\bar{z}_{1})
\mathbb{Y}({\mathbf{1}}; z_{2},\bar{z}_{2})w\rangle{\nonumber \\}&&= \langle w', \mathbb{Y}(\mathbb{Y}^{L, R}
(\rho(u^L\otimes {\mathbf{1}}^R); z_{1}-z_{2}, \bar{z}_{1}-\bar{z}_{2})
\rho({\mathbf{1}}^L\otimes {\mathbf{1}}^R); z_{2},\bar{z}_{2})w\rangle{\nonumber \\}&&=\langle w', \mathbb{Y}(\rho((Y^L
(u^L, z_{1}-z_{2})
{\mathbf{1}}^L)\otimes {\mathbf{1}}^R); z_{2},\bar{z}_{2})w\rangle\end{aligned}$$ Since the right-hand side of (\[asso-K-K-alg-2\]) is independent of $\bar{z}_{1}$, so is the left-hand side. Thus we see that $\mathbb{Y}(\rho(u^L\otimes {\mathbf{1}}^R), z, \bar{z})$ depends only on $z$ for all $u^L\in V^L$ and we shall also denote it by $Y^L(u^L, z)$. (Since it acts on $F$, there should be no confusion with the vertex operator $Y^L(u^L, z)$ acting on $V^{L}$.) So $Y^L(u^L, z)$ is a series in powers of $z$. But $Y^L(u^L, z)$ is also single valued. So by (\[expansion\]), there exists $u_{n}^{L}\in {\mbox{\rm End}\ }F$ for $n\in {\mathbb{Z}}$ such that ${\mbox{\rm wt}^{L}\;}u_n^L = {\mbox{\rm wt}\ }u^L-n-1$, ${\mbox{\rm wt}^{R}\;}u_n^L =0$ and $$Y^{L}(u^L, z)=\sum_{n\in {\mathbb{Z}}}u_{n}^{L}z^{-n-1}.$$ Similarly, $\mathbb{Y}(\rho({\mathbf{1}}^L\otimes u^R); z, \bar{z})$ depends only on $\bar{z}$ and will also be denoted by $Y^R(u^R, \bar{z})$. (There should also be no confusion with the vertex operator $Y^R(u^R, z)$ acting on $V^{R}$.) For $u^{R}\in V^{R}$, there exists $u_{n}^{R}\in {\mbox{\rm End}\ }F$ for $n\in {\mathbb{Z}}$ such that ${\mbox{\rm wt}^{R}\;}u_n^R = {\mbox{\rm wt}\ }u^R -n-1$, ${\mbox{\rm wt}^{L}\;}u_n^R =0$ and $$Y^{R}(u^R, z)=\sum_{n\in {\mathbb{Z}}}u_{n}^{R}z^{-n-1}.$$ We also have the formal vertex operator maps, denoted using the same notations $Y^L$ and $Y^R$, associated to $Y^L$ and $Y^R$ given by $$\begin{aligned}
Y^L(u^L, x) &=& \sum_{n\in {\mathbb{Z}}} u_n^L x^{-n-1}, \\
Y^R(u^R,\bar{x}) &=& \sum_{n\in {\mathbb{Z}}} u_n^R \bar{x}^{-n-1} \end{aligned}$$ for $u^L\in V^L$ and $u^R\in V^R$.
For $u^L\in V^L, u^R\in V^R$ and $w\in F, w'\in F'$, $$\label{L-R}
\langle w', Y^L(u^L,z_{1})Y^R(u^R, \bar{z}_{2})w\rangle
=\langle w', \mathbb{Y}(\rho(u^L\otimes {\mathbf{1}}^{R});
z_{1}, \bar{z}_{1})\mathbb{Y}(\rho({\mathbf{1}}^{L}\otimes u^R);
z_{2}, \bar{z}_{2})w\rangle$$ is absolutely convergent when $|z_{1}|>|z_{2}|>0$, and $$\label{R-L}
\langle w', Y^R(u^R, \bar{z}_{2})Y^L(u^L,z_{1})w\rangle
=\langle w', \mathbb{Y}(\rho({\mathbf{1}}^{L}\otimes u^R); z_{2}, \bar{z}_{2})
\mathbb{Y}(\rho(u^L\otimes {\mathbf{1}}^{R});
z_{1}, \bar{z}_{1})w\rangle$$ is absolutely convergent when $|z_{2}|>|z_{1}|>0$. They are both analytic in $z_{1}$ and $\bar{z}_{2}$. By the convergence property for full field algebras, both side of (\[L-R\]) and (\[R-L\]) can be extended to a same smooth function on $\{ (z_{1}, \bar{z}_{2}) \in ({\mathbb{C}}^{\times})^2 | z_{1}\neq z_{2} \}$. Since the complement of the union of the sets of convergence of (\[L-R\]) and (\[R-L\]) in $\{ (z_{1}, \bar{z}_{2}) \in ({\mathbb{C}}^{\times})^2 | z_{1}\neq z_{2} \}$ is of lower dimension, by the properties of analytic functions, it is clear that the extended smooth function is actually analytic on $\{ (z_{1}, \bar{z}_{2}) \in ({\mathbb{C}}^{\times})^2 | z_{1}\neq z_{2} \}$.
By associativity, we have $$\label{L-R-assoc}
\langle w', Y^L(u^L,z_{1})Y^R(u^R, \bar{z}_{2})w\rangle
=\langle w', \mathbb{Y}(\rho((Y^L(u^L, z_{1}-z_{2}){\mathbf{1}}^L) \otimes u^R); z_{2},
\bar{z}_{2})w\rangle$$ when $|z_{1}|>|z_{2}|>|z_{1}-z_{2}|>0$. The right-hand side of (\[L-R-assoc\]) has a well-defined limit as $z_{1}$ goes to $z_{2}$. Therefore (\[L-R\]) and (\[R-L\]) can be further extended to a single analytic function on $\{ (z_{1},\bar{z}_{2}) \in ({\mathbb{C}}^{\times})^2 \}$. This absence of singularity further implies that the left-hand sides of (\[L-R\]) and (\[R-L\]) are absolutely convergent and are equal for all $z_{1}, \bar{z}_{2} \in {\mathbb{C}}^{\times}$. Let $z_{1}=z_{2}=z$ in (\[L-R\]), (\[R-L\]) and (\[L-R-assoc\]). Use the discussion above and the creation property for the vertex operator map $Y^{L}$, we obtain $$\mathbb{Y}(\rho(u^L\otimes u^R); z, \bar{z}) = Y^L(u^L, z)Y^R(u^R, \bar{z}) =
Y^R(u^R, \bar{z})Y^L(u^L, z),$$ or equivalently, in terms of formal vertex operator, $$\label{Y-cl-x-x}
\mathbb{Y}_f(\rho(u^L\otimes u^R); x, \bar{x})= Y^{L}(u^L, x)Y^R(u^R, \bar{x})
=Y^R(u^R, \bar{x})Y^L(u^L, x)$$ for all $u^L\in V^L$ and $u^R\in V^R$. In particular, we have $[u_m^L, u_n^R]=0$ for all $u^L \in V^L$ and $u^R\in V^R$.
Since $F$ is lower truncated, we have $$\label{formal-Y-cl}
\mathbb{Y}_f(\rho(u^L\otimes u^R); x, x)v \in ({\mbox{\rm End}}\; F)((x)).$$ for $u^L\in V^L$, $u^R\in V^R$ and $v\in F$.
The associativity (\[asso-K-K-alg-1\]) together with (\[Y-cl-x-x\]) and (\[formal-Y-cl\]) implies the associativity for the vertex operator map $\mathbb{Y}_f(\rho(\cdot); x, x)\cdot$. Together with the identity property this associativity implies that $F$ is a module for the vertex operator algebra $V^L\otimes V^R$.
Next we show that $\mathbb{Y}_f(\cdot; x, x)$ is an intertwining operator of type ${{{F}\choose {FF}}}$.
Since For given $u, v\in F$, we have $\mathbb{Y}_k(u) w=0$ for sufficient large $k$, the lower-truncation property of $\mathbb{Y}_f(\cdot, x, x)$ holds. For $u\in F$, We also have $$\begin{aligned}
\mathbb{Y}_f( (D^L + D^R) u; x, x)
&=&\mathbb{Y}_f( (D^L + D^R) u; x, \bar{x})|_{\bar{x}=x}{\nonumber \\}& =& \left.\left(\left(\frac{\partial}{\partial x}+
\frac{\partial}{\partial\bar{x}}\right)
\mathbb{Y}_f( u; x, \bar{x})\right)\right|_{\bar{x}=x}{\nonumber \\}&=& \frac{d}{dx} \mathbb{Y}_f(u;x,x),\end{aligned}$$ proving the $D$-derivative property of $\mathbb{Y}_f(\cdot; x, x)$.
Now, we prove the Jacobi identity for $\mathbb{Y}_f(\cdot; x, x)$. For any fixed $r\in {\mathbb{R}}$, using the associativity for the full vertex operator map $\mathbb{Y}$ twice, we obtain $$\begin{aligned}
\label{asso-equ}
\lefteqn{\langle w', Y^{L}(u^L, z_{1})Y^{R}(u^{R}, \bar{z}_{2})
\mathbb{Y}(u; r, r)w\rangle} {\nonumber \\}&& = \langle w', \mathbb{Y}(\rho(u^L\otimes {\mathbf{1}}^{L});
z_{1}, \bar{z}_{1})\mathbb{Y}(\rho({\mathbf{1}}^{L}\otimes u^{R}); z_{2}, \bar{z}_{2})
\mathbb{Y}(u; r, r)w\rangle{\nonumber \\}&&=\langle w',
\mathbb{Y}(\mathbb{Y}(\rho(u^L\otimes {\mathbf{1}}^{L});
z_{1}-r, \bar{z}_{1}-r)\mathbb{Y}(\rho({\mathbf{1}}^{L}\otimes u^{R});
z_{2}-r, \bar{z}_{2}-r)
u; r, r)w\rangle{\nonumber \\}&&=\langle w',
\mathbb{Y}(Y^{L}(u^L,
z_{1}-r)Y^{R}(u^{R}, \bar{z}_{2}-r)
u; r, r)w\rangle{\nonumber \\}\end{aligned}$$ when $|z_{1}|, |z_{2}|>r>|z_{1}-r|, |z_{2}-r|>0$ for all $u^L\in V^L$, $u^R\in V^R$, $u, w\in F$ and $w'\in F'$. By the commutativity for the full vertex operator map $\mathbb{Y}$, $$\label{L-R-Y-cl}
\langle w', Y^{L}(u^L, z_{1})Y^{R}(u^{R}, \bar{z}_{2})
\mathbb{Y}(u; r, r)w\rangle$$ and $$\label{Y-cl-L-R}
\langle w',
\mathbb{Y}(u; r, r)Y^{L}(u^L, z_{1})Y^{R}(u^{R}, \bar{z}_{2})
w\rangle$$ are absolutely convergent in the regions $|z_{1}|, |z_{2}|>r>0$ and $r>|z_{1}|, |z_{2}|>0$, respectively, to the correlation function $$\label{4-pt-fn}
\langle w', m_{4}(u^L\otimes {\mathbf{1}}^{L},
{\mathbf{1}}^{R}\otimes u^R, u, w; z_{1}, \bar{z}_{1}, z_{2}, \bar{z}_{2},
r, r, 0, 0).$$
By our discussion above, we know that the right-hand side of (\[asso-equ\]), (\[L-R-Y-cl\]) and (\[Y-cl-L-R\]) are all analytic in $z_{1}$ and $\bar{z}_{2}$ and that we can take $z_{1}=\bar{z}_{2}$ in the right-hand side of (\[asso-equ\]), (\[L-R-Y-cl\]) and (\[Y-cl-L-R\]). Thus after taking $z_{1}=\bar{z}_{2}$, the right-hand side of (\[asso-equ\]), (\[L-R-Y-cl\]) and (\[Y-cl-L-R\]) are analytic in $z=z_{1}=\bar{z}_{2}$. Since the right-hand side of (\[asso-equ\]), (\[L-R-Y-cl\]) and (\[Y-cl-L-R\]) are the expansions of (\[4-pt-fn\]) in the regions $r>|z_{1}-r|, |z_{2}-r|>0$, $|z_{1}|, |z_{2}|>r>0$ and $r>|z_{1}|, |z_{2}|>0$, respectively, we see that we can also let $z_{1}=\bar{z}_{2}$ in (\[4-pt-fn\]) and the result is also analytic in $z=z_{1}=\bar{z}_{2}$. Thus we have proved that $$\begin{aligned}
&\langle w',
\mathbb{Y}(\mathbb{Y}_{f}(\rho(u^L\otimes u^{R}); z-r, z-r)
u; r, r)w\rangle,&\\
&\langle w',\mathbb{Y}_{f}(\rho(u^L\otimes u^{R}); z, z)
\mathbb{Y}(u; r, r)w\rangle,&\\
&\langle w',
\mathbb{Y}(u; r, r)\mathbb{Y}_{f}(\rho(u^L\otimes u^{R}); z, z)
w\rangle\end{aligned}$$ are absolutely convergent to $$\langle w', m_{4}(\rho(u^L\otimes {\mathbf{1}}^{L}),
\rho({\mathbf{1}}^{R}\otimes u^R), u, w; z, \bar{z}, \bar{z}, z,
r, r, 0, 0)$$ which is in fact analytic in $z$. Using the Cauchy formula for contour integrals, we obtain the Cauchy-Jacobi identity $$\begin{aligned}
\label{cauchy-jacobi}
\lefteqn{{\mbox{\rm Res}}_{z=\infty}f(z)\langle w',\mathbb{Y}_{f}(\rho(u^L\otimes u^{R});
z, z)
\mathbb{Y}(u; r, r)w\rangle}{\nonumber \\}&&-{\mbox{\rm Res}}_{z=0}f(z)\langle w',
\mathbb{Y}(u; r, r)\mathbb{Y}_{f}(\rho(u^L\otimes u^{R}); z, z)
w\rangle{\nonumber \\}&&={\mbox{\rm Res}}_{z=r}f(z)\langle w',
\mathbb{Y}(\mathbb{Y}_{f}(\rho(u^L\otimes u^{R}); z-r, z-r)
u; r, r)w\rangle,\end{aligned}$$ where $f(z)$ is a rational function of $z$ with the only possible poles at $z=0, r, \infty$. Since $w$ and $w'$ are arbitrary, this Cauchy-Jacobi identity gives us identities for the components of the vertex operator $\mathbb{Y}_{f}(u; x, x)$. These identities are the component form of the Jacobi identity for $\mathbb{Y}_{f}(u; x, x)$.
Let $c^{L}, c^{R}\in {\mathbb{C}}$. A [*conformal full field algebra of central charges $(c^{L}, c^{R})$*]{} is a grading-restricted ${\mathbb{R}}\times {\mathbb{R}}$-graded full field algebra $(F, m, {\mathbf{1}}, D^{L}, D^{R})$ equipped with elements $\omega^{L}$ and $\omega^{R}$ called [*left conformal element*]{} and [*right conformal element*]{}, respectively, satisfying the following conditions:
1. The formal full vertex operators $\mathbb{Y}_{f}(\omega^{L}; x, \bar{x})$ and $\mathbb{Y}_{f}(\omega^{R}; x, \bar{x})$ are Laurent series in $x$ and $\bar{x}$, respectively, that is, $$\begin{aligned}
\mathbb{Y}_{f}(\omega^{L}; x, \bar{x})&=&\sum_{n\in {\mathbb{Z}}}L^{L}(n)x^{-n-2},\\
\mathbb{Y}_{f}(\omega^{R}; x, \bar{x})&=&\sum_{n\in {\mathbb{Z}}}L^{R}(n)\bar{x}^{-n-2}.\\\end{aligned}$$
2. The [*Virasoro relations*]{}: For $m, n\in {\mathbb{Z}}$, $$\begin{aligned}
[L^{L}(m), L^{L}(n)]
&=&(m-n)L^{L}(m+n)+\frac{c^{L}}{12}(m^{3}-m)\delta_{m+n, 0},\\
{[L^{R}(m), L^{R}(n)]}
&=&(m-n)L^{R}(m+n)+\frac{c^{R}}{12}(m^{3}-m)\delta_{m+n, 0},\\
{[L^{L}(m), L^{R}(n)]}&=&0.\end{aligned}$$
3. $d^{L}=L^{L}(0)$, $d^{R}=L^{R}(0)$, $D^{L}=L^{L}(-1)$ and $D^{R}=L^{R}(-1)$.
We shall denote the conformal full field algebra by $(F, m, {\mathbf{1}}, \omega^{L}, \omega^{R})$ or simply by $F$.
We have:
Let $(F, m, {\mathbf{1}}, \omega^{L}, \omega^{R})$ be a conformal full field algebra. Then the following [*commutator formula for Virasoro operators and formal full vertex operators*]{} hold: For $u\in F$, $$\begin{aligned}
\lefteqn{[\mathbb{Y}_{f}(\omega^{L}; x_{1}, \bar{x}_{1}),
\mathbb{Y}_{f}(u; x_{2}, \bar{x}_{2})]}{\nonumber \\}&&={\mbox{\rm Res}}_{x_{0}}x_{2}^{-1}\delta\left(\frac{x_{1}-x_{0}}{x_{2}}\right)
\mathbb{Y}_{f}(\mathbb{Y}_{f}(\omega^{L}; x_{0}, \bar{x}_{0})u; x_{2},
\bar{x}_{2}),\label{omega-commu-l}\\
\lefteqn{[\mathbb{Y}_{f}(\omega^{R}; x_{1}, \bar{x}_{1}),
\mathbb{Y}_{f}(u; x_{2}, \bar{x}_{2})]}{\nonumber \\}&&={\mbox{\rm Res}}_{\bar{x}_{0}}\bar{x}_{2}^{-1}
\delta\left(\frac{\bar{x}_{1}-\bar{x}_{0}}{\bar{x}_{2}}\right)
\mathbb{Y}_{f}(\mathbb{Y}_{f}(\omega^{R}; x_{0}, \bar{x}_{0})u; x_{2},
\bar{x}_{2}).\label{omega-commu-r}\end{aligned}$$
[[*Proof.*]{}]{}For any $v'\in F'$, $u, v\in F$, we consider $$\label{m3-omega}
\langle v', m_{3}(\omega^{L}, u, v; z_{1}, \bar{z}_{1}, z_{2}, \bar{z}_{2},
0, 0)\rangle.$$ Using the convergence property and the permutation property for conformal full field algebras, we know that it is equal to $$\begin{aligned}
&\langle v', \mathbb{Y}(\omega^{L}; z_{1}, \bar{z}_{1})
\mathbb{Y}(u, z_{2}, \bar{z}_{2})v\rangle,&\label{omega-prod1}\\
&\langle v', \mathbb{Y}(u; z_{2}, \bar{z}_{2})
\mathbb{Y}(\omega^{L}, z_{1}, \bar{z}_{1})v\rangle,&\label{omega-prod2}\end{aligned}$$ in the regions $|z_{1}|>|z_{2}|>0$, $|z_{2}|>|z_{1}|>0$, respectively. By the definition of conformal full field algebra, we know that for any fixed $z_{2}\ne 0$, (\[omega-prod1\]) and (\[omega-prod2\]) are analytic as functions of $z_{1}$ in the regions $|z_{1}|>|z_{2}|>0$ and $|z_{2}|>|z_{1}|>0$, respectively. So (\[m3-omega\]) is analytic as a function of $z_{1}$ in the regions $|z_{1}|>|z_{2}|>0$ and $|z_{2}|>|z_{1}|>0$. But we know that (\[m3-omega\]) is smooth as a function of $z_{1}$ in ${\mathbb{C}}\setminus \{z_{2}, 0\}$. Thus (\[m3-omega\]) must be analytic in ${\mathbb{C}}\setminus \{z_{2}, 0\}$.
We know that (\[m3-omega\]) is equal to (\[omega-prod1\]), (\[omega-prod2\]) and $$\langle v', \mathbb{Y}(\mathbb{Y}(\omega^{L}; z_{1}-z_{2},
\bar{z}_{1}-\bar{z}_{2})u; z_{2}, \bar{z}_{2})v\rangle$$ in the regions $|z_{1}|>|z_{2}|>0$, $|z_{2}|>|z_{1}|>0$ and $z_{2}|>|z_{1}-z_{2}|>0$, respectively. Since $F$ is lower truncated, using the Virasoro relation, we see that $\mathbb{Y}_{f}(\omega^{L}; x, \bar{x})u$ and $\mathbb{Y}_{f}(\omega^{L}; x, \bar{x})v$ have only finitely many terms in negative powers of $x$. Also using the lower-truncation property of $F$ and the Virasoro relation, we see that for any $w\in F$, $\langle v', \mathbb{Y}_{f}(\omega^{L}; x, \bar{x})w\rangle$ has only finitely many terms in positive powers of $x$. Using these facts, we see that the singularities $z_{1}=z_{2}, 0, \infty$ of (\[m3-omega\]) are all poles. Using the Cauchy formula, we obtain the component form (\[omega-commu-l\]).
Similarly, we can prove (\[omega-commu-r\]).
The following is clear from the definition and Theorem \[mod-int-op\]:
Let $(V^{L}, Y^{L}, {\mathbf{1}}^{L}, \omega^{L})$ and $(V^{L}, Y^{L}, {\mathbf{1}}^{L},
\omega^{L})$ be vertex operator algebras of central charges $c^{L}$ and $c^{R}$, respectively. A full field algebra $(F, m,
\rho)$ over $V^{L}\otimes V^{R}$ equipped with the left and right conformal elements $\rho(\omega^{L}\otimes {\mathbf{1}}^{R})$ and $\rho({\mathbf{1}}^{L}\otimes \omega^{R})$ is a conformal full field algebra.
In view of this proposition, we shall call the conformal full field algebra in the proposition above, that is, a full field algebra $(F, m,
\rho)$ over $V^{L}\otimes V^{R}$ equipped with the left and right conformal elements $\rho(\omega^{L}\otimes {\mathbf{1}}^{R})$ and $\rho({\mathbf{1}}^{L}\otimes \omega^{R})$, a [*conformal full field algebra over $V^{L}\otimes V^{R}$*]{} and denote it by $(F, m, \rho)$ or simply by $F$.
Intertwining operator algebras and full field algebras
======================================================
Let $V^{L}$ and $V^{R}$ be vertex operator algebras. In the preceding section, we have shown that a conformal full field algebra $(F, m, \rho)$ over $V^{L}\otimes V^{R}$ is a module for the vertex operator algebra $V^{L}\otimes V^{R}$ and the $\mathbb{Y}_{f}(\cdot; x, x)$ is an intertwining operator of type ${{{F}\choose {FF}}}$. This result suggests a method to construct conformal full field algebras from intertwining operator algebras, which are algebras of intertwining operators for vertex operator algebras and were introduced and studied in [@H1], [@H2], [@H3], [@H3.5], [@H4] and [@H5] by the first author.
Let $V$ be a vertex operator algebra and for a $V$-module $W$, let $C_{1}(W)$ be the subspace of $V$ spanned by $u_{-1}w$ for $u\in V_{+}=\coprod_{n\in {\mathbb{Z}}_{+}}V_{(n)}$ and $w\in W$.
We consider the following conditions for a vertex operator algebra $V$:
1. Every ${\mathbb{C}}$-graded generalized $V$-module is a direct sum of ${\mathbb{C}}$-graded irreducible $V$-modules.
2. There are only finitely many inequivalent ${\mathbb{C}}$-graded irreducible $V$-modules and they are all ${\mathbb{R}}$-graded.
3. Every ${\mathbb{R}}$-graded irreducible $V$-module $W$ satisfies the $C_{1}$-cofiniteness condition, that is, $\dim W/C_{1}(W)<\infty$.
In this section, we fix vertex operator algebras $(V^L, Y^{L}, {\mathbf{1}}^{L}, \omega^{L})$ and $(V^{R}, Y^{R}, {\mathbf{1}}^{R}, \omega^{R})$ satisfying these conditions. Let $\mathcal{A}^L$ and $\mathcal{A}^R$ be the sets of equivalent classes of irreducible modules for $V^L$ and for $V^R$, respectively. Let $\{W^{L; a}\;|\;a\in {\mathcal{A}}^{L}\}$ be a complete set of representatives of the equivalence classes in ${\mathcal{A}}^{L}$ and $\{W^{R; b}\;|\;b\in {\mathcal{A}}^{R}\}$ a complete set of representatives of the equivalence classes in ${\mathcal{A}}^{R}$.
\[rational\] The vertex operator algebra $V^L\otimes V^R$ also satisfy the conditions above.
[[*Proof.*]{}]{}Let $W$ be a generalized $V^L\otimes V^R$-module. Then $W$ is a generalized $V^{L}$-module. So there exist vector spaces $M^{a}$ for $a\in {\mathcal{A}}^{L}$ such that $W$ is equivalent to the generalized $V^{L}$-module $\coprod_{a\in {\mathcal{A}}^{L}}(W^{L;a}\otimes M^{a})$. Since $W$ is also a generalized $V^{R}$-module, $M^{a}$ must be $V^{R}$-modules. So they can be written as direct sums of irreducible $V^{R}$-modules $W^{R;b}$, $b\in {\mathcal{A}}^{R}$. So $W$ is equivalent to $\coprod_{a\in {\mathcal{A}}^{L}, b\in {\mathcal{A}}^{R}}N_{ab}(W^{L;b}\otimes W^{R;a})$ where $N_{ab}\in {\mathbb{N}}$ for $a\in {\mathcal{A}}^{L}$, $b\in {\mathcal{A}}^{R}$. By Proposition 4.7.2 of [@FHL], $W^{L;a}\otimes W^{R;b}$ are irreducible $V^L\otimes V^R$-modules. So $V^L\otimes V^R$ satisfies Condition 1. The second condition follows from Theorem 4.7.4 of [@FHL]. The $C_{1}$-cofiniteness follows immediately from the fact that $$C_{1}(W^{L; a})\otimes W^{R; b}\oplus W^{L; a}\otimes
C_{1}(W^{R; b})\subset
C_{1}(W^{L; a}\otimes W^{R; b}).$$
This result immediately gives:
\[mod-decomp\] Let $F$ be a module for the vertex operator algebra $V^{L}\otimes V^{R}$. Then as a module for the vertex operator algebra $V^{L}\otimes V^{R}$, $F$ is isomorphic to $$\label{F-exp}
\coprod_{a\in \mathcal{A}^{L}}\coprod_{b\in \mathcal{A}^{R}} \coprod_{m_{ab}=1}
^{h_{ab}}(W^{L; a})^{(m_{ab})} \otimes (W^{R; b})^{(m_{ab})}$$
Let $F$ be a module for the vertex operator algebra $V^{L}\otimes V^{R}$ and let $\gamma$ be an isomorphism from (\[F-exp\]) to $F$. Then there exist operators $L^{L}(0)$ and $L^{R}(0)$ on $F$ given by $$\begin{aligned}
L^{L}(0)\rho(w^{L}\otimes w^{R})&=&\rho((L^{L}(0)w^{L})\otimes w^{R}),\\
L^{R}(0)\rho(w^{L}\otimes w^{R})&=&\rho(w^{L}\otimes (L^{R}(0)w^{R}))\end{aligned}$$ for $w^{L}\in (W^{L; a})^{(m_{ab})}$ and $w^{R}\in (W^{R; b})^{(m_{ab})}$. Clearly $L^{L}(0)$ and $L^{R}(0)$ commute with each other.
Let ${\mathcal{Y}}$ be an intertwining operator of type ${{{F}\choose {FF}}}$ and $\gamma$ an isomorphism from (\[F-exp\]) to $F$. Let $$\begin{aligned}
\mathbb{Y}^{\mathcal{Y}}: (F\otimes F)\times {\mathbb{C}}^{\times}&\to&
\overline{F}{\nonumber \\}(u\otimes v, z)&\mapsto& \mathbb{Y}^{\mathcal{Y}}(u; z, \bar{z})v\end{aligned}$$ and $$\begin{aligned}
\mathbb{Y}_{f}^{\mathcal{Y}}: F\otimes F&\to&
F\{x, \bar{x}\}{\nonumber \\}u\otimes v&\mapsto& \mathbb{Y}_{f}^{\mathcal{Y}}(u; x, \bar{x})v\end{aligned}$$ be linear maps given by $$\mathbb{Y}^{\mathcal{Y}}(u; z, \bar{z})v=z^{L^L(0)}
\bar{z}^{L^R(0))} \mathcal{Y}(u, 1)
z^{-L^L(0)} \bar{z}^{-L^R(0)}$$ and $$\mathbb{Y}_{f}^{\mathcal{Y}}(u; x, \bar{x})v=x^{L^L(0)}
\bar{x}^{L^R(0)} \mathcal{Y}(u, 1)
x^{-L^L(0)} \bar{x}^{-L^R(0)},$$ respectively, for $u\in F$. We call $\mathbb{Y}^{\mathcal{Y}}$ and $\mathbb{Y}_{f}^{\mathcal{Y}}$ the [*splitting*]{} and [*formal splitting*]{} of ${\mathcal{Y}}$, respectively.
\[int-decomp\] Let ${\mathcal{Y}}$ be an intertwining operator of type ${{{F}\choose {FF}}}$, $\mathbb{Y}^{\mathcal{Y}}$ and $\mathbb{Y}_{f}^{\mathcal{Y}}$, the splitting and formal splitting of ${\mathcal{Y}}$, respectively, and $\gamma$ an isomorphism from (\[F-exp\]) to $F$. Then for any $a_{1}, a_{2}\in {\mathcal{A}}^{L}$, $b_{1}, b_{2}\in
{\mathcal{A}}^{R}$, $1\le m_{a_{1}b_{1}}\le h_{a_{1}b_{1}}$ and $1\le m_{a_{2}b_{2}}\le
h_{a_{2}b_{2}}$, there exist intertwining operators ${\mathcal{Y}}_{a_{1}a_{2}}^{L;m_{a_{3}b_{3}};a_{3}}$ and ${\mathcal{Y}}_{b_{1}b_{2}}^{R; m_{a_{3}b_{3}}; b_{3}}$ for $a_{3}\in {\mathcal{A}}^{L}$, $b_{3}\in {\mathcal{A}}^{R}$ and $m_{a_{3}b_{3}}=1, \dots,
h_{a_{3}b_{3}}$ of types ${{{(W^{L;a_{3}})^{(m_{a_{3}b_{3}})}}\choose {(W^{L;a_{1}})^{(m_{a_{1}b_{1}})}\;(W^{L;a_{2}})^{(m_{a_{2}b_{2}})}}}}$ and ${{{(W^{R;b_{3}})^{(m_{a_{3}b_{3}})}}\choose {(W^{R;b_{1}})^{(m_{a_{1}b_{1}})}\;(W^{R;b_{2}})^{(m_{a_{2}b_{2}})}}}}$, respectively, such that for $u^L\otimes u^R \in (W^{L; a_{1}})^{(m_{a_{1}b_{1}})}\otimes
(W^{R; b_{1}})^{(m_{a_{1}b_{1}})}$ and $v^L\otimes v^R \in (W^{L; a_{2}})^{(m_{a_{2}b_{2}})}\otimes
(W^{R; b_{2}})^{(m_{a_{2}b_{2}})}$, we have $$\begin{aligned}
\label{ch-ach}
\lefteqn{\mathbb{Y}^{\mathcal{Y}}(\gamma(u^L\otimes u^R); z,\bar{z})
\gamma(v^L \otimes v^R)}{\nonumber \\}&& =
\sum_{a_{3}\in {\mathcal{A}}^{L}}\sum_{b_{3}\in {\mathcal{A}}^{R}}
\sum_{m_{a_{3}b_{3}}=1}^{h_{a_{3}b_{3}}}\gamma(
{\mathcal{Y}}_{a_{1}a_{2}}^{L;m_{a_{3}b_{3}};a_{3}}(u^L, z)v^L \otimes
{\mathcal{Y}}_{b_{1}b_{2}}^{R; m_{a_{3}b_{3}}; b_{3}}(u^R,\bar{z})v^R).{\nonumber \\}\end{aligned}$$ Similarly, for the formal full vertex operator, we have $$\begin{aligned}
\label{ch-ach-f}
\lefteqn{\mathbb{Y}_{f}^{\mathcal{Y}}(\gamma(u^L\otimes u^R); x,\bar{x})
\gamma(v^L \otimes v^R)}{\nonumber \\}&& =
\sum_{a_{3}\in {\mathcal{A}}^{L}}\sum_{b_{3}\in {\mathcal{A}}^{R}}
\sum_{m_{a_{3}b_{3}}=1}^{h_{a_{3}b_{3}}}\gamma(
{\mathcal{Y}}_{a_{1}a_{2}}^{L;m_{a_{3}b_{3}};a_{3}}(u^L, x)v^L \otimes
{\mathcal{Y}}_{b_{1}b_{2}}^{R; m_{a_{3}b_{3}}; b_{3}}(u^R,\bar{x})v^R).{\nonumber \\}\end{aligned}$$
[[*Proof.*]{}]{} Since $\mathbb{Y}^{{\mathcal{Y}}}$ restricted to $$\gamma((W^{L; a_{1}})^{(m_{a_{1}b_{1}})}\otimes
(W^{R; b_{1}})^{(m_{a_{1}b_{1}})})\otimes
\gamma((W^{L; a_{2}})^{(m_{a_{2}b_{2}})}\otimes
(W^{R; b_{2}})^{(m_{a_{2}b_{2}})})$$ is an intertwining operator of type $${{{F}\choose {\gamma((W^{L; a_{1}})^{(m_{a_{1}b_{1}})}\otimes
(W^{R; b_{1}})^{(m_{a_{1}b_{1}})})\;\;
\gamma((W^{L; a_{2}})^{(m_{a_{2}b_{2}})}\otimes
(W^{R; b_{2}})^{(m_{a_{2}b_{2}})})}}},$$ it was proved in [@DMZ] that (\[ch-ach\]) is true when $z=\bar{z}=r>0$. Then we have $$\begin{aligned}
\label{Y-cl-z}
\lefteqn{\mathbb{Y}^{{\mathcal{Y}}}(\gamma(u^L\otimes u^R); z, \bar{z})
\gamma(v^L \otimes v^R)}{\nonumber \\}&&= z^{L^L(0)}
\bar{z}^{L^R(0)} \mathcal{Y}(\gamma(u^L\otimes u^R), 1)
z^{-L^L(0)} \bar{z}^{-L^R(0)} \gamma(v^L \otimes v^R) {\nonumber \\}&&= \sum_{a_{3}\in {\mathcal{A}}^{L}}\sum_{b_{3}\in {\mathcal{A}}^{R}}
\sum_{m_{a_{3}b_{3}}=1}^{h_{a_{3}b_{3}}}\gamma(
(z^{L^L(0)}{\mathcal{Y}}_{a_{1}a_{2}}^{L;m_{a_{3}b_{3}};a_{3}}(u^L, 1)
\bar{z}^{-L^L(0)}v^{L}){\nonumber \\}&&\hspace{12em}\otimes (z^{L^R(0)} {\mathcal{Y}}_{b_{1}b_{2}}^{R; m_{a_{3}b_{3}}; b_{3}}
(u^R, 1) \bar{z}^{-L^R(0)}v^{R})) {\nonumber \\}&&= \sum_{a_{3}\in {\mathcal{A}}^{L}}\sum_{b_{3}\in {\mathcal{A}}^{R}}
\sum_{m_{a_{3}b_{3}}=1}^{h_{a_{3}b_{3}}}\gamma(
{\mathcal{Y}}_{a_{1}a_{2}}^{L;m_{a_{3}b_{3}};a_{3}}
(u^{L}, z)v^{L}\otimes {\mathcal{Y}}_{b_{1}b_{2}}^{R; m_{a_{3}b_{3}}; b_{3}}
(u^{R}, \bar{z})v^{R}). \end{aligned}$$ The proof of (\[ch-ach-f\]) is completely the same.
\[r-cfa-decomp\] Let $(F, m, \rho)$ be a conformal full field algebra over $V^L\otimes V^R$. Then as a module for the vertex operator algebra $V^{L}\otimes V^{R}$, $F$ is isomorphic to (\[F-exp\]). Moreover, if $\gamma$ is an isomorphism from (\[F-exp\]) to $F$, then for any $a_{1}, a_{2}\in {\mathcal{A}}^{L}$, $b_{1}, b_{2}\in
{\mathcal{A}}^{R}$, $1\le m_{a_{1}b_{1}}\le h_{a_{1}b_{1}}$ and $1\le m_{a_{2}b_{2}}\le
h_{a_{2}b_{2}}$, there exist intertwining operators ${\mathcal{Y}}_{a_{1}a_{2}}^{L;m_{a_{3}b_{3}};a_{3}}$ and ${\mathcal{Y}}_{b_{1}b_{2}}^{R; m_{a_{3}b_{3}}; b_{3}}$ for $a_{3}\in {\mathcal{A}}^{L}$, $b_{3}\in {\mathcal{A}}^{R}$ and $m_{a_{3}b_{3}}=1, \dots,
h_{a_{3}b_{3}}$ of types ${{{(W^{L;a_{3}})^{(m_{a_{3}b_{3}})}}\choose {(W^{L;a_{1}})^{(m_{a_{1}b_{1}})}\;(W^{L;a_{2}})^{(m_{a_{2}b_{2}})}}}}$ and ${{{(W^{R;b_{3}})^{(m_{a_{3}b_{3}})}}\choose {(W^{R;b_{1}})^{(m_{a_{1}b_{1}})}\;(W^{R;b_{2}})^{(m_{a_{2}b_{2}})}}}}$, respectively, such that for $u^L\otimes u^R \in (W^{L; a_{1}})^{(m_{a_{1}b_{1}})}\otimes
(W^{R; b_{1}})^{(m_{a_{1}b_{1}})}$ and $v^L\otimes v^R \in (W^{L; a_{2}})^{(m_{a_{2}b_{2}})}\otimes
(W^{R; b_{2}})^{(m_{a_{2}b_{2}})}$, the formulas (\[ch-ach\]) and (\[ch-ach-f\]) hold when $\mathbb{Y}^{\mathcal{Y}}$ and $\mathbb{Y}^{\mathcal{Y}}_{f}$ are replaced by $\mathbb{Y}$ and $\mathbb{Y}_{f}$, respectively.
[[*Proof.*]{}]{}The first conclusion follows immediately from Corollary \[mod-decomp\]. Now if we consider the intertwining operator $\mathbb{Y}_{f}(\cdot; x, x)$, then the second conclusion follows immediately from Proposition \[int-decomp\].
For either the map $\mathbb{Y}_{f}^{{\mathcal{Y}}}$ in Proposition \[int-decomp\] or the formal full vertex operator map $\mathbb{Y}_{f}$ for a conformal full field algebra over $V^L\otimes V^R$, we can substitute $z$ and $\zeta$ for the formal variables $x$ and $\bar{x}$ in $\mathbb{Y}_{f}^{{\mathcal{Y}}}(\cdot; x,\bar{x})$ or $\mathbb{Y}_{f}(\cdot; x,\bar{x})$ (that is, substitute $e^{r\log z}$ and $e^{s\;\overline{\log \zeta}}$ for $x^{r}$ and $\bar{x}^{s}$, respectively, for $r, s\in {\mathbb{R}}$) to obtain $\mathbb{Y}_{\rm an}^{{\mathcal{Y}}}(\cdot; z, \zeta)$ (called [ *analytic splitting of ${\mathcal{Y}}$*]{}) or $\mathbb{Y}_{\rm an}(\cdot; z, \zeta)$. Then by (\[ch-ach-f\]), we have:
\[fvo-decomp\] For the analytic splitting $\mathbb{Y}_{\rm an}^{{\mathcal{Y}}}$ of ${\mathcal{Y}}$ in Proposition \[int-decomp\], we have $$\begin{aligned}
\label{ch-ach-z-zeta}
\lefteqn{\mathbb{Y}_{\rm an}^{{\mathcal{Y}}}(\gamma(u^L\otimes u^R); z,\zeta)
\gamma(v^L \otimes v^R)}{\nonumber \\}&& =
\sum_{a_{3}\in {\mathcal{A}}^{L}}\sum_{b_{3}\in {\mathcal{A}}^{R}}
\sum_{m_{a_{3}b_{3}}=1}^{h_{a_{3}b_{3}}}\gamma(
{\mathcal{Y}}_{a_{1}a_{2}}^{L;m_{a_{3}b_{3}};a_{3}}(u^L, z)v^L \otimes
{\mathcal{Y}}_{b_{1}b_{2}}^{R; m_{a_{3}b_{3}}; b_{3}}(u^R,\zeta)v^R){\nonumber \\}\end{aligned}$$ for $u^L\otimes u^R \in (W^{L; a_{1}})^{(m_{a_{1}b_{1}})}\otimes
(W^{R; b_{1}})^{(m_{a_{1}b_{1}})}$ and $v^L\otimes v^R \in (W^{L; a_{2}})^{(m_{a_{2}b_{2}})}\otimes
(W^{R; b_{2}})^{(m_{a_{2}b_{2}})}$. The same is also true for the analytic full vertex operator map $\mathbb{Y}_{\rm an}$ for a conformal full field algebra over $V^L\otimes V^R$.
This corollary allows us to treat the left and right variables $z$ and $\bar{z}$ in $\mathbb{Y}^{{\mathcal{Y}}}(\cdot; z, \bar{z})$ or $\mathbb{Y}(\cdot; z, \bar{z})$ independently. In particular, we have the following strong versions of associativity and commutativity for conformal full field algebra over $V^L\otimes V^R$:
\[anal-assoc\] Let $(F, m, \rho)$ be a conformal full field algebra over $V^L\otimes V^R$. Then for $u,v,w\in F$ and $w'\in F'$, $$\begin{aligned}
\label{asso-z-zeta}
\lefteqn{\langle w', \mathbb{Y}_{\rm an}(u; z_{1},\zeta_{1})
\mathbb{Y}_{\rm an}(v; z_{2}, \zeta_{2})w\rangle} {\nonumber \\}&&= \langle w',
\mathbb{Y}_{\rm an}(\mathbb{Y}_{\rm an}(u; z_{1}-z_{2}, \zeta_{1}-\zeta_{2})v;
z_{2}, \zeta_{2})w\rangle\end{aligned}$$ when $|z_{1}|>|z_{2}|>|z_{1}-z_{2}|>0$ and $|\zeta_{1}|>|\zeta_{2}|>|\zeta_{1}-\zeta_{2}|>0$.
[[*Proof.*]{}]{}Using (\[ch-ach-z-zeta\]) and the convergence result proved by the first author in [@H5] for vertex operator algebras satisfying the conditions assumed for $V^L$ and $V^R$ in the beginning of this section, the left-hand side of (\[asso-z-zeta\]) converges absolutely when $|z_{1}|>|z_{2}|>0$ and $|\zeta_{1}|>|\zeta_{1}|>0$, and the right-hand side of (\[asso-z-zeta\]) converges absolutely when $|z_{2}|>|z_{1}-z_{2}|>0$ and $|\zeta_{2}|>|\zeta_{1}-\zeta_{2}|>0$. By the associativity (\[asso-1\]), (\[asso-z-zeta\]) is true when $\zeta_{1}=\bar{z}_{1}$ and $\zeta_{2}=\bar{z}_{2}$ for all $u,v,w\in F$ and $w'\in F'$. In particular, replacing $u$ by $(L^L(-1))^k (L^R(-1))^l u$, $v$ by $(L^L(-1))^m L^R(-1))^n v$, for $k,l, m, n\in {\mathbb{N}}$ and using the $L^{L}(-1)$- and $L^{R}(-1)$-derivative properties, we obtain $$\begin{aligned}
\label{asso-z-zeta-1}
\lefteqn{\frac{\partial^k}{\partial z_{1}^k}
\frac{\partial^l}{\partial \zeta_{1}^l}
\frac{\partial^m}{\partial z_{2}^m} \frac{\partial^n}{\partial \zeta_{2}^n}
\langle w', \mathbb{Y}_{\rm an}(u; z_{1},\zeta_{1})
\mathbb{Y}_{\rm an}(v; z_{2}, \zeta_{2})w\rangle
{\bigg\vert}_{\zeta_{1}=\bar{z}_{1}, \zeta_{2}=\bar{z}_{2}}
} {\nonumber \\}&&= \frac{\partial^k}{\partial z_{1}^k} \frac{\partial^l}{\partial \zeta_{1}^l}
\frac{\partial^m}{\partial z_{2}^m} \frac{\partial^n}{\partial \zeta_{2}^n}
\langle w',
\mathbb{Y}_{\rm an}(\mathbb{Y}_{\rm an}(u; z_{1}-z_{2}, \zeta_{1}-\zeta_{2})v;
z_{2}, \zeta_{2})w\rangle{\bigg\vert}_{\zeta_{1}=\bar{z}_{1}, \zeta_{2}=\bar{z}_{2}}{\nonumber \\}\end{aligned}$$ for all $k,l, m, n\in {\mathbb{N}}$, when $|z_{1}|>|z_{2}|>0$ and $|z_{2}|>|z_{1}-z_{2}|>0$. We know that both sides of (\[asso-z-zeta\]) give branches of some multivalued analytic functions in the region given by $|z_{1}|>|z_{2}|>0$, $|\zeta_{1}|>|\zeta_{1}|>0$, $|z_{2}|>|z_{1}-z_{2}|>0$ and $|\zeta_{2}|>|\zeta_{1}-\zeta_{2}|>0$. From (\[asso-z-zeta-1\]), we know that the power series expansions of these branches are equal in the neighborhood of those points satisfying $\zeta_{1}=\bar{z}_{1}$, $\zeta_{2}=\bar{z}_{2}$. Thus (\[asso-z-zeta-1\]) holds in the region $|z_{1}|>|z_{2}|>0$, $|\zeta_{1}|>|\zeta_{1}|>0$, $|z_{2}|>|z_{1}-z_{2}|>0$ and $|\zeta_{2}|>|\zeta_{1}-\zeta_{2}|>0$.
\[comm-0\] Let $(F, m, \rho)$ be a conformal full field algebra over $V^L\otimes V^R$. Then for $u,v,w\in F$ and $w'\in F'$, $$\label{prod-12}
\langle w', \mathbb{Y}_{\rm an}(u; z_{1}, \zeta_{1})
\mathbb{Y}_{\rm an}(v; z_{2}, \zeta_{2} )w\rangle$$ and $$\label{prod-21}
\langle w', \mathbb{Y}_{\rm an}(v; z_{2}, \zeta_{2})
\mathbb{Y}_{\rm an}(u; z_{1}, \zeta_{1} )w\rangle$$ are absolutely convergent when $|z_{1}|>|z_{2}|>0$, $|\zeta_{1}|>|\zeta_{2}|>0$ and when $|z_{2}|>|z_{1}|>0$, $|\zeta_{2}|>|\zeta_{1}|>0$, respectively, and can both be analytically extended to a same multivalued analytic function of $(z_{1}, z_{2}; \zeta_{1}, \zeta_{2})$ for $(z_{1}, z_{2}; \zeta_{1}, \zeta_{2})\in M^{2}\times M^{2}$, where $M^{2}=\{(z_{1}, z_{2})\in ({\mathbb{C}}^{\times})^{2}\;|\;z_{1}\ne z_{2}\}$.
[[*Proof.*]{}]{}The convergence and the existence of analytic extensions follow immediately from Corollary \[fvo-decomp\] and the convergence and the existence of analytic extensions of products of intertwining operators for the vertex operator algebras $V^{L}$ and $V^{R}$.
By Proposition \[commu-general\], we know that these two multivalued analytic functions obtained by analytically extending (\[prod-12\]) and (\[prod-21\]) have equal values at points of the form $(z_{1}, \bar{z}_{1},
\dots, z_{n}, \bar{z}_{n})$ for $(z_{1}, \dots, z_{n})\in
\mathbb{F}_{n}({\mathbb{C}})$. Using the $L^{L}(-1)$- and $L^{R}(-1)$-conjugation properties for full vertex operators, we see that these two analytic functions are actually the same, that is, they are analytic extensions of each other.
For $(z_{1}, \dots, z_{n}), (\zeta_{1}, \dots, \zeta_{n})\in
\mathbb{F}_{n}({\mathbb{C}})$, we denote the corresponding elements of $\mathbb{F}_{n}({\mathbb{C}})\times \mathbb{F}_{n}({\mathbb{C}})$ by $(z_{1}, \zeta_{1}, \dots, z_{n}, \zeta_{n})$ instead of $(z_{1}, \dots, z_{n}, \zeta_{1}, \dots, \zeta_{n})$. We have the following analyticity of the correlation functions:
Let $(F, m, \rho)$ be a conformal full field algebra over $V^L\otimes V^R$. For any $n\in {\mathbb{Z}}_{+}$ and $u_{1}, \dots, u_{n}$, there exists a multivalued analytic function of $(z_{1}, \zeta_{1}, \dots,
z_{n}, \zeta_{n})\in \mathbb{F}_{n}({\mathbb{C}})\times \mathbb{F}_{n}({\mathbb{C}})$ such that for $(z_{1}, \dots,
z_{n})\in \mathbb{F}_{n}({\mathbb{C}})$, the values $$m_{n}(u_{1},
\dots, u_{n}; z_{1}, \bar{z}_{1}, \dots,
z_{n}, \bar{z}_{n})$$ of the correlation function is a value of this multivalued analytic function above at the point $(z_{1}, \bar{z}_{1}, \dots,
z_{n}, \bar{z}_{n})$. Moreover, these multivalued analytic functions are determined uniquely by the products of analytic full vertex operators in their regions of convergence.
[[*Proof.*]{}]{}The proof of this result is basically the same as the proof of the generalized rationality for intertwining operator algebras in [@H4]. We have proved the above strong versions of associativity and commutativity for analytic full vertex operators. Using these strong versions of associativity and commutativity, we see that the multivalued analytic functions in various regions obtained from all kinds of products and iterates of analytic full vertex operators are analytic extensions of each other. Thus we have such a global multivalued analytic function. Clearly these multivalued analytic functions are determined uniquely by the products of analytic full vertex operators in their regions of convergence.
By the results above, we see that for a conformal full field algebra over $V^L\otimes V^R$, the correlation function maps are determined uniquely by the products of analytic full vertex operators in their regions of convergence, and thus are determined uniquely by the full vertex operator map. In view of this fact, we shall use also $(F, \mathbb{Y}, \rho)$ to denote a conformal full field algebra over $V^L\otimes V^R$.
We shall use $$\label{anal-ext}
E(m)_{n}(u_{1}, \dots, u_{n}; z_{1}, \zeta_{1}, \dots,
z_{n}, \zeta_{n})$$ to denote the analytic extension obtained in the proposition above together with the prefered values $$m_{n}(u_{1},
\dots, u_{n}; z_{1}, \bar{z}_{1}, \dots,
z_{n}, \bar{z}_{n})$$ at the special points of the form $(z_{1}, \bar{z}_{1}, \dots,
z_{n}, \bar{z}_{n})$. For $u_{1}, \dots, u_{n}\in F$ and a path $$\begin{aligned}
\gamma: [0, 1]&\to &\mathbb{F}_{n}({\mathbb{C}})\times
\mathbb{F}_{n}({\mathbb{C}}){\nonumber \\}t&\mapsto& (z_{1}(t), \zeta_{1}(t), \dots, z_{n}(t), \zeta_{n}(t))\end{aligned}$$ starting from a point of the form $(z_{1}, \bar{z}_{1}, \dots,
z_{n}, \bar{z}_{n})$, we shall use $$E(m)_{n}(u_{1}, \dots, u_{n}; z_{1}(t), \zeta_{1}(t), \dots,
z_{n}(t), \zeta_{n}(t))$$ to denote the value of (\[anal-ext\]) at the point $\gamma(t)$ obtained by analytically extend the preferred value of (\[anal-ext\]) at the starting point $\gamma(0)$ of $\gamma$ along the path $\gamma$ to the point $\gamma(t)$.
\[anal-perm-cor\] Let $(F, \mathbb{Y}, \rho)$ be a conformal full field algebra over $V^L\otimes V^R$. Let $$\begin{aligned}
\gamma: [0, 1]&\to &\mathbb{F}_{n}({\mathbb{C}})\times
\mathbb{F}_{n}({\mathbb{C}}){\nonumber \\}t&\mapsto& (z_{1}(t), \zeta_{1}(t), \dots, z_{n}(t), \zeta_{n}(t))\end{aligned}$$ be a path starting from a point of the form $(z_{1}, \bar{z}_{1}, \dots,
z_{n}, \bar{z}_{n})$. Then we have the following [*permutation property*]{}: For $u_{1}, \dots, u_{n}\in F$ and $\sigma\in S_{n}$, $$\begin{aligned}
\label{anal-perm}
\lefteqn{E(m)_{n}(u_{1}, \dots, u_{n}; z_{1}(t), \zeta_{1}(t), \dots,
z_{n}(t), \zeta_{n}(t))}{\nonumber \\}&&=E(m)_{n}(u_{\sigma(1)}, \dots, u_{\sigma(n)}; z_{\sigma(1)}(t),
\zeta_{\sigma(1)}(t), \dots,
z_{\sigma(n)}(t), \zeta_{\sigma(n)}(t)).\end{aligned}$$
[[*Proof.*]{}]{}This follows immediately from the permutation property for full field algebras and the uniqueness of analytic extensions.
\[comm\] Let $(F, \mathbb{Y}, \rho)$ be a conformal full field algebra over $V^L\otimes V^R$. Let $r_{1}, r_{2}\in {\mathbb{R}}$ satisfying $r_{2}>r_{1}>0$. Then for $u,v,w\in F$ and $w'\in F'$, $$\label{comm-cft-1-2}
\langle w', \mathbb{Y}_{\rm an}(v; r_{2}, r_{2})
\mathbb{Y}_{\rm an}(u; r_{1}, r_{1} )w\rangle,$$ can be obtained by analytically extending the analytic function (which is a branch of a multivalued function) $$\label{comm-cft-2-1}
\langle w', \mathbb{Y}_{\rm an}(u; z_{1}, \zeta_{1})
\mathbb{Y}_{\rm an}(v; z_{2}, \zeta_{2} )w\rangle ,$$ defined near the point $z_{1}=\zeta_{1}=r_{2}$, $z_{2}=\zeta_{2}=r_{1}$, in the region $|z_{1}|>|z_{2}|>0$ and $|\zeta_{1}|
>|\zeta_{2}|>0$, along the path given by $$\begin{aligned}
[0, 1]&\to &M^{2}\times M^{2}{\nonumber \\}t&\mapsto& ((z_{1}(t), z_{2}(t)), (\zeta_{1}(t), \zeta_{2}(t))),\end{aligned}$$ where $$\begin{aligned}
z_{1}(t)&=&\frac{r_{1}+r_{2}}{2}+e^{i\pi t}\frac{r_{2}-r_{1}}{2},{\nonumber \\}z_{2}(t)&=&\frac{r_{1}+r_{2}}{2}-e^{i\pi t}\frac{r_{2}-r_{1}}{2},{\nonumber \\}\zeta_{1}(t)&=&\frac{r_{1}+r_{2}}{2}+e^{-i\pi t}\frac{r_{2}-r_{1}}{2},{\nonumber \\}\zeta_{2}(t)&=&\frac{r_{1}+r_{2}}{2}-e^{-i\pi t}\frac{r_{2}-r_{1}}{2},\end{aligned}$$ to the region $|z_{2}|>|z_{1}|>0$ and $|\zeta_{2}|
>|\zeta_{1}|>0$ and then evaluated at $z_{1}=\zeta_{1}=r_{1}$ and $z_{2}=\zeta_{2}=r_{2}$.
[[*Proof.*]{}]{}By Proposition \[comm-0\], we know that (\[comm-cft-1-2\]) can indeed be obtained from (\[comm-cft-2-1\]) by analytic extension. What we need to show now is that the analytic extension along the path given above gives precisely (\[comm-cft-1-2\]).
Since $\mathbb{Y}(\cdot; z, \bar{z})
=\mathbb{Y}_{\rm an}(\cdot; z, \zeta)|_{\zeta=\bar{z}}$ and $\zeta_{1}(t)=\overline{z_{1}(t)}$ and $\zeta_{2}(t)=\overline{z_{2}(t)}$, we see that (\[comm-cft-1-2\]) is equal to $\langle w', m_{3}(v, u, w; r_{2}, r_{2}, r_{1}, r_{1}, 0, 0)\rangle$ and that $$\label{comm-cft-2-1-1}
\langle w', \mathbb{Y}_{\rm an}(u; z_{1}(t), \zeta_{1}(t))
\mathbb{Y}_{\rm an}(v; z_{2}(t), \zeta_{2}(t))w\rangle$$ when $|z_{2}(t)|>|z_{1}(t)|>0$ and $$\label{comm-cft-2-1-2}
\langle w', \mathbb{Y}_{\rm an}(v; z_{2}(t), \zeta_{2}(t))
\mathbb{Y}_{\rm an}(u; z_{1}(t), \zeta_{1}(t))w\rangle$$ when $|z_{1}(t)|>|z_{2}(t)|>0$ are equal to $$\langle w', E(m)_{3}(u, v, w; z_{1}(t), \zeta_{1}(t),
z_{2}(t), \zeta_{2}(t), 0, 0)\rangle$$ and $$\langle w', E(m)_{3}(v, u, w; z_{2}(t), \zeta_{2}(t), z_{1}(t),
\zeta_{1}(t), 0, 0)\rangle,$$ respectively. By the permutation property of full field algebras and Corollary \[anal-perm-cor\]), we see that (\[comm-cft-1-2\]), (\[comm-cft-2-1-1\]) and (\[comm-cft-2-1-2\]) are equal to $$\langle w', m_{3}(u, v, w;
r_{1}, r_{1}, r_{2}, r_{2}, 0, 0)\rangle,$$ $$\langle w', E(m)_{3}(u, v, w; z_{1}(t), \zeta_{1}(t),
z_{2}(t), \zeta_{2}(t), 0, 0)\rangle$$ and $$\langle w', E(m)_{3}(u, v, w; z_{1}(t), \zeta_{1}(t), z_{2}(t),
\zeta_{2}(t), 0, 0)\rangle,$$ respectively. From this fact, we see that indeed the analytic extension of (\[comm-cft-2-1\]) near the point $z_{1}=\zeta_{1}=r_{2}$, $z_{2}=\zeta_{2}=r_{1}$, along the path given above gives (\[comm-cft-1-2\]).
This result can also be proved directly using the associativity (Proposition \[anal-assoc\]) and the skew-symmetry (\[skew-formal\]) (see [@K] for details).
\[r-conf-alg-2-prop\] A conformal full field algebra over $V^L\otimes V^R$ is equivalent to a module $F$ for the vertex operator algebra $V^L\otimes V^R$ equipped with an intertwining operator ${\mathcal{Y}}$ of type ${{{F}\choose {FF}}}$ and an injective linear map $\rho: V^L\otimes V^R \to F$, satisfying the following conditions:
1. The [*identity property*]{}: ${\mathcal{Y}}(\rho({\mathbf{1}}^L\otimes {\mathbf{1}}^R), x)=I_{F}$.
2. The [*creation property*]{}: For $u\in F$, $\lim_{x\rightarrow 0} {\mathcal{Y}}(u, x)\rho({\mathbf{1}}^L\otimes {\mathbf{1}}^R)=u$.
3. The [*associativity*]{}: The equality (\[asso-z-zeta\]) holds when $|z_{1}|>|z_{2}|>0$ and $|\zeta_{1}|>|\zeta_{2}|>0$.
4. The [*single-valuedness property*]{}: $$\label{sing-val-1}
e^{2\pi i (L^L(0)-L^R(0))} = I_{F}.$$
5. The [*skew symmetry*]{}: $$\label{skew-1-1}
\mathbb{Y}^{{\mathcal{Y}}}(u; 1, 1)v = e^{L^L(-1)+L^R(-1)}
\mathbb{Y}^{{\mathcal{Y}}}(v; e^{\pi i}, e^{-\pi i})u.$$
[[*Proof.*]{}]{} If $(F, \mathbb{Y}, \rho)$ is a conformal full field algebra over $V^L\otimes V^R$, then the results in Section 1 shows that $F$ is a $V^L\otimes V^R$-module, $\mathbb{Y}_{f}(\cdot; x, x)$ is an intertwining operator of type ${{{F}\choose {FF}}}$ and the five conditions are all satisfied. We now prove the converse.
Let $F$ be a module for $V^L\otimes V^R$, ${\mathcal{Y}}$ an intertwining operator of type ${{{F}\choose {FF}}}$ and $\rho: V^L\otimes V^R \to F$ an injective linear map, satisfying the five conditions above. We take the splitting $\mathbb{Y}^{{\mathcal{Y}}}$ of ${\mathcal{Y}}$ to be the full vertex operator map. For simplicity, we shall denote $\mathbb{Y}^{{\mathcal{Y}}}$ simply by $\mathbb{Y}$. We now want to construct the maps $m_{n}$ for $n\in {\mathbb{N}}$ and to verify the convergence property.
Using (\[ch-ach-z-zeta\]) and the convergence property of the intertwining operators for the vertex operator algebras $V^{L}$ and $V^{R}$, we know that for $u_{1}, \dots, u_{n}\in F$ and $w'\in F'$, $$\label{prod}
\langle w', \mathbb{Y}(u_{1}; z_{1}, \zeta_{1})\cdots
\mathbb{Y}(u_{n}; z_{n}, \zeta_{n}){\mathbf{1}}\rangle$$ is absolutely convergent when $|z_{1}|>\cdots>|z_{n}|>0$, $|\zeta_{1}|>\cdots>|\zeta_{n}|>0$, and can be analytically extended to a (possibly multivalued) analytic function of $z_{1}, \dots, z_{n}, \zeta_{1}, \dots, \zeta_{n}$ in the region given by $z_{i}\ne z_{j}$, $z_{i}\ne 0$, $\zeta_{i}\ne \zeta_{j}$, $\zeta_{i}\ne 0$. We use $$\label{correl-fn-ex}
E(m)_{n}(w', u_{1}, \dots, u_{n};
z_{1}, \zeta_{1}, \dots, z_{n}, \zeta_{n})$$ to denote this function. This is a function of $z_{1}, \zeta_{1},
\dots, z_{n}, \zeta_{n}$ where $(z_{1}, \dots, z_{n})$, $(\zeta_{1}, \dots, \zeta_{n})\in
\mathbb{F}_{n}({\mathbb{C}})$. So we can view this function as a function on $\mathbb{F}_{n}({\mathbb{C}})\times \mathbb{F}_{n}({\mathbb{C}})$. In general, this function is multivalued. Using analytic extension, a value of this function at a point $\mathbb{P}_{1}
\in \mathbb{F}_{n}({\mathbb{C}})\times \mathbb{F}_{n}({\mathbb{C}})$ and a path $\gamma$ in $\mathbb{F}_{n}({\mathbb{C}})\times \mathbb{F}_{n}({\mathbb{C}})$ from $\mathbb{P}_{1}$ to $\mathbb{P}_{2}\in
\mathbb{F}_{n}({\mathbb{C}})\times \mathbb{F}_{n}({\mathbb{C}})$, determines uniquely a value of the function at the point $\mathbb{P}_{2}$. Moreover, this value depends only on the homotopy class of the path $\gamma$. We shall call the value of the function (\[correl-fn-ex\]) at $\mathbb{P}_{2}$ obtained this way the [*value of (\[correl-fn-ex\]) at $\mathbb{P}_{2}$ obtained by analytically extending the value of (\[correl-fn-ex\]) at $\mathbb{P}_{1}$ along $\gamma$*]{}.
We choose the correlation function $$\label{correl-fn}
\langle w', m_{n}(u_{1}, \dots, u_{n};
z_{1}, \bar{z}_{1}, \dots, z_{n}, \bar{z}_{n})\rangle$$ as follows: For $z_{1}=n, \dots, z_{n}=1$, we define (\[correl-fn\]) to be (\[prod\]) with $z_{1}=\zeta_{1}=n, \dots,
z_{n}=\zeta_{n}=1$. For general $(z_{1}, \dots, z_{n})\in \mathbb{F}_{n}({\mathbb{C}})$, we choose a path $\gamma$ from $(n, \dots, 1)$ to $(z_{1}, \dots, z_{n})$. Then we have a path $\gamma\times \overline{\gamma}$ from $$((n, \dots, 1), (n, \dots, 1))\in
\mathbb{F}_{n}({\mathbb{C}})\times \mathbb{F}_{n}({\mathbb{C}})$$ to $$((z_{1}, \dots, z_{n}), (\bar{z}_{1}, \dots, \bar{z}_{n}))\in
\mathbb{F}_{n}({\mathbb{C}})\times \mathbb{F}_{n}({\mathbb{C}}).$$ We define (\[correl-fn\]) at to be the value of (\[correl-fn-ex\]) obtained by analytically extending the value of (\[correl-fn-ex\]) at $((n, \dots, 1), (n, \dots, 1))$ along $\gamma\times \overline{\gamma}$.
The first thing we have to prove is that the correlation function we just defined is indeed independent of the path $\gamma$. To prove this fact, we need only prove that if $\gamma$ is a loop in $\mathbb{F}_{n}({\mathbb{C}})$ based at $(n, \dots, 1)$ , then the value of (\[correl-fn-ex\]) at $((n, \dots, 1), (n, \dots, 1))$ obtained by analytically extending the value (\[correl-fn\]) of (\[correl-fn-ex\]) at $((n, \dots, 1), (n, \dots, 1))$ along the loop $\gamma\times \overline{\gamma}$ is equal to the original value (\[correl-fn\]) of (\[correl-fn-ex\]) at $((n, \dots, 1), (n, \dots, 1))$. In other words, we need only prove that the [*monodromy along the path $\gamma\times \overline{\gamma}$ is trivial*]{}. Note that the group of the homotopy classes of based loops in $\mathbb{F}_n({\mathbb{C}})$ , that is, the fundamental group of $\mathbb{F}_{n}({\mathbb{C}})$, is the pure braid group of $n$ strands (see [@Bi]). This group is generated by the homotopy classes of the loops given by fixing $z_{1}, \dots, z_{j-1}$, $z_{j+1}, \dots, z_{n}$ to be $n, \dots, n-(j-2)$, $n-j, \dots, 1$, respectively, and moving $z_{j}$ starting from $z_{j}=n-(j-1)$ around $z_{i}=n-(i-1)$ once (but not around other points above) in the counter clockwise direction, for $i\ne j$, $i, j=1, \dots, n$. Hence we need only prove that the monodromy along the path $\gamma\times \overline{\gamma}$ is trivial for (the homotopy class of) such a loop $\gamma$.
We now prove that the monodromy along the path $\gamma\times \bar{\gamma}$ is trivial for (the homotopy classes of) such a loop $\gamma$. Let $r$ be a positive real number satisfying $n-(i-1)>r>n-i$. Note that $r$ satisfies $r>n-(i-1)-r>0$. We know that $$\begin{aligned}
\label{prod-1}
\lefteqn{\langle w', \mathbb{Y}(u_{1}; n, n)\cdots
\mathbb{Y}(u_{i}; n-(i-1), n-(i-1))\mathbb{Y}(u_{j}; r, r)\cdot}{\nonumber \\}&&\quad\quad\cdot
\mathbb{Y}(u_{i+1}; n-i, n-i)\cdots
\mathbb{Y}(u_{i}; n-(j-2), n-(j-2))\cdot{\nonumber \\}&&\quad\quad\quad\quad\cdot
\mathbb{Y}(u_{i+1}; n-j, n-j))\cdots
\mathbb{Y}(u_{n}; 1, 1){\mathbf{1}}\rangle\end{aligned}$$ can be obtained by analytically extending the value $$\langle w', \mathbb{Y}(u_{1}; n, n)\cdots
\mathbb{Y}(u_{n}; 1, 1){\mathbf{1}}\rangle.$$along a path from $((n, \dots, 1), (n, \dots, 1))$ to $$\label{point}
((n, \dots, n-(j-2), r, n-j, \dots,
1), (n, \dots, n-(j-2), r, n-j, \dots,
1)).$$ Such a path can always be taken to be of the form $\gamma_{0}\times \bar{\gamma}_{0}$ where $\gamma_{0}$ is a path in $\mathbb{F}_{n}({\mathbb{C}})$ from $(n, \dots, 1)$ to $(n, \dots, n-(j-2), r,
n-j, \dots, 1)$. This path $\gamma_{0}$ induces an isomorphism from the fundamental group of $\mathbb{F}_{n}({\mathbb{C}})$ based at $(n, \dots, 1)$ to that based at $(n, \dots, n-(j-2), r,
n-j, \dots, 1)$. It is clear that the monodromy along a loop based at $(n, \dots, 1)$ is trivial if and only if the monodromy along the corresponding loop based at $(n, \dots, n-(j-2), r,
n-j, \dots, 1)$ is trivial. So we need only prove that the monodromy along a loop of the form $\gamma\times \bar{\gamma}$ is trivial where $\gamma$ is a loop based at $(n, \dots, n-(j-2), r,
n-j, \dots, 1)$ given by fixing $z_{1}, \dots, z_{j-1}$, $z_{j+1}, \dots, z_{n}$ to be $n, \dots, n-(j-2)$, $n-j, \dots, 1$, respectively, and moving $z_{j}$ starting from $z_{j}=r$ around $z_{i}=n-(i-1)$ once (but not around other points above) in the counter clockwise direction. By the definition of $\gamma_{0}$, the value of (\[correl-fn-ex\]) at the point (\[point\]) obtained by analytically extending the value (\[correl-fn\]) of (\[correl-fn-ex\]) at the point $((n, \dots, 1), (n, \dots, 1))$ along $\gamma_{0}$ is (\[prod-1\]). Since we also have $r>n-(i-1)-r>0$, by associativity, (\[prod-1\]) is equal to $$\begin{aligned}
\lefteqn{\langle w', \mathbb{Y}(u_{1}; n, n)\cdots
\mathbb{Y}(u_{i-1}; n-(i-2), n-(i-2))\cdot}{\nonumber \\}&&\quad\cdot
\mathbb{Y}(\mathbb{Y}(u_{i}; n-(i-1)-r, n-(i-1)-r)u_{j};
r, r)\cdot{\nonumber \\}&&\quad\quad\quad\cdot
\mathbb{Y}(u_{i+1}; n-i, n-i))\cdots
\mathbb{Y}(u_{i}; n-(j-2), n-(j-2))\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad \cdot
\mathbb{Y}(u_{i+1}; n-j, n-j))\cdots
\mathbb{Y}(u_{n}; r_{n}, r_{n}){\mathbf{1}}\rangle.\end{aligned}$$ Now let $\gamma: [0, 1]\to \mathbb{F}_{n}({\mathbb{C}})$ be the loop given by $$t\mapsto (n, \dots, n-(i-2),
r+e^{2\pi i t}(n-(i-1)-r), n-i, \dots,
n-(j-2), r, n-j, \dots, 1).$$Then the value of (\[correl-fn-ex\]) at (\[point\]) obtained by analytically extending the original value (\[prod-1\]) of (\[correl-fn-ex\]) at the point (\[point\]) along $\gamma$ is $$\begin{aligned}
\label{iter-2}
\lefteqn{\langle w', \mathbb{Y}(u_{1}; n, n)\cdots
\mathbb{Y}(u_{i-1}; n-(i-2), n-(i-2))\cdot}{\nonumber \\}&&\quad\cdot
\mathbb{Y}(\mathbb{Y}(u_{i}; e^{2\pi i}(n-(i-1)-r),
e^{-2\pi i}(n-(i-1)-r))u_{j}; r, r)\cdot{\nonumber \\}&&\quad\quad\quad\cdot \mathbb{Y}(u_{i+1}; n-i, n-i)
\cdots
\mathbb{Y}(u_{i}; n-(j-2), n-(j-2))\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\cdot
\mathbb{Y}(u_{i+1}; n-j, n-j))\cdots
\mathbb{Y}(u_{n}; r_{n}, r_{n}){\mathbf{1}}\rangle.\end{aligned}$$ But by the $L^{L}(0)$- and $L^{R}(0)$-conjugation properties and the siungle-valuedness property, we have $$\begin{aligned}
\label{single-v}
\lefteqn{\mathbb{Y}(u_{i}; e^{2\pi i}(n-(i-1)-r),
e^{-2\pi i}(n-(i-1)-r))}{\nonumber \\}&&=e^{2\pi i(L^{L}(0)-L^{R}(0))}
\mathbb{Y}(e^{-2\pi i(L^{L}(0)-L^{R}(0))}u_{i}; n-(i-1)-r,
n-(i-1)-r)\cdot{\nonumber \\}&&\quad\quad\cdot e^{-2\pi i(L^{L}(0)-L^{R}(0))}{\nonumber \\}&&=\mathbb{Y}(u_{i}; n-(i-1)-r,
n-(i-1)-r).\end{aligned}$$ Using (\[single-v\]) and the associativity again, we see that (\[iter-2\]) is equal to (\[prod-1\]). Thus the analytic extension along this loop indeed gives trivial monodromy.
Now the correlation functions and thus the maps $m_{n}$ for $n\in {\mathbb{N}}$ are defined. The only remaining thing to be shown is the convergence property. We need to show that for any $k\in {\mathbb{Z}}_+$, $l_1, \dots, l_k\in {\mathbb{Z}}_+$, $(z_1, \dots, z_k)\in \mathbb{F}_{n}({\mathbb{C}})$, $(z_1^{(i)}, \dots, z_{l_{i}}^{(i)}) \in
\mathbb{F}_{l_{i}}({\mathbb{C}})$, $i=1, \dots, k$, the series (\[series-case-2-0\]) converges absolutely to (\[sum\]) when $|z_p^{(i)}| + |z_q^{(j)}|< |z^{(0)}_i-z^{(0)}_j|$ for $i\ne j$, $i, j=1, \dots, k$, $p=1, \dots, l_i$ and $q=1, \dots, l_j$.
We use induction on $k$. We first prove the special case in which $k=2$ and $z_{2}^{(0)}=\bar{z}_{2}^{(0)}=0$. The case $k=1$ is in fact a special case. By the definition of $\mathbb{Y}$, (\[series-case-2-0\]) becomes $$\begin{aligned}
\label{conv-pf-0}
\lefteqn{\sum_{p_1,q_1, p_{2}, q_{2}}
\mathbb{Y}(P_{p_1,q_1}m_{l_1}(u_1^{(1)}, \dots, u_{l_1}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)},
\dots, z_{l_1}^{(1)}, \bar{z}_{l_1}^{(1)});
z_1^{(0)}, \bar{z}_1^{(0)})\cdot}{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\cdot
P_{p_2,q_2}m_{l_2}(u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)})).\end{aligned}$$ We use induction on $l_{1}$. When $l_{1}=1$, (\[conv-pf-0\]) becomes $$\begin{aligned}
\label{conv-pf-0.1}
\lefteqn{\sum_{p_1,q_1, p_{2}, q_{2}}
\mathbb{Y}(P_{p_1,q_1}\mathbb{Y}(u_1^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)}){\mathbf{1}};
z_1^{(0)}, \bar{z}_1^{(0)})\cdot}{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\cdot
P_{p_2,q_2}m_{l_2}(u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)})).\end{aligned}$$ Using the construction of $\mathbb{Y}$ in terms of intertwining operators, the properties of intertwining operators and noticing that our condition $|z_{1}^{(1)}|+|z_{j}^{(2)}|<|z_{1}^{(0)}|$ implies $|z_{1}^{(1)}|<|z_{1}^{(0)}|$ and $|z_{1}^{(1)}+z_{1}^{(0)}|>|z_{j}^{(2)}|$, we know that $$\begin{aligned}
\label{conv-pf-0.2}
\lefteqn{\sum_{p_1,q_1}
\mathbb{Y}(P_{p_1,q_1}\mathbb{Y}(u_1^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)}){\mathbf{1}};
z_1^{(0)}, \bar{z}_1^{(0)})\cdot}{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\cdot
P_{p_2,q_2}m_{l_2}(u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)}))\end{aligned}$$ is absolutely convergent to $$\begin{aligned}
\label{conv-pf-0.3}
\lefteqn{\mathbb{Y}(u_1^{(1)};
z_1^{(1)}+z_1^{(0)}, \bar{z}_1^{(1)}+\bar{z}_1^{(0)})
\mathbb{Y}({\mathbf{1}};
z_1^{(0)}, \bar{z}_1^{(0)})\cdot}{\nonumber \\}&&\quad\quad\quad\quad\quad\cdot
P_{p_2,q_2}m_{l_2}(u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)})){\nonumber \\}&&=\mathbb{Y}(u_1^{(1)};
z_1^{(1)}+z_1^{(0)}, \bar{z}_1^{(1)}+\bar{z}_1^{(0)})
\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\cdot
P_{p_2,q_2}m_{l_2}(u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)})).\end{aligned}$$ Then by the construction of the correlation function maps and, in particular, by the fact that the correlation functions are values of multivalued analytic functions at certain particular points, we know that the right-hand side of (\[conv-pf-0.1\]) is absolutely convergent to $$m_{1+l_{2}}(u_1^{(1)}, u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(1)}+z_1^{(0)}, \bar{z}_1^{(1)}+\bar{z}_1^{(0)},
z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)}).$$ Here we have used the fact that if a certain iterated sum of a series in powers of these complex variables is convergent to an analytic functions in the region above, then the multisum must also be absolutely convergent.
Now we assume that for $l_{1}<l$, the conclusion holds. We want to prove the conclusion for the case $l_{1}=l$. We first assume that $|z_{1}^{(1)}|, \dots, |z_{l}^{(1)}|$ are all different from each other. Then in particular there exists $t$ such that $|z_{t}^{(1)}|>|z_{1}^{(1)}|, \dots, \widehat{|z_{t}^{(1)}|},
\dots, |z_{l}^{(1)}|$, where and also below we use $\hat{~}$ to denote that the item under $\hat{~}$ is missing. Then (\[conv-pf-0\]) in this case is equal to $$\begin{aligned}
\label{conv-pf-0.4}
\lefteqn{\sum_{p_1,q_1, p_{2}, q_{2}}\sum_{r, s}
\mathbb{Y}(P_{p_1,q_1}\mathbb{Y}(u_t^{(1)};
z_t^{(1)}, \bar{z}_t^{(1)})\cdot}{\nonumber \\}&&\cdot
P_{r, s}
m_{l-1}(u_1^{(1)}, \dots, \widehat{u_{t}^{(1)}}, \dots,
u_{l}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)}, \dots, \widehat{z_{t}^{(1)}},
\widehat{\bar{z}_{t}^{(1)}},
\dots, z_{l}^{(1)}, \bar{z}_{l}^{(1)}); {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
z_1^{(0)}, \bar{z}_1^{(0)})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\cdot
P_{p_2,q_2}m_{l_2}(u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)})).\end{aligned}$$ Using the construction of $\mathbb{Y}$ in terms of intertwining operators and the properties of intertwining operators, and noticing that our condition $|z_{t}^{(1)}|+|z_{j}^{(2)}|<|z_{1}^{(0)}|$ implies $|z_{t}^{(1)}|<|z_{1}^{(0)}|$ and $|z_{t}^{(1)}+z_{1}^{(0)}|>|z_{j}^{(2)}|$, we know that $$\begin{aligned}
\label{conv-pf-0.5}
\lefteqn{\sum_{p_1,q_1}
\mathbb{Y}(P_{p_1,q_1}\mathbb{Y}(u_t^{(1)};
z_t^{(1)}, \bar{z}_t^{(1)})\cdot}{\nonumber \\}&&\cdot
P_{r, s}
m_{l-1}(u_1^{(1)}, \dots, \widehat{u_{t}^{(1)}}, \dots,
u_{l}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)}, \dots, \widehat{z_{t}^{(1)}},
\widehat{\bar{z}_{t}^{(1)}},
\dots, z_{l}^{(1)}, \bar{z}_{l}^{(1)}); {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
z_1^{(0)}, \bar{z}_1^{(0)})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\cdot
P_{p_2,q_2}m_{l_2}(u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)}))\end{aligned}$$ is convergent absolutely and, when $|z_t^{(1)}+z_{1}^{(0)}|>|z_{1}^{(0)}|$, it is absolutely convergent to $$\begin{aligned}
\label{conv-pf-0.6}
\lefteqn{\mathbb{Y}(u_t^{(1)};
z_t^{(1)}+z_{1}^{(0)}, \bar{z}_t^{(1)}+\bar{z}_{1}^{(0)})\cdot}{\nonumber \\}&&\cdot \mathbb{Y}(P_{r, s}
m_{l-1}(u_1^{(1)}, \dots, \widehat{u_{t}^{(1)}}, \dots,
u_{l}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)}, \dots, \widehat{z_{t}^{(1)}},
\widehat{\bar{z}_{t}^{(1)}},
\dots, z_{l}^{(1)}, \bar{z}_{l}^{(1)}); {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
z_1^{(0)}, \bar{z}_1^{(0)})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\cdot
P_{p_2,q_2}m_{l_2}(u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)})).\end{aligned}$$ By the induction assumption, $$\begin{aligned}
\label{conv-pf-0.7}
\lefteqn{\sum_{r, s, p_{2}, q_{2}}\mathbb{Y}(u_t^{(1)};
z_t^{(1)}+z_{1}^{(0)}, \bar{z}_t^{(1)}+\bar{z}_{1}^{(0)})\cdot}{\nonumber \\}&&\cdot \mathbb{Y}(P_{r, s}
m_{l-1}(u_1^{(1)}, \dots, \widehat{u_{t}^{(1)}}, \dots,
u_{l}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)}, \dots, \widehat{z_{t}^{(1)}},
\widehat{\bar{z}_{t}^{(1)}},
\dots, z_{l}^{(1)}, \bar{z}_{l}^{(1)}); {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
z_1^{(0)}, \bar{z}_1^{(0)})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\cdot
P_{p_2,q_2}m_{l_2}(u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)}))\end{aligned}$$ is absolutely convergent to $$\begin{aligned}
\label{conv-pf-0.8}
\lefteqn{\mathbb{Y}(u_t^{(1)};
z_t^{(1)}+z_{1}^{(0)}, \bar{z}_t^{(1)}+\bar{z}_{1}^{(0)})\cdot}{\nonumber \\}&&\cdot
m_{l+l_{2}-1}(u_1^{(1)}, \dots, \widehat{u_{t}^{(1)}}, \dots,
u_{l}^{(1)}, u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(1)}+z_1^{(0)}, \bar{z}_1^{(1)}+\bar{z}_1^{(0)}, \dots, {\nonumber \\}&&\quad
\widehat{z_{t}^{(1)}+z_1^{(0)}},
\widehat{\bar{z}_{t}^{(1)}+\bar{z}_1^{(0)}},
\dots, z_{l}^{(1)}+z_1^{(0)}, \bar{z}_{l}^{(1)}+\bar{z}_1^{(0)}z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)}){\nonumber \\}&&=m_{l+l_{2}}(u_1^{(1)}, \dots,
u_{l}^{(1)}, u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(1)}+z_1^{(0)}, \bar{z}_1^{(1)}+\bar{z}_1^{(0)}, \dots, {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
z_{l}^{(1)}+z_1^{(0)}, \bar{z}_{l}^{(1)}+\bar{z}_1^{(0)}z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)}).\end{aligned}$$ We know that the right-hand side of (\[conv-pf-0.7\]) is a value of the multivalued analytic function $$\begin{aligned}
\lefteqn{E(m)_{l+l_{2}}(u_1^{(1)}, \dots,
u_{l}^{(1)}, u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(1)}+z_1^{(0)}, \zeta_1^{(1)}+\zeta_1^{(0)}, \dots,} {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
z_{l}^{(1)}+z_1^{(0)}, \zeta_{l}^{(1)}+\zeta_1^{(0)}z_1^{(2)}, \zeta_1^{(2)},
\dots, z_{l_2}^{(2)}, \zeta_{l_2}^{(2)})\end{aligned}$$ at the points satisfying $\zeta_1^{(0)}=\bar{z}_1^{(0)}$, $\zeta_{p}^{(i)}=\bar{z}_{p}^{(i)}$ for $p=1, \dots, l_{i}$, $i=1, 2$. Since both the sum of (\[conv-pf-0.7\]) and the right-hand side of (\[conv-pf-0.8\]) are values of multivalued analytic functions in the same region and we have proved that their values are equal when $|z_t^{(1)}+z_{1}^{(0)}|>|z_{1}^{(0)}|$, (\[conv-pf-0.7\]) must be convergent absolutely to the right-hand side of (\[conv-pf-0.8\]) even when $|z_t^{(1)}+z_{1}^{(0)}|>|z_{1}^{(0)}|$ is not satisfied. By the properties of analytic functions, we know that (\[conv-pf-0.4\]) as a sum in a different order is also convergent absolutely to the right-hand side of (\[conv-pf-0.8\]).
Now we discuss the case that some of $|z_{1}^{(1)}|, \dots, |z_{l}^{(1)}|$ are equal. Let $N(z_{1}^{(1)}, \dots, z_{l}^{(1)})$ be the subset of $\{z_{1}^{(1)}, \dots, z_{l}^{(1)}\}$ consisting of those elements whose absolute values are equal to the absolute values of some other elements of $\{z_{1}^{(1)}, \dots, z_{l}^{(1)}\}$. We use induction on the number of elements of $N(z_{1}^{(1)}, \dots, z_{l}^{(1)})$. When the number is $0$, this is the case discussed above. Now assume that when the number is equal to $n$, the conclusion holds. When this number is equal to $n+1$, let $\epsilon$ be a complex number such that the number of elements of $N(z_{1}^{(1)}+\epsilon, \dots, z_{l}^{(1)}+\epsilon)$ is $n$ and $|z_{p}^{(1)}+\epsilon|+|z_{q}^{(2)}+\epsilon|<|z_1^{(0)}|$ for $p=1, \dots, l$ and $q=1, \dots, l_{2}$. Note that we can always find such an $\epsilon$ and we can take such an $\epsilon$ with $|\epsilon|$ to be arbitrarily small.
By induction assumption, $$\begin{aligned}
\label{conv-pf-0.9}
\lefteqn{\sum_{p_1,q_1, p_{2}, q_{2}}
\mathbb{Y}(P_{p_1,q_1}m_{l}(u_1^{(1)}, \dots, u_{l}^{(1)};
z_1^{(1)}+\epsilon, \bar{z}_1^{(1)}+\bar{\epsilon},
\dots, z_{l}^{(1)}+\epsilon, \bar{z}_{l}^{(1)}+\bar{\epsilon});} {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
z_1^{(0)}, \bar{z}_1^{(0)})\cdot{\nonumber \\}&&\quad\quad\quad\cdot
P_{p_2,q_2}m_{l_2}(u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(2)}+\epsilon, \bar{z}_1^{(2)}+\bar{\epsilon},
\dots, z_{l_2}^{(2)}+\epsilon, \bar{z}_{l_2}^{(2)}+\bar{\epsilon}){\nonumber \\}&&\end{aligned}$$ is absolutely convergent to $$\begin{aligned}
\label{conv-pf-0.10}
\lefteqn{m_{l+l_{2}}(u_1^{(1)}, \dots,
u_{l}^{(1)}, u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(1)}+z_1^{(0)}+\epsilon,
\bar{z}_1^{(1)}+\bar{z}_1^{(0)}+\bar{\epsilon}, \dots,} {\nonumber \\}&&\quad\quad
z_{l}^{(1)}+z_1^{(0)}+\epsilon,
\bar{z}_{l}^{(1)}+\bar{z}_1^{(0)}+\bar{\epsilon}z_1^{(2)}+\epsilon, \bar{z}_1^{(2)}+\bar{\epsilon},
\dots, z_{l_2}^{(2)}+\epsilon, \bar{z}_{l_2}^{(2)}+\bar{\epsilon}).{\nonumber \\}&&\end{aligned}$$ We have $$\begin{aligned}
\label{conv-pf-0.115}
\lefteqn{\sum_{r_1,s_1, r_{2}, s_{2}}\sum_{p_1,q_1, p_{2}, q_{2}}
\langle e^{-\epsilon L^{L}(1)-\bar{\epsilon}L^{R}(1)}u',
\mathbb{Y}(P_{r_{1}, s_{1}}e^{\epsilon L^{L}(-1)+\bar{\epsilon}L^{R}(-1)}
\cdot}{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\cdot
P_{p_1,q_1}m_{l}(u_1^{(1)}, \dots, u_{l}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)},
\dots, z_{l}^{(1)}, \bar{z}_{l}^{(1)});
z_1^{(0)}, \bar{z}_1^{(0)})\cdot{\nonumber \\}&&\quad\quad\;\;\cdot
P_{r_{1}, s_{1}}e^{\epsilon L^{L}(-1)+\bar{\epsilon}L^{R}(-1)}
P_{p_2,q_2}m_{l_2}(u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)})\rangle{\nonumber \\}&&=\sum_{r_1,s_1, r_{2}, s_{2}}
\langle e^{-\epsilon L^{L}(1)-\bar{\epsilon}L^{R}(1)}u', {\nonumber \\}&&\quad\quad\quad
\mathbb{Y}(P_{r_{1}, s_{1}}m_{l}(u_1^{(1)}, \dots, u_{l}^{(1)};
z_1^{(1)}+\epsilon, \bar{z}_1^{(1)}+\bar{\epsilon},
\dots, z_{l}^{(1)}+\epsilon, \bar{z}_{l}^{(1)}+\bar{\epsilon}); {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
z_1^{(0)}, \bar{z}_1^{(0)})\cdot{\nonumber \\}&&\quad\quad\quad\quad\cdot
P_{r_{1}, s_{1}}m_{l_2}(u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(2)}+\epsilon, \bar{z}_1^{(2)}+\bar{\epsilon},
\dots, z_{l_2}^{(2)}+\epsilon, \bar{z}_{l_2}^{(2)}+\bar{\epsilon})\rangle.{\nonumber \\}&&\end{aligned}$$ So the right-hand side and thus also the left-hand side of (\[conv-pf-0.115\]) is absolutely convergent to $$\begin{aligned}
\label{conv-pf-0.12}
\lefteqn{\langle e^{-\epsilon L^{L}(1)-\bar{\epsilon}L^{R}(1)}u',
m_{l+l_{2}}(u_1^{(1)}, \dots,
u_{l}^{(1)}, u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(1)}+z_1^{(0)}+\epsilon, } {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad
\bar{z}_1^{(1)}+\bar{z}_1^{(0)}+\bar{\epsilon}, \dots,
z_{l}^{(1)}+z_1^{(0)}+\epsilon,
\bar{z}_{l}^{(1)}+\bar{z}_1^{(0)}+\bar{\epsilon}, {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
z_1^{(2)}+\epsilon, \bar{z}_1^{(2)}+\bar{\epsilon},
\dots, z_{l_2}^{(2)}+\epsilon, \bar{z}_{l_2}^{(2)}+\bar{\epsilon})\rangle{\nonumber \\}&&=\langle u', e^{-\epsilon L^{L}(-1)-\bar{\epsilon}L^{R}(-1)}
m_{l+l_{2}}(u_1^{(1)}, \dots,
u_{l}^{(1)}, u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(1)}+z_1^{(0)}+\epsilon, {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad
\bar{z}_1^{(1)}+\bar{z}_1^{(0)}+\bar{\epsilon}, \dots,
z_{l}^{(1)}+z_1^{(0)}+\epsilon,
\bar{z}_{l}^{(1)}+\bar{z}_1^{(0)}+\bar{\epsilon}, {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
z_1^{(2)}+\epsilon, \bar{z}_1^{(2)}+\bar{\epsilon},
\dots, z_{l_2}^{(2)}+\epsilon, \bar{z}_{l_2}^{(2)}+\bar{\epsilon})\rangle{\nonumber \\}&&=\langle u',
m_{l+l_{2}}(u_1^{(1)}, \dots,
u_{l}^{(1)}, u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(1)}+z_1^{(0)},
\bar{z}_1^{(1)}+\bar{z}_1^{(0)}, \dots, {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
z_{l}^{(1)}+z_1^{(0)},
\bar{z}_{l}^{(1)}+\bar{z}_1^{(0)}z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)})\rangle.{\nonumber \\}\end{aligned}$$ Since the left-hand side of (\[conv-pf-0.12\]) is a value of a multivalued analytic function, any of its expansion must be absolutely convergent. In particular, the left-hand side of (\[conv-pf-0.115\]) as an expansion of the left-hand side of (\[conv-pf-0.12\]) is absolutely convergent. Thus we can exchange the order of the two summation signs such that the resulting series is still absolutely convergent to the left-hand side of (\[conv-pf-0.12\]) and thus to the right-hand side of (\[conv-pf-0.12\]).
But for $u'\in F'$, $$\begin{aligned}
\label{conv-pf-0.11}
\lefteqn{\sum_{p_1,q_1, p_{2}, q_{2}}
\langle u', \mathbb{Y}(P_{p_1,q_1}m_{l}(u_1^{(1)}, \dots, u_{l}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)},
\dots, z_{l}^{(1)}, \bar{z}_{l}^{(1)}); } {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
z_1^{(0)}, \bar{z}_1^{(0)})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\cdot
P_{p_2,q_2}m_{l_2}(u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)})\rangle{\nonumber \\}&&=\sum_{p_1,q_1, p_{2}, q_{2}}
\langle e^{-\epsilon L^{L}(1)-\bar{\epsilon}L^{R}(1)}u',
e^{\epsilon L^{L}(-1)+\bar{\epsilon}L^{R}(-1)}\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\cdot
\mathbb{Y}(P_{p_1,q_1}m_{l}(u_1^{(1)}, \dots, u_{l}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)},
\dots, z_{l}^{(1)}, \bar{z}_{l}^{(1)});
z_1^{(0)}, \bar{z}_1^{(0)})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\cdot
P_{p_2,q_2}m_{l_2}(u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)})\rangle{\nonumber \\}&&=\sum_{p_1,q_1, p_{2}, q_{2}}
\langle e^{-\epsilon L^{L}(1)-\bar{\epsilon}L^{R}(1)}u',
\mathbb{Y}(e^{\epsilon L^{L}(-1)+\bar{\epsilon}L^{R}(-1)}\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\cdot
P_{p_1,q_1}m_{l}(u_1^{(1)}, \dots, u_{l}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)},
\dots, z_{l}^{(1)}, \bar{z}_{l}^{(1)});
z_1^{(0)}, \bar{z}_1^{(0)})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\cdot
e^{\epsilon L^{L}(-1)+\bar{\epsilon}L^{R}(-1)}
P_{p_2,q_2}m_{l_2}(u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)})\rangle{\nonumber \\}&&=\sum_{p_1,q_1, p_{2}, q_{2}}\sum_{r_1,s_1, r_{2}, s_{2}}
\langle e^{-\epsilon L^{L}(1)-\bar{\epsilon}L^{R}(1)}u',
\mathbb{Y}(P_{r_{1}, s_{1}}e^{\epsilon L^{L}(-1)+\bar{\epsilon}L^{R}(-1)}
\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\cdot
P_{p_1,q_1}m_{l}(u_1^{(1)}, \dots, u_{l}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)},
\dots, z_{l}^{(1)}, \bar{z}_{l}^{(1)});
z_1^{(0)}, \bar{z}_1^{(0)})\cdot{\nonumber \\}&&\quad\quad\;\;\cdot
P_{r_{1}, s_{1}}e^{\epsilon L^{L}(-1)+\bar{\epsilon}L^{R}(-1)}
P_{p_2,q_2}m_{l_2}(u_1^{(2)}, \dots, u_{l_2}^{(2)};
z_1^{(2)}, \bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)})\rangle.{\nonumber \\}&&\end{aligned}$$ We have shown that the right-hand side of (\[conv-pf-0.11\]) is absolute convergent to the right-hand side of (\[conv-pf-0.12\]). Thus the left-hand side of (\[conv-pf-0.11\]) is also absolute convergent to the right-hand side of (\[conv-pf-0.12\]). So we have proved the convergence property when the number of elements of $N(z_{1}^{(1)}, \dots, z_{l}^{(1)})$ is $n+1$. Thus the convergence property is proved when some of $|z_{1}^{(1)}|, \dots, |z_{l}^{(1)}|$ are equal.
By induction principle, we have proved the convergence property in this special case.
We now assume that when $k<K$, (\[series-case-2-0\]) converges absolutely to (\[sum\]) when $z_{p}^{(0)}\ne
z_{q}^{(0)}$ for $p, q=1, \dots, K$, $z_{p}^{(i)}\ne
z_{q}^{(i)}$ for $p, q=1, \dots, l_{i}$ and $i=1, \dots, K$ $1\leq p, q\leq l_i$, and $|z_p^{(i)}| + |z_q^{(j)}|< |z^{(0)}_i-z^{(0)}_j|$ for $p=1, \dots, l_i$, $q=1, \dots, l_j$, $i, j=1, \dots, K$, $i\ne j$. Now we consider the case $k=K$. We first consider the case that $z_p^{(i)}\in {\mathbb{R}}_{+}\cup \{0\}$ for $p=1, \dots, l_{i}$ and $i=0, \dots, K$ and $z_{1}^{(0)}>\cdots >z_{K}^{(0)}$. By the definition of the correlation function maps, we know that (\[series-case-2-0\]) in this case is equal to $$\begin{aligned}
\label{conv-pf-1}
\lefteqn{\sum_{p_1,q_1, \dots, p_K, q_K}\sum_{r, s}
\mathbb{Y}(P_{p_1,q_1}m_{l_1}(u_1^{(1)}, \dots, u_{l_1}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)},
\dots, z_{l_1}^{(1)}, \bar{z}_{l_1}^{(1)});
z_1^{(0)}, \bar{z}_1^{(0)})\cdot} {\nonumber \\}&&\cdot
P_{r, s}m_{K-1}(P_{p_{2}, q_{2}}m_{l_{2}}(u_1^{(2)},\dots, u_{l_2}^{(2)};
z_1^{(2)},
\bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)}), \dots, {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad
P_{p_K,q_K}m_{l_K}(u_1^{(K)},\dots, u_{l_K}^{(K)}; z_1^{(K)},
\bar{z}_1^{(K)},
\dots, z_{l_K}^{(K)}, \bar{z}_{l_K}^{(K)}); {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad \quad\quad\quad
z_2^{(0)}, \bar{z}_2^{(0)},
\dots,
z_K^{(0)}, \bar{z}_K^{(0)}).{\nonumber \\}&&\end{aligned}$$ Using the induction assumption, we have $$\begin{aligned}
\label{conv-pf-2}
\lefteqn{\sum_{r, s, p_1,q_1}\sum_{p_{2}, q_{2}, \dots, p_K, q_K}
\mathbb{Y}(P_{p_1,q_1}m_{l_1}(u_1^{(1)}, \dots, u_{l_1}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)},
\dots, z_{l_1}^{(1)}, \bar{z}_{l_1}^{(1)});
z_1^{(0)}, \bar{z}_1^{(0)})\cdot} {\nonumber \\}&&\cdot
P_{r, s}m_{K-1}(P_{p_{2}, q_{2}}m_{l_{2}}(u_1^{(2)},\dots, u_{l_2}^{(2)};
z_1^{(2)},
\bar{z}_1^{(2)},
\dots, z_{l_2}^{(2)}, \bar{z}_{l_2}^{(2)}), \dots, {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad
P_{p_K,q_K}m_{l_K}(u_1^{(K)},\dots, u_{l_K}^{(K)}; z_1^{(K)},
\bar{z}_1^{(K)},
\dots, z_{l_K}^{(K)}, \bar{z}_{l_K}^{(K)}); {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad \quad\quad\quad
z_2^{(0)}, \bar{z}_2^{(0)},
\dots,
z_K^{(0)}, \bar{z}_K^{(0)}){\nonumber \\}&&=\sum_{r, s, p_1,q_1}
\mathbb{Y}(P_{p_1,q_1}m_{l_1}(u_1^{(1)}, \dots, u_{l_1}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)},
\dots, z_{l_1}^{(1)}, \bar{z}_{l_1}^{(1)});
z_1^{(0)}, \bar{z}_1^{(0)})\cdot {\nonumber \\}&&\quad\quad\cdot
P_{r, s}m_{l_{2}+\cdots+l_{K}}(u_1^{(2)},\dots, u_{l_2}^{(2)},
\dots, u_1^{(K)},\dots, u_{l_K}^{(K)};
z_1^{(2)}+z_2^{(0)}, {\nonumber \\}&&\quad \quad \quad\quad\quad\quad\quad\quad\quad
\bar{z}_1^{(2)}+\bar{z}_2^{(0)},
\dots,
z_{l_2}^{(2)}+z_2^{(0)}, \bar{z}_{l_2}^{(2)}+\bar{z}_2^{(0)}, \dots, {\nonumber \\}&&\quad \quad \quad\quad\quad\quad\quad
z_1^{(K)}+z_K^{(0)},
\bar{z}_1^{(K)}+\bar{z}_K^{(0)},
\dots, z_{l_K}^{(K)}+z_K^{(0)}, \bar{z}_{l_K}^{(K)}+\bar{z}_K^{(0)}).{\nonumber \\}&&\end{aligned}$$ Since $z_p^{(1)} + z_q^{(i)}< z^{(0)}_1-z^{(0)}_i$ for $p=1, \dots, l_{1}$, $q=1, \dots, l_{i}$ and $i=2, \dots, K$, we have $z_p^{(1)} + (z_q^{(i)}+z^{(0)}_i)< z^{(0)}_1-0$. Thus by the special case we proved above, the right-hand side of (\[conv-pf-2\]) is absolutely convergent to $$\begin{aligned}
\label{conv-pf-2.5}
\lefteqn{m_{l_{1}+\cdots +l_{K}}(u_1^{(1)},\dots, u_{l_1}^{(1)},
\dots, u_1^{(K)},\dots, u_{l_K}^{(K)};
z_1^{(1)}+z_1^{(0)}, }{\nonumber \\}&&\quad \quad \quad\quad\quad\quad
\bar{z}_1^{(1)}+\bar{z}_1^{(0)},
\dots,
z_{l_1}^{(1)}+z_1^{(0)}, \bar{z}_{l_1}^{(1)}+\bar{z}_1^{(0)}, \dots,
z_1^{(K)}+z_K^{(0)}, {\nonumber \\}&&\quad \quad \quad\quad\quad\quad\quad\quad
\bar{z}_1^{(K)}+\bar{z}_K^{(0)},
\dots, z_{l_K}^{(K)}+z_K^{(0)}, \bar{z}_{l_K}^{(K)}+\bar{z}_K^{(0)}).\end{aligned}$$
Note that (\[conv-pf-2.5\]) is a value of the $\overline{F}$-valued multivalued analytic function $$\begin{aligned}
\label{conv-pf-3}
\lefteqn{E(m)_{l_{1}+\cdots +l_{K}}(u_1^{(1)},\dots, u_{l_1}^{(1)},
\dots, u_1^{(K)},\dots, u_{l_K}^{(K)};
z_1^{(1)}+z_1^{(0)},} {\nonumber \\}&&\quad \quad \quad\quad\quad\quad
\zeta_1^{(1)}+\zeta_1^{(0)},
\dots,
z_{l_1}^{(1)}+z_1^{(0)}, \zeta_{l_1}^{(1)}+\zeta_1^{(0)}, \dots,
z_1^{(K)}+z_K^{(0)}, {\nonumber \\}&&\quad \quad \quad\quad\quad\quad\quad\quad
\zeta_1^{(K)}+\zeta_K^{(0)},
\dots, z_{l_K}^{(K)}+z_K^{(0)}, \zeta_{l_K}^{(K)}+\zeta_K^{(0)}).\end{aligned}$$ at the point $z_{i}^{(j)}=z_{i}^{(j)}$, $\zeta_{i}^{(j)}=\bar{z}_{i}^{(j)}$. Thus its expansions, no matter in which ways, must be convergent absolutely. In particular, (\[conv-pf-1\]) as one expansion of (\[conv-pf-2.5\]) must be convergent absolutely to (\[conv-pf-2.5\]), proving the convergence in this special case of the case $k=K$.
We know that for a series in powers of several variables, if it is absolutely convergent when these variables are equal to some real numbers, then it is also convergent when the variables are equal to complex numbers whose absolute values are equal to these real numbers. Using this property, we see that $$\begin{aligned}
\label{conv-pf-4}
\lefteqn{\sum_{p_1,q_1, \dots, p_K, q_K}\sum_{r, s}
\mathbb{Y}(P_{p_1,q_1}E(m)_{l_1}(u_1^{(1)}, \dots, u_{l_1}^{(1)};
z_1^{(1)}, \zeta_1^{(1)},
\dots, z_{l_1}^{(1)}, \zeta_{l_1}^{(1)});
z_1^{(0)}, \zeta_1^{(0)})\cdot} {\nonumber \\}&&\cdot
P_{r, s}E(m)_{K-1}(P_{p_{2}, q_{2}}E(m)_{l_{2}}(u_1^{(2)},\dots, u_{l_2}^{(2)};
z_1^{(2)},
\zeta_1^{(2)},
\dots, z_{l_2}^{(2)}, \zeta_{l_2}^{(2)}), \dots, {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad
P_{p_K,q_K}E(m)_{l_K}(u_1^{(K)},\dots, u_{l_K}^{(K)}; z_1^{(K)},
\zeta_1^{(K)},
\dots, z_{l_K}^{(K)}, \zeta_{l_K}^{(K)}); {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad \quad\quad\quad
z_2^{(0)}, \zeta_2^{(0)},
\dots,
z_K^{(0)}, \zeta_K^{(0)}).{\nonumber \\}&&\end{aligned}$$ is convergent absolutely to a branch of (\[conv-pf-3\]) when $z_{p}^{(0)}\ne
z_{q}^{(0)}$, $\zeta_{p}^{(0)}\ne
\zeta_{q}^{(0)}$ for $p, q=1, \dots, K$, $z_{p}^{(i)}\ne
z_{q}^{(i)}$, $\zeta_{p}^{(i)}\ne
\zeta_{q}^{(i)}$ for $p, q=1, \dots, l_{i}$, $i=1, \dots, K$ $|z_p^{(i)}| + |z_q^{(j)}|< ||z^{(0)}_i|-|z^{(0)}_j||$, $|\zeta_p^{(i)}| + |\zeta_q^{(j)}|< ||\zeta^{(0)}_i|-|\zeta^{(0)}_j||$, for $p=1, \dots, l_i$, $q=1, \dots, l_j$, $i, j=1, \dots, K$, $i\ne j$, and $|z_{1}^{(0)}|>\cdots >|z_{K}^{(0)}|$, $|\zeta_{1}^{(0)}|>\cdots >|\zeta_{K}^{(0)}|$. Using the permutation property for the correlation functions, we obtain that (\[conv-pf-4\]) is convergent absolutely to a branch of (\[conv-pf-3\]) when $|z_{\sigma(1)}^{(0)}|> \cdots> |z_{\sigma(K)}^{(0)}|$ and $|\zeta_{\sigma(1)}^{(0)}|> \cdots> |\zeta_{\sigma(K)}^{(0)}|$ for some $\sigma\in S_{K}$.
Now for fixed $z_{1}^{(0)}, \dots, z_{K}^{(0)}$, $\zeta_{1}^{(0)}, \dots, \zeta_{K}^{(0)}$ satisfying $|z_{\sigma(1)}^{(0)}|> \cdots> |z_{\sigma(K)}^{(0)}|$ and $|\zeta_{\sigma(1)}^{(0)}|> \cdots> |\zeta_{\sigma(K)}^{(0)}|$ for some $\sigma\in S_{K}$, any branch of (\[conv-pf-3\]) can be expanded as a series in powers of $z_{p}^{(i)}$ and $\zeta_{p}^{(i)}$, $p=1, \dots, l_{i}$, $i=1, \dots, K$ in the region $$\begin{aligned}
\label{conv-pf-5}
\lefteqn{\{(z_{1}^{(1)}, \dots, z_{l_{K}}^{(K)},
\zeta_{1}^{(1)}, \dots, \zeta_{l_{K}}^{(K)})\;|\;z_{p}^{(i)}\ne
z_{q}^{(i)}, \zeta_{p}^{(i)}\ne
\zeta_{q}^{(i)}}{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad \;\mbox{\rm for}\; p, q=1, \dots, l_{i},
i=1, \dots, K,{\nonumber \\}&&\quad |z_p^{(i)}| + |z_q^{(j)}|< |z^{(0)}_i-z^{(0)}_j|,
|\zeta_p^{(i)}| + |\zeta_q^{(j)}|< |\zeta^{(0)}_i-\zeta^{(0)}_j|,{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad
\;\mbox{\rm for}\;
p=1, \dots, l_i, q=1, \dots, l_j, i, j=1, \dots, K,
i\ne j\}.{\nonumber \\}&&\end{aligned}$$ But in the region $$\begin{aligned}
\label{conv-pf-6}
\lefteqn{\{(z_{1}^{(1)}, \dots, z_{l_{K}}^{(K)},
\zeta_{1}^{(1)}, \dots, \zeta_{l_{K}}^{(K)})\;|\;z_{p}^{(i)}\ne
z_{q}^{(i)}, \zeta_{p}^{(i)}\ne
\zeta_{q}^{(i)}}{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad \;\mbox{\rm for}\; p, q=1, \dots, l_{i},
i=1, \dots, K,{\nonumber \\}&&\quad |z_p^{(i)}| + |z_q^{(j)}|< ||z^{(0)}_i|-|z^{(0)}_j||,
|\zeta_p^{(i)}| + |\zeta_q^{(j)}|< ||\zeta^{(0)}_i|-|\zeta^{(0)}_j||,{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad
\;\mbox{\rm for}\;
p=1, \dots, l_i, q=1, \dots, l_j, i, j=1, \dots, K,
i\ne j\},{\nonumber \\}&&\end{aligned}$$ we have proved that one branch of (\[conv-pf-3\]) can be expanded as the series (\[conv-pf-4\]), which can be further expanded as a series in powers of $z_{p}^{(i)}$ and $\zeta_{p}^{(i)}$, $p=1, \dots, l_{i}$, $i=1, \dots, K$ in this region. Since the region (\[conv-pf-6\]) is contained in the region (\[conv-pf-5\]) and the coefficients of the expansion can be determined completely using the values of the branch in the region (\[conv-pf-6\]), we see that the restriction to the region (\[conv-pf-6\]) of the expansion in the region (\[conv-pf-5\]) is the same as the expansion in the region (\[conv-pf-5\]). Thus, the series (\[conv-pf-4\]) is convergent absolutely to a branch of (\[conv-pf-3\]) in the region (\[conv-pf-5\]).
In the region (\[conv-pf-5\]), when $\zeta_{p}^{(0)}=z_{p}^{(0)}\in {\mathbb{R}}$ for $p=1, \dots, K$, $\zeta_{p}^{(i)}=z_{p}^{(i)}\in {\mathbb{R}}$ for $p=1, \dots, l_{i}$, $i=1, \dots, K$, we have proved that (\[series-case-2-0\]) in this case is convergent absolutely to the right-hand side of (\[conv-pf-2\]). Thus in the region (\[conv-pf-5\]), (\[series-case-2-0\]) with $k=K$ is convergent absolutely to the right-hand side of (\[conv-pf-2\]), the value of a branch of (\[conv-pf-3\]).
Finally we consider the case that some of $|z_{1}^{(0)}|, \dots,
|z_{K}^{(0)}|$ are equal. Recall the subset $N(z_{1}^{(0)}, \dots, z_{K}^{(0)})$ of $\{z_{1}^{(0)}, \dots, z_{K}^{(0)}\}$ consisting of those elements whose absolute values are equal to the absolute values of some other elements of $\{z_{1}^{(0)}, \dots, z_{K}^{(0)}\}$. We use induction on the number of elements of $N(z_{1}^{(0)}, \dots, z_{K}^{(0)})$. When the number is $0$, this is the case discussed above. Now assume that when the number is equal to $n$, the conclusion holds. When this number is equal to $n+1$, let $\epsilon$ be a complex number such that the number of elements of $N(z_{1}^{(0)}+\epsilon, \dots, z_{K}^{(0)}+\epsilon)$ is $n$ and that the other conditions are still satisfied. Note that we can always find such an $\epsilon$ and we can take such an $\epsilon$ with $|\epsilon|$ to be arbitrary small. By induction assumption, $$\begin{aligned}
\lefteqn{\sum_{p_1,q_1, \dots, p_K, q_K}
m_K(P_{p_1,q_1}m_{l_1}(u_1^{(1)},\dots, u_{l_1}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)},
\dots, z_{l_1}^{(1)}, \bar{z}_{l_1}^{(1)}),
\dots,} {\nonumber \\}&&\quad\quad\quad\quad\quad\quad
P_{p_K,q_K}m_{l_K}(u_1^{(K)},\dots, u_{l_K}^{(K)}; z_1^{(K)},
\bar{z}_1^{(K)},
\dots, z_{l_K}^{(K)}, \bar{z}_{l_K}^{(K)}); {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad
z_1^{(0)}+\epsilon, \bar{z}_1^{(0)}+\bar{\epsilon}, \dots,
z_K^{(0)}+\epsilon, \bar{z}_K^{(0)}+\bar{\epsilon})\end{aligned}$$ is absolutely convergent to $$\begin{aligned}
\label{conv-pf-8}
\lefteqn{
m_{l_{1}+\cdots +l_{K}}(u_1^{(1)},\dots, u_{l_1}^{(1)},
\dots, u_1^{(K)},\dots, u_{l_K}^{(K)};
z_1^{(1)}+z_1^{(0)}+\epsilon, }{\nonumber \\}&&\quad
\bar{z}_1^{(1)}+\bar{z}_1^{(0)}+\bar{\epsilon},
\dots,
z_{l_1}^{(1)}+z_1^{(0)}+\epsilon,
\bar{z}_{l_1}^{(1)}+\bar{z}_1^{(0)}+\bar{\epsilon}, \dots,
z_1^{(K)}+z_K^{(0)}+\epsilon, {\nonumber \\}&&\quad \quad \quad\quad\quad\quad\quad\quad\quad
\bar{z}_1^{(K)}+\bar{z}_K^{(0)}+\bar{\epsilon},
\dots, z_{l_K}^{(K)}+z_K^{(0)}+\epsilon,
\bar{z}_{l_K}^{(K)}+\bar{z}_K^{(0)}+\bar{\epsilon}).\end{aligned}$$ Thus for $u'\in F$, $$\begin{aligned}
\label{conv-pf-9}
\lefteqn{\sum_{p_1,q_1, \dots, p_K, q_K}
\langle u', m_K(P_{p_1,q_1}m_{l_1}(u_1^{(1)},\dots, u_{l_1}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)},
\dots, z_{l_1}^{(1)}, \bar{z}_{l_1}^{(1)}),
\dots,} {\nonumber \\}&&\quad \quad\quad\quad\quad\quad\quad
P_{p_K,q_K}m_{l_K}(u_1^{(K)},\dots, u_{l_K}^{(K)}; z_1^{(K)},
\bar{z}_1^{(K)},
\dots, z_{l_K}^{(K)}, \bar{z}_{l_K}^{(K)}); {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad\quad\quad
z_1^{(0)}, \bar{z}_1^{(0)}, \dots,
z_K^{(0)}, \bar{z}_K^{(0)})\rangle{\nonumber \\}&&=\sum_{p_1,q_1, \dots, p_K, q_K}
\langle e^{\epsilon L^{L}(1)+\bar{\epsilon}L^{R}(1)}u',
e^{-\epsilon L^{L}(-1)-\bar{\epsilon}L^{R}(-1)}\cdot{\nonumber \\}&&\quad\quad\quad\quad\cdot
m_K(P_{p_1,q_1}m_{l_1}(u_1^{(1)},\dots, u_{l_1}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)},
\dots, z_{l_1}^{(1)}, \bar{z}_{l_1}^{(1)}),
\dots, {\nonumber \\}&&\quad \quad\quad\quad\quad\quad\quad
P_{p_K,q_K}m_{l_K}(u_1^{(K)},\dots, u_{l_K}^{(K)}; z_1^{(K)},
\bar{z}_1^{(K)},
\dots, z_{l_K}^{(K)}, \bar{z}_{l_K}^{(K)}); {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad\quad\quad\quad\quad
z_1^{(0)}, \bar{z}_1^{(0)}, \dots,
z_K^{(0)}, \bar{z}_K^{(0)})\rangle{\nonumber \\}&&=\sum_{p_1,q_1, \dots, p_K, q_K}
\langle e^{\epsilon L^{L}(1)+\bar{\epsilon}L^{R}(1)}u', {\nonumber \\}&&\quad\quad\quad\quad
m_K(P_{p_1,q_1}m_{l_1}(u_1^{(1)},\dots, u_{l_1}^{(1)};
z_1^{(1)}, \bar{z}_1^{(1)},
\dots, z_{l_1}^{(1)}, \bar{z}_{l_1}^{(1)}),
\dots, {\nonumber \\}&&\quad \quad\quad\quad\quad\quad\quad
P_{p_K,q_K}m_{l_K}(u_1^{(K)},\dots, u_{l_K}^{(K)}; z_1^{(K)},
\bar{z}_1^{(K)},
\dots, z_{l_K}^{(K)}, \bar{z}_{l_K}^{(K)}); {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
z_1^{(0)}+\epsilon, \bar{z}_1^{(0)}+\bar{\epsilon},, \dots,
z_K^{(0)}+\epsilon,, \bar{z}_K^{(0)}+\bar{\epsilon})\rangle.{\nonumber \\}&&\end{aligned}$$ is absolutely convergent to $$\begin{aligned}
\label{conv-pf-10}
\lefteqn{\langle e^{\epsilon L^{L}(1)+\bar{\epsilon}L^{R}(1)}u',
m_{l_{1}+\cdots +l_{K}}(u_1^{(1)},\dots, u_{l_1}^{(1)},
\dots, u_1^{(K)},\dots, u_{l_K}^{(K)};
z_1^{(1)}+z_1^{(0)}+\epsilon, }{\nonumber \\}&&\quad
\bar{z}_1^{(1)}+\bar{z}_1^{(0)}+\bar{\epsilon},
\dots,
z_{l_1}^{(1)}+z_1^{(0)}+\epsilon,
\bar{z}_{l_1}^{(1)}+\bar{z}_1^{(0)}+\bar{\epsilon}, \dots,
z_1^{(K)}+z_K^{(0)}+\epsilon, {\nonumber \\}&&\quad \quad \quad\quad\quad\quad\quad\quad\quad
\bar{z}_1^{(K)}+\bar{z}_K^{(0)}+\bar{\epsilon},
\dots, z_{l_K}^{(K)}+z_K^{(0)}+\epsilon,
\bar{z}_{l_K}^{(K)}+\bar{z}_K^{(0)}+\bar{\epsilon})\rangle{\nonumber \\}&&=\langle u', m_{l_{1}+\cdots +l_{K}}(u_1^{(1)},\dots, u_{l_1}^{(1)},
\dots, u_1^{(K)},\dots, u_{l_K}^{(K)};
z_1^{(1)}+z_1^{(0)}, {\nonumber \\}&&\quad \quad \quad\quad\quad\quad
\bar{z}_1^{(1)}+\bar{z}_1^{(0)},
\dots,
z_{l_1}^{(1)}+z_1^{(0)}, \bar{z}_{l_1}^{(1)}+\bar{z}_1^{(0)}, \dots,
z_1^{(K)}+z_K^{(0)}, {\nonumber \\}&&\quad \quad \quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\bar{z}_1^{(K)}+\bar{z}_K^{(0)},
\dots, z_{l_K}^{(K)}+z_K^{(0)}, \bar{z}_{l_K}^{(K)}+\bar{z}_K^{(0)})
\rangle.\end{aligned}$$ Since $u'$ is arbitrary, we have proved that (\[series-case-2-0\]) with $k=K$ is convergent absolutely to the right-hand side of (\[conv-pf-2\]) in the case that the number of elements of $N(z_{1}^{(0)}, \dots,
z_{K}^{(0)})$ is $n+1$. Thus we have proved this conclusion in the case that some of $|z_{1}^{(0)}|, \dots,
|z_{K}^{(0)}|$ are equal.
By the principle of induction, the convergence property is proved.
[Although the definition of full field algebra in Definition \[ffa\] is very general, it is not easy to verify all the axioms directly. Theorem \[r-conf-alg-2-prop\] gives an equivalent definition of conformal full field algebra over $V^L\otimes V^R$ and the axioms in this definition are much easier to verify than those in Definitions \[ffa\], \[RR-grading-ffa\] and \[g-r-RR-g-ffa\]. In our construction of full field algebras in the next section, we shall use this definition to verify the structure we construct is indeed a full field algebra. ]{}
A construction of full field algebras with nondegenerate invariant bilinear forms
=================================================================================
Let $V$ be a simple vertex operator algebra and $C_{2}(V)$ the subspace of $V$ spanned by $u_{-2}v$ for $u, v\in V$. In this section, we assume that $V$ satisfies the following conditions:
1. $V_{(n)}=0$ for $n<0$, $V_{(0)}=\mathbb{C}\mathbf{1}$ and $W_{(0)}=0$ for any irreducible $V$-module $W$ which is not equivalent to $V$.
2. Every $\mathbb{N}$-gradable weak $V$-module is completely reducible.
3. $V$ is $C_{2}$-cofinite, that is, $\dim V/C_{2}(V)<\infty$.
(Note that by results of Li [@L] and Abe, Buhl and Dong [@ABD], Conditions 2 and 3 can be replaced by a single condition that every weak $V$-module is completely reducible.)
Since $V$ satisfies the conditions above, all the results in [@H7] can be used. We shall use all the notations, conventions and choices used in this paper. In particular, we use the following notations and choices: $\mathcal{A}$ is the (finite) set of equivalence classes of irreducible $V$-modules; $e$ is the equivalence class containing $V$; $^{\prime}:
\mathcal{A}\to \mathcal{A}$ is the map induced from the functor given by taking contragredient modules; for $a\in \mathcal{A}$, $W^{a}$ is a representative of $a$; $(\cdot, \cdot)$ is the nondegenerate bilinear form on $V$ normalized by $(\mathbf{1}, \mathbf{1})=1$; for $a_{1}, a_{2}, a_{3}\in \mathcal{A}$, $\mathcal{V}_{a_{1}a_{2}}^{a_{3}}$ are the spaces of intertwining operators of type ${{{W^{a_{3}}}\choose {W^{a_{1}}W^{a_{2}}}}}$; $\sigma_{12}$ and $\sigma_{23}$ are actions of $(12)$ and $(23)$ on $\mathcal{V}$ and they generate an action of $S_{3}$ on $$\mathcal{V}=\coprod_{a_{1}, a_{2}, a_{3}\in
\mathcal{A}}\mathcal{V}_{a_{1}a_{2}}^{a_{3}};$$ for any bases ${\mathcal{Y}}_{a_{1}a_{2}; i}^{a_{3}; (p)}$, $i=1, \dots,
N_{a_{1}a_{2}}^{a_{3}}$, $p=1, 2, 3, 4, 5, 6, \dots$ and $a_{1}, a_{2},
a_{3}\in \mathcal{A}$, of $\mathcal{V}_{a_{1}a_{2}}^{a_{3}}$, $$F({\mathcal{Y}}_{a_{1}a_{5}; i}^{a_{4}; (1)}\otimes {\mathcal{Y}}_{a_{2}a_{3}; j}^{a_{5}; (2)};
{\mathcal{Y}}_{a_{6}a_{3}; l}^{a_{4}; (3)}
\otimes {\mathcal{Y}}_{a_{1}a_{2}; k}^{a_{6}; (4)}) \in {\mathbb{C}}$$ are matrix elements of the fusing isomorphism; for $a\in {\mathcal{A}}$, ${\mathcal{Y}}_{ea; 1}^{a}$, ${\mathcal{Y}}_{ae; 1}^{a}$ and ${\mathcal{Y}}_{aa'; 1}^{e}$ are bases of ${\mathcal{V}}_{ea}^{a}$, ${\mathcal{V}}_{ae}^{a}$ and ${\mathcal{V}}_{aa'}^{e}$ chosen in [@H7]; for $a\in {\mathcal{A}}$, there exists $h_{a}\in {\mathbb{Q}}$ such that $W^{a}=\coprod_{n\in h_{a}+{\mathbb{N}}}W_{(n)}$.
For $a_{1}, a_{2}, a_{3}\in {\mathcal{A}}$, we now want to introduce a pairing between $\mathcal{V}_{a_{1}a_{2}}^{a_{3}}$ and $\mathcal{V}_{a'_{1}a'_{2}}^{a'_{3}}$.
For $a\in {\mathcal{A}}$, $w_{a}\in W^{a}$ and $w'_{a}\in (W^{a})'$, we shall use $\tilde{w}_{a}$ and $\tilde{w}'_{a}$ to denote $e^{-L(1)}w_{a}$ and $e^{-L(1)}w'_{a}$, respectively. Then we have $$\langle e^{L(1)}\tilde{w}'_{a}, e^{L(1)}\tilde{w}_{a}\rangle
=\langle w'_{a}, w_{a}\rangle.$$ We have:
\[res\] For $a\in {\mathcal{A}}$, $w_{a}\in W^{a}$ and $w'_{a}\in (W^{a})'$, $${\mbox{\rm Res}}_{z=0}z^{-1}{\mathcal{Y}}_{a'a;1}^{e}(z^{L(0)}e^{\pi i(L(0)-h_{a})}
\tilde{w}'_{a}, z)
z^{L(0)}\tilde{w}_{a}=\langle w'_{a}, w_{a}\rangle{\mathbf{1}}.$$
[[*Proof.*]{}]{}Since $V_{(0)}={\mathbb{C}}{\mathbf{1}}$ and ${\mathcal{Y}}_{a'a;1}^{e}=\sigma_{23}({\mathcal{Y}}_{a'e; 1}^{a'})$, $$\begin{aligned}
\lefteqn{{\mbox{\rm Res}}_{z=0}z^{-1}{\mathcal{Y}}_{a'a;1}^{e}(z^{L(0)}
e^{\pi i(L(0)-h_{a})}\tilde{w}'_{a}, z)
z^{L(0)}\tilde{w}_{a}}{\nonumber \\}&&=(\mathbf{1},
{\mathcal{Y}}_{a'a;1}^{e}(1^{L(0)}e^{\pi i(L(0)-h_{a})}\tilde{w}'_{a}, 1)
\tilde{w}_{a}){\mathbf{1}}{\nonumber \\}&&=(\mathbf{1},
\sigma_{23}({\mathcal{Y}}_{a'e;1}^{a'})(e^{\pi i(L(0)-h_{a})}\tilde{w}'_{a_{i}}, 1)
\tilde{w}_{a}){\mathbf{1}}{\nonumber \\}&&=\langle e^{\pi ih_{a}}{\mathcal{Y}}_{a'e;1}^{a'}(e^{L(1)}e^{-\pi iL(0)}
e^{\pi i(L(0)-h_{a})}
\tilde{w}'_{a}, 1)
\mathbf{1},
\tilde{w}_{a}\rangle {\mathbf{1}}{\nonumber \\}&&=\langle {\mathcal{Y}}_{a'e;1}^{a'}(e^{L(1)}\tilde{w}'_{a}, 1)
{\mathbf{1}}, \tilde{w}_{a}\rangle{\mathbf{1}}{\nonumber \\}&&=\langle e^{L(-1)}e^{L(1)}\tilde{w}'_{a}, \tilde{w}_{a}\rangle{\mathbf{1}}{\nonumber \\}&&=\langle e^{L(1)}\tilde{w}'_{a}, e^{L(1)}\tilde{w}_{a}\rangle{\mathbf{1}}{\nonumber \\}&&=\langle w'_{a}, w_{a}\rangle {\mathbf{1}}.\end{aligned}$$
For a single-valued branch $f_{1}(z_{1}, z_{2})$ of a multivalued analytic function in a region $A$, we use $E(f_{1}(z_{1}, z_{2}))$ to denote the multivalued analytic extension together with the preferred branch $f_{1}(z_{1}, z_{2})$. Let $w_{1}=w_{1}(z_{1}, z_{2})$ and $w_{2}=w_{2}(z_{1}, z_{2})$ be a change of variables and $f_{2}(z_{1}, z_{2})$ a branch of $E(f_{1}(z_{1}, z_{2}))$ in a region $B$ containing $w_{1}(z_{1}, z_{2})=0$ and $w_{2}(z_{1}, z_{2})=0$ such that $A\cap B\ne \emptyset$ and $f_{1}(z_{1}, z_{2})=f(z_{1}, z_{2})$ for $(z_{1}, z_{2})\in A\cap B$. Then we use $${\mbox{\rm Res}}_{w_{1}=0\;|\;w_{2}}E(f_{1}(z_{1}, z_{2}))$$ to denote the coefficient of $w_{1}^{-1}$ in the expansion of $f_{2}(z_{1}, z_{2})$ as a series in powers of $w_{1}$ whose coefficients are analytic functions of $w_{2}$. By definition, we have $$\label{residue}
{\mbox{\rm Res}}_{w_{1}=0\;|\;C_{1}w_{2}+C_{2}}E(f_{1}(z_{1}, z_{2}))
={\mbox{\rm Res}}_{w_{1}=0\;|\;w_{2}}E(f_{1}(z_{1}, z_{2}))$$ for any $C_{1}\in {\mathbb{C}}^{\times},
C_{2}\in {\mathbb{C}}$ independent of $z_{1}$ and $z_{2}$. We have:
\[pairing\] For $a_{1}, a_{2}, a_{3}\in {\mathcal{A}}$, $w_{a_{1}}\in W^{a_{1}}$, $w_{a_{2}}\in W^{a_{2}}$, $w_{a'_{1}}\in (W^{a_{1}})'$, $w'_{a_{2}}\in (W^{a_{1}})'$, ${\mathcal{Y}}_{1}\in
{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}$ and ${\mathcal{Y}}_{2}\in {\mathcal{V}}_{a'_{1}a'_{2}}^{a'_{3}}$, there exists a constant $\langle {\mathcal{Y}}_{1},
{\mathcal{Y}}_{2}\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}}\in {\mathbb{C}}$ such that $$\begin{aligned}
\label{inner}
\lefteqn{{\mbox{\rm Res}}_{1-z_{1}-z_{2}=0\;|\;z_{2}}(1-z_{1}-z_{2})^{-1}
E(\langle e^{L(1)}{\mathcal{Y}}_{2}((1-z_{1}-z_{2})^{L(0)}
\tilde{w}'_{a_{1}}, z_{1})
\tilde{w}'_{a_{2}}, }{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
e^{L(1)}{\mathcal{Y}}_{1}((1-z_{1}-z_{2})^{L(0)}\tilde{w}_{a_{1}}, z_{2})
\tilde{w}_{a_{2}}\rangle){\nonumber \\}&&=\langle w'_{a_{1}}, w_{a_{1}}\rangle\langle w'_{a_{2}}, w_{a_{2}}\rangle
\langle {\mathcal{Y}}_{1},
{\mathcal{Y}}_{2}\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}}.\end{aligned}$$ Explicitly, for any bases $\{{\mathcal{Y}}_{a_{1}a_{2}; i}^{a_{3}; (1)}\;|\;
i=1, \dots, N_{a_{1}a_{2}}^{a_{3}}\}$ and $\{{\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}; (2)}\;|\;
i=1, \dots, N_{a'_{1}a'_{2}}^{a'_{3}}\}$ of ${\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}$ and ${\mathcal{V}}_{a'_{1}a'_{2}}^{a'_{3}}$, respectively, and for $m, n, k, l\in {\mathbb{Z}}_{+}$, $i=1, \dots, N_{a_{1}a_{2}}^{a_{3}}$ and $j=1, \dots, N_{a'_{1}a'_{2}}^{a'_{3}}$, we have $$\begin{aligned}
\label{inner-fusing}
\langle {\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}; (1)}, {\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}; (2)}
\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}}&=&
F(\sigma_{23}({\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}; (2)})
\otimes {\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}; (1)}; {\mathcal{Y}}_{ea_{2}; 1}^{a_{2}}\otimes
{\mathcal{Y}}_{a'_{1}a_{1}; 1}^{e}){\nonumber \\}&=&F(\sigma_{23}({\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}; (1)})
\otimes {\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}; (2)}; {\mathcal{Y}}_{ea'_{2}; 1}^{a'_{2}}\otimes
{\mathcal{Y}}_{a_{1}a'_{1}; 1}^{e}).\end{aligned}$$
[[*Proof.*]{}]{}We prove (\[inner\]) in the case ${\mathcal{Y}}_{1}={\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}; (1)}$ and ${\mathcal{Y}}_{2}={\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}; (2)}$ for $i=1, \dots, N_{a'_{1}a'_{2}}^{a'_{3}}$ and $j=1, \dots, N_{a_{1}a_{2}}^{a_{3}}$, respectively, or equivalently, we prove (\[inner-fusing\]). The general case follows immediately from the bilinearity in ${\mathcal{Y}}_{1}$ and ${\mathcal{Y}}_{2}$ of the right-hand side of (\[inner\]).
For $a_{1}, a_{2}\in {\mathcal{A}}$, $a_{1}, a_{2}\ne e$, let $\{{\mathcal{Y}}_{a_{1}a_{2}; i}^{a_{2}}\;|\;i=1, \dots, N_{a_{1}a_{2}}^{a_{2}}\}$ and $\{{\mathcal{Y}}_{a'_{1}a_{1}; i}^{a_{2}}\;|\;i=1, \dots, N_{a'_{1}a_{1}}^{a_{2}}\}$ be an arbitrary basis of ${\mathcal{V}}_{a_{1}a_{2}}^{a_{2}}$ and ${\mathcal{V}}_{a'_{1}a_{1}}^{a_{2}}$, respectively.
For $w_{a_{1}}\in W^{a_{1}}$, $w_{a_{2}}\in W^{a_{2}}$, $w_{a'_{1}}\in (W^{a_{1}})'$, $w'_{a_{2}}\in (W^{a_{1}})'$, we have $$\begin{aligned}
\lefteqn{{\mbox{\rm Res}}_{1-z_{1}-z_{2}=0\;|\;z_{2}}(1-z_{1}-z_{2})^{-1}
E(\langle e^{L(1)}
{\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}; (2)}
((1-z_{1}-z_{2})^{L(0)}\tilde{w}'_{a_{1}}, z_{1})
\tilde{w}'_{a_{2}}, }{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
e^{L(1)}{\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}; (1)}
((1-z_{1}-z_{2})^{L(0)}\tilde{w}_{a_{1}}, z_{2})
\tilde{w}_{a_{2}}\rangle){\nonumber \\}&&={\mbox{\rm Res}}_{1-z_{1}-z_{2}=0\;|\;z_{2}}(1-z_{1}-z_{2})^{-1}\cdot{\nonumber \\}&&\quad\quad\cdot
E(\langle {\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}; (2)}
(e^{(1-z_{1})L(1)}(1-z_{1})^{-2L(0)}\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\cdot
(1-z_{1}-z_{2})^{L(0)}\tilde{w}'_{a_{1}}, (1-z_{1})^{-1})
e^{L(-1)}e^{L(1)}\tilde{w}'_{a_{2}}, {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad \quad\quad\quad\quad\quad
{\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}; (1)}((1-z_{1}-z_{2})^{L(0)}
\tilde{w}_{a_{1}}, z_{2})
\tilde{w}_{a_{2}}\rangle){\nonumber \\}&&={\mbox{\rm Res}}_{1-z_{1}-z_{2}=0\;|\;z_{2}}(1-z_{1}-z_{2})^{-1}{\nonumber \\}&&\quad E(\langle
e^{-L(1)}e^{L(1)}\tilde{w}'_{a_{2}},
\sigma_{23}({\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}; (2)})((1-z_{1}-z_{2})^{L(0)}
e^{\pi i(L(0)-h_{a'_{1}})}\tilde{w}'_{a_{1}}, 1-z_{1})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\cdot
{\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}; (1)}((1-z_{1}-z_{2})^{L(0)}
\tilde{w}_{a_{1}}, z_{2})
\tilde{w}_{a_{2}}\rangle){\nonumber \\}&&=\sum_{a_{4}\in {\mathcal{A}}}\sum_{p=1}^{N_{a_{4}a_{2}}^{a_{2}}}
\sum_{q=1}^{N_{a'_{1}a_{1}}^{a_{4}}}
F(\sigma_{23}({\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}; (2)})
\otimes {\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}; (1)}; {\mathcal{Y}}_{a_{4}a_{2}; p}^{a_{2}}\otimes
{\mathcal{Y}}_{a'_{1}a_{1}; q}^{a_{4}})\cdot {\nonumber \\}&&\quad\cdot{\mbox{\rm Res}}_{1-z_{1}-z_{2}=0\;|\;z_{2}}(1-z_{1}-z_{2})^{-1}\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\cdot E(\langle
e^{L(-1)}e^{L(1)}\tilde{w}'_{a_{2}},
{\mathcal{Y}}_{a_{4}a_{2}; p}^{a_{2}}({\mathcal{Y}}_{a'_{1}a_{1}; q}^{a_{4}}
((1-z_{1}-z_{2})^{L(0)}\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\cdot e^{\pi i(L(0)-h_{a'_{1}})}
\tilde{w}'_{a_{1}}, 1-z_{1}-z_{2})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad
\cdot (1-z_{1}-z_{2})^{L(0)}
\tilde{w}_{a_{1}}, z_{2})
\tilde{w}_{a_{2}}\rangle){\nonumber \\}&&=F(\sigma_{23}({\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}; (2)})
\otimes {\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}; (1)}; {\mathcal{Y}}_{ea_{2}; p}^{a_{2}}\otimes
{\mathcal{Y}}_{a'_{1}a_{1}; 1}^{e})\cdot {\nonumber \\}&&\quad\cdot{\mbox{\rm Res}}_{1-z_{1}-z_{2}=0\;|\;z_{2}}(1-z_{1}-z_{2})^{-1}\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\cdot E(\langle
e^{L(-1)}e^{L(1)}\tilde{w}'_{a_{2}},
{\mathcal{Y}}_{ea_{2}; 1}^{a_{2}}({\mathcal{Y}}_{a'_{1}a_{1}; 1}^{e}
((1-z_{1}-z_{2})^{L(0)}\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\cdot
e^{\pi i(L(0)-h_{a'_{1}})}
\tilde{w}'_{a_{1}}, 1-z_{1}-z_{2})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\quad\quad
\cdot
(1-z_{1}-z_{2})^{L(0)}\tilde{w}_{a_{1}}, z_{2})
\tilde{w}_{a_{2}}\rangle){\nonumber \\}&&=F(\sigma_{23}({\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}; (2)})
\otimes {\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}; (1)}; {\mathcal{Y}}_{ea_{2}; p}^{a_{2}}\otimes
{\mathcal{Y}}_{a'_{1}a_{1}; 1}^{e})\cdot {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\cdot\langle
e^{L(-1)}e^{L(1)}\tilde{w}'_{a_{2}},
{\mathcal{Y}}_{ea_{2}; 1}^{a_{2}}(\langle w'_{a_{1}}, w_{a_{1}}\rangle {\mathbf{1}}, z_{2})
\tilde{w}_{a_{2}}\rangle{\nonumber \\}&&=\langle w'_{a_{1}}, w_{a_{1}}\rangle
F(\sigma_{23}({\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}; (2)})
\otimes {\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}; (1)}; {\mathcal{Y}}_{ea_{2}; p}^{a_{2}}\otimes
{\mathcal{Y}}_{a'_{1}a_{1}; 1}^{e})\langle
e^{L(1)}\tilde{w}'_{a_{2}},
e^{L(1)}\tilde{w}_{a_{2}}\rangle{\nonumber \\}&&=\langle w'_{a_{1}}, w_{a_{1}}\rangle
\langle w'_{a_{2}}, w_{a_{2}}\rangle
F(\sigma_{23}({\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}; (2)})
\otimes {\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}; (1)}; {\mathcal{Y}}_{ea_{2}; p}^{a_{2}}\otimes
{\mathcal{Y}}_{a'_{1}a_{1}; 1}^{e}),\end{aligned}$$ where we have used the fact that $W_{(0)}^{a_{4}}=0$ for $a_{4}\ne e$. This proves (\[inner\]) and also the first equality in (\[inner-fusing\]). The second equality in (\[inner-fusing\]) can be proved similarly or can be simply obtained using the first equality in (\[inner-fusing\]) and symmetry.
Clearly, $\langle {\mathcal{Y}}_{1},
{\mathcal{Y}}_{2}\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}}$ is bilinear in ${\mathcal{Y}}_{1}$ and ${\mathcal{Y}}_{2}$. Thus we have a pairing $\langle \cdot, \cdot\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}}:
{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}\otimes {\mathcal{V}}_{a'_{1}a'_{2}}^{a'_{3}}
\to {\mathbb{C}}$.
We need the following lemma:
\[zero\] For $a_{1}, a_{2}, a_{3}, a_{4}, a_{5}\in {\mathcal{A}}$, ${\mathcal{Y}}_{1}\in
{\mathcal{V}}_{a'_{1}a'_{2}}^{a'_{3}}$, ${\mathcal{Y}}_{2}\in {\mathcal{V}}_{a_{4}a_{5}}^{a_{3}}$, $w'_{a_{1}}\in (W^{a_{1}})'$, $w'_{a_{2}}\in (W^{a_{2}})'$, $w_{a_{4}}\in W^{a_{1}}$, $w_{a_{4}}
\in W^{a_{5}}$, if $a_{1}\ne a_{4}$ or $a_{2}\ne a_{5}$, then $$\begin{aligned}
\lefteqn{{\mbox{\rm Res}}_{1-z_{1}-z_{2}=0\;|\;z_{2}}(1-z_{1}-z_{2})^{-1}E(\langle e^{L(1)}
{\mathcal{Y}}_{1}((1-z_{1}-z_{2})^{L(0)}\tilde{w}'_{a_{1}}, z_{1})
\tilde{w}'_{a_{2}}, }{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
e^{L(1)}{\mathcal{Y}}_{2}((1-z_{1}-z_{2})^{L(0)}\tilde{w}_{a_{4}}, z_{2})
\tilde{w}_{a_{5}}\rangle)=0.\end{aligned}$$
[[*Proof.*]{}]{}Using the $L(1)$- and $L(-1)$-conjugation formulas for intertwining operators, the definition of $\sigma_{23}$ and the associativity of intertwining operators, we know that there exist a $V$-module $W$ and intertwining operators ${\mathcal{Y}}_{3}$ and ${\mathcal{Y}}_{4}$ of types ${{{W^{a_{2}}}\choose {WW^{a_{5}}}}}$ and ${{{W}\choose {W^{a'_{1}}W^{a_{4}}}}}$, respectively, such that $$\begin{aligned}
\label{zero-1}
\lefteqn{{\mbox{\rm Res}}_{1-z_{1}-z_{2}=0\;|\;z_{2}}(1-z_{1}-z_{2})^{-1}E(\langle e^{L(1)}
{\mathcal{Y}}_{1}((1-z_{1}-z_{2})^{L(0)}\tilde{w}'_{a_{1}}, z_{1})
\tilde{w}'_{a_{2}}, }{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
e^{L(1)}{\mathcal{Y}}_{2}((1-z_{1}-z_{2})^{L(0)}\tilde{w}_{a_{4}}, z_{2})
\tilde{w}_{a_{5}}\rangle){\nonumber \\}&&={\mbox{\rm Res}}_{1-z_{1}-z_{2}=0\;|\;z_{2}}(1-z_{1}-z_{2})^{-1}\cdot{\nonumber \\}&&\quad\quad\cdot E(\langle
e^{L(-1)}e^{L(1)}\tilde{w}'_{a_{2}},
\sigma_{23}({\mathcal{Y}}_{1})((1-z_{1}-z_{2})^{L(0)}\tilde{w}'_{a_{1}}, 1-z_{1})
\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\cdot
{\mathcal{Y}}_{2}((1-z_{1}-z_{2})^{L(0)}\tilde{w}_{a_{4}}, z_{2})
\tilde{w}_{a_{5}}\rangle){\nonumber \\}&&={\mbox{\rm Res}}_{1-z_{1}-z_{2}=0\;|\;z_{2}}(1-z_{1}-z_{2})^{-1}\cdot{\nonumber \\}&&\quad\quad\cdot E(\langle
e^{L(-1)}e^{L(1)}\tilde{w}'_{a_{2}},
{\mathcal{Y}}_{3}({\mathcal{Y}}_{4}((1-z_{1}-z_{2})^{L(0)}\tilde{w}'_{a_{1}}, 1-z_{1}-z_{2})
\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\cdot
(1-z_{1}-z_{2})^{L(0)}\tilde{w}_{a_{4}}, z_{2})
\tilde{w}_{a_{5}}\rangle).\end{aligned}$$ If $a_{1}\ne a_{4}$, $W^{a_{4}}$ is not equivalent to $W^{a_{1}}$. Thus ${\mathcal{V}}_{a'_{1}a_{4}}^{e}=0$. So it is possible to find such a $V$-module $W$ which does not contain a summand equivalent to $V$. By the assumption on $V$, we have $W_{(0)}=0$. So the right-hand side of (\[zero-1\]) is $0$, proving the lemma in this case. If $a_{1}= a_{4}$, ${\mathcal{V}}_{a'_{1}a_{1}}^{e}$ is one-dimensional. We can choose $W$ to contain one and only one copy of $V$. If $a_{2}\ne a_{5}$, any intertwining operator of type ${{{a_{2}}\choose {ea_{5}}}}$ (that is, type ${{{W^{a_{2}}}\choose {VW^{a_{5}}}}}$) must be $0$. So ${\mathcal{Y}}_{3}({\mathbf{1}}, z_{2})=0$. Since $W_{(0)}={\mathbb{C}}{\mathbf{1}}$, there exists $\lambda\in {\mathbb{C}}$ such that the right-hand side of (\[zero-1\]) is equal to $$\lambda\langle
e^{L(-1)}e^{L(1)}\tilde{w}'_{a_{2}},
{\mathcal{Y}}_{3}({\mathbf{1}}, z_{2})
\tilde{w}_{a_{2}}\rangle=0,$$ proving the lemma in the case $a_{2}\ne a_{5}$.
As in [@H7], we now choose a canonical basis of ${\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}$ for $a_{1}, a_{2}, a_{3}\in
{\mathcal{A}}$ when one of $a_{1}, a_{2}, a_{3}$ is $e$: For $a\in {\mathcal{A}}$, we choose ${\mathcal{Y}}_{ea; 1}^{a}$ to be the vertex operator $Y_{W^{a}}$ defining the module structure on $W^{a}$ and we choose ${\mathcal{Y}}_{ae; 1}^{a}$ to be the intertwining operator defined using the action of $\sigma_{12}$, or equivalently the skew-symmetry in this case, $$\begin{aligned}
{\mathcal{Y}}_{ae; 1}^{a}(w_{a}, x)u&=&\sigma_{12}({\mathcal{Y}}_{ea; 1}^{a})(w_{a}, x)u{\nonumber \\}&=&e^{xL(-1)}{\mathcal{Y}}_{ea; 1}^{a}(u, -x)w_{a}{\nonumber \\}&=&e^{xL(-1)}Y_{W^{a}}(u, -x)w_{a}\end{aligned}$$ for $u\in V$ and $w_{a}\in W^{a}$. Since $V'$ as a $V$-module is isomorphic to $V$, we have $e'=e$. From [@FHL], we know that there is a nondegenerate invariant bilinear form $(\cdot, \cdot)$ on $V$ such that $(\mathbf{1},
\mathbf{1})=1$. We choose ${\mathcal{Y}}_{aa'; 1}^{e}={\mathcal{Y}}_{aa'; 1}^{e'}$ to be the intertwining operator defined using the action of $\sigma_{23}$ by $${\mathcal{Y}}_{aa'; 1}^{e'}=\sigma_{23}({\mathcal{Y}}_{ae; 1}^{a}),$$ that is, $$(u, {\mathcal{Y}}_{aa'; 1}^{e}(w_{a}, x)w_{a'})
=e^{\pi i h_{a}}\langle {\mathcal{Y}}_{ae; 1}^{a}(e^{xL(1)}(e^{-\pi i}x^{-2})^{L(0)}w_{a}, x^{-1})u,
w_{a'}\rangle$$ for $u\in V$, $w_{a}\in W^{a}$ and $w_{a'}\in W^{a'}$. Since the actions of $\sigma_{12}$ and $\sigma_{23}$ generate the action of $S_{3}$ on $\mathcal{V}$, we have $${\mathcal{Y}}_{a'a; 1}^{e}=\sigma_{12}({\mathcal{Y}}_{aa'; 1}^{e})$$ for any $a\in \mathcal{A}$.
The pairing $\langle \cdot, \cdot\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}}:
{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}\otimes {\mathcal{V}}_{a'_{1}a'_{2}}^{a'_{3}}
\to {\mathbb{C}}$ is nondegenerate. In particular, $N_{a'_{1}a'_{2}}^{a'_{3}}
=N_{a_{1}a_{2}}^{a_{3}}.$
[[*Proof.*]{}]{} For $a_{1}, a_{2}, a_{3}\in
{\mathcal{A}}$ such that one of $a_{1}, a_{2}, a_{3}$ is $e$, we have a canonical basis $\mathcal{Y}_{a_{1}a_{2}; 1}^{a_{3}}$ given above. For $a_{1}, a_{2}, a_{3}\ne e$, let $\mathcal{Y}_{a_{1}a_{2}; i}^{a_{3}}$, $i=1, \dots,
N_{a_{1}a_{2}}^{a_{3}}$, be an arbitrary basis of $\mathcal{V}_{a_{1}a_{2}}^{a_{3}}$.
For $a_{1}, a_{2}, a_{3}\in
{\mathcal{A}}$, let $$\begin{aligned}
{\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}; (1)}
&=&\sigma_{123}({\mathcal{Y}}_{a_{2}a_{3}'; j}^{a_{1}'}),
\\
{\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}; (2)}
&=&\sigma_{23}({\mathcal{Y}}_{a'_{1}a_{3}; i}^{a_{2}}).\end{aligned}$$ Then the first equality of (\[inner-fusing\]) gives $$\begin{aligned}
\label{inner-fusing-1}
\langle \sigma_{23}({\mathcal{Y}}_{a'_{1}a_{3}; i}^{a_{2}}),
\sigma_{123}({\mathcal{Y}}_{a_{2}a_{3}'; j}^{a_{1}'})
\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}}
&=&\langle {\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}; (2)},
{\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}; (1)}
\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}}{\nonumber \\}&=&F({\mathcal{Y}}_{a'_{1}a_{3}; i}^{a_{2}}
\otimes \sigma_{123}({\mathcal{Y}}_{a_{2}a_{3}'; j}^{a_{1}'});
{\mathcal{Y}}_{ea_{2}; 1}^{a_{2}}\otimes
{\mathcal{Y}}_{a'_{1}a_{1}; 1}^{e}).{\nonumber \\}&&\end{aligned}$$ In [@H7], the first author proved the following formula ((4.9) in [@H7]): $$\begin{aligned}
\label{formula1}
\lefteqn{\sum_{k=1}^{N_{a'_{1}a_{3}}^{a_{2}}}
F({\mathcal{Y}}_{a_{2}e; 1}^{a_{2}}\otimes {\mathcal{Y}}_{a'_{3}a_{3}; 1}^{e};
{\mathcal{Y}}_{a'_{1}a_{3}; k}^{a_{2}}\otimes {\mathcal{Y}}_{a_{2}a'_{3}; j}^{a'_{1}})\cdot}{\nonumber \\}&&\quad\quad\quad\quad\quad\cdot
F({\mathcal{Y}}_{a_{1}'a_{3}; k}^{a_{2}}\otimes
\sigma_{123}({\mathcal{Y}}_{a_{2}a'_{3}; i}^{a'_{1}});
{\mathcal{Y}}_{ea_{2}; 1}^{a_{2}}\otimes {\mathcal{Y}}_{a'_{1}a_{1}; 1}^{e}){\nonumber \\}&&=\delta_{ij}
F({\mathcal{Y}}_{a_{2}e; 1}^{a_{2}} \otimes {\mathcal{Y}}_{a'_{2}a_{2}; 1}^{e};
{\mathcal{Y}}_{ea_{2}; 1}^{a_{2}}\otimes {\mathcal{Y}}_{a_{2}a'_{2}; 1}^{e}).\end{aligned}$$ In the same paper [@H7], the first author also proved that $$F({\mathcal{Y}}_{a_{2}e; 1}^{a_{2}} \otimes {\mathcal{Y}}_{a'_{2}a_{2}; 1}^{e};
{\mathcal{Y}}_{ea_{2}; 1}^{a_{2}}\otimes {\mathcal{Y}}_{a_{2}a'_{2}; 1}^{e})\ne 0.$$ Thus from (\[inner-fusing-1\]) and (\[formula1\]), we see that the matrix $$\label{bl-form-mat}
(\alpha_{ij})=(\langle \sigma_{23}({\mathcal{Y}}_{a'_{1}a_{3}; i}^{a_{2}}),
\sigma_{123}({\mathcal{Y}}_{a_{2}a_{3}'; j}^{a_{1}'})
\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}})$$ is left invertible. Note that when $a_{1}, a_{2}, a_{3}\ne e$, ${\mathcal{Y}}_{a'_{1}a_{3}; i}^{a_{2}}$ and ${\mathcal{Y}}_{a_{2}a_{3}'; j}^{a_{1}'}$ in (\[bl-form-mat\]) are arbitrary bases of $\mathcal{V}_{a'_{1}a_{3}}^{a_{2}}$ and $\mathcal{V}_{a_{2}a_{3}'}^{a_{1}'}$, respectively.
We now show that (\[bl-form-mat\]) is also right invertible. By definition, the bilinear form $\langle\cdot,
\cdot\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}}$ is symmetric in the sense that $$\langle {\mathcal{Y}}_{1}, {\mathcal{Y}}_{2}\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}}
=\langle {\mathcal{Y}}_{2}, {\mathcal{Y}}_{1}\rangle_{{\mathcal{V}}_{a'_{1}a'_{2}}^{a'_{3}}}$$ for $a_{1}, a_{2}, a_{3}\in
{\mathcal{A}}$. So $$\begin{aligned}
\label{nondeg-1}
\lefteqn{\langle \sigma_{23}({\mathcal{Y}}_{a'_{1}a_{3}; i}^{a_{2}}),
\sigma_{123}({\mathcal{Y}}_{a_{2}a_{3}'; j}^{a_{1}'})
\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}}}{\nonumber \\}&&=\langle \sigma_{123}({\mathcal{Y}}_{a_{2}a_{3}'; j}^{a_{1}'}),
\sigma_{23}({\mathcal{Y}}_{a'_{1}a_{3}; i}^{a_{2}})
\rangle_{{\mathcal{V}}_{a'_{1}a'_{2}}^{a'_{3}}}{\nonumber \\}&&=\langle \sigma_{23}(\sigma_{13}({\mathcal{Y}}_{a_{2}a_{3}'; j}^{a_{1}'})),
\sigma_{123}(\sigma_{13}({\mathcal{Y}}_{a'_{1}a_{3}; i}^{a_{2}}))
\rangle_{{\mathcal{V}}_{a'_{1}a'_{2}}^{a'_{3}}}.\end{aligned}$$ Note that for $a_{1}, a_{2}, a_{3}\in {\mathcal{A}}$, $\sigma_{13}({\mathcal{Y}}_{a_{3}'a_{2}; i}^{a_{1}'})$ is a basis of $\mathcal{V}_{a_{1}a_{2}}^{a_{3}}$ such that when one of the elements $a_{1}, a_{2}, a_{3}\in {\mathcal{A}}$ is $e$, these basis elements are equal to the special ones we chosen above. Thus by the result we obtained above, the matrix $$(\beta_{ij})=(\langle \sigma_{23}(\sigma_{13}({\mathcal{Y}}_{a_{2}a_{3}'; i}^{a_{1}'})),
\sigma_{123}(\sigma_{13}({\mathcal{Y}}_{a'_{1}a_{3}; j}^{a_{2}}))
\rangle_{{\mathcal{V}}_{a'_{1}a'_{2}}^{a'_{3}}})$$ must be left invertible. So the transpose of $(\beta_{ij})$, that is, the matrix $$(\gamma_{kl})
=(\langle \sigma_{23}(\sigma_{13}({\mathcal{Y}}_{a_{2}a_{3}'; l}^{a_{1}'})),
\sigma_{123}(\sigma_{13}({\mathcal{Y}}_{a'_{1}a_{3}; k}^{a_{2}}))
\rangle_{{\mathcal{V}}_{a'_{1}a'_{2}}^{a'_{3}}}),$$ is right invertible. By (\[nondeg-1\]), we see that (\[bl-form-mat\]) is also right invertible.
Now we have shown that the matrix (\[bl-form-mat\]) is in fact invertible. This is equivalent to the nondegeneracy of the bilinear form. It also implies $N_{a'_{1}a'_{2}}^{a'_{3}}
=N_{a_{1}a_{2}}^{a_{3}}.$
For $a\in {\mathcal{A}}$, let $$F_{a}=F({\mathcal{Y}}_{ae; 1}^{a} \otimes {\mathcal{Y}}_{a'a; 1}^{e};
{\mathcal{Y}}_{ea; 1}^{a}\otimes {\mathcal{Y}}_{aa'; 1}^{e})\ne 0.$$ Then by (3.12) in [@H7], $F_{a'}=F_{a}$ for $a\in {\mathcal{A}}$.
If for $a\in {\mathcal{A}}$, ${\mathcal{Y}}_{ea; 1}^{a}$, ${\mathcal{Y}}_{ae; 1}^{a}$, ${\mathcal{Y}}_{aa'; 1}^{e}$ are the canonical bases of ${\mathcal{V}}_{ea}^{a}$, ${\mathcal{V}}_{ae}^{a}$, ${\mathcal{V}}_{aa'}^{e}$, respectively, chosen in [@H7] and above, then their dual bases in ${\mathcal{V}}_{ea'}^{a'}$, ${\mathcal{V}}_{a'e}^{a'}$, ${\mathcal{V}}_{a'a}^{e}$ with respect to the pairing $\langle\cdot,
\cdot\rangle_{{\mathcal{V}}_{ea}^{a}}$, $\langle\cdot,
\cdot\rangle_{{\mathcal{V}}_{ae}^{a}}$, $\langle\cdot,
\cdot\rangle_{{\mathcal{V}}_{aa'}^{e}}$, respectively, are equal to ${\mathcal{Y}}_{ea'; 1}^{a'}$, ${\mathcal{Y}}_{a'e; 1}^{a}$, $\frac{{\mathcal{Y}}_{a'a; 1}^{e}}{F_{a}}$.
[[*Proof.*]{}]{}This result follows immediately from the definition of the canonical bases in [@H7].
We have:
\[fusing\] For $a_{1}, a_{2}, a_{3}\in {\mathcal{A}}$, let $\{{\mathcal{Y}}_{a_{1}a_{2}; i}^{a_{3}}\;|\;i=1, \dots, N_{a_{1}a_{2}}^{a_{3}}\}$ be bases of ${\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}$ and let $\{{\mathcal{Y}}_{a'_{1}a'_{2}; i}^{\prime; a'_{3}}\;|\;i=1, \dots,
N_{a'_{1}a'_{2}}^{a'_{3}}\}$ be the dual bases of $\{{\mathcal{Y}}_{a_{1}a_{2}; i}^{a_{3}}\;|\;i=1, \dots, N_{a_{1}a_{2}}^{a_{3}}\}$ with respect to the pairing $\langle\cdot,
\cdot\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}}$. Assume that for $a\in {\mathcal{A}}$, ${\mathcal{Y}}_{ea; 1}^{a}$, ${\mathcal{Y}}_{ae; 1}^{a}$, ${\mathcal{Y}}_{aa'; 1}^{e}$ are the canonical bases of ${\mathcal{V}}_{ea}^{a}$, ${\mathcal{V}}_{ae}^{a}$, ${\mathcal{V}}_{aa'}^{e}$, respectively, we have chosen. Then for $a_{1}, a_{2}, a_{3}, a_{4}\in {\mathcal{A}}$, $$\begin{aligned}
\lefteqn{\sum_{a_{5}\in {\mathcal{A}}}\sum_{p=1}^{N_{a_{1}a_{5}}^{a_{4}}}
\sum_{q=1}^{N_{a_{2}a_{3}}^{a_{5}}}
F({\mathcal{Y}}_{a_{1}a_{5}; p}^{a_{4}}\otimes {\mathcal{Y}}_{a_{2}a_{3}; q}^{a_{5}};
{\mathcal{Y}}_{a_{6}a_{3}; m}^{a_{4}}\otimes {\mathcal{Y}}_{a_{1}a_{2}; k}^{a_{6}})\cdot}{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\cdot
F({\mathcal{Y}}_{a'_{1}a'_{5}; p}^{\prime; a'_{4}}\otimes {\mathcal{Y}}_{a'_{2}a'_{3}; q}^{\prime;
a_{5}};
{\mathcal{Y}}_{a'_{7}a'_{3}; n}^{\prime; a'_{4}}\otimes {\mathcal{Y}}_{a'_{1}a'_{2}; l}^{\prime;
a'_{7}}){\nonumber \\}&&=\delta_{a_{6}a_{7}}\delta_{mn}\delta_{kl}.\end{aligned}$$
[[*Proof.*]{}]{} For $a_{1}, a_{2}, a_{3}, a_{4}\in {\mathcal{A}}$, $w_{a_{i}}\in W^{a_{i}}$ and $w'_{a_{i}}\in (W^{a_{i}})'$ satisfying $\langle w'_{a_{i}}, w_{a_{i}}\rangle=1$ for $i=1, 2, 3$, using (\[inner\]) and Lemma \[zero\], we have $$\begin{aligned}
\label{fusing-1}
\lefteqn{{\mbox{\rm Res}}_{1-z_{1}-z_{3}=0\;|\; z_{3}}{\mbox{\rm Res}}_{1-z_{2}-z_{4}=0\;|\;z_{4}}
(1-z_{1}-z_{3})^{-1}(1-z_{2}-z_{4})^{-1}}{\nonumber \\}&&\quad\quad
E(\langle e^{L(1)}{\mathcal{Y}}_{a'_{1}a'_{5}; k}^{\prime; a'_{4}}
((1-z_{1}-z_{3})^{L(0)}\tilde{w}'_{a_{1}}, z_{1})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad \quad\quad\quad \quad\quad\quad \cdot
{\mathcal{Y}}_{a'_{2}a'_{3};l}^{\prime; a'_{5}}
((1-z_{2}-z_{4})^{L(0)}\tilde{w}'_{a_{2}; q}, z_{2})\tilde{w}'_{a_{3}},{\nonumber \\}&&\quad\quad\quad
e^{L(1)}{\mathcal{Y}}_{a_{1}a_{6}; m}^{a'_{4}}((1-z_{1}-z_{3})^{L(0)}\tilde{w}_{a_{1}}, z_{3})
\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad \quad\quad\quad \quad\quad\quad \cdot
{\mathcal{Y}}_{a_{2}a_{3};n}^{a_{6}}
((1-z_{2}-z_{4})^{L(0)}\tilde{w}_{a_{2}; q}, z_{4})\tilde{w}_{a_{3}}\rangle){\nonumber \\}&&=\delta_{a_{5}a_{6}}\delta_{km}{\mbox{\rm Res}}_{1-z_{2}-z_{4}=0\;|\;z_{4}}
(1-z_{2}-z_{4})^{-1}{\nonumber \\}&&\quad\quad
E(\langle{\mathcal{Y}}_{a'_{2}a'_{3};l}^{\prime; a'_{5}}
((1-z_{2}-z_{4})^{L(0)}\tilde{w}'_{a_{2}}, z_{2})
\tilde{w}'_{a_{3}},{\nonumber \\}&&\quad\quad\quad\quad\quad\quad \quad\quad\quad \quad\quad\quad
{\mathcal{Y}}_{a_{2}a_{3};n}^{a_{5}}
((1-z_{2}-z_{4})^{L(0)}\tilde{w}_{a_{2}}, z_{4})\tilde{w}_{a_{3}}\rangle){\nonumber \\}&&=\delta_{a_{5}a_{6}}\delta_{km}\delta_{ln}.\end{aligned}$$
On the other hand, by the associativity of intertwining operators and Lemma \[zero\], we have $$\begin{aligned}
\label{fusing-2}
\lefteqn{{\mbox{\rm Res}}_{1-z_{1}-z_{3}=0\;|\;z_{3}}{\mbox{\rm Res}}_{1-z_{2}-z_{4}=0\;|\;z_{4}}
(1-z_{1}-z_{3})^{-1}(1-z_{2}-z_{4})^{-1}}{\nonumber \\}&&\quad\quad E(\langle e^{L(1)}{\mathcal{Y}}_{a'_{1}a'_{5}; k}^{\prime; a'_{4}}
((1-z_{1}-z_{3})^{L(0)}\tilde{w}'_{a_{1}}, z_{1})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad \quad\quad\quad \quad\quad\quad \cdot
{\mathcal{Y}}_{a'_{2}a'_{3};l}^{\prime; a'_{5}}((1-z_{2}-z_{4})^{L(0)}
\tilde{w}'_{a_{2}}, z_{2})\tilde{w}'_{a_{3}},
{\nonumber \\}&&\quad\quad\quad
e^{L(1)}{\mathcal{Y}}_{a_{1}a_{6}; m}^{a'_{4}}
((1-z_{1}-z_{3})^{L(0)}\tilde{w}_{a_{1}}, z_{3})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad \quad\quad\quad \quad\quad\quad \cdot
{\mathcal{Y}}_{a_{2}a_{3};n}^{a_{6}}((1-z_{2}-z_{4})^{L(0)}
\tilde{w}_{a_{2}}, z_{4})\tilde{w}_{a_{3}}\rangle){\nonumber \\}&&=\sum_{a_{7}, a_{8}\in {\mathcal{A}}}\sum_{i, j, s, t}
F({\mathcal{Y}}_{a'_{1}a'_{5}; k}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{2}a'_{3};l}^{\prime; a'_{5}};
{\mathcal{Y}}_{a'_{7}a'_{3};i}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{1}a'_{2};j}^{\prime; a'_{7}})\cdot {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad
\cdot
F({\mathcal{Y}}_{a_{1}a_{6}; m}^{a_{4}}\otimes {\mathcal{Y}}_{a_{2}a_{3};n}^{a_{6}};
{\mathcal{Y}}_{a_{8}a_{3};s}^{a_{4}}\otimes {\mathcal{Y}}_{a_{1}a_{2};t}^{a_{8}})\cdot{\nonumber \\}&&\quad\quad\cdot{\mbox{\rm Res}}_{1-z_{1}-z_{3}=0\;|\;z_{3}}{\mbox{\rm Res}}_{1-z_{2}-z_{4}=0\;|\;
z_{4}}
(1-z_{1}-z_{3})^{-1}(1-z_{2}-z_{4})^{-1} {\nonumber \\}&&\quad\quad\quad
E(\langle e^{L(1)}{\mathcal{Y}}_{a'_{7}a'_{3}; i}^{\prime; a'_{4}}(
{\mathcal{Y}}_{a'_{1}a'_{2};j}^{\prime; a'_{7}}((1-z_{1}-z_{3})^{L(0)}
\tilde{w}'_{a_{1}}, z_{1}-z_{2})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad \quad\quad\quad \quad\quad\quad
\quad\quad \cdot
(1-z_{2}-z_{4})^{L(0)}\tilde{w}'_{a_{2}}, z_{2})\tilde{w}'_{a_{3}}, {\nonumber \\}&&\quad\quad\quad\quad
e^{L(1)}{\mathcal{Y}}_{a_{8}a_{3}; s}^{a_{4}}
({\mathcal{Y}}_{a_{1}a_{2};t}^{a_{8}}((1-z_{1}-z_{3})^{L(0)}\tilde{w}_{a_{1}},
z_{3}-z_{4})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad \quad\quad\quad \quad\quad\quad
\quad\quad \cdot
(1-z_{2}-z_{4})^{L(0)}\tilde{w}_{a_{2}}, z_{4})\tilde{w}_{a_{3}}\rangle){\nonumber \\}&&=\sum_{a_{7}, a_{8}\in {\mathcal{A}}}\sum_{i, j, s, t}
F({\mathcal{Y}}_{a'_{1}a'_{5}; k}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{2}a'_{3};l}^{\prime; a'_{5}};
{\mathcal{Y}}_{a'_{7}a'_{3};i}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{1}a'_{2};j}^{\prime; a'_{7}})\cdot {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad
\cdot
F({\mathcal{Y}}_{a_{1}a_{6}; m}^{a_{4}}\otimes {\mathcal{Y}}_{a_{2}a_{3};n}^{a_{6}};
{\mathcal{Y}}_{a_{8}a_{3};s}^{a_{4}}\otimes {\mathcal{Y}}_{a_{1}a_{2};t}^{a_{8}})\cdot{\nonumber \\}&&\quad\quad\cdot{\mbox{\rm Res}}_{1-z_{2}-z_{4}=0\;|\;z_{4}}
{\mbox{\rm Res}}_{1-z_{1}-z_{3}=0\;|\;z_{3}}
(1-z_{1}-z_{3})^{-1}(1-z_{2}-z_{4})^{-1} {\nonumber \\}&&\quad\quad\quad
E\Biggl(\Biggl\langle e^{L(1)}{\mathcal{Y}}_{a'_{7}a'_{3}; i}^{\prime; a'_{4}}\Biggl(
(1-z_{2}-z_{4})^{L(0)}
{\mathcal{Y}}_{a'_{1}a'_{2};j}^{\prime; a'_{7}}
\Biggl(\Biggl(\frac{1-z_{1}-z_{3}}{1-z_{2}-z_{4}}\Biggr)^{L(0)}
\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \cdot
\tilde{w}'_{a_{1}}, \frac{z_{1}-z_{2}}{1-z_{2}-z_{4}}\Biggr)
\tilde{w}'_{a_{2}}, z_{2}\Biggr)\tilde{w}'_{a_{3}}, {\nonumber \\}&&\quad\quad\quad\quad
e^{L(1)}{\mathcal{Y}}_{a_{8}a_{3}; s}^{a_{4}}
\Biggl((1-z_{2}-z_{4})^{L(0)}
{\mathcal{Y}}_{a_{1}a_{2};t}^{a_{8}}
\Biggl(\Biggl(\frac{1-z_{1}-z_{3}}{1-z_{2}-z_{4}}\Biggr)^{L(0)}\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \cdot
\tilde{w}_{a_{1}},
\frac{z_{3}-z_{4}}{1-z_{2}-z_{4}}\Biggr)
\tilde{w}_{a_{2}}, z_{4}\Biggr)\tilde{w}_{a_{3}}\Biggr\rangle\Biggr).\end{aligned}$$
We now change the variables $z_{1}$ and $z_{3}$ to $$z_{5}=\frac{z_{1}-z_{2}}{1-z_{2}-z_{4}}$$ and $$z_{6}=\frac{z_{3}-z_{4}}{1-z_{2}-z_{4}}.$$ Then $$\begin{aligned}
1-z_{5}-z_{6}&=&\frac{1-z_{1}-z_{3}}{1-z_{2}-z_{4}},\\
z_{3}&=&(1-z_{2}-z_{4})z_{6}+z_{4}.\end{aligned}$$ For any branch $f(z_{1}, z_{2}, z_{3}, z_{4})$ of a multivalued analytic function of $z_{1}$, $z_{2}$, $z_{3}$ and $z_{4}$ on a suitable region $A$ such that it is equal to the restriction to $A\cap B$ of a branch of the same analytic function on a region $B$ containing the point $1-z_{1}-z_{3}=0$, by definition, we have $$\begin{aligned}
\label{residue2}
\lefteqn{{\mbox{\rm Res}}_{1-z_{1}-z_{3}=0\;|\;z_{3}}E(f(z_{1}, z_{2}, z_{3}, z_{4}))}{\nonumber \\}&&={\mbox{\rm Res}}_{1-z_{5}-z_{6}=0\;|\;(1-z_{2}-z_{4})z_{6}+z_{4}}
E(f(z_{1}, z_{2}, z_{3}, z_{4}))\frac{1-z_{1}-z_{3}}{1-z_{5}-z_{6}}.\end{aligned}$$ By (\[residue\]), we have $$\begin{aligned}
\label{residue3}
\lefteqn{{\mbox{\rm Res}}_{1-z_{5}-z_{6}=0\;|\;(1-z_{2}-z_{4})z_{6}+z_{4}}
E(f(z_{1}, z_{2}, z_{3}, z_{4}))\frac{1-z_{1}-z_{3}}{1-z_{5}-z_{6}}}{\nonumber \\}&&={\mbox{\rm Res}}_{1-z_{5}-z_{6}=0\;|\;z_{6}}
E(f(z_{1}, z_{2}, z_{3}, z_{4}))\frac{1-z_{1}-z_{3}}{1-z_{5}-z_{6}}.\end{aligned}$$ From (\[residue2\]) and (\[residue3\]), we obtain $$\begin{aligned}
\label{residue4}
\lefteqn{{\mbox{\rm Res}}_{1-z_{1}-z_{3}=0\;|\;z_{3}}E(f(z_{1}, z_{2}, z_{3}, z_{4}))}{\nonumber \\}&&={\mbox{\rm Res}}_{1-z_{5}-z_{6}=0\;|\;z_{6}}
E(f(z_{1}, z_{2}, z_{3}, z_{4}))\frac{1-z_{1}-z_{3}}{1-z_{5}-z_{6}}.\end{aligned}$$
Using (\[residue4\]), the definition of the pairings $\langle\cdot,
\cdot\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}}$, Lemma \[zero\], and the fact that for $a_{1}, a_{2}, a_{3}\in {\mathcal{A}}$, $\{{\mathcal{Y}}_{a'_{1}a'_{2}; i}^{\prime; a'_{3}}\;|\;i=1, \dots,
N_{a'_{1}a'_{2}}^{a'_{3}}\}$ are the dual bases of $\{{\mathcal{Y}}_{a_{1}a_{2}; i}^{a_{3}}\;|\;i=1, \dots, N_{a_{1}a_{2}}^{a_{3}}\}$ with respect to the pairing $\langle\cdot,
\cdot\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}}$, we see that the right-hand side of (\[fusing-2\]) is equal to $$\begin{aligned}
\label{fusing-3}
\lefteqn{\sum_{a_{7}, a_{8}\in {\mathcal{A}}}\sum_{i, j, s, t}
F({\mathcal{Y}}_{a'_{1}a'_{5}; k}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{2}a'_{3};l}^{\prime; a'_{5}};
{\mathcal{Y}}_{a'_{7}a'_{3};i}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{1}a'_{2};j}^{\prime; a'_{7}})\cdot} {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad
\cdot
F({\mathcal{Y}}_{a_{1}a_{6}; m}^{a_{4}}\otimes {\mathcal{Y}}_{a_{2}a_{3};n}^{a_{6}};
{\mathcal{Y}}_{a_{8}a_{3};s}^{a_{4}}\otimes {\mathcal{Y}}_{a_{1}a_{2};t}^{a_{8}})\cdot{\nonumber \\}&&\quad\quad\cdot{\mbox{\rm Res}}_{1-z_{2}-z_{4}=0\;|\;z_{4}}
{\mbox{\rm Res}}_{1-z_{5}-z_{6}=0\;|\;z_{6}}
(1-z_{5}-z_{6})^{-1}(1-z_{2}-z_{4})^{-1} {\nonumber \\}&&\quad\quad\quad
E(\langle e^{L(1)}{\mathcal{Y}}_{a'_{7}a'_{3}; i}^{\prime; a'_{4}}(
(1-z_{2}-z_{4})^{L(0)}
\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \cdot
{\mathcal{Y}}_{a'_{1}a'_{2};j}^{\prime; a'_{7}}
((1-z_{5}-z_{6})^{L(0)}
\tilde{w}'_{a_{1}}, z_{5})
\tilde{w}'_{a_{2}}, z_{2})\tilde{w}'_{a_{3}}, {\nonumber \\}&&\quad\quad\quad\quad
e^{L(1)}{\mathcal{Y}}_{a_{8}a_{3}; s}^{a_{4}}
((1-z_{2}-z_{4})^{L(0)}
\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \cdot
{\mathcal{Y}}_{a_{1}a_{2};t}^{a_{8}}
((1-z_{5}-z_{6})^{L(0)}
\tilde{w}_{a_{1}},
z_{6})
\tilde{w}_{a_{2}}, z_{4})\tilde{w}_{a_{3}}\rangle){\nonumber \\}&&=\sum_{a_{7}, a_{8}\in {\mathcal{A}}}\sum_{i, j, s, t}
F({\mathcal{Y}}_{a'_{1}a'_{5}; k}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{2}a'_{3};l}^{\prime; a'_{5}};
{\mathcal{Y}}_{a'_{7}a'_{3};i}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{1}a'_{2};j}^{\prime; a'_{7}})\cdot {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad
\cdot
F({\mathcal{Y}}_{a_{1}a_{6}; m}^{a_{4}}\otimes {\mathcal{Y}}_{a_{2}a_{3};n}^{a_{6}};
{\mathcal{Y}}_{a_{7}a_{3};s}^{a_{4}}\otimes {\mathcal{Y}}_{a_{1}a_{2};t}^{a_{7}})\cdot{\nonumber \\}&&\quad\quad\cdot
{\mbox{\rm Res}}_{1-z_{5}-z_{6}=0\;|\;z_{6}}
{\mbox{\rm Res}}_{1-z_{2}-z_{4}=0\;|\;z_{4}}
(1-z_{5}-z_{6})^{-1}(1-z_{2}-z_{4})^{-1} {\nonumber \\}&&\quad\quad\quad
E(\langle e^{L(1)}{\mathcal{Y}}_{a'_{7}a'_{3}; i}^{\prime; a'_{4}}(
(1-z_{2}-z_{4})^{L(0)}
\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \cdot
{\mathcal{Y}}_{a'_{1}a'_{2};j}^{\prime; a'_{7}}
((1-z_{5}-z_{6})^{L(0)}
\tilde{w}'_{a_{1}}, z_{5})
\tilde{w}'_{a_{2}}, z_{2})\tilde{w}'_{a_{3}}, {\nonumber \\}&&\quad\quad\quad\quad
e^{L(1)}{\mathcal{Y}}_{a_{8}a_{3}; s}^{a_{4}}
((1-z_{2}-z_{4})^{L(0)}
\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad \cdot
{\mathcal{Y}}_{a_{1}a_{2};t}^{a_{7}}
((1-z_{5}-z_{6})^{L(0)}
\tilde{w}_{a_{1}},
z_{6})
\tilde{w}_{a_{2}}, z_{4})\tilde{w}_{a_{3}}\rangle){\nonumber \\}&&=\sum_{a_{7}\in {\mathcal{A}}}\sum_{i, j, s, t}
F({\mathcal{Y}}_{a'_{1}a'_{5}; k}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{2}a'_{3};l}^{\prime; a'_{5}};
{\mathcal{Y}}_{a'_{7}a'_{3};i}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{1}a'_{2};j}^{\prime; a'_{7}})\cdot {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad
\cdot
F({\mathcal{Y}}_{a_{1}a_{6}; m}^{a_{4}}\otimes {\mathcal{Y}}_{a_{2}a_{3};n}^{a_{6}};
{\mathcal{Y}}_{a_{7}a_{3};s}^{a_{4}}\otimes {\mathcal{Y}}_{a_{1}a_{2};t}^{a_{7}})\cdot{\nonumber \\}&&\quad\quad\cdot{\mbox{\rm Res}}_{1-z_{5}-z_{6}=0\;|\;z_{6}}
(1-z_{5}-z_{6})^{-1} \delta_{is}{\nonumber \\}&&\quad\quad\quad
E(\langle e^{L(1)}{\mathcal{Y}}_{a'_{1}a'_{2};j}^{\prime; a'_{7}}
((1-z_{5}-z_{6})^{L(0)}
\tilde{w}'_{a_{1}}, z_{5})
\tilde{w}'_{a_{2}}, {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
e^{L(1)}{\mathcal{Y}}_{a_{1}a_{2};t}^{a_{8}}
((1-z_{5}-z_{6})^{L(0)}
\tilde{w}_{a_{1}},
z_{6})
\tilde{w}_{a_{2}}\rangle){\nonumber \\}&&=\sum_{a_{7}\in {\mathcal{A}}}\sum_{i, j, s, t}
F({\mathcal{Y}}_{a'_{1}a'_{5}; k}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{2}a'_{3};l}^{\prime; a'_{5}};
{\mathcal{Y}}_{a'_{7}a'_{3};i}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{1}a'_{2};j}^{\prime; a'_{7}})\cdot {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad
\cdot
F({\mathcal{Y}}_{a_{1}a_{6}; m}^{a_{4}}\otimes {\mathcal{Y}}_{a_{2}a_{3};n}^{a_{6}};
{\mathcal{Y}}_{a_{8}a_{3};s}^{a_{4}}\otimes {\mathcal{Y}}_{a_{1}a_{2};t}^{a_{8}})\delta_{is}\delta_{jt}{\nonumber \\}&&=\sum_{a_{7}\in {\mathcal{A}}}\sum_{i, j}
F({\mathcal{Y}}_{a'_{1}a'_{5}; k}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{2}a'_{3};l}^{\prime; a'_{5}};
{\mathcal{Y}}_{a'_{7}a'_{3};i}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{1}a'_{2};j}^{\prime; a'_{7}})\cdot {\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad
\cdot
F({\mathcal{Y}}_{a_{1}a_{6}; m}^{a_{4}}\otimes {\mathcal{Y}}_{a_{2}a_{3};n}^{a_{6}};
{\mathcal{Y}}_{a_{7}a_{3};i}^{a_{4}}\otimes {\mathcal{Y}}_{a_{1}a_{2};j}^{a_{7}}).\end{aligned}$$ From (\[fusing-1\])–(\[fusing-3\]), we see that the right inverse of the matrix with entries $$F({\mathcal{Y}}_{a'_{1}a'_{5}; k}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{2}a'_{3};l}^{\prime; a'_{5}};
{\mathcal{Y}}_{a'_{7}a'_{3};i}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{1}a'_{2};j}^{\prime; a'_{7}})$$ is the transpose of the matrix with entries $$F({\mathcal{Y}}_{a_{1}a_{6}; m}^{a_{4}}\otimes {\mathcal{Y}}_{a_{2}a_{3};n}^{a_{6}};
{\mathcal{Y}}_{a_{7}a_{3};i}^{a_{4}}\otimes {\mathcal{Y}}_{a_{1}a_{2};j}^{a_{7}}).$$ Since for square matrices, right inverses are also left inverses, the proposition is proved.
For $a\in {\mathcal{A}}$, we use $\sqrt{F_{a}}$ to denote the square root $\sqrt{|F_{a}|}e^{\frac{i\arg F_{a}}{2}}$ of $F_{a}$. For $a_{1}, a_{2}, a_{3}\in {\mathcal{A}}$, consider the modified pairings $$\frac{\sqrt{F_{a_{3}}}}{\sqrt{F_{a_{1}}}\sqrt{F_{a_{2}}}}
\langle \cdot, \cdot\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}}.$$ These pairings give a nondegenerate bilinear form $(\cdot, \cdot)_{{\mathcal{V}}}$ on ${\mathcal{V}}$. For any $\sigma\in S_{3}$, $\{\sigma({\mathcal{Y}}_{a_{1}a_{2}; i}^{a_{3}})\;
|\;i=\dots, N_{a_{1}a_{2}}^{a_{3}}\}$ is a basis of $\sigma({\mathcal{V}}_{a_{1}a_{2}}^{a_{3}})$.
We have:
\[skew\] The nondegenerate bilinear form $(\cdot, \cdot)_{{\mathcal{V}}}$ is invariant with respect to the action of $S_{3}$ on ${\mathcal{V}}$, that is, for $a_{1}, a_{2}, a_{3}\in {\mathcal{A}}$, $\sigma\in S_{3}$, ${\mathcal{Y}}_{1}\in
{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}$ and ${\mathcal{Y}}_{2}\in {\mathcal{V}}_{a'_{1}a'_{2}}^{a'_{3}}$, $$(\sigma({\mathcal{Y}}_{1}), \sigma({\mathcal{Y}}_{2}))_{{\mathcal{V}}}=
({\mathcal{Y}}_{1}, {\mathcal{Y}}_{2})_{{\mathcal{V}}}.$$ Equivalently, for $a_{1}, a_{2}, a_{3}\in {\mathcal{A}}$, $$\left\{\left.\frac{\sqrt{F_{a_{1}}}\sqrt{F_{a_{2}}}
\sqrt{F_{a_{\sigma^{-1}(3)}}}}
{\sqrt{F_{a_{\sigma^{-1}(1)}}}\sqrt{F_{a_{\sigma^{-1}(2)}}}\sqrt{F_{a_{3}}}}
\sigma({\mathcal{Y}}_{a'_{1}a'_{2}; i}^{\prime; a'_{3}})\;\right|\;i=\dots,
N_{a_{1}a_{2}}^{a_{3}}\right\}$$ is the dual basis of $\{\sigma({\mathcal{Y}}_{a_{1}a_{2}; i}^{a_{3}})\;
|\;i=\dots, N_{a_{1}a_{2}}^{a_{3}}\}$.
[[*Proof.*]{}]{}The equivalence of the first conclusion and the second conclusion is clear.
We first prove the result for $\sigma=\sigma_{12}$. In this case, we need to show that $\{\sigma_{12}({\mathcal{Y}}_{a'_{1}a'_{2}; j}^{a'_{3}})\;|\;
i=\dots, N_{a_{1}a_{2}}^{a_{3}}\}$ is the dual basis of $\{\sigma_{12}({\mathcal{Y}}_{a_{1}a_{2}; i}^{a_{3}})\;
|\;i=\dots, N_{a_{1}a_{2}}^{a_{3}}\}$. For $i, j=1, \dots, N_{a_{1}a_{2}}^{a_{3}}$, by (\[inner-fusing\]), we have $$\label{skew-1}
\langle \sigma_{12}({\mathcal{Y}}_{a'_{1}a'_{2}; i}^{\prime; a'_{3}}),
\sigma_{12}({\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}})\rangle=
F(\sigma_{23}(\sigma_{12}({\mathcal{Y}}_{a'_{1}a'_{2}; i}^{\prime; a'_{3}}))
\otimes \sigma_{12}({\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}}); {\mathcal{Y}}_{ea_{1}; 1}^{a_{1}}\otimes
{\mathcal{Y}}_{a'_{2}a_{2}; 1}^{e}).$$ By Proposition 3.4 in [@H7], the right-hand side of (\[skew\]) is equal to $$\begin{aligned}
\label{skew-2}
&F(\sigma_{132}(\sigma_{12}({\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}}))
\otimes \sigma_{123}(\sigma_{23}(\sigma_{12}
({\mathcal{Y}}_{a'_{1}a'_{2}; i}^{\prime; a'_{3}})));
\sigma_{123}({\mathcal{Y}}_{a'_{2}a_{2}; 1}^{e})\otimes
\sigma_{132}({\mathcal{Y}}_{ea_{1}; 1}^{a_{1}}))&{\nonumber \\}&=F(\sigma_{23}({\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}})
\otimes {\mathcal{Y}}_{a'_{1}a'_{2}; i}^{\prime; a'_{3}};
{\mathcal{Y}}_{ea'_{2}; 1}^{a'_{2}}\otimes
{\mathcal{Y}}_{a_{1}a'_{1}; 1}^{e})&{\nonumber \\}&\;\;=F(\sigma_{23}({\mathcal{Y}}_{a''_{1}a''_{2}; j}^{a''_{3}})
\otimes {\mathcal{Y}}_{a'_{1}a'_{2}; i}^{\prime; a'_{3}};
{\mathcal{Y}}_{ea'_{2}; 1}^{a'_{2}}\otimes
{\mathcal{Y}}_{a''_{1}a'_{1}; 1}^{e}).&\end{aligned}$$ By (\[inner-fusing\]) again, the right-hand side of (\[skew-2\]) is equal to $$\langle {\mathcal{Y}}_{a''_{1}a''_{2}; j}^{a''_{3}},
{\mathcal{Y}}_{a'_{1}a'_{2}; i}^{\prime; a'_{3}}\rangle
=\delta_{ij},$$ proving the case of $\sigma=\sigma_{12}$.
Next we prove the result for $\sigma=\sigma_{23}$. We need to find the relation between the matrices $$\langle \sigma_{23}({\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}}),
\sigma_{23}({\mathcal{Y}}_{a_{1}a_{2}'; j}^{a_{3}})
\rangle_{{\mathcal{V}}_{a_{1}a'_{3}}^{a'_{2}}}$$ and $$\langle {\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}},
{\mathcal{Y}}_{a_{1}a_{2}'; j}^{a_{3}}\rangle_{{\mathcal{V}}_{a_{1}a_{2}}^{a_{3}}}.$$ By definition, we need to find the relation between the matrices $$\label{form-matrix}
F({\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}}
\otimes \sigma_{23}({\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}});
{\mathcal{Y}}_{ea'_{3}; 1}^{a'_{3}}\otimes
{\mathcal{Y}}_{a'_{1}a_{1}; 1}^{e})$$ and $$\label{form-matrix-1}
F(\sigma_{23}({\mathcal{Y}}_{a'_{1}a'_{2}; i}^{a'_{3}; (2)})
\otimes {\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}; (1)}; {\mathcal{Y}}_{ea_{2}; 1}^{a_{2}}\otimes
{\mathcal{Y}}_{a'_{1}a_{1}; 1}^{e}).$$ From (4.9) in [@H7] (or (\[formula1\])), we see that the inverse of the matrix (\[form-matrix\]) is $$\label{form-matrix-inv}
\frac{F({\mathcal{Y}}_{a'_{3}e; 1}^{a'_{3}}\otimes {\mathcal{Y}}_{a_{2}a'_{2}; 1}^{e};
{\mathcal{Y}}_{a'_{1}a'_{2}; k}^{a'_{3}}\otimes
\sigma_{132}(\sigma_{23}({\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}})))}
{F_{a'_{3}}}.$$ By Proposition 3.4 in [@H7] and the fact $F_{a'}=F_{a}$ for $a\in {\mathcal{A}}$, (\[form-matrix-inv\]) is equal to $$\begin{aligned}
\lefteqn{\frac{F({\mathcal{Y}}_{a'_{2}e; 1}^{a'_{2}}\otimes {\mathcal{Y}}_{a_{3}a'_{3}; 1}^{e};
\sigma_{23}({\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}}) \otimes
\sigma_{132}({\mathcal{Y}}_{a'_{1}a'_{2}; k}^{a'_{3}}))}
{F_{a_{3}}}}{\nonumber \\}&&=\frac{F_{a_{2}}}{F_{a_{3}}}
\frac{F({\mathcal{Y}}_{a'_{2}e; 1}^{a'_{2}}\otimes {\mathcal{Y}}_{a_{3}a'_{3}; 1}^{e};
\sigma_{23}({\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}}) \otimes
\sigma_{132}({\mathcal{Y}}_{a'_{1}a'_{2}; k}^{a'_{3}}))}
{F_{a'_{2}}},\end{aligned}$$ which by (4.9) in [@H7] (or (\[formula1\])) again is equal to $\frac{F_{a_{2}}}{F_{a_{3}}}$ times the inverse of (\[form-matrix-1\]). So the inverse of the matrix (\[form-matrix\]) is equal to $\frac{F_{a_{2}}}{F_{a_{3}}}$ times the inverse of (\[form-matrix-1\]). Thus the matrix (\[form-matrix\]) is equal to $\frac{F_{a_{3}}}{F_{a_{2}}}$ times the matrix (\[form-matrix-1\]), or equivalently, $$(\sigma_{23}({\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}}),
\sigma_{23}({\mathcal{Y}}_{a'_{1}a'_{2}; k}^{a'_{3}}))_{{\mathcal{V}}}
=({\mathcal{Y}}_{a_{1}a_{2}; j}^{a_{3}},
{\mathcal{Y}}_{a'_{1}a'_{2}; k}^{a'_{3}})_{{\mathcal{V}}}.$$
Since $S_{3}$ is generated by $\sigma_{12}$ and $\sigma_{23}$, the conclusion of the proposition follows.
We are ready to construct a full field algebra using the bases of intertwining operators we have chosen. Let $$F=\oplus_{a\in {\mathcal{A}}} W^{a}\otimes W^{a'}.$$ For $w_{a_{1}}\in W^{a_{1}}$, $w_{a_{2}}\in W^{a_{2}}$, $w_{a'_{1}}\in W^{a'_{1}}$ and $w_{a'_{2}}\in W^{a'_{2}}$, we define $$\begin{aligned}
\lefteqn{\mathbb{Y}((w_{a_{1}}\otimes w_{a'_{1}}), z, \zeta)
(w_{a_{2}}\otimes w_{a'_{2}})}{\nonumber \\}&&=\sum_{a_{3}\in {\mathcal{A}}}\sum_{p=1}^{N_{a_{1}a_{2}}^{a_{3}}}
{\mathcal{Y}}_{a_{1}a_{2}; p}^{a_{3}}(w_{a_{1}}, z) w_{a_{2}}
\otimes {\mathcal{Y}}_{a'_{1}a'_{2}; p}^{\prime; a'_{3}}
(w_{a'_{1}}, \zeta)w_{a'_{2}}.\end{aligned}$$
The quadruple $(F, \mathbb{Y}, {\mathbf{1}}\otimes {\mathbf{1}}, \omega\otimes {\mathbf{1}},
{\mathbf{1}}\otimes \omega)$ is a conformal full field algebra over $V\otimes V$.
[[*Proof.*]{}]{} The identity property, the creation property and the single-valuedness property are clear. We prove the associativity and the skew-symmetry here.
We prove associativity first. For $a_{1}, a_{2}\in {\mathcal{A}}$, $w_{a_{1}}\in W^{a_{1}}$, $w_{a_{2}}\in W^{a_{2}}$, $w'_{a_{1}}\in (W^{a_{1}})'$, $w'_{a_{2}}\in (W^{a_{2}})'$, using the associativity of intertwining operators and Proposition \[fusing\], we have $$\begin{aligned}
\lefteqn{{\mathbb{Y}}((w_{a_{1}}\otimes w'_{a_{1}}), z_{1}, \zeta_{1})
{\mathbb{Y}}((w_{a_{2}}\otimes w'_{a_{2}}), z_{2}, \zeta_{2})}{\nonumber \\}&&=\sum_{a_{3}, a_{4}, a_{5}\in{\mathcal{A}}}\sum_{p=1}^{N_{a_{1}a_{5}}^{a_{4}}}
\sum_{q=1}^{N_{a_{2}a_{3}}^{a_{5}}}({\mathcal{Y}}_{a_{1}a_{5}; p}^{a_{4}}
(w_{a_{1}}, z_{1})
{\mathcal{Y}}_{a_{2}a_{3}; q}^{a_{5}}(w_{a_{2}}, z_{2})){\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\otimes
({\mathcal{Y}}_{a'_{1}a'_{5}; p}^{\prime; a'_{4}}(w'_{a_{1}}, \zeta_{1})
{\mathcal{Y}}_{a'_{2}a'_{3}; q}^{\prime; a'_{5}}(w'_{a_{2}}, \zeta_{2})){\nonumber \\}&&=\sum_{a_{3}, a_{4}, a_{5}, a_{6},a_{7}\in{\mathcal{A}}}
\sum_{p=1}^{N_{a_{1}a_{5}}^{a_{4}}}
\sum_{q=1}^{N_{a_{2}a_{3}}^{a_{5}}}\sum_{m=1}^{N_{a_{6}a_{3}}^{a_{5}}}
\sum_{n=1}^{N_{a_{1}a_{2}}^{a_{6}}}\sum_{k=1}^{N_{a'_{7}a'_{3}}^{a'_{5}}}
\sum_{l=1}^{N_{a'_{1}a'_{2}}^{a'_{7}}}{\nonumber \\}&&\quad\quad\quad\quad\quad \cdot
F({\mathcal{Y}}_{a_{1}a_{5}; p}^{a_{4}}\otimes {\mathcal{Y}}_{a_{2}a_{3}; q}^{a_{5}};
{\mathcal{Y}}_{a_{6}a_{3}; m}^{a_{4}}\otimes {\mathcal{Y}}_{a_{1}a_{2}; n}^{a_{6}})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\cdot
F({\mathcal{Y}}_{a'_{1}a'_{5}; p}^{\prime; a'_{4}}\otimes
{\mathcal{Y}}_{a'_{2}a'_{3}; q}^{\prime; a'_{5}};
{\mathcal{Y}}_{a'_{7}a'_{3}; k}^{a'_{4}}\otimes {\mathcal{Y}}_{a'_{1}a'_{2}; l}^{a'_{7}})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad
\cdot (({\mathcal{Y}}_{a_{6}a_{3}; m}^{a_{4}}
({\mathcal{Y}}_{a_{1}a_{2}; n}^{a_{6}}(w_{a_{1}}, z_{1}-z_{2})
w_{a_{2}}, z_{2})){\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\otimes
({\mathcal{Y}}_{a'_{7}a'_{3}; k}^{\prime; a'_{4}}
({\mathcal{Y}}_{a'_{1}a'_{2}; l}^{\prime; a'_{7}}(w'_{a_{1}},
\zeta_{1}-\zeta_{2})w'_{a_{2}}, \zeta_{2}))){\nonumber \\}&&=\sum_{a_{3}, a_{4}, a_{6}, a_{7}\in{\mathcal{A}}}\sum_{m=1}^{N_{a_{6}a_{3}}^{a_{5}}}
\sum_{n=1}^{N_{a_{1}a_{2}}^{a_{6}}}\sum_{k=1}^{N_{a'_{7}a'_{3}}^{a'_{5}}}
\sum_{l=1}^{N_{a'_{1}a'_{2}}^{a'_{7}}}
\delta_{a_{6}a_{7}}\delta_{mk}\delta_{nl}\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad
\cdot
({\mathcal{Y}}_{a_{6}a_{3}; m}^{a_{4}}({\mathcal{Y}}_{a_{1}a_{2}; n}^{a_{6}}(w_{a_{1}}, z_{1}-z_{2})
w_{a_{2}}, z_{2})){\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\otimes
({\mathcal{Y}}_{a'_{7}a'_{3}; k}^{\prime; a'_{4}}
({\mathcal{Y}}_{a'_{1}a'_{2}; l}^{\prime; a'_{7}}(w'_{a_{1}},
\zeta_{1}-\zeta_{2})w'_{a_{2}}, \zeta_{2})){\nonumber \\}&&=\sum_{a_{3}, a_{4}, a_{6}\in{\mathcal{A}}}\sum_{m=1}^{N_{a_{6}a_{3}}^{a_{5}}}
\sum_{n=1}^{N_{a_{1}a_{2}}^{a_{6}}}
({\mathcal{Y}}_{a_{6}a_{3}; m}^{a_{4}}({\mathcal{Y}}_{a_{1}a_{2}; n}^{a_{6}}(w_{a_{1}}, z_{1}-z_{2})
w_{a_{2}}, z_{2})){\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\otimes
({\mathcal{Y}}_{a'_{6}a'_{3}; m}^{\prime; a'_{4}}
({\mathcal{Y}}_{a'_{1}a'_{2}; n}^{\prime; a'_{6}}(w'_{a_{1}},
\zeta_{1}-\zeta_{2})w'_{a_{2}}, \zeta_{2})){\nonumber \\}&&={\mathbb{Y}}({\mathbb{Y}}((w_{a_{1}}\otimes w'_{a_{1}}), z_{1}-z_{2}, \zeta_{1}-\zeta_{2})
(w_{a_{2}}\otimes w'_{a_{2}}), z_{2}, \zeta_{2}).\end{aligned}$$
We now prove the skew-symmetry. By Proposition \[skew\], $\{\sigma_{12}({\mathcal{Y}}_{a'_{1}a'_{2}; i}^{\prime; a'_{3}})\;|\; i=1, \dots,
N_{a_{1}a_{2}}^{a_{3}}\}$ is the dual basis of $\{\sigma_{12}({\mathcal{Y}}_{a_{1}a_{2}; i}^{a_{3}})\;|\; i=1, \dots,
N_{a_{1}a_{2}}^{a_{3}}\}$. Thus for $a_{1}, a_{2}\in {\mathcal{A}}$, $w_{a_{1}}\in W^{a_{1}}$, $w_{a_{2}}\in W^{a_{2}}$, $w'_{a_{1}}\in (W^{a_{1}})'$, $w'_{a_{2}}\in (W^{a_{2}})'$, we have $$\begin{aligned}
\lefteqn{{\mathbb{Y}}((w_{a_{2}}\otimes w'_{a_{2}}), z, \zeta)
(w_{a_{1}}\otimes w'_{a_{1}})}{\nonumber \\}&&=\sum_{a_{3}\in {\mathcal{A}}}\sum_{p=1}^{N_{a_{2}a_{1}}^{a_{3}}}
\sigma_{12}({\mathcal{Y}}_{a_{1}a_{2}; p}^{a_{3}})(w_{a_{2}}, z) w_{a_{1}}
\otimes \sigma_{12}({\mathcal{Y}}_{a'_{1}a'_{2}; p}^{\prime; a'_{3}})
(w_{a'_{2}}, \zeta)w_{a'_{1}}{\nonumber \\}&&=\sum_{a_{3}\in {\mathcal{A}}}\sum_{p=1}^{N_{a_{1}a_{2}}^{a_{3}}}
e^{-\pi i\Delta({\mathcal{Y}}_{a_{1}a_{2}; p}^{a_{3}})}
e^{zL(-1)}{\mathcal{Y}}_{a_{1}a_{2}; p}^{a_{3}}(w_{a_{1}}, e^{\pi i}z) w_{a_{2}}{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\otimes e^{\pi i\Delta({\mathcal{Y}}_{a'_{1}a'_{2}; p}^{a'_{3}})}
e^{\zeta L(-1)}{\mathcal{Y}}_{a'_{1}a'_{2}; p}^{\prime; a'_{3}})
(w_{a'_{1}}, e^{-\pi i}\zeta)w_{a'_{2}}{\nonumber \\}&&=(e^{zL(-1)}\otimes e^{\zeta L(-1)})\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\sum_{a_{3}\in {\mathcal{A}}}
\sum_{p=1}^{N_{a_{1}a_{2}}^{a_{3}}}
{\mathcal{Y}}_{a_{1}a_{2}; p}^{a_{3}}(w_{a_{1}}, e^{\pi i}z) w_{a_{2}}
\otimes {\mathcal{Y}}_{a'_{1}a'_{2}; p}^{\prime; a'_{3}}
(w_{a'_{1}}, e^{-\pi i}\zeta)w_{a'_{2}}{\nonumber \\}&&=(e^{zL(-1)}\otimes e^{\zeta L(-1)}){\mathbb{Y}}((w_{a_{1}}\otimes w'_{a_{1}}),
e^{\pi i}z, e^{-\pi i}\zeta)(w_{a_{2}}\otimes w'_{a_{2}}).\end{aligned}$$
[A nondegenerate bilinear form $(\cdot, \cdot)$ on a conformal full field algebra $$(F, m, {\mathbf{1}}, \omega^{L}, \omega^{R})$$ is said to be [*invariant*]{} if for $u, v, w\in F$, $$\begin{aligned}
\lefteqn{(\mathbb{Y}(u; z, \bar{z})v, w)}{\nonumber \\}&&=(v, \mathbb{Y}(e^{zL^{L}(1)+\bar{z}L^{R}(1)}
e^{\pi iL^{L}(0)-\pi iL^{R}(0)}z^{-2L^{L}(0)}\bar{z}^{-2L^{R}(0)}
u; z^{-1}, \bar{z}^{-1})w).\end{aligned}$$]{}
The conformal full field algebra $F$ we constructed above has a natural nondegenerate bilinear form $(\cdot, \cdot)_{F}: F\otimes F\to {\mathbb{C}}$ given by $$((w_{a_{1}}\otimes w'_{a_{1}}), (w_{a_{2}}\otimes w'_{a_{2}}))_{F}
=\left\{\begin{array}{ll}0&a_{1}\ne a'_{2}\\
F_{a_{1}}\langle w_{a_{1}}, w_{a_{2}}\rangle\langle w'_{a_{1}}, w'_{a_{2}}
\rangle&a_{1}=a'_{2}
\end{array}\right.$$ for $a_{1}, a_{2}\in {\mathcal{A}}$, $w_{a_{1}}\in W^{a_{1}}$, $w_{a_{2}}\in W^{a_{2}}$, $w'_{a_{1}}\in (W^{a_{1}})'$, $w'_{a_{2}}\in (W^{a_{2}})'$. We have:
The nondegenerate bilinear form $(\cdot, \cdot)_{F}$ is invariant.
[[*Proof.*]{}]{}For $a_{1}, a_{2}, a_{3}\in {\mathcal{A}}$, $w_{a_{1}}\in W^{a_{1}}$, $w_{a_{2}}\in W^{a_{2}}$, $w_{a_{3}}\in W^{a_{3}}$, $w'_{a_{1}}\in (W^{a_{1}})'$, $w'_{a_{2}}\in (W^{a_{2}})'$, $w'_{a_{3}}\in (W^{a_{3}})'$, using Proposition \[skew\] for the case $\sigma=\sigma_{23}$, we have $$\begin{aligned}
\lefteqn{({\mathcal{Y}}((w_{a_{1}}\otimes w'_{a_{1}}), z, \zeta)(w_{a_{2}}
\otimes w'_{a_{2}}),
(w_{a_{3}}\otimes w'_{a_{3}}))_{F}}{\nonumber \\}&&=\sum_{a_{4}\in {\mathcal{A}}}\sum_{p=1}^{N_{a_{1}a_{2}}^{a_{4}}}
(({\mathcal{Y}}_{a_{1}a_{2}; p}^{a_{4}}(w_{a_{1}}, z) w_{a_{2}}
\otimes {\mathcal{Y}}_{a'_{1}a'_{2}; p}^{\prime; a'_{4}}
(w'_{a_{1}}, \zeta)w'_{a_{2}}), (w_{a_{3}}\otimes w'_{a_{3}}))_{F}{\nonumber \\}&&=\sum_{p=1}^{N_{a_{1}a_{2}}^{a_{3}}}
F_{a_{3}}
\langle {\mathcal{Y}}_{a_{1}a_{2}; p}^{a_{3}}(w_{a_{1}}, z) w_{a_{2}}, w_{a_{3}}\rangle
\langle {\mathcal{Y}}_{a'_{1}a'_{2}; p}^{\prime; a'_{3}}
(w'_{a_{1}}, \zeta)w'_{a_{2}}, w'_{a_{3}}\rangle{\nonumber \\}&&=\sum_{p=1}^{N_{a_{1}a_{2}}^{a_{3}}}
F_{a_{3}}\langle \sigma_{23}(\sigma_{23}({\mathcal{Y}}_{a_{1}a_{2}; p}^{a_{3}})
(w_{a_{1}}, z) w_{a_{2}}, w_{a_{3}}\rangle
\langle \sigma_{23}(\sigma_{23}({\mathcal{Y}}_{a'_{1}a'_{2}; p}^{\prime; a'_{3}})
(w'_{a_{1}}, \zeta)w'_{a_{2}}, w'_{a_{3}}\rangle{\nonumber \\}&&=\sum_{p=1}^{N_{a_{1}a_{2}}^{a_{3}}}
F_{a_{3}}\langle w_{a_{2}}, e^{\pi ih_{a_{1}}}\sigma_{23}({\mathcal{Y}}_{a_{1}a_{2}; p}^{a_{3}})
(e^{zL(1)}e^{-\pi i L(0)}z^{-2L(0)}w_{a_{1}}, z^{-1})w_{a_{3}}\rangle\cdot{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\cdot
\langle w'_{a_{2}}, e^{-\pi ih_{a_{1}}}
\sigma_{23}({\mathcal{Y}}_{a'_{1}a'_{2}; p}^{\prime; a'_{3}})
(e^{\zeta L(1)}e^{\pi i L(0)}\zeta^{-2L(0)}w'_{a_{1}}, \zeta^{-1})
w'_{a_{3}}\rangle{\nonumber \\}&&=\sum_{p=1}^{N_{a_{1}a_{2}}^{a_{3}}}
F_{a_{2}}\langle w_{a_{2}}, \sigma_{23}({\mathcal{Y}}_{a_{1}a_{2}; p}^{a_{3}})
(e^{zL(1)}e^{-\pi i L(0)}z^{-2L(0)}w_{a_{1}}, z^{-1})w_{a_{3}}\rangle\cdot{\nonumber \\}&&\quad\quad\quad\quad\cdot
\left\langle w'_{a_{2}}, \left(\frac{F_{a_{3}}}{F_{a_{2}}}
\sigma_{23}({\mathcal{Y}}_{a'_{1}a'_{2}; p}^{\prime; a'_{3}})\right)
(e^{\zeta L(1)}e^{\pi i L(0)}\zeta^{-2L(0)}w'_{a_{1}}, \zeta^{-1})
w'_{a_{3}}\right\rangle{\nonumber \\}&&=\sum_{p=1}^{N_{a_{1}a_{2}}^{a_{3}}}
((w_{a_{2}}\otimes w'_{a_{2}}), (\sigma_{23}({\mathcal{Y}}_{a_{1}a_{2}; p}^{a_{3}})
(e^{zL(1)}e^{-\pi i L(0)}z^{-2L(0)}w_{a_{1}}, z^{-1})w_{a_{3}}{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\otimes
\left(\frac{F_{a_{3}}}{F_{a_{2}}}
\sigma_{23}({\mathcal{Y}}_{a'_{1}a'_{2}; p}^{\prime; a'_{3}})\right)
(e^{\zeta L(1)}e^{\pi i L(0)}\zeta^{-2L(0)}w'_{a_{1}}, \zeta^{-1})
w'_{a_{3}}))_{F}{\nonumber \\}&&=\sum_{a_{4}\in {\mathcal{A}}}\sum_{p=1}^{N_{a_{1}a_{4}}^{a_{3}}}
((w_{a_{2}}\otimes w'_{a_{2}}), (\sigma_{23}({\mathcal{Y}}_{a_{1}a_{4}; p}^{a_{3}})
(e^{zL(1)}e^{-\pi i L(0)}z^{-2L(0)}w_{a_{1}}, z^{-1})w_{a_{3}}{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad\otimes
\left(\frac{F_{a_{3}}}{F_{a_{2}}}
\sigma_{23}({\mathcal{Y}}_{a'_{1}a'_{4}; p}^{\prime; a'_{3}})\right)
(e^{\zeta L(1)}e^{\pi i L(0)}\zeta^{-2L(0)}w'_{a_{1}}, \zeta^{-1})
w'_{a_{3}}))_{F}{\nonumber \\}&&=((w_{a_{2}}\otimes w'_{a_{2}}), {\mathcal{Y}}((e^{zL(1)}
e^{-\pi i L(0)}z^{-2L(0)}w_{a_{1}}
{\nonumber \\}&&\quad\quad\quad\quad\quad\quad\quad
\otimes (e^{\zeta L(1)}e^{\pi i L(0)}\zeta^{-2L(0)}w'_{a_{1}}), z^{-1},
\zeta^{-1})
(w_{a_{3}}\otimes w'_{a_{3}}))_{F},\end{aligned}$$ proving the invariance.
[KWak2]{}
T. Abe, G. Buhl and C. Dong, Rationality, regularity and $C_2$-cofiniteness. [*Trans. Amer. Math. Soc.*]{} [**356**]{} (2004), no. 8, 3391-3402.
B. Bakalov and A. Kirillov, Jr., [*Lectures on tensor categories and modular functors*]{}, University Lecture Series, Vol. 21, Amer. Math. Soc., Providence, RI, 2001.
A. A. Belavin, A. M. Polyakov and A. B. Zamolodchikov, Infinite conformal symmetries in two-dimensional quantum field theory, (1984), 333–380.
J. S. Birman, Braids, links, and mapping class groups, Annals of Mathematics Studies, Vol. 82, Princeton University Press, Princeton, 1974.
R. E. Borcherds, Vertex algebras, Kac-Moody algebras, and the Monster, (1986), 3068–3071.
C. Dong, G. Mason and Y. Zhu, Discrete series of the Virasoro algebra and the moonshine module, in [*Algebraic Groups and Their Generalizations: Quantum and infinite-dimensional Methods, Proc. 1991 American Math. Soc. Summer Research Institute*]{}, ed. by W. J. Haboush and B. J. Parshall, Proc. Symp. Pure Math. [**56**]{}, Part 2, Amer. Math. Soc., Providence, 1994, 295-316.
G. Felder, J. Fröhlich, J. Fuchs and C. Schweigert, Correlation functions and boundary conditions in rational conformal field theory and three-dimensional topology, [*Compositio Math.*]{} [**131**]{} (2002), 189–237.
I. B. Frenkel, Y.-Z. Huang and J. Lepowsky, On axiomatic approaches to vertex operator algebras and modules, preprint, 1989; [*Memoirs Amer. Math. Soc.*]{} [**104**]{}, 1993.
I. B. Frenkel, J. Lepowsky, and A. Meurman, [*Vertex operator algebras and the Monster*]{}, Pure and Appl. Math., Vol. 134, Academic Press, New York, 1988.
J. Fjelstad, J. Fuchs, I. Runkel and C. Schweigert, TFT construction of RCFT correlators V: Proof of modular invariance and factorisation, to appear; hep-th/0509040.
J. Fuchs, I. Runkel and C. Schweigert, Conformal correlation functions, Frobenius algebras and triangulations, [*Nucl. Phys.*]{} [**B624**]{} (2002), 452–468.
J. Fuchs, I. Runkel and C. Schweigert, TFT construction of RCFT correlators I: Partition functions, [*Nucl. Phys.*]{} [**B646**]{} (2002), 353–497.
J. Fuchs, I. Runkel and C. Schweigert, TFT construction of RCFT correlators, IV: Structure constants and correlation functions, [*Nucl. Phys.*]{} [**B715**]{} (2005), 539–638.
Y.-Z. Huang, A theory of tensor products for module categories for a vertex operator algebra, IV, [*J. Pure Appl. Alg.*]{} [**100**]{} (1995), 173-216.
Y.-Z. Huang, Virasoro vertex operator algebras, (nonmeromorphic) operator product expansion and the tensor product theory, [*J. Alg.*]{} [**182**]{} (1996), 201–234.
Y.-Z. Huang, Intertwining operator algebras, genus-zero modular functors and genus-zero conformal field theories, in: [*Operads: Proceedings of Renaissance Conferences*]{}, ed. J.-L. Loday, J. Stasheff, and A. A. Voronov, Contemporary Math., Vol. 202, Amer. Math. Soc., Providence, 1997, 335–355.
Y.-Z. Huang, Genus-zero modular functors and intertwining operator algebras, [*Internat, J. Math.*]{} [**9**]{} (1998), 845-863.
Y.-Z. Huang, Generalized rationality and a “Jacobi identity” for intertwining operator algebras, [*Selecta Math. (N.S.)*]{}, [**6**]{} (2000), 225–267.
Y.-Z. Huang, Vertex operator algebras, the Verlinde conjecture and modular tensor categories, [*Proc. Natl. Acad. Sci. USA*]{} [**102**]{} (2005), 5352–5356.
Y.-Z. Huang, Differential equations and intertwining operators, [*Comm. Contemp. Math.*]{} [**7**]{} (2005), 375–400.
Y.-Z. Huang, Differential equations, duality and modular invariance, [*Comm. Contemp. Math.*]{}[**7**]{} (2005), 649-706.
Y.-Z. Huang, Vertex operator algebras and the Verlinde conjecture, to appear; math.QA/0406291.
Y.-Z. Huang, Rigidity and modularity of vertex tensor categories, to appear; math.QA/0502533.
A. Kapustin and D. Orlov, Vertex algebras, mirror symmetry, and D-branes: The case of complex tori, [*Comm. Math. Phys.*]{} [**233**]{} (2003), 79–136.
L. Kong, A mathematical study of open-closed conformal field theories, Ph.D. thesis, Rutgers University, 2005.
H.S. Li, Some finiteness properties of regular vertex operator algebras, [*J. Algebra*]{}, [**212**]{}, 1999, 495-514.
G. Moore and N. Seiberg, Polynomial equations for rational conformal field theories, [*Phys. Lett.*]{} [**B212**]{} (1988), 451–460.
G. Moore and N. Seiberg, Classical and quantum conformal field theory, [*Comm. Math. Phys.*]{} [**123**]{} (1989), 177–254.
G. Moore and N. Seiberg, Naturality in conformal field theory, [*Nucl. Phys.*]{} [**B313**]{} (1989), 16–40.
M. Rosellen, OPE-algebras, Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn, Bonn, 2002, Bonner Mathematische Schriften \[Bonn Mathematical Publications\], Vol. 352, Universität Bonn, Mathematisches Institut, Bonn, 2002.
M. Rosellen, OPE-algebras and their modules. [*Int. Math. Res. Not.*]{} [**2005**]{} (2005), 433-447.
G. Segal, The definition of conformal field theory, in: [*Differential geometrical methods in theoretical physics (Como, 1987)*]{}, NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., 250, Kluwer Acad. Publ., Dordrecht, 1988, 165–171.
G. Segal, Two-dimensional conformal field theories and modular functors, in: [*Proceedings of the IXth International Congress on Mathematical Physics, Swansea, 1988*]{}, Hilger, Bristol, 1989, 22–37.
G. Segal, The definition of conformal field theory, preprint, 1988; also in: [*Topology, geometry and quantum field theory*]{}, ed. U. Tillmann, London Math. Soc. Lect. Note Ser., Vol. 308. Cambridge University Press, Cambridge, 2004, 421–577.
H. Tsukada, String path integral realization of vertex operator algebras, [*Mem. Amer. Math. Soc.*]{} [**91**]{}, 1991.
V. G. Turaev, [*Quantum invariants of knots and $3$-manifolds*]{}, de Gruyter Studies in Math., Vol. 18, Walter de Gruyter, Berlin, 1994.
[Department of Mathematics, Rutgers University, 110 Frelinghuysen Rd., Piscataway, NJ 08854-8019]{}
[*E-mail address*]{}: [email protected]
[Department of Mathematics, Rutgers University, 110 Frelinghuysen Rd., Piscataway, NJ 08854-8019]{}
and
[Max Planck Institute for Mathematics in the Sciences, Inselstrasse 22, D-04103, Leipzig, Germany]{} (current address)
[*E-mail address*]{}: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Equal-arm interferometric detectors of gravitational radiation allow phase measurements many orders of magnitude below the intrinsic phase stability of the laser injecting light into their arms. This is because the noise in the laser light is common to both arms, experiencing exactly the same delay, and thus cancels when it is differenced at the photo detector. In this situation, much lower level secondary noises then set overall performance. If, however, the two arms have different lengths (as will necessarily be the case with space-borne interferometers), the laser noise experiences different delays in the two arms and will hence not directly cancel at the detector. In order to solve this problem, a technique involving heterodyne interferometry with unequal arm lengths and independent phase-difference readouts has been proposed. It relies on properly time-shifting and linearly combining independent Doppler measurements, and for this reason it has been called Time-Delay Interferometry (or TDI). This article provides an overview of the theory and mathematical foundations of TDI as it will be implemented by the forthcoming space-based interferometers such as the Laser Interferometer Space Antenna (LISA) mission. We have purposely left out from this first version of our “Living Review” article on TDI all the results of more practical and experimental nature, as well as all the aspects of TDI that the data analysts will need to account for when analyzing the LISA TDI data combinations. Our forthcoming “second edition” of this review paper will include these topics.'
author:
- Massimo Tinto
- 'Sanjeev V. Dhurandhar'
title: 'Time-Delay Interferometry'
---
Introduction {#Introduction}
============
Breakthroughs in modern technology have made possible the construction of extremely large interferometers both on ground and in space for the detection and observation of gravitational waves (GW). Several ground based detectors are being constructed or are already operational around the globe. These are the LIGO and VIRGO interferometers, which have arm-lengths of 4 km and 3 km respectively, and the GEO and TAMA interferometers with arm-lengths of 600 m and 300 m respectively. These detectors will operate in the high frequency range of GW of $\sim 1$ Hz to a few kHz. A natural limit occurs on decreasing the lower frequency cut-off of $10$ Hz because it is not practical to increase the arm-lengths on ground and also because of the gravity gradient noise which is difficult to eliminate below $10$ Hz. However, VIRGO and future detectors such as the advanced LIGO, the proposed LCGT in Japan and the large European detector plan to go to substantially below $10$ Hz. Thus, in any case, the ground based interferometers will not be sensitive below the limiting frequency of $1$ Hz. But on the other hand, there exist in the cosmos, interesting astrophysical GW sources which emit GW below this frequency such as the galactic binaries, massive and super-massive black-hole binaries etc. If we wish to observe these sources, we need to go to lower frequencies. The solution is to build an interferometer in space, where such noises will be absent and allow the detection of GW in the low frequency regime. LISA - [*Laser Interferometric Space Antenna* ]{} - is a proposed mission which will use coherent laser beams exchanged between three identical spacecraft forming a giant (almost) equilateral triangle of side $5 \times 10^6$ kilometers to observe and detect low frequency cosmic GW. The ground based detectors and LISA complement each other in the observation of GW in an essential way, analogous to the way optical, radio, X-ray, $\gamma$-ray etc., observations do for the electromagnetic spectrum. As these detectors begin to operate, a new era of gravitational astronomy is on the horizon and a radically different view of the universe is expected to emerge.
The astrophysical sources that LISA could observe include galactic binaries, extra-galactic super-massive black-hole binaries and coalescences, and stochastic GW background from the early universe. Coalescing binaries are one of the important sources in the LISA frequency band. These include galactic and extra galactic stellar mass binaries, and massive and super massive black-hole binaries. The frequency of the GW emitted by such a system is twice its orbital frequency. Population synthesis studies indicate a large number of stellar mass binaries in the frequency range below 2-3 mHz [@PBDH; @NYPZ]. In the lower frequency range ($ \leq 1$ mHz) there are a large number of such sources in each of the frequency bins. Since GW detectors are omni-directional, it is impossible to resolve an individual source. These sources effectively form a stochastic GW background referred to as [*binary confusion noise*]{}.
Massive black-hole binaries are interesting both from the astrophysical and theoretical points of view. Coalescences of massive blackholes from different galaxies after their merger during growth of the present galaxies would provide unique new information on galaxy formation. Coalescence of binaries involving intermediate mass blackholes could help understand the formation and growth of massive blackholes. The super massive black hole binaries are strong emitters of GW and these spectacular events can be detectable beyond red-shift of $z=1$. These systems would help to determine the cosmological parameters independently. And, just as the cosmic microwave background is left over from the Big Bang, so too should there be a background of gravitational waves. Unlike electromagnetic waves, gravitational waves do not interact with matter after a few Planck times after the Big Bang, so they do not thermalize. Their spectrum today, therefore, is simply a red-shifted version of the spectrum they formed with, which would throw light on the physical conditions at the epoch of the early universe.
Interferometric, non-resonant, detectors of gravitational radiation (with frequency content $0 < f < f_u$) use a coherent train of electromagnetic waves (of nominal frequency $\nu_0 \gg f_u$) folded into several beams, and at one or more points where these intersect, monitor relative fluctuations of frequency or phase (homodyne detection). The observed low frequency fluctuations are due to several causes: (a) frequency variations of the source of the electromagnetic signal about $\nu_0$, (b) relative motions of the electromagnetic source and the mirrors (or amplifying transponders) that do the folding, (c) temporal variations of the index of refraction along the beams, and, (d) according to general relativity, to any time-variable gravitational fields present, such as the transverse-traceless metric curvature of a passing plane gravitational wave train. To observe gravitational waves in this way, it is thus necessary to control, or monitor, the other sources of relative frequency fluctuations, and, in the data analysis, to use optimal algorithms based on the different characteristic interferometer responses to gravitational waves (the signal) and to the other sources (the noise) [@TE95]. By comparing phases of electromagnetic beams referenced to the same frequency generator and propagated along non-parallel equal-length arms, frequency fluctuations of the frequency reference can be removed and gravitational wave signals at levels many orders of magnitude lower can be detected.
In the present single-spacecraft Doppler tracking observations, for instance, many of the noise sources can be either reduced or calibrated by implementing appropriate microwave frequency links and by using specialized electronics [@Cass], so the fundamental limitation is imposed by the frequency (time-keeping) fluctuations inherent to the reference clock that controls the microwave system. Hydrogen maser clocks, currently used in Doppler tracking experiments, achieve their best performance at about $1000$ seconds integration time, with a fractional frequency stability of a few parts in $10^{-16}$. This is the reason why these one-arm interferometers in space (which have one Doppler readout and a “3-pulse” response to gravitational waves [@EW75]) are most sensitive to millihertz gravitational waves. This integration time is also comparable to the microwave propagation (or “storage”) time $2 L/c$ to spacecraft en route to the outer solar system (for example $L \simeq 5 - 8 \ {\rm
AU}$ for the Cassini spacecraft) [@T02].
Next-generation low-frequency interferometric gravitational wave detectors in solar orbits, such as the Laser Interferometer Space Antenna (LISA) mission [@LISA98], have been proposed to achieve greater sensitivity to millihertz gravitational waves. Since the armlengths of these space-based interferometers can differ by few percent, the direct recombination of the two beams at a photo detector will not however effectively remove the laser frequency noise. This is because the frequency fluctuations of the laser will be delayed by different amounts within the two unequal length arms. In order to cancel the laser frequency noise, the time-varying Doppler data must be recorded and post-processed to allow for arm-length differences [@TA99]. The data streams will have temporal structure, which can be described as due to many-pulse responses to $\delta$-function excitations, depending on time-of-flight delays in the response functions of the instrumental Doppler noises and in the response to incident plane-parallel, transverse, and traceless gravitational waves.
LISA will consists of three spacecraft orbiting the sun. Each spacecraft will be equipped with two lasers sending beams to the other two (${\sim}0.03$ AU away) while simultaneously measuring the beat frequencies between the local laser and the laser beams received from the other two spacecraft. The analysis of TDI presented in this article will assume a successful prior removal of any first-order Doppler beat notes due to relative motions [@TEA02], giving six residual Doppler time series as the raw data of a [*stationary*]{} time delay space interferometer. Following [@T98], [@AET99], [@DNV02], we will regard LISA not as constituting one or more conventional Michelson interferometers, but rather, in a symmetrical way, a closed array of six one-arm delay lines between the test masses. In this way, during the course of the article, we will show that it is possible to synthesize new data combinations that cancel laser frequency noises, and estimate achievable sensitivities of these combinations in terms of the separate and relatively simple single arm-responses both to gravitational wave and instrumental noise (cf. [@T98], [@AET99], [@DNV02]).
In contrast to Earth-based interferometers, which operate in the long-wavelength limit (LWL) (arm lengths $<<$ gravitational wavelength ${\sim}c/f_0$, where $f_0$ is a characteristic frequency of the GW), LISA will ${\it not}$ operate in the LWL over much of its frequency band. When the physical scale of a free mass optical interferometer intended to detect gravitational waves is comparable to or larger than the GW wavelength, time delays in the response of the instrument to the waves, and travel times along beams in the instrument, cannot be ignored and must be allowed for in computing the detector response used for data interpretation. It is convenient to formulate the instrumental responses in terms of observed differential frequency shifts $-$ for short, Doppler shifts $-$ rather than in terms of phase shifts usually used in interferometry, although of course these data, as functions of time, are interconvertible.
This first review article on TDI is organized as follows. In Section \[SECII\] we provide an overview of the physical and historical motivations of TDI. In Section \[SECIII\] we summarize the one-arm Doppler transfer functions of an optical beam between two carefully shielded test masses inside each spacecraft resulting from (i) frequency fluctuations of the lasers used in transmission and reception, (ii) fluctuations due to non-inertial motions of the spacecraft, (iii) beam-pointing fluctuations and shot noise [@ETA00]. Among these, the dominant noise is from the frequency fluctuations of the lasers and is several orders (perhaps 7 or 8) above the other noises. This noise must be very precisely removed from the data in order to achieve the GW sensitivity at the level set by the remaining Doppler noise sources which are at a much lower level and which constitute the noise floor after the laser frequency noise is suppressed. We show that this can be accomplished by shifting and linearly combining the twelve one-way Doppler data LISA will measure. The actual procedure can easily be understood in terms of properly defined time-delay operators that act on the one-way Doppler measurements. We develop a formalism involving the algebra of the time-delay operators which is based on the theory of rings and modules and computational commutative algebra. We show that the space of all possible interferometric combinations cancelling the laser frequency noise is a module over the polynomial ring in which the time-delay operators play the role of the indeterminates. The module, in the literature, is called the Module of Syzygies [@DNV02]. We show that the module can be generated from [*four*]{} generators, so that any data combination cancelling the laser frequency noise is simply a linear combination formed from these generators. We would like to emphasize that this is the mathematical structure underlying TDI in LISA.
In Section \[SECIV\] specific interferometric combinations are then derived, and their physical interpretations are discussed. The expressions for the Sagnac interferometric combinations, ($\alpha,
\beta, \gamma, \zeta$) are first obtained; in particular, the symmetric Sagnac combination $\zeta$, for which each raw data set needs to be delayed by only a ${\it single}$ arm transit time, distinguishes itself against all the other TDI combinations by having a higher order response to gravitational radiation in the LWL when the spacecraft separations are equal. We then express the Unequal-arm Michelson combinations, ($X, Y, Z$), in terms of the $\alpha$, $\beta$, $\gamma$, and $\zeta$ combinations with further transit time delays. One of these interferometric data combinations would still be available if the links between one pair of spacecraft were lost. Other TDI combinations, which rely on only four of the possible six inter-spacecraft Doppler measurements (denoted $P$, $E$ and $U$) are also presented. They would of course be quite useful in case of potential loss of any two inter-spacecraft Doppler measurements.
Time-Delay Interferometry so formulated presumes the spacecraft-to-spacecraft light-travel-times to be constant in time, and independent from being up- or down-links. Reduction of data from moving interferometric laser arrays in solar orbit will in fact encounter non-symmetric up- and downlink light time differences that are significant, and need to be accounted for in order to exactly cancel the laser frequency fluctuations [@S03; @CH03; @STEA03]. In Section \[SECV\] we show that, by introducing a set of non-commuting time-delay operators, there exists a quite general procedure for deriving generalized TDI combinations that account for the effects of time-dependence of the arms. Using this approach it is possible to derive “flex-free” expression for the unequal-arm Michelson combinations $X_1$, and obtain the generalized expressions for all the TDI combinations [@TEA04].
In Section \[SECVI\] we address the question of maximization of the LISA signal-to-noise-ratio (SNR) to any gravitational wave signal present in its data. This is done by treating the SNR as a functional over the space of all possible TDI combinations. As a simple application of the general formula we have derived, we apply our results to the case of sinusoidal signals randomly polarized and randomly distributed on the celestial sphere. We find that the standard LISA sensitivity figure derived for a single Michelson Interferometer [@ETA00; @PTLA02; @NPDV03_1] can be improved by a factor of $\sqrt{2}$ in the low-part of the frequency band, and by more than $\sqrt{3}$ in the remaining part of the accessible band. Further, we also show that if the location of the GW source is known, then as the source appears to move in the LISA reference frame, it is possible to optimally track the source, by appropriately changing the data combinations during the course of its trajectory [@PTLA02], [@NDPV03_2]. As an example of such type of source, we consider known binaries within our own galaxy.
This first version of our “Living Review” article on TDI does not include all the results of more practical and experimental nature, as well as all the aspects of TDI that the data analysts will need to account for when analyzing the LISA TDI data combinations. Our forthcoming “second edition” of this review paper will include these topics. It is worth mentioning that, as of today, the LISA project has endorsed TDI as its baseline technique for achieving the desired sensitivity to gravitational radiation. Several experimental verifications and tests of TDI are being, and will be, performed at the NASA and ESA LISA laboratories. Although significant theoretical and experimental work has already been done for understanding and overcoming practical problems related to the implementation of TDI, more work on both sides of the Atlantic is still needed. Results of this undergoing effort will be included in the second edition of this living document.
Physical and historical motivations of TDI {#SECII}
==========================================
Equal-arm interferometer detectors of gravitational waves can observe gravitational radiation by cancelling the laser frequency fluctuations affecting the light injected into their arms. This is done by comparing phases of split beams propagated along the equal (but non-parallel) arms of the detector. The laser frequency fluctuations affecting the two beams experience the same delay within the two equal-length arms and cancel out at the photodetector where relative phases are measured. This way gravitational wave signals of dimensionless amplitude less than $10^{-20}$ can be observed when using lasers whose frequency stability can be as large as roughly a few parts in $10^{-13}$.
If the arms of the interferometer have different lengths, however, the exact cancellation of the laser frequency fluctuations, say $C (t)$, will no longer take place at the photodetector. In fact, the larger the difference between the two arms, the larger will be the magnitude of the laser frequency fluctuations affecting the detector response. If $L_1$ and $L_2$ are the lengths of the two arms, it is easy to see that the amount of laser relative frequency fluctuations remaining in the response is equal to (units in which the speed of light $c = 1$) $$\Delta C (t) = C(t - 2L_1) - C(t - 2L_2) \ .
\label{DC}$$ In the case of a space-based interferometer such as LISA, whose lasers are expected to display relative frequency fluctuations equal to about $10^{-13}/\sqrt{Hz}$ in the millihertz band, and whose arms will differ by a few percent [@LISA98], equation (\[DC\]) implies the following expression for the amplitude of the Fourier components of the uncanceled laser frequency fluctuations (an over imposed tilde denotes the operation of Fourier transform) $$|{\widetilde {\Delta C}} (f)| \simeq |{\widetilde {C}} (f)| \
4 \pi f |(L_1 - L_2)| \ .
\label{FDC}$$ At $f = 10^{-3}$ Hz, for instance, and assuming $|L_1 - L_2| \simeq
0.5 \ \ {\rm sec}$, the uncanceled fluctuations from the laser are equal to $6.3 \times 10^{-16}/\sqrt{\rm Hz}$. Since the LISA sensitivity goal is about $10^{-20}/\sqrt{\rm Hz}$ in this part of the frequency band, it is clear that an alternative experimental approach for canceling the laser frequency fluctuations is needed.
A first attempt to solve this problem was presented in [@FB84; @FBHHV85; @FBHHSV89], and the scheme proposed there can be understood through Figure \[2unequalarms\]. In this idealized model the two beams exiting the two arms are not made to interfere at a common photodetector. Rather, each is made to interfere with the incoming light from the laser at a photodetector, decoupling in this way the phase fluctuations experienced by the two beams in the two arms. Now two Doppler measurements are available in digital form, and the problem now becomes one of identifying an algorithm for digitally cancelling the laser frequency fluctuations from a resulting new data combination.
![Light from a laser is split into two beams, each injected into an arm formed by pairs of free-falling mirrors. Since the length of the two arms, $L_1$ and $L_2$, are different, now the light beams from the two arms are not recombined at one photo detector. Instead each is separately made to interfere with the light that is injected into the arms. Two distinct photo detectors are now used, and phase (or frequency) fluctuations are then monitored and recorded there.[]{data-label="2unequalarms"}](Fig1.eps){width="7.0"}
The algorithm they first proposed, and refined subsequently in [@GHTF96], required processing the two Doppler measurements, say $y_1 (t)$ and $y_2 (t)$, in the Fourier domain. If we denote with $h_1
(t)$, $h_2 (t)$ the gravitational wave signals entering into the Doppler data $y_1$, $y_2$ respectively, and with $n_1$, $n_2$ any other remaining noise affecting $y_1$ and $y_2$ respectively, then the expressions for the Doppler observables $y_1$, $y_2$ can be written in the following form $$\begin{aligned}
y_1(t) & = & C(t - 2L_1) - C(t) + h_1(t) + n_1(t) \ , \
\label{2Doppler1}
\\
y_2(t) & = & C(t - 2L_2) - C(t) + h_2(t) + n_2(t) \ .
\label{2Doppler2}\end{aligned}$$ From Eqs. (\[2Doppler1\], \[2Doppler2\]) it is important to note the characteristic time signature of the random process $C(t)$ in the Doppler responses $y_1$ , $y_2$. The time signature of the noise $C(t)$ in $y_1(t)$, for instance, can be understood by observing that the frequency of the signal received at time $t$ contains laser frequency fluctuations transmitted $2L_1$ seconds earlier. By subtracting from the frequency of the received signal the frequency of the signal transmitted at time $t$, we also subtract the frequency fluctuations $C(t)$ with the net result shown in Eq. (\[2Doppler1\]).
The algorithm for cancelling the laser noise in the Fourier domain suggested in [@FB84] works as follow. If we take an infinitely long Fourier transform of the data $y_1$, the resulting expression of $y_1$ in the Fourier domain becomes (see eq. (\[2Doppler1\])) $${\widetilde y}_1 (f) = {\widetilde C} (f) \ [e^{4 \pi i f L_1} - 1] +
{\widetilde h}_1 (f) + {\widetilde n}_1 (f) \ .
\label{2DF1}$$ If the arm length $L_1$ is known exactly, we can use the ${\widetilde
y}_1$ data to estimate the laser frequency fluctuations ${\widetilde C} (f)$. This can be done by dividing ${\widetilde y}_1$ by the transfer function of the laser noise $C$ into the observable $y_1$ itself. By then further multiplying ${\widetilde y}_1/[e^{4 \pi i f
L_1} - 1]$ by the transfer function of the laser noise into the other observable ${\widetilde y}_2$, i.e., $ [e^{4 \pi i f L_2} - 1]$, and then subtract the resulting expression from ${\widetilde y}_2$ one accomplishes the cancellation of the laser frequency fluctuations.
The problem with this procedure is the underlying assumption of being able to take an infinitely long Fourier transform of the data. Even if one neglects the variation in time of the LISA arms, by taking a finite length Fourier transform of, say, $y_1 (t)$ over a time interval $T$, the resulting transfer function of the laser noise $C$ into $y_1$ no longer will be equal to $[e^{4 \pi i f L_1} - 1]$. This can be seen by writing the expression of the finite length Fourier transform of $y_1$ in the following way $${\widetilde{y}}^{T}_1 \equiv \int_{-T}^{+T} y_1(t) \ e^{2 \pi i f t} \
dt = \int_{-\infty}^{+\infty} y_1(t) \ H(t) \ e^{2 \pi i f t} \ dt \ ,
\label{FiniteF}$$ where we have denoted with $H(t)$ the function that is equal to $1$ in the interval $[-T, +T]$, and zero everywhere else. Equation (\[FiniteF\]) implies that the finite-length Fourier transform ${\widetilde{y}}^{T}_1$ of $y_1(t)$ is equal to the convolution in the Fourier domain of the infinitely long Fourier transform of $y_1 (t)$, ${\widetilde y}_1$, with the Fourier transform of $H(t)$ [@JW68] (i.e. the “Sinc Function” of width $1/T$). The key point here is that we can no longer use the transfer function $[e^{4 \pi i f L_i} -
1] \ , \ i=1, 2$, for estimating the laser noise fluctuations from one of the measured Doppler data, without retaining residual laser noise into the combination of the two Doppler data $y_1$, $y_2$ valid in the case of infinite integration time. The amount of residual laser noise remaining in the Fourier-based combination described above, as a function of the integration time $T$ and type of “window function” used, was derived in the appendix of [@TA99]. There it was shown that, in order to suppress the residual laser noise below the LISA sensitivity level identified by secondary noises (such as proof-mass and optical path noises) with the use of the Fourier-based algorithm an integration time of about six months was needed.
A solution to this problem was suggested in [@TA99], which works entirely in the time-domain. From Eqs.(\[2Doppler1\], \[2Doppler2\]) we may notice that, by taking the difference of the two Doppler data $y_1(t)$, $y_2(t)$, the frequency fluctuations of the laser now enter into this new data set in the following way $$y_1(t) - y_2(t) = C(t - 2L_1) - C(t - 2L_2) + h_1(t) - h_2(t) + n_1(t)
- n_2(t) \ .
\label{y1-y2}$$ If we now compare how the laser frequency fluctuations enter into Eq. (\[y1-y2\]) against how they appear in Eqs. (\[2Doppler1\], \[2Doppler2\]) we can further make the following observation. If we time-shift the data $y_1(t)$ by the round trip light time in arm $2$, $y_1(t - 2L_2)$, and subtract from it the data $y_2(t)$ after it has been time-shifted by the round trip light time in arm $1$, $y_2(t - 2L_1)$, we obtain the following data set $$\begin{aligned}
y_1(t - 2L_2) - y_2(t - 2L_1) & = & C(t - 2L_1) - C(t - 2L_2) + h_1(t -
2L_2) - h_2(t - 2L_1)
\nonumber
\\
&& + n_1(t - 2L_2) - n_2(t - 2L_1) \ .
\label{y1d-y2d}\end{aligned}$$ In other words, the laser frequency fluctuations enter into $y_1(t) -
y_2(t)$, and $y_1(t - 2L_2) - y_2(t - 2L_1)$ with the same time structure. This implies that, by subtracting Eq. (\[y1d-y2d\]) from Eq. (\[y1-y2\]) we can generate a new data set that does not contain the laser frequency fluctuations $C(t)$ $$X \equiv [y_1(t) - y_2(t)] - [y_1(t - 2L_2) - y_2(t - 2L_1)] \ .
\label{XTA99}$$ The expression above of the $X$ combination shows that it is possible to cancel the laser frequency noise in the time domain by properly time-shifting and linearly combining Doppler measurements recorded by different Doppler readouts. This in essence is what Time-Delay Interferometry (TDI) amounts to. In the following sections we will further elaborate and generalize TDI to the realistic LISA configuration.
Time-Delay Interferometry {#SECIII}
=========================
The description of TDI for LISA is greatly simplified if we adopt the notation shown in Figure \[Geometry\] , where the overall geometry of the LISA detector is defined. There are three spacecraft, six optical benches, six lasers, six proof-masses and twelve photodetectors. There are also six phase difference data going clock-wise and counter-clockwise around the LISA triangle. For the moment we will make the simplifying assumption that the array is stationary, i.e. the back and forth optical paths between pairs of spacecraft are simply equal to their relative distances [@S03; @CH03; @STEA03; @TEA04].
Several notations have been used in this context. The double index notation recently employed in [@STEA03] where six quantities are involved is self-evident. However, when algebraic manipulations are involved the following notation seems more convenient to use. The spacecraft are labeled $1$, $2$, $3$ and their separating distances are denoted $L_1$, $L_2$, $L_3$, with $L_i$ being opposite spacecraft $i$. We orient the vertices $1, 2, 3$ clockwise in figure \[Geometry\]. Unit vectors between spacecraft are $\hat n_i$, oriented as indicated in figure \[Geometry\]. We index the phase difference data to be analyzed as follows: The beam arriving at spacecraft $i$ has subscript $i$ and is primed or unprimed depending on whether the beam is traveling clockwise or counter-clockwise (the sense defined here with reference to figure \[Geometry\]) around the LISA triangle respectively. Thus, as seen from the figure, $s_{1}$ is the phase difference time series measured at reception at spacecraft $1$ with transmission from spacecraft $2$ (along $L_3$).
![Schematic LISA configuration. The spacecraft are labeled 1, 2, and 3. The optical paths are denoted by $L_i$, $L_i'$ where the index $i$ corresponds to the opposite spacecraft. The unit vectors $\hat{{\bf n}}_i$ point between pairs of spacecraft, with the orientation indicated. \[Geometry\]](Fig2.eps){width="4.0in"}
Similarly, $s_{1}'$ is the phase difference series derived from reception at spacecraft $1$ with transmission from spacecraft $3$. The other four one-way phase difference time series from signals exchanged between the spacecraft are obtained by cyclic permutation of the indices: $1$ $\to$ $2$ $\to$ $3$ $\to$ $1$. We also adopt a notation for delayed data streams, which will be convenient later for algebraic manipulations. We define the three time-delay operators $\D_i, ~i = 1, 2, 3$ where for any data stream $x(t)$, \_i x(t) = x(t - L\_i) , where $L_i, i = 1, 2, 3$ are the light travel times along the three arms of the LISA triangle (the speed of light $c$ is assumed to be unity in this article). Thus, for example, $\D_2 s_{1} (t) =
s_{1}(t - L_2)$, $\D_2 \D_3 s_{1}(t) = s_{1}(t - L_2 - L_3) = \D_3
\D_2 s_{1}(t)$, etc. Note that the operators commute here. This is because the arm-lengths have been assumed to be constant in time. If the $L_i$ are functions of time then the operators no longer commute [@CH03; @TEA04], as will be described in section IV. Six more phase difference series result from laser beams exchanged between adjacent optical benches within each spacecraft; these are similarly indexed as $\tau_{i}, \tau_i', i
= 1, 2, 3$. The proof-mass-plus-optical-bench assemblies for LISA spacecraft number $1$ are shown schematically in figure 2. The photo receivers that generate the data $s_{1}$, $s_{1}'$, $\tau_{1}$, and $\tau_{1}'$ at spacecraft $1$ are shown. The phase fluctuations from the six lasers, which need to be canceled, can be represented by six random processes $p_{i}, p_{i}'$, where $p_{i}, p_{i}'$ are the phases of the lasers in spacecraft $i$ on the left and right optical benches respectively as shown in the figure. Note that this notation is in the same spirit as in the references [@TEA02; @STEA03] in which moving spacecraft arrays have been analyzed.
We extend the cyclic terminology so that at vertex $i$ ($i = 1, 2, 3$) the random displacement vectors of the two proof masses are respectively denoted by $\vec \delta_{i}(t), \vec \delta_{i}'(t)$, and the random displacements (perhaps several orders of magnitude greater) of their optical benches are correspondingly denoted by $\vec
\Delta_{i}(t), \vec \Delta_{i}'(t)$ where the primed and unprimed indices correspond to the right and left optical benches respectively. As pointed out in [@ETA00], the analysis does assume that pairs of optical benches are rigidly connected, i.e. $\vec
\Delta_{i} \neq \vec \Delta_{i}'$, in general. The present LISA design shows optical fibers transmitting signals both ways between adjacent benches. We ignore time-delay effects for these signals and will simply denote by $\mu_i(t)$ the phase fluctuations upon transmission through the fibers of the laser beams with frequencies $\nu_{i}$, and $\nu_{i}'$. The $\mu_i (t)$ phase shifts within a given spacecraft might not be the same for large frequency differences $\nu_{i} - \nu_{i}'$. For the envisioned frequency differences (a few hundred megahertz), however, the remaining fluctuations due to the optical fiber can be neglected [@ETA00]. It is also assumed that the phase noise added by the fibers is independent of the direction of light propagation through them. For ease of presentation, in what follows we will assume the center frequencies of the lasers to be the same, and denote this frequency by $\nu_0$.
The laser phase noise in $s_{3}'$ is therefore equal to $\D_1 p_{2}(t)
- p_{3}'(t)$. Similarly, since $s_{2}$ is the phase shift measured on arrival at spacecraft $2$ along arm $1$ of a signal transmitted from spacecraft $3$, the laser phase noises enter into it with the following time signature: $\D_1 p_{3}'(t) - p_{2}(t)$. Figure 2 endeavors to make the detailed light paths for these observations clear. An outgoing light beam transmitted to a distant spacecraft is routed from the laser on the local optical bench using mirrors and beam splitters; this beam does not interact with the local proof mass. Conversely, an [*[incoming]{}*]{} light beam from a distant spacecraft is bounced off the local proof mass before being reflected onto the photo receiver where it is mixed with light from the laser on that same optical bench. The inter-spacecraft phase data are denoted $s_{1}$ and $s_{1}'$ in figure 2.
{width="5.0"}
Beams between adjacent optical benches within a single spacecraft are bounced off proof masses in the opposite way. Light to be [ *[transmitted]{}*]{} from the laser on an optical bench is [*[first]{}*]{} bounced off the proof mass it encloses and then directed to the other optical bench. Upon reception it does [*not*]{} interact with the proof mass there, but is directly mixed with local laser light, and again down converted. These data are denoted $\tau_{1}$ and $\tau_{1}'$ in figure 2.
The expressions for the $s_{i}, s_{i}'$ and $\tau_{i}, \tau_{i}'$ phase measurements can now be developed from figures 1 and 2, and they are for the particular LISA configuration in which all the lasers have the same nominal frequency $\nu_0$, and the spacecraft are stationary with respect to each other. Consider the $s_{1}' (t)$ process (equation (\[eq:tre\]) below). The photo receiver on the right bench of spacecraft $1$, which (in the spacecraft frame) experiences a time-varying displacement $\vec \Delta_{1}'$, measures the phase difference $s_{1}'$ by first mixing the beam from the distant optical bench $3$ in direction $\hat n_2$, and laser phase noise $p_{3}$ and optical bench motion $\vec \Delta_{3}$ that have been delayed by propagation along $L_2$, after one bounce off the proof mass ($\vec
\delta_{1}'$), with the local laser light (with phase noise $p_{1}'$). Since for this simplified configuration no frequency offsets are present, there is of course no need for any heterodyne conversion [@TEA02].
In equation (\[eq:due\]) the $\tau_{1}$ measurement results from light originating at the right-bench laser ($p_{1}'$, $\vec \Delta_{1}'$), bounced once off the right proof mass ($\vec \delta_{1}'$), and directed through the fiber (incurring phase shift $\mu_1(t)$), to the left bench, where it is mixed with laser light ($p_{1}$). Similarly the right bench records the phase differences $s_{1}'$ and $\tau_{1}'$. The laser noises, the gravitational wave signals, the optical path noises, and proof-mass and bench noises, enter into the four data streams recorded at vertex $1$ according to the following expressions [@ETA00]:
$$\begin{aligned}
s_{1} & = & s^{\rm gw}_{1} + s^{\rm opt. \ path}_{1} +
\D_3 p_{2}' - p_{1} + \nu_{0} \ \left[\ - 2 {\hat n_3} \cdot
{\vec \delta_{1}} + {\hat n_3} \cdot {\vec \Delta_{1}} +
{\hat n_3} \cdot \D_3 {\vec \Delta_{2}'} \right] \ ,
\label{eq:uno}
\\
\tau_{1} & = & p_{1}' - p_{1} -
\ 2 \ \nu_{0} \ {\hat n_2} \cdot ({\vec \delta_{1}'} - {\vec
\Delta_{1}'}) + \mu_1 \ .
\label{eq:due}
\\
s_{1}' & = & s^{'{\rm gw}}_{1} + s^{'{\rm opt. \ path}}_{1} +
\D_2 p_{3} - p_{1}' + \nu_{0} \ \left[ 2 {\hat n_2} \cdot {\vec \delta_{1}'} -
{\hat n_2} \cdot {\vec \Delta_{1}'} - {\hat n_2} \cdot \D_2 {\vec
\Delta_{3}} \right] \ ,
\label{eq:tre}
\\
\tau_{1}' & = & p_{1} - p_{1}' + 2 \ \nu_{0} \ {\hat n_3} \cdot
({\vec \delta_{1}} - {\vec \Delta_{1}}) + \mu_1 \ .
\label{eq:quattro}\end{aligned}$$
Eight other relations, for the readouts at vertices 2 and 3, are given by cyclic permutation of the indices in equations (\[eq:uno\])-(\[eq:quattro\]).
The gravitational wave phase signal components, $s^{\rm gw}_{i},
s^{'{\rm gw}}_{i} \ , \ i = 1, 2, 3$, in equations (\[eq:uno\]) and (\[eq:tre\]) are given by integrating with respect to time the equations (1), and (2) of reference [@AET99], which relate metric perturbations to optical frequency shifts. The optical path phase noise contributions, $s^{\rm opt. \ path}_{i}, s^{'{\rm opt. \
path}}_{i}$, which include shot noise from the low signal-to-noise ratio (SNR) in the links between the distant spacecraft, can be derived from the corresponding term given in [@ETA00]. The $\tau_{i}, \tau_{i}'$ measurements will be made with high SNR so that for them the shot noise is negligible.
Albegraic approach to cancelling laser and optical bench noises {#SECIV}
===============================================================
In ground based detectors the arms are chosen to be of equal length so that the laser light experiences identical delay in each arm of the interferometer. This arrangement precisely cancels the laser frequency/phase noise at the photodetector. The required sensitivity of the instrument can thus only be achieved by near exact cancellation of the laser frequency noise. However, in LISA it is impossible to achieve equal distances between spacecraft and the laser noise cannot be cancelled in this way. It is possible to combine the recorded data linearly with suitable time-delays corresponding to the three arm-lengths of the giant triangular interferometer so that the laser phase noise is cancelled. Here we present a [*systematic method*]{} based on modules over polynomial rings which guarantees all the data combinations that cancel both the laser phase and the optical bench motion noises.
We first consider the simpler case, where we ignore the optical-bench motion noise and consider only the laser phase noise. We do this because the algebra is somewhat simpler and the method is easy to apply. The simplification amounts to physically considering each spacecraft rigidly carrying the assembly of lasers, beam-splitters and photodetectors. The two lasers on each spacecraft could be considered to be locked, so effectively there would be only one laser on each spacecraft. This mathematically amounts to setting $\vec \Delta_i =
\vec \Delta_i'= 0$ and $p_i = p_i'$. The scheme we describe here for laser phase noise can be extended in a straight forward way to include optical bench motion noise, which we address in the last part of this section.
The data combinations, when only the laser phase noise is considered, consist of the six suitably delayed data streams (inter-spacecraft), the delays being integer multiples of the light travel times between spacecraft, which can be conveniently expressed in terms of polynomials in the three delay operators $\D_1, \D_2, \D_3$. The laser noise cancellation condition puts three constraints on the six polynomials of the delay operators corresponding to the six data streams. The problem therefore consists of finding six tuples of polynomials which satisfy the laser noise cancellation constraints. These polynomial tuples form a module [^1] called in the literature, the [*module of syzygies*]{}. There exist standard methods for obtaining the module, by that we mean, methods for obtaining the generators of the module so that the linear combinations of the generators generate the entire module. The procedure first consists of obtaining a Gröbner basis for the ideal generated by the coefficients appearing in the constraints. This ideal is in the polynomial ring in the variables $\D_1, \D_2, \D_3$ over the domain of rational numbers (or integers if one gets rid of the denominators). To obtain the Gröbner basis for the ideal, one may use the Buchberger algorithm or use an application such as Mathematica [@Wolf02]. From the Gröbner basis there is a standard way to obtain a generating set for the required module. This procedure has been described in the literature [@becker; @KR]. We thus obtain seven generators for the module. However, the method does not guarantee a minimal set and we find that a generating set of 4 polynomial six tuples suffice to generate the required module. Alternatively, we can obtain generating sets by using the software Macaulay 2.
The importance of obtaining more data combinations is evident: they provide the necessary redundancy - different data combinations produce different transfer functions for GW and the system noises so specific data combinations could be optimal for given astrophysical source parameters in the context of maximizing SNR, detection probability, improving parameter estimates etc.
Cancellation of laser phase noise
---------------------------------
We now only have six data streams: $s_i$ and $s_i'$ where, $i =
1,2,3$. These can be regarded as 3 component vectors $\bs$ and $\bs'$ respectively. The six data streams with terms containing [*only*]{} the laser frequency noise are: s\_1 &=& \_3 p\_2 - p\_1 ,\
s\_1’ &=& \_2 p\_3 - p\_1 , \[beams\] and their cyclic permutations.
Note that we have excluded intentionally from the data, additional phase fluctuations due to the GW signal, and noises such as the optical-path noise, proof-mass noise etc. Since our immediate goal is to cancel the laser frequency noise we have only kept the relevant terms. Combining the streams for cancelling the laser frequency noise will introduce transfer functions for the other noises and the GW signal. This is important and will be discussed subsequently in the article.
[*The goal of the analysis is to add suitably delayed beams together so that the laser frequency noise terms add up to zero.*]{} This amounts to seeking data combinations that cancel the laser frequency noise. In the notation/formalism that we have invoked, the delay is obtained by applying the operators $\D_k$ to the beams $s_i$ and $s_i'$. A delay of $k_1 L_1 + k_2 L_2 + k_3 L_3$ is represented by the operator $\D_1^{k_1} \D_2^{k_2} \D_3^{k_3}$ acting on the data, where $k_1, k_2$ and $k_3$ are integers. In general a polynomial in $\D_k$, which is a polynomial in three variables, applied to say $s_1$ combines the same data stream $s_1(t)$ with different time-delays of the form $k_1 L_1 + k_2 L_2 + k_3 L_3$. This notation conveniently rephrases the problem. One must find six polynomials say $q_i (\D_1,
\D_2, \D_3), \ q_i' (\D_1, \D_2, \D_3), \ i=1,2,3 $ such that: \_[i = 1]{}\^3 q\_i s\_i + q\_i’ s\_i’ = 0 . \[cncl\] The zero on the R.H.S. of the above equation signifies zero laser phase noise.
It is useful to express Eq. (\[beams\]) in matrix form. This allows us to obtain a matrix operator equation whose solutions are $\bq$ and $\bq'$ where $q_i$ and $q_i'$ are written as column vectors. We can similarly express $s_i, s_i', p_i$ as column vectors $\bs, \bs', \bp$ respectively. In matrix form Eq. (\[beams\]) become: = \^T , ’ = , where, $\bD$ is a $3 \times 3$ matrix given by, = (
[ccc]{} -1 & 0 & \_2\
\_3 & -1 & 0\
0 & \_1 & -1
) . \[Dmat\] The exponent ‘$T$’ represents the transpose of the matrix. Eq. (\[cncl\]) becomes: \^T + ’\^T ’ = (\^T \^T + ’\^T ) = 0 . where we have taken care to put $\bp$ on the right-hand-side of the operators. Since the above equation must be satisfied for an arbitrary vector $\bp$, we obtain a matrix equation for the polynomials $(\bq,
\bq')$: \^T \^T + ’ = 0 . \[opeq\] Note that since the $\D_k$ commute, the order in writing these operators is unimportant. In mathematical terms, the polynomials form a commutative ring.
Cancellation of laser phase noise in the unequal arm interferometer
-------------------------------------------------------------------
The use of commutative algebra is very conveniently illustrated with the help of the simpler example of the unequal arm interferometer. Here there are only two arms instead of three as we have for LISA and the mathematics is much simpler and so it easy to see both physically and mathematically how commutative algebra can be applied to this problem of laser phase noise cancellation. The procedure is well known for the unequal arm interferometer, but here we will describe the same method but in terms of the delay opertors that we have introduced.
Let $\phi (t)$ denote the laser phase noise entering the laser cavity as shown in Fig. \[XU\]. Consider this light $\phi (t)$ making a round trip around arm 1 whose length we take to be $L_1$. If we interfere this phase with the incoming light we get the phase $\phi_1
(t)$, where, \_1 (t) = (t - 2 L\_1) - (t) (\_1\^2 - 1) (t) . The second expression we have written in terms of the delay operators. This makes the procedure transparent as we shall see. We can do the same for the arm 2 to get another phase $\phi_2 (t)$, where, \_2 (t) = (t - 2 L\_2) - (t) (\_2\^2 - 1) (t) . Clearly, if $L_1 \neq L_2$, then the difference in phase $\phi_2 (t) - \phi_1 (t)$ is not zero and the laser phase noise does not cancel out. However, if one further delays the phases $\phi_1 (t)$ and $\phi_2 (t)$ and constructs the following combination: X(t) = \[\_2 (t - 2 L\_1) - \_2 (t)\] - \[\_1 (t - 2 L\_2) - \_1 (t)\] , \[X\] then the laser phase noise does cancel out. We have already encountered this combination at the end of section II, Eq.(\[XTA99\]), when it was first proposed by Tinto and Armstrong in reference [@TA99]. The cancellation of laser frequency noise becomes obvious from the operator algebra in the following way. In the operator notation: X(t) &=& (\_1\^2 - 1) \_2 (t) - (\_2\^2 - 1) \_1 (t)\
&=& \[(\_1\^2 - 1)(\_2\^2 - 1) - (\_2\^2 - 1)(\_1\^2 - 1)\] (t)\
&=& 0 . From this one immediately sees that just the commutativity of the operators has been used to cancel the laser phase noise. The basic idea was to compute the lowest common multiple (L.C.M.) of the polynomials $\D_1^2 - 1$ and $\D_2^2 - 1$ (in this case the L.C.M. is just the product, because the polynomials are relatively prime) and use this fact to construct $X(t)$ in which the laser phase noise is cancelled. The operation is shown physically in Fig. \[XU\].
![Schematic diagram of the unequal-arm Michelson interferometer. The beam shown corresponds to the term $(\D_1^2 -
1)(\D_2^2 - 1) \phi_ (t)$ in $X(t)$ which is first sent around arm 1 followed by arm 2. The second beam (not shown) is first sent around arm 2 and then through arm 1. The difference in these two beams constitutes $X(t)$.[]{data-label="XU"}](Fig4.eps){width="3in"}
The notions of commutativity of polynomials, L.C.M. etc. belong to the field of commutative algebra. In fact we will be using the notion of a Gröbner basis which is in a sense the generalization of the notion of the greatest common divisor or in short gcd. Since LISA has three spacecraft and six inter-spacecraft beams, the problem of the unequal arm interferometer only gets technically more complex; in principle the problem is the same as in this simpler case. Thus the simple operations which were performed here to obtain a laser noise free combination $X(t)$ are not sufficient and more sophisticated methods need to be adopted from the field of commutative algebra. We address this problem in the forthcoming text.
The Module of Syzygies
----------------------
Eq. (\[opeq\]) has non-trivial solutions. Several solutions have been exhibited in [@AET99; @ETA00]. We merely mention these solutions here; in the next section we will discuss them in detail. The solution $\zeta$ is given by $-\bq^T = \bq'^T = (\D_1, \D_2,
\D_3)$. The solution $\alpha$ is described by $\bq^T = -(1, \D_3, \D_1
\D_3)$ and $\bq'^T = (1, \D_1 \D_2, \D_2)$. The solutions $\beta$ and $\gamma$ are obtained from $\alpha$ by cyclically permuting the indices of $\D_k, \bq$ and $\bq'$. These solutions are important, because they consist of polynomials with lowest possible degrees and thus are simple. Other solutions containing higher degree polynomials can be generated conveniently from these solutions. Since the system of equations is linear, linear combinations of these solutions are also solutions to Eq. (\[opeq\]).
However, it is important to realize that we do not have a vector space here. Three independent constraints on a six tuple do not produce a space which is necessarily generated by three basis elements. This conclusion would follow if the solutions formed a vector space but they do not. The polynomial six-tuple $\bq, \bq'$ can be multiplied by polynomials in $\D_1, \D_2, \D_3$ (scalars) which do not form a field. So that the inverse in general does not exist within the ring of polynomials. We therefore have a module over the ring of polynomials in the three variables $\D_1, \D_2, \D_3$. First we present the general methodology for obtaining the solutions to (\[opeq\]) and then apply it to equations (\[opeq\]).
There are three linear constraints on the polynomials given by the equations (\[opeq\]). Since the equations are linear the solutions space is a submodule of the module of six-tuples of polynomials. The module of six-tuples is a free module, i.e. it has six basis elements that not only generate the module but are linearly independent. A natural choice of the basis is $f_m = (0, ..., 1, ...,0)$ with 1 in the $m$-th place and 0 everywhere else; $m$ runs from 1 to 6. The definitions of generation (spanning) and linear independence are the same as that for vector spaces. A free module is essentially like a vector space. But our interest lies in its submodule which need not be free and need not have just three generators as it would seem if we were dealing with vector spaces.
The problem at hand is of finding the generators of this submodule i.e. any element of the submodule should be expressible as a linear combination of the generating set. In this way the generators are capable of spanning the full submodule or generating the submodule. In order to achieve our goal, we rewrite the Eq. (\[opeq\]) explicitly component-wise: q\_1 + q\_1’ - \_3 q\_2’ - \_2 q\_3 &=& 0 ,\
q\_2 + q\_2’ - \_1 q\_3’ - \_3 q\_1 &=& 0 ,\
q\_3 + q\_3’ - \_2 q\_1’ - \_1 q\_2 &=& 0 . \[lneq\]
The first step is to use Gaussian elimination to obtain $q_1$ and $q_2$ in terms of $q_3, q_1', q_2', q_3'$:
q\_1 &=& - q\_1’ + \_3 q\_2’ + \_2 q\_3 ,\
q\_2 &=& - q\_2’ + \_1 q\_3’ + \_3 q\_1\
&=& - \_3 q\_1’ - (1 - \_3\^2) q\_2’ + \_1 q\_3’ + \_2 \_3 q\_3 , \[gauss\] and then substitute these values in the third equation to obtain a linear implicit relation between $q_3, q_1', q_2', q_3'$. We then have: (1 - \_1 \_2 \_3) q\_3 + (\_1 \_3 - \_2) q\_1’ + \_1 (1 - \_3\^2) q\_2’ + (1 - \_1\^2) q\_3’ = 0 . \[implct\] Obtaining solutions to Eq. (\[implct\]) amounts to solving the problem since the the remaining polynomials $q_1, q_2$ have been expressed in terms of $q_3, q_1', q_2', q_3'$ in (\[gauss\]). Note that we cannot carry on the Gaussian elimination process any further, because none of the polynomial coefficients appearing in Eq.(\[implct\]) have an inverse in the ring.
We will assume that the polynomials have rational coefficients [ *i.e*]{} the coefficients belong to ${\cal Q}$ the field of the rational numbers. The set of polynomials form a ring - the polynomial ring in three variables which we denote by ${\cal R}={\cal Q}[\D_1, \D_2,
\D_3]$. The polynomial vector $(q_3, q_1', q_2', q_3')\, \in \, {\cal
R}^4$. The set of solutions to (\[implct\]) is just the [ *kernel*]{} of the homomorphism $\varphi:\, {\cal R}^4 \rightarrow {\cal
R}$ where the polynomial vector $(q_3, q_1', q_2', q_3')$ is mapped to the polynomial $(1 - \D_1 \D_2 \D_3) q_3 + (\D_1 \D_3 -
\D_2) q_1' + \D_1 (1 - \D_3^2) q_2' + (1 - \D_1^2) q_3' $. Thus the solution space $ ker \varphi$ is a submodule of ${\cal R}^4$. It is called the [*module of syzygies*]{} in the literature. The generators of this module can be obtained from standard methods available in the literature. We briefly outline the method given in the books by Becker et al. [@becker] and Kreuzer and Robbiano [@KR] below. The details have been included in appendix A.
Gröbner Basis
-------------
The first step is to obtain the Gröbner basis for the ideal $\cU$ generated by the coefficients in Eq. (\[implct\]): u\_1 = 1 - \_1 \_2 \_3, u\_2 = \_1 \_3 - \_2, u\_3 = \_1 (1 - \_3\^2), u\_4 = 1 - \_1\^2 . \[idgen\] The ideal $\cU$ consists of linear combinations of the form $\sum v_i
u_i$ where $v_i$, $i = 1, ..., 4$ are polynomials in the ring ${\cal
R}$. There can be several sets of generators for $\cU$. A Gröbner basis is a set of generators which is ‘small’ in a specific sense.
There are several ways to look at the theory of Gröbner basis. One way is, suppose we are given polynomials $g_1, g_2, ..., g_m$ in one variable over say ${\cal Q}$ and we would like to know whether another polynomial $f$ belongs to the ideal generated by the $g$’s. A good way to decide the issue would be to first compute the gcd (greatest common divisor) $g$ of $g_1, g_2, ..., g_m$ and check whether $f$ is a multiple of $g$. One can achieve this by doing the long division of $f$ by $g$ and checking whether the remainder is zero. All this is possible because ${\cal Q}[x]$ is a Euclidean domain and also a principle ideal domain (PID) wherein any ideal is generated by a single element. Therefore we have essentially just one polynomial - the gcd - which generates the ideal generated by $g_1, g_2, \dots ,
g_m$. The ring of integers or the ring of polynomials in one variable over any field are examples of PIDs whose ideals are generated by single elements. However, when we consider more general rings (not PIDs) like the one we are dealing with here, we do not have a single gcd but a set of several polynomials which generates an ideal in general. A Gröbner basis of an ideal can be thought of as a generalization of the gcd. In the univariate case, the Gröbner basis reduces to the gcd.
Gröbner basis theory generalizes these ideas to multivariate polynomials which are neither Euclidean rings nor PIDs. Since there is in general not a single generator for an ideal, Gröbner basis theory comes up with the idea of dividing a polynomial with a [*set*]{} of polynomials, the set of generators of the ideal, so that by successive divisions by the polynomials in this generating set of the given polynomial, the remainder becomes zero. Clearly, every generating set of polynomials need not possess this property. Those special generating sets that do possess this property (and they exist!) are called Gröbner bases. In order for a division to be carried out in a sensible manner, an order must be put on the ring of polynomials, so that the final remainder after every division is strictly smaller than each of the divisors in the generating set. A natural order exists on the ring of integers or on the polynomial ring ${\cal Q}(x)$; the degree of the polynomial decides the order in ${\cal Q}(x)$. However, even for polynomials in two variables there is no natural order apriori (Is $x^2 + y$ greater or smaller than $x + y^2$?). But one can, by hand as it were, put an order on such a ring by saying $x >>
y$, where $ >> $ is an order, called the lexicographical order. We follow this type of order, $\D_1 >> \D_2 >> \D_3$ and ordering polynomials by considering their highest degree terms. It is possible to put different orderings on a given ring which then produce different Gröbner bases. Clearly, a Gröbner basis must have ‘small’ elements so that division is possible and every element of the ideal when divided by the Gröbner basis elements leaves zero remainder, [*i.e.*]{} every element modulo the Gröbner basis reduces to zero.
In the literature, there exists a well-known algorithm called the the Buchberger algorithm which may be used to obtain the Gröbner basis for a given set of polynomials in the ring. So a Gröbner basis of $\cU$ can be obtained from the generators $u_i$ given in Eq. (\[idgen\]) using this algorithm. It is essentially again a generalization of the usual long division that we perform on univariate polynomials. More conveniently, we prefer to use the well known application ‘Mathematica’. Mathematica yields a 3 element Gröbner basis ${\cal G}$ for $\cU$: = {\_3\^2 - 1, \_2\^2 - 1, \_1 - \_2 \_3 } . One can easily check that all the $u_i$ of Eq. (\[idgen\]) are linear combinations of the polynomials in ${\cal G}$ and hence ${\cal G}$ generates $\cU$. One also observes that the elements look ‘small’ in the order mentioned above. However, one can satisfy oneself that ${\cal G}$ is a Gröbner basis by using the standard methods available in the literature. One method consists of computing the S-polynomials (see Appendix A) for all the pairs of the Gröbner basis elements and checking whether these reduce to zero modulo ${\cal G}$.
This Gröbner basis of the ideal $\cU$ is then used to obtain the generators for the module of syzygies. Note that although the Gröbner basis depends on the order we choose among the $\D_k$, the module itself is [*independent*]{} of the order [@becker].
Generating Set for the Module of Syzygies {#sec:genset}
-----------------------------------------
The generating set for the module is obtained by further following the procedure in the literature [@becker; @KR]. The details are given in Appendix A, specifically for our case. We obtain 7 generators for the module. These generators do not form a minimal set and there are relations between them; in fact this method does not guarantee a minimum set of generators. These generators can be expressed as linear combinations of $\alpha, \beta, \gamma, \zeta$ and also in terms of $X^{(1)}, X^{(2)},
X^{(3)}, X^{(4)}$ given below in Eq. (\[gen6\]). The importance in obtaining the 7 generators is that the standard theorems guarantee that these 7 generators do in fact generate the required module. Therefore, from this proven set of generators we can check whether a particular set is in fact a generating set. We present several generating sets below.
Alternatively, we may use a software package called [*Macaulay 2*]{} which directly calculates the generators given the the equations (\[lneq\]). Using [*Macaulay 2*]{}, we obtain six generators. Again, Macaulay’s algorithm does not yield a minimal set; we can express the last two generators in terms of the first four. Below we list this smaller set of four generators in the order $X = (q_1, q_2, q_3, q_1', q_2', q_3')$: X\^[(1)]{} &=& (\_2 - \_1 \_3, 0, 1 - \_3\^2, 0, \_2 \_3 - \_1, \_3\^2 - 1) ,\
X\^[(2)]{} &=& (- \_1, - \_2, - \_3, \_1, \_2, \_3) ,\
X\^[(3)]{} &=& (-1, - \_3, - \_1 \_3, 1, \_1 \_2, \_2) ,\
X\^[(4)]{} &=& (- \_1 \_2, -1, - \_1, \_3, 1, \_2 \_3) . \[gen6\] Note that the last three generators are just $X^{(2)} = \zeta, X^{(3)} =
\alpha, X^{(4)} = \beta$. An extra generator $X^{(1)}$ is needed to generate all the solutions.
Another set of generators which may be useful for further work is a Gröbner basis of a module. The concept of a Gröbner basis of an ideal can be extended to that of a Gröbner basis of a submodule of $(K[x_1, x_2, ..., x_n])^m$ where $K$ is a field, since a module over the polynomial ring can be considered as generalization of an ideal in a polynomial ring. Just as in the case of an ideal, a Gröbner basis for a module is a generating set with special properties. For the module under consideration we obtain a Gröbner basis using : G\^[(1)]{} &=& (- \_1, - \_2, - \_3, \_1, \_2, \_3) ,\
G\^[(2)]{} &=& (\_2-\_1 \_3, 0, 1 - \_3\^2, 0, \_2 \_3 - \_1, \_3\^2 - 1) ,\
G\^[(3)]{} &=& (- \_1 \_2, -1, - \_1, \_3, 1, \_2 \_3) ,\
G\^[(4)]{} &=& (-1, -\_3, -\_1 \_3, 1, \_1 \_2, \_2) ,\
G\^[(5)]{} &=& (\_3 (1 - \_1\^2), \_3\^2 - 1, 0, 0, 1 - \_1\^2, \_1 (\_3\^2 - 1)) . \[grbn6\] Note that in this Gröbner basis $G^{(1)} = \zeta = X^{(2)}, G^{(2)} =
X^{(1)}, G^{(3)} = \beta = X^{(4)}, G^{(4)} = \alpha = X^{(3)}$. Only $G^{(5)}$ is the new generator.
Another set of generators are just $\alpha, \beta, \gamma$ and $\zeta$. This can be checked using [*Macaulay 2*]{} or one can relate $\alpha,
\beta, \gamma$ and $\zeta$ to the generators $X^{(A)}, A = 1, 2, 3, 4$ by polynomial matrices. In Appendix B, we express the 7 generators we obtained following the literature, in terms of $\alpha, \beta, \gamma$ and $\zeta$. Also we express $\alpha, \beta, \gamma$ and $\zeta$ in terms of $X^{(A)}$. This proves that all these sets generate the required module of syzygies.
The question now arises as to which set of generators we should choose which facilitates further analysis. The analysis is simplified if we choose a smaller number of generators. Also we would prefer low degree polynomials to appear in the generators so as to avoid cancellation of leading terms in the polynomials. By these two criteria we may choose, $X^{(A)}$ or $\alpha, \beta, \gamma, \zeta$. However, $\alpha, \beta,
\gamma, \zeta$ possess the additional property that this set is left invariant under a cyclic permutation of indices $1,2,3$. It is found that this set is more convenient to use because of this symmetry.
Canceling optical bench motion noise
------------------------------------
There are now twelve Doppler data streams which have to be combined in an appropriate manner in order to cancel the noise from the laser as well as from the motion of the optical benches. As in the previous case of cancelling laser phase noise, here too, we keep the relevant terms only, namely, those terms containing laser phase noise and optical bench motion noise. We then have the following expressions for the four data streams on spacecraft $1$: s\_1 &=& \_3 \[p\_2’ + \_0 \_3 \_2’ \] - \[ p\_1 - \_0 \_3 \_1 \] , \[s1\]\
s\_1’ &=& \_2 \[p\_3 - \_0 \_2 \_3 \] - \[p\_1’ + \_0 \_2 \_1’ \] , \[s1’\]\
\_1 &=& p\_1’ - p\_1 + 2 \_0 \_2 \_1’ + \_1 , \[tau1\]\
\_1’ &=& p\_1 - p\_1’ - 2 \_0 \_3 \_1 + \_1 \[tau1’\] . The other eight data streams on spacecraft 2 and 3 are obtained by cyclic permutations of the indices in the above equations. In order to simplify the derivation of the expressions cancelling the optical bench noises, we note that by subtracting eq. (\[tau1’\]) from eq. (\[tau1\]), we can rewriting the resulting expression (and those obtained from it by permutation of the spacecraft indices) in the following form, $$z_1 \equiv \frac{1}{2} (\tau_1 - \tau_1') = \phi_1' - \phi_1 \, ,
\label{z1}$$ where $\phi_1'$, $\phi_1$ are defined as, \_1’ && p\_1’ + \_0 \_2 \_1’ ,\
\_1 && p\_1 - \_0 \_3 \_1 , The importance in defining these combinations is that the expressions for the data streams $s_i$, $s_i'$ simplify into the following form, s\_1 &=& \_3 \_2’ - \_1 ,\
s\_1’ &=& \_2 \_3 - \_1’ . If we now combine the $s_i$, $s_i'$, and $z_i$ in the following way, $$\begin{aligned}
\eta_{1} & \equiv & s_{1} - \D_3 z_2
= \D_3 \phi_2 - \phi_1
\ , \
\eta_{1'} \equiv s_{1'} + z_1
= \D_{2} \phi_3 - \phi_1 \ ,
\label{eq:novea}
\\
\eta_{2} & \equiv & s_{2} - \D_1 z_3
= \D_1 \phi_3 - \phi_2
\ , \
\eta_{2'} \equiv s_{2'} + z_2
= \D_{3} \phi_1 - \phi_2 \ ,
\label{eq:noveb}
\\
\eta_{3} & \equiv & s_{3} - \D_2 z_1
= \D_2 \phi_1 - \phi_3
\ , \
\eta_{3'} \equiv s_{3'} + z_3
= \D_{1} \phi_2 - \phi_3 \ ,
\label{eq:novec}\end{aligned}$$ we have just reduced the problem of cancelling of six laser and six optical bench noises to the equivalent problem of removing the three random processes, $\phi_1$, $\phi_2$, and $\phi_3$, from the six linear combinations, $\eta_i$, $\eta_i'$, of the one-way measurements $s_i$, $s_i'$, and $z_i$. By comparing the equations above to equation (\[beams\]) for the simpler configuration with only three lasers, analyzed in the previous section, we see that they are identical in form.
Physical Interpretation of the TDI combinations
-----------------------------------------------
It is important to notice that the four interferometric combinations ($\alpha, \beta, \gamma, \zeta$), which can be used as a basis for generating the entire TDI space, are actually synthesized Sagnac interferometers. This can be seen by rewriting the expression for $\alpha$, for instance, in the following form, $$\alpha = [\eta_{1'} + \D_{2} \eta_{3'} + \D_{1} \D_{2'} \eta_{2'}]
-
[\eta_{1} + \D_{3} \eta_{2} + \D_{1} \D_{3} \eta_{2}]
\, ,
\label{alpha_Phys}$$ and noticing that the first square bracket on the right-hand side of equation (\[alpha\_Phys\]) contains a combination of one-way measurements describing a light beam propagating clockwise around the array, while the other terms in the second square-bracket give the equivalent of another beam propagating counter-clockwise around the constellation.
Contrary to $\alpha$, $\beta$, and $\gamma$, $\zeta$ can not be visualized as the difference(or interference) of two synthesized beams. However, it should still be regarded as a Sagnac combination since there exists a time-delay relationship between it and $\alpha$, $\beta$, and $\gamma$ [@AET99], $$\zeta - \D_1 \D_2 \D_3 \zeta = \D_1 \alpha - \D_2 \D_3 \alpha +
\D_2 \alpha - \D_3 \D_1 \beta + \D_3 \gamma - \D_1 \D_2 \gamma \, .
\label{zeta_all}$$ As a consequence of the time-structure of this relationship, $\zeta$ has been called the [*Symmetrized Sagnac*]{} combination.
By using the four generators, it is possible to construct several other interferometric combinations, such as the unequal-arm Michelson ($X, Y, Z$), the Beacons ($P, Q, R$), the Monitors ($E, F, G$), and the Relays ($U, V, W$). Contrary to the Sagnac combinations, these only use four of the six data combinations $\eta_i$, $\eta_i'$. For this reason they have obvious utility in the event of selected subsystem failures [@ETA00].
These observables can be written in terms of the Sagnac observables ($\alpha, \beta, \gamma, \zeta$) in the following way, $$\begin{aligned}
\D_1 X & = & \D_2 \D_3 \alpha - \D_2 \beta - \ D_3 \gamma + \zeta \ ,
\no \\
P & = & \zeta - \D_1 \alpha \ ,
\no \\
E & = & \D_1 - \D_1 \zeta \ ,
\no \\
U & = & \D_1 \gamma - \beta \ ,
\label{relationships}\end{aligned}$$ as it is easy to verify by substituting the expressions for the Sagnac combinations into the above equations. Their physical interpretations are schematically shown in Figure \[All\].
![Schematic diagrams of the unequal-arm Michelson, Monitor, Beacon, and Relay combinations. These TDI combinations rely only on four of the six one-way Doppler measurements, as illustrated here.[]{data-label="All"}](Fig5.eps){width="6in"}
In the case of the combination $X$, in particular, by writing it in the following form [@AET99], $$X = [(\eta_1' + \D_{2'}\eta_3) + \D_{2'} \D_2 (\eta_{1} + \D_3 \eta_2')]
- [(\eta_{1} + \D_3 \eta_2') + \D_3 \D_{3'} (\eta_1' + \D_{2'} \eta_{3})] \ ,
\label{Xcomb}$$ one can notice (as pointed out in [@Summers] and [@STEA03]) that this combination can be visualized as the difference of two sums of phase measurements, each corresponding to a specific light path from a laser onboard spacecraft $1$ having phase noise $\phi_1$. The first square-bracket term in equation (\[Xcomb\]) represents a synthesized light-beam transmitted from spacecraft $1$ and made to bounce once at spacecraft $2$ and $3$ respectively. The second square-bracket term instead corresponds to another beam also originating from the same laser, experiencing the same overall delay as the first beam, but bouncing off spacecraft $3$ first and then spacecraft $2$. When they are recombined they will cancel the laser phase fluctuations exactly, having both experienced the same total delay (assuming stationary spacecraft). The $X$ combinations should therefore be regarded as the response of a zero-area Sagnac interferometer.
Time-Delay Interferometry with moving spacecraft {#SECV}
================================================
The rotational motion of the LISA array results in a difference of the light travel times in the two directions around a Sagnac circuit [@S03],[@CH03]. Two time delays along each arm must be used, say $L_i^{'}$ and $L_i$ for clockwise or counterclockwise propagation as they enter in any of the TDI combinations. Furthermore, since $L_i$ and $L_i^{'}$ not only differ from one another but can be time dependent (they “flex”), it was shown that the “first generation” TDI combinations do not completely cancel the laser phase noise (at least with present laser stability requirements), which can enter at a level above the secondary noises. For LISA, and assuming $\dot L_i \simeq
10 {\rm m/sec}$ [@Folkner], the estimated magnitude of the remaining frequency fluctuations from the laser can be about $30$ times larger than the level set by the secondary noise sources in the center of the frequency band. In order to solve this potential problem, it has been shown that there exist new TDI combinations that are immune to first order shearing (flexing, or constant rate of change of delay times). These combinations can be derived by using the time-delay operators formalism introduced in the previous section, although one has to keep in mind that now these operators no longer commute [@TEA04].
In order to derive the new, “flex-free” TDI combinations we will start by taking specific combinations of the one-way data entering in each of the expressions derived in the previous section. These combinations are chosen in such a way so as to retain only one of the three noises $\phi_i , i=1, 2, 3$ if possible. In this way we can then implement an iterative procedure based on the use of these basic combinations and of time-delay operators, to cancel the laser noises after dropping terms that are quadratic in $\dot{L}/c$ or linear in the accelerations. This iterative time-delay method, to first order in the velocity, is illustrated abstractly as follows. Given a function of time $\Psi = \Psi(t)$, time delay by $L_i$ is now denoted either with the standard comma notation [@AET99] or by applying the delay operator $\D_{i}$ introduced in the previous section $$\D_{i} \Psi = \Psi_{,i} \equiv \Psi(t - L_i(t)) \ .
\label{eq:25}$$ We then impose a second time delay $L_j(t)$: $$\begin{aligned}
\D_{j} \D_{i} \Psi = \Psi_{;ij} & \equiv & \Psi(t - L_j(t) - L_i(t - L_j(t)))
\nonumber \\
& \simeq & \Psi(t - L_j(t) - L_i(t) + \dot L_i(t) L_j)
\nonumber \\
& \simeq & \Psi_{,ij} + \dot \Psi_{,ij} \dot L_i L_j \ .
\label{eq:26}\end{aligned}$$
A third time delay $L_k(t)$ gives: $$\begin{aligned}
\D_{k} \D_{j} \D_{i} \Psi = \Psi_{;ijk} & = &
\Psi(t - L_k(t) - L_j(t - L_k(t)) - L_i(t - L_k(t) - L_j(t - L_k(t))))
\nonumber \\
& & \simeq \Psi_{,ijk} + \dot \Psi_{,ijk} [\dot L_i (L_j + L_k) + \dot L_j L_k] \ ,
\label{eq:27}\end{aligned}$$ and so on, recursively; each delay generates a first-order correction proportional to its rate of change times the sum of all delays coming after it in the subscripts. Commas have now been replaced with semicolons [@STEA03], to remind us that we consider moving arrays. When the sum of these corrections to the terms of a data combination vanishes, the combination is called flex-free.
Also, note that each delay operator, $\D_{i}$, has a unique inverse, $D^{-1}_{i}$, whose expression can be derived by requiring that $D^{-1}_{i} \D_{i} = I$, and neglecting quadratic and higher order velocity terms. Its action on a time series $\Psi (t)$ is $$D^{-1}_{i} \Psi (t) \equiv \Psi (t + L_i (t + L_i)) \ .
\label{eq:28}$$ Note that this is not like an advance operator one might expect, since it advances not by $L_i (t)$ but rather $L_i (t + L_i)$.
The Unequal-Arm Michelson
-------------------------
The unequal-arm Michelson combination relies on the four measurements $\eta_{1}$, $\eta_{1'}$, $\eta_{2'}$, and $\eta_{3}$. Note that the two combinations $\eta_{1} + \eta_{2',3}$, $\eta_{1'} + \eta_{3,2'}$ represent the two synthesized two-way data measured onboard spacecraft $1$, and can be written in the following form $$\begin{aligned}
\eta_{1} + \eta_{2',3} & = & \left(\D_{3}\D_{3'} - I\right) \ \phi_1 \ ,
\label{eq:diecia}
\\
\eta_{1'} + \eta_{3,2'} & = & \left(\D_{2'}\D_{2} - I\right) \ \phi_1 \ ,
\label{eq:diecib}\end{aligned}$$ where $I$ is the identity operator. Since in the stationary case any pairs of these operators commute, i.e. $\D_i
\D_{j'} - \D_{j'} \D_i = 0$, from equations (\[eq:diecia\], \[eq:diecib\]) it is easy to derive the following expression for the unequal-arm interferometric combination, $X$, which eliminates, $\phi_1$ $$X = \left[\D_{2'}\D_{2} - I\right] (\eta_{1} + \eta_{2',3}) -
\left[\left(\D_{3}\D_{3'} - I\right)\right] (\eta_{1'} + \eta_{3,2'}).
\label{eq:undici}$$ If, on the other hand, the time-delays depend on time, the expression of the unequal-arm Michelson combination above no longer cancels $\phi_1$. In order to derive the new expression for the unequal-arm interferometer that accounts for “flexing”, let us first consider the following two combinations of the one-way measurements entering into the $X$ observable given in equation (\[eq:undici\]): $$\begin{aligned}
\left[(\eta_{1'} + \eta_{3;2'}) + (\eta_{1} + \eta_{2;3})_{;22'}\right] & = &
\left[D_{2'}D_{2}D_{3}D_{3'} - I \right] \phi_1 \ ,
\label{eq:ventinovea}
\\
\left[(\eta_{1} + \eta_{2';3}) + (\eta_{1'} + \eta_{3;2'})_{;3'3}\right] & = &
\left[D_{3}D_{3'}D_{2'}D_{2} - I \right] \phi_1 \ .
\label{eq:ventinoveb}\end{aligned}$$ Using equations (\[eq:ventinovea\], \[eq:ventinoveb\]) we can use the delay technique again to finally derive the following expression for the new unequal-arm Michelson combination $X_1$ that accounts for the flexing effect, $$\begin{aligned}
X_1 & = & \left[D_{2}D_{2'}D_{3'}D_{3} - I \right] \
\left[(\eta_{21} + \eta_{12;3'}) + (\eta_{31} +
\eta_{13;2})_{;33'}\right]
\nonumber
\\
& & -
\left[D_{3'}D_{3}D_{2}D_{2'} - I \right] \
\left[(\eta_{31} + \eta_{13;2}) + (\eta_{21} + \eta_{12;3'})_{;2'2}\right].
\label{eq:trenta}\end{aligned}$$
As usual, $X_2$ and $X_3$ are obtained by cyclic permutation of the spacecraft indices. This expression is readily shown to be laser-noise-free to first order of spacecraft separation velocities $\dot L_i$: it is “flex-free”.
The Sagnac Combinations
-----------------------
In the above subsection we have used the same symbol $X$ for the unequal-arm Michelson combination for both the rotating (i.e. constant delay times) and stationary cases. This emphasizes that, for this TDI combination (and, as we will see below, also for all the combinations including only four links) the forms of the equations do not change going from systems at rest to the rotating case. One needs only distinguish between the time-of-flight variations in the clockwise and counter-clockwise senses (primed and unprimed delays).
In the case of the Sagnac variables, ($\alpha, \beta, \gamma, \zeta$), however, this is not the case as it is easy to understand on simple physical grounds. In the case of $\alpha$ for instance, light originating from spacecraft 1 is simultaneously sent around the array on clockwise and counterclockwise loops, and the two returning beams are then recombined. If the array is rotating, the two beams experience a different delay (the Sagnac effect), preventing the noise $\phi_1$ from cancelling in the $\alpha$ combination.
In order to find the solution to this problem let us first rewrite $\alpha$ in such a way to explicitly emphasize what it does: attempts to remove the same fluctuations affecting two beams that have been made to propagated clockwise and counter-clockwise around the array, $$\alpha = [\eta_{1'} + \D_{2'} \eta_{3'} + \D_{1'} \D_{2'} \eta_{2'}]
-
[\eta_{1} + \D_{3} \eta_{2} + \D_{1} \D_{3} \eta_{2}]
\, ,
\label{alpha}$$ where we have accounted for clockwise and counterclockwise light delays. It is straightforward to verify that this combination no longer cancels the laser and optical bench noises. If, however, we expand the two terms inside the square-brackets on the right-hand-side of equation (\[alpha\]) we find that they are equal to: $$[\eta_{1'} + \D_{2'} \eta_{3'} + \D_{1'} \D_{2'} \eta_{2'}] =
[\D_{2'}\D_{1'}\D_{3'} - I] \phi_1
\nonumber$$ $$[\eta_{1} + \D_{3} \eta_{2} + \D_{1} \D_{3} \eta_{2}] =
[\D_{3}\D_{1}\D_{2} - I] \phi_1 \, .
\label{alpha_int}$$ If we now apply our iterative scheme to the combinations given in equation (\[alpha\_int\]) we finally get the expression for the Sagnac combination, $\alpha_1$, that is unaffected by laser noise in presence of rotation, $$\alpha_1 = [\D_{3}\D_{1}\D_{2} - I] \
[\eta_{1'} + \D_{2'} \eta_{3'} + \D_{1'} \D_{2'} \eta_{2'}]
- [\D_{2'}\D_{1'}\D_{3'} - I]
[\eta_{1} + \D_{3} \eta_{2} + \D_{1} \D_{3} \eta_{2}] \ .
\label{alpha1}$$ If the delay-times are also time-dependent, we find that the residual laser noise remaining into the combination $\alpha_1$ is actually equal to $$\begin{aligned}
\dot \phi_{1,1231'2'3'}
[(\dot L_1 + \dot L_2 + \dot L_3) (L_1^{'} + L_2^{'} + L_3^{'})
- (\dot L_1^{'} + \dot L_2^{'} + \dot L_3^{'}) (L_1 + L_2 + L_3)] \ .\end{aligned}$$ Fortunately, although first order in the relative velocities, the residual is small, as it involves the difference of the clockwise and counterclockwise rates of change of the propagation delays on the [*[same]{}*]{} circuit. For LISA, the remaining laser phase noises in $\alpha_i$, $i = 1, 2, 3$, are several orders of magnitude below the secondary noises.
In the case of $\zeta$, however, the rotation of the array breaks the symmetry and therefore its uniqueness. However, there still exist three generalized TDI laser-noise-free data combinations that have properties very similar to $\zeta$, and which can be used for the same scientific purposes [@TAE01]. These combinations, which we call ($\zeta_1, \zeta_2, \zeta_3$), can be derived by applying again our time-delay operator approach.
Let us consider the following combination of the $\eta_{i}$, $\eta_{i'}$ measurements, each being delayed only once [@AET99]: $$\begin{aligned}
\eta_{3,3} - \eta_{3',3} + \eta_{1,1'} & = & \left[D_{3} D_{2} -
D_{1'}\right] \phi_1 \ ,
\label{eq:ventiduea}
\\
\eta_{1',1} - \eta_{2,2'} + \eta_{2',2'} & = & \left[D_{3'} D_{2'} -
D_{1}\right] \phi_1 \ ,
\label{eq:ventidueb}\end{aligned}$$ where we have used the commutativity property of the delay operators in order to cancel the $\phi_2$ and $\phi_3$ terms. Since both sides of the two equations above contain only the $\phi_1$ noise, $\zeta_1$ is found by the following expression: $$\zeta_1 = \left[D_{3'}D_{2'} - D_{1}\right] \left(\eta_{31,1'} - \eta_{32,2} + \eta_{12,2}\right)
- \left[D_{2}D_{3} - D_{1'}\right]\left(\eta_{13,3'} - \eta_{23,3'} +
\eta_{21,1}\right) \ .
\label{eq:ventitre}$$ If the light-times in the arms are equal in the clockwise and counterclockwise senses (e.g. no rotation) there is no distinction between primed and unprimed delay times. In this case, $\zeta_1$ is related to our original symmetric Sagnac $\zeta$ by $\zeta_1 =
\zeta_{,23} - \zeta_{,1}$. Thus for the practical LISA case (arm length difference $< 1 \%$), the SNR of $\zeta_1$ will be the same as the SNR of $\zeta$.
If the delay-times also change with time, the perfect cancellation of the laser noises is no longer achieved in the ($\zeta_1, \zeta_2,
\zeta_3$) combinations. However, it has been shown in [@TEA04] that the magnitude of the residual laser noises in these combinations are significantly smaller than the LISA secondary system noises, making their effects entirely negligible.
The expressions for the Monitor, Beacon, and Relay combinations, accounting for the rotation and flexing of the LISA array, have been derived in the literature [@TEA04] by applying the time-delay iterative procedure highlighted in this section. The interested reader is referred to that paper for details.
A mathematical formulation of the “second generation” TDI, which generalizes the one presented in Section \[SECIV\] for the stationary LISA, still needs to be derived. In the case when only the Sagnac effect is considered (and the delay-times remain constant in time) the mathematical formulation of Section \[SECIV\] can be extended in a straight forward way where now the six time-delays $\D_{i}$ and $\D_{i}'$ must be taken into account. The polynomial ring is now in these six variables and the corresponding module of syzygies can be constructed over this enlarged polynomial ring [@NV04]. However, when the arms are allowed to flex, that is, the operators themselves are functions of time, the operators no longer commute. One must then resort to non-commutative Gröbner basis methods. We will investigate this mathematical problem in the near future.
Optimal LISA Sensitivity {#SECVI}
========================
All the above interferometric combinations have been shown to individually have rather different sensitivities [@ETA00], as a consequence of their different responses to gravitational radiation and system noises. Since LISA has the capability of observing a gravitational wave signal with many different interferometric combinations (all having different antenna patterns and noises), we should no longer regard LISA as a single detector system but rather as an array of gravitational wave detectors working in coincidence. This suggests that the presently adopted LISA sensitivity could be improved by [*optimally*]{} combining elements of the TDI space.
Before proceeding with this idea, however, let us consider again the so called “second generation” TDI Sagnac observables: ($\alpha_1,
\alpha_2, \alpha_3$). The expressions of the gravitational wave signal and the secondary noise sources entering into $\alpha_1$ will in general be different from those entering into $\alpha$, the corresponding Sagnac observable derived under the assumption of a stationary LISA array [@AET99], [@ETA00]. However, the other remaining, secondary noises in LISA are so much smaller, and the rotation and systematic velocities in LISA are so intrinsically small, that index permutation may still be done for them [@TEA04]. It is therefore easy to derive the following relationship between the signal and secondary noises in $\alpha_1$, and those entering into the stationary TDI combination $\alpha$ [@STEA03], [@TEA04]: $$\alpha_1 (t) \simeq \alpha (t) - \alpha (t - L_1 - L_2 - L_3) \ ,
\label{eq:duebis}$$ where $L_i \ , \ \ i=1, 2, 3$ are the unequal-arm lengths of the stationary LISA array. Equation (\[eq:duebis\]) implies that any data analysis procedure and algorithm that will be implemented for the second-generation TDI combinations can actually be derived by considering the corresponding “first generation” TDI combinations. For this reason, from now on we will focus our attention on the gravitational wave responses of the first-generation TDI observables ($\alpha, \beta, \gamma, \zeta$).
As a consequence of these considerations, we can still regard ($\alpha,
\beta, \gamma, \zeta$) as the generators of the TDI space, and write the most general expression for an element of the TDI space, $\eta
(f)$, as a linear combination of the Fourier transforms of the four generators (${\widetilde {\alpha}}, {\widetilde {\beta}}, {\widetilde
{\gamma}}, {\widetilde {\zeta}}$) $$\eta(f) \equiv a_1 (f, {\vec \lambda}) \ {\widetilde{\alpha}} (f) \ +
\ a_2 (f, {\vec \lambda}) \ {\widetilde{\beta}} (f) \ +
\ a_3 (f, {\vec \lambda}) \ {\widetilde{\gamma}} (f) \ +
\ a_4 (f, {\vec \lambda}) \ {\widetilde{\zeta}} (f) \ ,
\label{eq:otto}$$ where the $\{a_i (f, \vec \lambda)\}^4_{i=1}$ are arbitrary complex functions of the Fourier frequency $f$, and of a vector $\vec \lambda$ containing parameters characterizing the gravitational wave signal (source location in the sky, waveform parameters, etc.) and the noises affecting the four responses (noise levels, their correlations, etc.). For a given choice of the four functions $\{a_i \}^4_{i=1}$, $\eta$ gives an element of the functional space of interferometric combinations generated by ($\alpha, \beta, \gamma, \zeta$). Our goal is therefore to identify, for a given gravitational wave signal, the four functions $\{a_i \}^4_{i=1}$ that maximize the signal-to-noise ratio, ${\rm SNR}_{\eta}^2$, of the combination $\eta$, $${\rm SNR}_{\eta}^2 =
\int_{f_{l}}^{f_u}
\frac
{|a_1 \ {\widetilde \alpha_s} + a_2 \ {\widetilde \beta_s} + a_3 \
{\widetilde \gamma_s} + a_4 {\widetilde \zeta_s} |^2}
{\langle|a_1 \ {\widetilde \alpha_n} + a_2 \ {\widetilde \beta_n} +
a_3 \ {\widetilde \gamma_n} + a_4 {\widetilde \zeta_n} |^2 \rangle} \ df \ .
\label{eq:9bis}$$ In equation (\[eq:9bis\]) the subscripts $s$ and $n$ refer to the signal and the noise parts of (${\widetilde {\alpha}}, {\widetilde
{\beta}}, {\widetilde {\gamma}}, {\widetilde {\zeta}}$) respectively, the angle brackets represent noise ensemble averages, and the interval of integration ($f_l, f_u$) corresponds to the frequency band accessible by LISA.
Before proceeding with the maximization of the ${\rm SNR}_{\eta}^2$ we may notice from equation (\[zeta\_all\]) that the Fourier transform of the totally symmetric Sagnac combination, $\widetilde \zeta$, multiplied by the transfer function $1 - e^{2 \pi i f (L_1 + L_2 + L_3)}$ can be written as a linear combination of the Fourier transforms of the remaining three generators ($ {\widetilde {\alpha}}, {\widetilde
{\beta}}, {\widetilde {\gamma}}$). Since the signal-to-noise ratio of $\eta$ and $(1 - e^{2 \pi i f (L_1 + L_2 + L_3)}) \eta$ are equal, we may conclude that the optimization of the signal-to-noise ratio of $\eta$ can be performed only on the three observables $\alpha, \beta,
\gamma$. This implies the following redefined expression for ${\rm SNR}_{\eta}^2$: $${\rm SNR}_{\eta}^2 =
\int_{f_{l}}^{f_u}
\frac
{|a_1 \ {\widetilde \alpha_s} + a_2 \ {\widetilde \beta_s} + a_3 \
{\widetilde \gamma_s} |^2}
{\langle|a_1 \ {\widetilde \alpha_n} + a_2 \ {\widetilde \beta_n} +
a_3 \ {\widetilde \gamma_n} |^2 \rangle} \ df \ .
\label{eq:9}$$ The ${\rm SNR}_{\eta}^2$ can be regarded as a functional over the space of the three complex functions $\{ a_i \}^3_{i=1}$, and the particular set of complex functions that extremize it can of course be derived by solving the associated set of Euler-Lagrange equations.
In order to make the derivation of the optimal SNR easier, let us first denote by ${\bf x}^{(s)}$ and ${\bf x}^{(n)}$ the two vectors of the signals (${\widetilde{\alpha}}_s, {\widetilde{\beta}}_s,
{\widetilde{\gamma}}_s$) and the noises (${\widetilde{\alpha}}_n,
{\widetilde{\beta}}_n, {\widetilde{\gamma}}_n$) respectively. Let us also define $\bf a$ to be the vector of the three functions $\{a_i
\}^3_{i=1}$, and denote with ${\bf C}$ the hermitian, non-singular, correlation matrix of the vector random process ${\bf x}_n$, $$({\bf C})_{rt} \equiv \langle {\bf x}^{(n)}_{r} {\bf x}^{(n)*}_{t} \rangle \ .
\label{eq:14}$$ If we finally define $({\bf A})_{ij}$ to be the components of the hermitian matrix ${\bf x}^{(s)}_i {\bf x}^{(s)*}_j$, we can rewrite ${\rm SNR}_{\eta}^2$ in the following form, $${\rm SNR}_{\eta}^2 =
\int_{f_{l}}^{f_u}
\frac{{\bf a}_i {\bf A}_{ij} {\bf a}^*_j }
{{\bf a}_r {\bf C}_{rt} {\bf a}^*_t }
\ df \ ,
\label{eq:16}$$ where we have adopted the usual convention of summation over repeated indices. Since the noise correlation matrix ${\bf C}$ is non-singular, and the integrand is positive definite or null, the stationary values of the signal-to-noise ratio will be attained at the stationary values of the integrand, which are given by solving the following set of equations (and their complex conjugated expressions): $$\frac{\partial}{\partial {\bf a}_k} \ \left[
\frac
{{\bf a}_i {\bf A}_{ij} {\bf a}^*_j }
{{\bf a}_r {\bf C}_{rt} {\bf a}^*_t }
\right] = 0 \ \ \ , \ \ \ k = 1, 2, 3 \ .
\label{eq:16bis}$$ After taking the partial derivatives, equation (\[eq:16bis\]) can be rewritten in the following form, $$({\bf C}^{-1})_{ir}
({\bf A})_{rj} ({\bf a}^*)_j = \left[
\frac
{{\bf a}_p {\bf A}_{pq} {\bf a}^*_q }
{{\bf a}_l {\bf C}_{lm} {\bf a}^*_m }
\right]
\ ({\bf a}^*)_i \ \ \ , \ \ \ i = 1, 2, 3
\label{eq:16tris}$$ which tells us that the stationary values of the signal-to-noise ratio of $\eta$ are equal to the eigenvalues of the the matrix ${\bf C^{-1}
\cdot A}$. The result in equation (\[eq:16bis\]) is well known in the theory of quadratic forms, and it is called the Rayleigh’s principle [@Noble69], [@Selby64].
In order now to identify the eigenvalues of the matrix ${\bf C^{-1}
\cdot A}$, we first notice that the $3 \times 3$ matrix ${\bf A}$ has rank $1$. This implies that the matrix ${\bf
C}^{-1} \cdot {\bf A}$ has also rank $1$, as it is easy to verify. Therefore two of its three eigenvalues are equal to zero, while the remaining non-zero eigenvalue represents the solution we are looking for.
The analytic expression of the third eigenvalue can be obtained by using the property that the trace of the $3 \times 3$ matrix ${\bf
C}^{-1} \cdot {\bf A}$ is equal to the sum of its three eigenvalues, and in our case to the eigenvalue we are looking for. From these considerations we derive the following expression for the optimized signal-to-noise ratio ${{\rm SNR}_{\eta}^2}_{\rm opt.}$: $${{\rm SNR}_{\eta}^2}_{\rm opt.} = \int_{f_{l}}^{f_u}
{\bf x}^{(s)*}_i \ ({\bf C}^{-1})_{ij} \ {\bf x}^{(s)}_j \ df \ .
\label{eq:18}$$ We can summarize the results derived in this section, which are given by equations (\[eq:9\],\[eq:18\]), in the following way:
\(i) among all possible interferometric combinations LISA will be able to synthesize with its four generators $\alpha, \beta, \gamma, \zeta$, the particular combination giving maximum signal-to-noise ratio can be obtained by using only three of them, namely ($\alpha, \beta,
\gamma$);
\(ii) the expression of the optimal signal-to-noise ratio given by equation (\[eq:18\]) implies that LISA should be regarded as a network of three interferometer detectors of gravitational radiation (of responses ($\alpha, \beta, \gamma$)) working in coincidence [@Finn01; @NPDV03_1].
General application {#GA}
-------------------
As an application of equation (\[eq:18\]), here we calculate the sensitivity that LISA can reach when observing sinusoidal signals uniformly distributed on the celestial sphere and of random polarization. In order to calculate the optimal signal-to-noise ratio we will also need to use a specific expression for the noise correlation matrix ${\bf C}$. As a simplification, we will assume the LISA arm-lengths to be equal to its nominal value $L = 16.67 \ {\rm
sec.}$, the optical-path noises to be equal and uncorrelated to each other, and finally the noises due to the proof-mass noises to be also equal, uncorrelated to each other and to the optical-path noises. Under these assumptions the correlation matrix becomes real, its three diagonal elements are equal, and all the off-diagonal terms are equal to each other, as it is easy to verify by direct calculation [@ETA00]. The noise correlation matrix ${\bf C}$ is therefore uniquely identified by two real functions, $S_{\alpha}$ and $S_{\alpha
\beta}$, in the following way $${\bf C} = \left( \begin{array}{ccc}
S_{\alpha} & S_{\alpha \beta} & S_{\alpha \beta} \\
S_{\alpha \beta} & S_{\alpha} & S_{\alpha \beta} \\
S_{\alpha \beta} & S_{\alpha \beta} & S_{\alpha}
\end{array}
\label{eq:19}
\right).$$
The expression of the optimal signal-to-noise ratio assumes a rather simple form if we diagonalize this correlation matrix by properly “choosing a new basis”. There exists an orthogonal transformation of the generators ($ {\widetilde {\alpha}}, {\widetilde
{\beta}}, {\widetilde {\gamma}}$) which will transform the optimal signal-to-noise ratio into the sum of the signal-to-noise ratios of the “transformed” three interferometric combinations. The expressions of the three eigenvalues $\{\mu_i\}^3_{i=1}$ (which are real) of the noise correlation matrix ${\bf C}$ can easily be found by using the algebraic manipulator [*Mathematica*]{} [@Wolf02], and they are equal to $$\mu_1 = \mu_2 = S_{\alpha} - S_{\alpha \beta} \ \ \ , \ \ \
\mu_3 = S_{\alpha} + 2 \ S_{\alpha \beta} \ .
\label{eq:20}$$ Note that two of the three real eigenvalues, ($\mu_1, \mu_2$), are equal. This implies that the eigenvector associated to $\mu_3$ is orthogonal to the two-dimensional space generated by the eigenvalue $\mu_1$, while any chosen pair of eigenvectors corresponding to $\mu_1$ will not necessarily be orthogonal. This inconvenience can be avoided by choosing an arbitrary set of vectors in this two-dimensional space, and by ortho-normalizing them. After some simple algebra, we have derived the following three ortho-normalized eigenvectors $${\bf v_1} = \frac{1}{\sqrt{2}} \ (-1, 0, 1) \ \ \ , \ \ \
{\bf v_2} = \frac{1}{\sqrt{6}} \ (1, -2, 1) \ \ \ , \ \ \
{\bf v_3} = \frac{1}{\sqrt{3}} \ (1, 1, 1) \ .
\label{eq:21}$$ Equation (\[eq:21\]) implies the following three linear combinations of the generators (${\widetilde{\alpha}}, {\widetilde{\beta}},
{\widetilde{\gamma}}$) $$A = \frac{1}{\sqrt{2}} \ ({\widetilde{\gamma}} - {\widetilde{\alpha}}) \ \ \ , \ \ \
E = \frac{1}{\sqrt{6}} \ ({\widetilde{\alpha}} - 2 {\widetilde{\beta}} + {\widetilde{\gamma}}) \ \ \ , \ \ \
T = \frac{1}{\sqrt{3}} \ ({\widetilde{\alpha}} + {\widetilde{\beta}} +
{\widetilde{\gamma}}) \ ,
\label{eq:22}$$ where $A$, $E$, and $T$ are italicized to indicate that these are “orthogonal modes”. Although the expressions for the modes $A$ and $E$ depend on our particular choice for the two eigenvectors (${\bf
v_1}, {\bf v_2}$), it is clear from our earlier considerations that the value of the optimal signal-to-noise ratio is unaffected by such a choice. From equation (\[eq:22\]) it is also easy to verify that the noise correlation matrix of these three combinations is diagonal, and that its non-zero elements are indeed equal to the eigenvalues given in equation (\[eq:20\]).
In order to calculate the sensitivity corresponding to the expression of the optimal signal-to-noise ratio, we have proceeded similarly to what was done in [@AET99], [@ETA00], and described in more detail in [@WP02]. We assume an equal-arm LISA ($L = 16.67$ light seconds), and take the one-sided spectra of proof mass and aggregate optical-path-noises (on a single link), expressed as fractional frequency fluctuation spectra, to be ([@ETA00], [@LISA98]), $S^{proof\ mass}_y = 2.5 \times {10^{-48}} \ {\left[f/{1 Hz}
\right]}^{-2} \ {\rm Hz}^{-1}$ and $S^{optical \ path}_y = 1.8
\times {10^{-37}} \ {\left[f/1 Hz \right]}^2 \ {\rm Hz}^{-1}$, respectively. We also assume that aggregate optical path noise has the same transfer function as shot noise.
The optimum SNR is the square root of the sum of the squares of the SNRs of the three “orthogonal modes” ($A, E, T$). To compare with previous sensitivity curves of a single LISA Michelson interferometer, we construct the SNRs as a function of Fourier frequency for sinusoidal waves from sources uniformly distributed on the celestial sphere. To produce the SNR of each of the ($A, E, T$) modes we need the gravitational wave response and the noise response as a function of Fourier frequency. We build up the gravitational wave responses of the three modes ($A, E, T$) from the gravitational wave responses of ($\alpha, \beta, \gamma$). For $7000$ Fourier frequencies in the ${\sim}10^{-4}$ Hz to ${\sim}1$ Hz LISA band, we produce the Fourier transforms of the gravitational wave response of ($\alpha, \beta,
\gamma$) from the formulas in [@AET99], [@WP02]. The averaging over source directions (uniformly distributed on the celestial sphere) and polarization states (uniformly distributed on the Poincaré sphere) is performed via a Monte Carlo method. From the Fourier transforms of the ($\alpha, \beta, \gamma$) responses at each frequency, we construct the Fourier transforms of ($A, E, T$). We then square and average to compute the mean-squared responses of ($A, E, T$) at that frequency from $10^4$ realizations of (source position, polarization state) pairs.
We adopt the following terminology: we refer to a single element of the module as a data [*combination*]{}; while a function of the elements of the module, such as taking the maximum over several data combinations in the module or squaring and adding data combinations belonging to the module, is called as an [*observable*]{}. The important point to note is that the laser frequency noise is also suppressed for the observable although it may not be an element of the module.
The noise spectra of ($A, E, T$) are determined from the raw spectra of proof-mass and optical-path noises, and the transfer functions of these noises to ($A, E, T$). Using the transfer functions given in [@ETA00], the resulting spectra are equal to, $$\begin{aligned}
S_{A}(f) = S_{E}(f) & = & \ 16 \ \sin^2(\pi f L) \
[3 + 2 \cos(2 \pi f L) + \cos(4 \pi f L)] S^{proof\ mass}_y(f)
\nonumber \\
& + & \ 8 \ \sin^2(\pi f L) \ [2 + \cos(2 \pi f L)] \ S^{optical \
path}_y(f) \ ,\end{aligned}$$
$$\begin{aligned}
S_{T}(f) & = & \ 2 [1 + 2 \cos (2 \pi f L)]^2 \
[4 \ \sin^2 (\pi f L) S_y^{proof\ mass} + S_y^{optical\ path} (f)] \ .\end{aligned}$$
Let the amplitude of the sinusoidal gravitational wave be $h$. The SNR for, e.g. $A$, ${\rm SNR}_{A}$, at each frequency $f$ is equal to $h$ times the ratio of the root-mean-squared gravitational wave response at that frequency divided by $\sqrt{S_{A}(f) \ B}$, where $B$ is the bandwidth conventionally taken to be equal to $1$ cycle per year. Finally, if we take the reciprocal of ${\rm SNR}_{A}/h$ and multiply it by $5$ to get the conventional ${\rm SNR} = 5$ sensitivity criterion, we obtain the sensitivity curve for this combination which can then be compared against the corresponding sensitivity curve for the equal-arm Michelson Interferometer.
In Figure \[SENS\] we show the sensitivity curve for the LISA equal-arm Michelson response (${\rm SNR} = 5$) as a function of the Fourier frequency, and the sensitivity curve from the optimum weighting of the data described above: $5 h / \sqrt{{\rm SNR}_{A}^2 + {\rm SNR}_{E}^2 + {\rm SNR}_{T}^2}$. The SNRs were computed for a bandwidth of 1 cycle/year. Note that at frequencies where the LISA Michelson combination has best sensitivity, the improvement in signal-to-noise ratio provided by the optimal observable is slightly larger than $\sqrt{2}$.
In Figure \[SNRO\] we plot the ratio between the optimal SNR and the SNR of a single Michelson interferometer. In the long-wavelength limit, the SNR improvement is $\sqrt{2}$. For Fourier frequencies greater than or about equal to $1/L$, the SNR improvement is larger and varies with the frequency, showing an average value of about $\sqrt{3}$. In particular, for bands of frequencies centered on integer multiples of $1/L$, $SNR_{T}$ contributes strongly and the aggregate SNR in these bands can be greater than $2$.
In order to better understand the contribution from the three different combinations to the optimal combination of the three generators, in Figure \[OSNR\] we plot the signal-to-noise ratios of ($A, E,
T$) as well as the optimal signal-to-noise ratio. For an assumed $h =
10^{-23}$, the SNRs of the three modes are plotted versus frequency. For the equal-arm case computed here, the SNRs of $A$ and $E$ are equal across the band. In the long wavelength region of the band, modes $A$ and $E$ have SNRs much greater than mode $T$, where its contribution to the total SNR is negligible. At higher frequencies, however, the $T$ combination has SNR greater than or comparable to the other modes and can dominate the SNR improvement at selected frequencies. Some of these results have also been obtained in [@NPDV03_1].
Optimization of SNR for binaries with known direction but with unknown orientation of the orbital plane
-------------------------------------------------------------------------------------------------------
Binaries will be important sources for LISA and therefore the analysis of such sources is of major importance. One such class is of massive or super massive binaries whose individual masses could range from $10^3 M_{\odot}$ to $10^8 ~ M_{\odot}$ and which could be upto a few Gpc away. Another class of interest are known binaries within our own galaxy whose individual masses are of the order of a solar mass but are just at a distance of a few kpc or less. Here the focus will be on this latter class of binaries. It is assumed that the direction of the source is known, which is so for known binaries in our galaxy. However, even for such binaries, the inclination angle of the plane of the orbit of the binary is either poorly estimated or unknown. The optimization problem is now posed differently: the SNR is optimized [*after*]{} averaging over the polarizations of the binary signals, so the results obtained are optimal on the average, that is, the source is tracked with an observable which is optimal on the average [@NPDV03_1]. For computing the average, a uniform distribution for the direction of the orbital angular momentum of the binary is assumed.
When the binary masses are of the order of a solar mass and the signal typically has a frequency of a few mHz, the GW frequency of the binary may be taken to be constant over the period of observation, which is typically taken to be of the order of an year. A complete calculation of the signal matrix and the optimization procedure of SNR, is given in [@NDPV03_2]. Here we briefly mention the main points and the final results.
A source fixed in the Solar System Barycentric reference frame in the direction $(\theta_B, \phi_B)$ is considered. But as the LISA constellation moves along its heliocentric orbit, the apparent direction $(\theta_L, \phi_L)$ of the source in the LISA reference frame $(x_L, y_L, z_L)$ changes with time. The LISA reference frame $(x_L, y_L, z_L)$ has been defined in reference [@NDPV03_2] as follows: the origin lies at the center of the LISA triangle and the plane of LISA coincides with the $(x_L, y_L)$ plane with spacecraft 2 lying on the $x_L$ axis. Fig. (\[trk\]) displays this apparent motion for a source lying in the ecliptic plane, that is with $\theta_B = 90^{\circ}$ and $\phi_B = 0^{\circ}$. The source in the LISA reference frame describes a figure of 8. Optimizing the SNR amounts to tracking the source with an optimal observable as the source apparently moves in the LISA reference frame .
 \[trk\]
Since an average has been taken over the orientation of the orbital plane of the binary or equivalently over the polarizations, the signal matrix ${\bf A}$ is now of rank 2 instead of rank 1 as compared with the application in the previous subsection. The mutually orthogonal data combinations $A, E, T$ are convenient in carrying out the computations because in this case as well, they simultaneously diagonalize the signal and the noise covariance matrix. The optimization problem now reduces to an eigenvalue problem with the eigenvalues being the squares of the SNRs. There are two eigen-vectors which are labelled as $\vec{v}_{+, \times}$ belonging to two non-zero eigenvalues. The two SNRs are labelled as, ${\rm SNR}_+$ and ${\rm
SNR}_{\times}$ corresponding to the two orthogonal (thus statistically independent) eigen-vectors $\vec{v}_{+, \times}$. As was done in the previous subsection the two SNRs can be squared and added to yield a network SNR, which is defined through the equation:
\_[network]{}\^2 = [SNR]{}\_+\^2 + [SNR]{}\_\^2 . The corresponding observable is called as the network observable. The third eigenvalue is zero and the corresponding eigenvector orthogonal to $\vec {v}_{+}$ and $\vec{v}_{\times}$ gives zero signal.
The eigenvectors and the SNRs are functions of the apparent source direction parameters $(\theta_L, \phi_L)$ in the LISA reference frame, which in turn are functions of time. The eigenvectors optimally track the source as it moves in the LISA reference frame. Assuming an observation period of an year, the SNRs are integrated over this period of time. The sensitivities are computed according to the procedure described in the previous subsection \[GA\]. The results of these findings are displayed in Fig.\[fig:sens\].
![Sensitivity curves for the observables: Michelson, $\max[X, Y, Z]$, $\vec{v}_+$ and network for the source direction ($\theta_B = 90^{\circ}, \phi_B = 0^{\circ}$).](Fig10.eps "fig:") \[fig:sens\]
It shows the sensitivity curves of the following observables:
1. The Michelson combination $X$ (faint solid curve).
2. The observable obtained by taking the maximum sensitivity among $X, Y$ and $Z$ for each direction, where $Y$ and $Z$ are the Michelson observables corresponding to the remaining two pairs of arms of LISA [@AET99]. This maximum is denoted by $\max [X, Y,
Z]$ (dash-dotted curve) and is operationally given by switching the combinations $X, Y, Z$ so that the best sensitivity is achieved.
3. The eigen-combination $\vec{v}_+$ which has the best sensitivity among all data combinations (dashed curve).
4. The network observable (solid curve).
It is observed that the sensitivity over the band-width of LISA increases as one goes from (1) to (4). Also it is seen that the $\max [X, Y, Z]$ does not do much better than $X$. This is because for the source direction chosen $\theta_B = 90^{\circ}$, $X$ is reasonably well oriented and switching to $Y$ and $Z$ combinations does not improve the sensitivity significantly. However, the network and $\vec{v}_+$ observables show significant improvement in sensitivity over both $X$ and $\max [X, Y,
Z]$. This is the typical behavior and the sensitivity curves (except $X$) do not show much variations for other source directions and the plots are similar. Also it may be fair to compare the optimal sensitivities with $\max [X, Y, Z]$ rather than $X$. This comparison of sensitivities is shown in Fig.\[fig:SNRratio\], where the network and the eigen-combinations $\vec{v}_{+, \times}$ are compared with $\max [X, Y, Z]$.
![Ratios of the sensitivities of the observables network, $\vec{v}_{+, \times}$ with $\max [X, Y, Z]$ for the source direction $\theta_B = 90^{\circ}, \phi_B = 0^{\circ}$.](Fig11.eps "fig:"){width="50.00000%"} \[fig:SNRratio\]
Defining: $$\kappa_a (f) = \frac{{\rm SNR}_a(f)}{{\rm SNR}_{\max [X, Y, Z]}(f)} \, ,$$ where the subscript $a$ stands for network or $+, \times$ and ${\rm
SNR}_{\max [X, Y, Z]}$ is the SNR of the observable $\max [X, Y,
Z]$, the ratios of sensitivities are plotted over the LISA band-width. The improvement in sensitivity for the network observable is about 34$\%$ at low frequencies and rises to nearly 90$\%$ at about 20 mHz, while at the same time the $\vec{v}_+$ combination shows improvement of 12$\%$ at low frequencies rising to over 50$\%$ at about 20 mHz.
Concluding Remarks {#SECVII}
==================
In this article we have summarized the use of TDI for canceling the laser phase noise from heterodyne phase measurements performed by a constellation of three spacecraft tracking each other along arms of unequal length. Underlying the TDI technique is the mathematical structure of the theory of Gröbner basis and the algebra of modules over polynomial rings. These methods have been motivated and illustrated with the simple example of an unequal arm interferometer in order to give a physical insight of TDI. Here, these methods have been rigorously applied to the idealized case of a stationary LISA for deriving the generators of the module from which the entire TDI data set can be obtained; they can be extended in a straightforward way to more than three spacecraft for possible LISA follow-on missions. The stationary LISA case was used as a propaedeutical introduction to the physical motivation of TDI, and for further extending it to the realistic LISA configuration of free-falling spacecraft orbiting around the Sun. The TDI data combinations canceling laser phase noise in this general case are referred to as [*second generation TDI*]{}, and they contain twice as many terms as their corresponding [*first generation*]{} combinations valid for the stationary configuration.
As a data analysis application we have shown that it is possible to identify specific TDI combinations that will allow LISA to achieve optimal sensitivity to gravitational radiation [@PTLA02; @NPDV03_1; @NDPV03_2]. The resulting improvement in sensitivity over that of an unequal-arm Michelson Interferometer, in the case of monochromatic signals randomly distributed over the celestial sphere and of random polarization, is non negligible. We have found this to be equal to a factor of $\sqrt{2}$ in the low-part of the frequency band, and slightly more than $\sqrt{3}$ in the high-part of the LISA band. The SNR for binaries whose location in the sky is known but their polarization is not can also be optimized, and the degree of improvement depends on the location of the source in the sky.
As a final remark we would like to emphasize that this field of research, TDI, is still very young and evolving. Possible physical phenomena, yet unrecognized, might turn out to be important to account for within the TDI framework. The purpose of this review was to provide the basic mathematical tools needed for working on future TDI projects. We hope to have accomplished this goal, and that others will be stimulated to work in this new and fascinating field of research.
Acknowledgement {#acknowledgement .unnumbered}
===============
SVD acknowledges support from IFCPAR, Delhi, India under which the work was carried out in collaboration with J-Y Vinet. This research was performed at the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration.
Generators of the Module of Syzygies {#syz}
====================================
We require the 4-tuple solutions $(q_3, q_1', q_2', q_3')$ to the equation: (1 - xyz) q\_3 + (xz - y) q\_1’ + x(1 - z\^2) q\_2’ + (1 - x\^2) q\_3’ = 0 , \[cnstr1\] where for convenience we have substituted $x = \D_1, y = \D_2, z =
\D_3$. $q_3, q_1', q_2', q_3'$ are polynomials in $x, y, z$ with integral coefficients i.e. in Z\[x,y,z\].
We now follow the procedure in the book by Becker et al. [@becker].
Consider the ideal in $Z[x,y,z]$ (or ${\cal Q}[x,y,z]$ where ${\cal
Q}$ denotes the field of rational numbers), formed by taking linear combinations of the coefficients in Eq.(\[cnstr1\]) $f_1 = 1 - xyz,
f_2 = xz - y, f_3 = x(1 - z^2), f_4 = 1 - x^2$. A Gröbner basis for this ideal is: = {g\_1 = z\^2 - 1, g\_2 = y\^2 - 1, g\_3 = x - yz } . The above Gröbner basis is obtained using the function GroebnerBasis in Mathematica. One can check that both the $f_i, i = 1,2,3,4$ and $g_j, j =1,2,3$ generate the same ideal because we can express one generating set in terms of the other and vice-versa: f\_i = d\_[ij]{} g\_j , g\_j = c\_[ji]{} f\_i , where $d$ and $c$ are $4 \times 3$ and $3 \times 4$ polynomial matrices respectively, and are given by, d = (
[ccc]{} -1 & -z\^2 & -yz\
y & 0 & z\
-x & 0 & 0\
-1 & - z\^2 & -(x + yz)
) , c = (
[cccc]{} 0 & 0 & -x & z\^2 - 1\
-1 & -y & 0 & 0\
0 & z & 1 & 0
) . \[cd\] The generators of the 4-tuple module are given by the set $A \,
\bigcup B^*$ where $A$ and $B^*$ are the sets described below:
$A$ is the set of row vectors of the matrix $I - d \cdot c $ where the dot denotes the matrix product and I is the identity matrix, $4 \times
4$ in our case. Thus, a\_1 &=& (z\^2 - 1, 0, x - yz, 1 - z\^2) ,\
a\_2 &=& (0, z(1 - z\^2), xy - z, y(1 - z\^2)) ,\
a\_3 &=& (0, 0, 1 - x\^2, x(z\^2 - 1)) ,\
a\_4 &=& (-z\^2, xz, yz, z\^2) . We thus first get 4 generators. The additional generators are obtained by computing the S-polynomials of the Gröbner basis ${\cal
G}$. The S-polynomial of two polynomials $g_1, g_2$ is obtained by multiplying $g_1$ and $g_2$ by suitable terms and then adding, so that the highest terms cancel. For example in our case $g_1 = z^2 - 1$ and $g_2 = y^2 - 1$ and the highest terms are $z^2$ for $g_1$ and $y^2$ for $g_2$ . Multiply $g_1$ by $y^2$ and $g_2$ by $z^2$ and subtract. Thus, the S-polynomial $p_{12}$ of $g_1$ and $g_2$ is: p\_[12]{} = y\^2 g\_1 - z\^2 g\_2 = z\^2 - y\^2 . Note that order is defined ($x
>> y >> z$) and the $y^2 z^2$ term cancels. For the Gröbner basis of 3 elements we get 3 S-polynomials $p_{12}, p_{13}, p_{23}$. The $p_{ij}$ must now be re-expressed in terms of the Gröbner basis ${\cal
G}$. This gives a $3 \times 3$ matrix $b$. The final step is to transform to 4-tuples by multiplying $b$ by the matrix $c$ to obtain $b^* = b \cdot c$. The row vectors $b^*_i , i = 1, 2, 3 $ of $b^*$ form the set $B^*$: b\^\*\_1 &=& (z\^2 - 1, y(z\^2 - 1), x(1 - y\^2), (y\^2 -1)(z\^2 - 1)) ,\
b\^\*\_2 &=& (0, z(1 - z\^2), 1 - z\^2 - x(x - yz), (x - yz)(z\^2 - 1)) ,\
b\^\*\_3 &=& (-x + yz, z - xy, 1 - y\^2, 0) . Thus we obtain 3 more generators which gives us a total of 7 generators of the required module of syzygies.
Conversion between generating sets {#mat}
==================================
We list the three sets of generators and relations among them. We first list below $\a, \b, \g, \z$: &=& (-1, -z, -xz, 1, xy, y) ,\
&=& (-xy, -1, -x, z, 1, yz) ,\
&=& (-y, -yz, -1, xz, x, 1) ,\
&=& (-x, -y, -z, x, y, z) .
We now express the $a_i$ and $b^*_j$ in terms of $\a, \b, \g, \z$: a\_1 &=& - z ,\
a\_2 &=& - z ,\
a\_3 &=& - z + - x + xz ,\
a\_4 &=& z\
b\^\*\_1 &=& - y + yz + - z ,\
b\^\*\_2 &=& (1 - z\^2) - x + xz ,\
b\^\*\_3 &=& - y . Further we also list below $\a, \b, \g, \z$ in terms of $X^{(A)}$: &=& X\^[(3)]{} ,\
&=& X\^[(4)]{} ,\
&=& - X\^[(1)]{} + z X\^[(2)]{} ,\
&=& X\^[(2)]{} . This proves that since the $a_i, b^*_j$ generate the required module, the $\a, \b, \g, \z$ and $X^{(A)}, A = 1, 2, 3, 4$ also generate the same module.
The Gröbner basis is given in terms of the above generators as follows: $G^{(1)} = \z, G^{(2)}= X^{(1)}, G^{(3)} = \b, G^{(4)} = \a$ and $G^{(5)} = a_3$
P.L. Bender and D. Hils, *CQG*, **14**, 1439(1997).
G. Nelemans, L.R. Yungelson and S.F. Portegies Zwart, *A & A*, **375**, 890(2001).
M. Tinto, and F.B. Estabrook, [*Phys. Rev. D*]{}, [**52**]{}, 1749, (1995).
M. Tinto, [*Class. Quantum Grav.*]{}, [**19**]{}, 7, 1767 (2002).
F.B. Estabrook, and H.D. Wahlquist, [*Gen. Relativ. Gravit.*]{} [**6**]{}, 439 (1975).
M. Tinto, [*Class. Quantum Grav.*]{}, [**19**]{}, 1767 (2002)
LISA: (Laser Interferometer Space Antenna) [*An international project in the field of Fundamental Physics in Space*]{}, Pre-Phase A Report, [**MPQ 233**]{}, (Max-Planck-Institute für Quantenoptic, Garching bei München, 1998).
M. Tinto, and J.W. Armstrong [*Phys. Rev. D*]{}, [ **59**]{}, 102003 (1999).
M. Tinto, F.B. Estabrook, and J.W. Armstrong, [ *Phys. Rev. D*]{}, [**65**]{}, 082003 (2002).
M. Tinto, [*Phys. Rev. D*]{}, [**58**]{}, 102001, (1998).
J.W. Armstrong, and F.B. Estabrook, & M. Tinto, [ *ApJ*]{}, [**527**]{}, 814 (1999).
S.V. Dhurandhar, K. R. Nayak, and J.-Y. Vinet [ *Phys. Rev. D*]{}, [**65**]{}, 102002 (2002)
F.B. Estabrook, M. Tinto, and J.W. Armstrong, [ *Phys. Rev. D*]{}, [**62**]{}, 042002 (2000).
D.A. Shaddock, [*Phys. Rev. D*]{}, [**69** ]{}, 022001 (2004).
N.J. Cornish & R.W. Hellings, [*Class. Quantum Grav.*]{}, [**20**]{}, 4851, (2003).
D.A. Shaddock, M. Tinto, F.B. Estabrook, and J.W. Armstrong, [*Phys. Rev. D*]{}, [**68**]{}, 061303 (2003).
M. Tinto, F.B. Estabrook, and J.W. Armstrong, [ *Phys. Rev. D*]{}, [**69**]{}, 082001, (2004).
T. A. Prince, M. Tinto, S. L. Larson, and J.W. Armstrong, [*Phys.Rev. D*]{} [**66**]{}, 122002 (2002).
K. Rajesh Nayak, A. Pai, S. V. Dhurandhar, and J-Y. Vinet, *Class. Quantum Grav*, **20,** 1217(2003).
K. Rajesh Nayak, S. V. Dhurandhar, A. Pai, and J-Y. Vinet, [*Phys. Rev. D*]{}, [**68**]{}, 122001 (2003).
J.E. Faller and P.L. Bender, in: [*Precision Measurement and Fundamental Constants II*]{}, eds. B.N. Taylor and W.D. Phillips, NBS Spec. Pub. 617, (1984).
J.E. Faller, P.L. Bender, J.L. Hall, D. Hils and M.A. Vincent, in: [*Proceedings Colloquium Kilometric Optical Arrays in Space*]{}, ESA SP-126, Noordwijk, The Netherlands (1985).
J.E. Faller, PL. Bender, J.L. Hall, D. Hils, R.T. Stebbins and M.A. Vincent, in: [*Advances in Space Research*]{}, Proc. XXVII COSPAR, Symp. 15 on Relativistic Gravitation, Vol. 9 (1989).
G. Giampieri, R.W. Hellings, M. Tinto, and J.E. Faller, [*Optics Communications*]{}, [**123**]{}, 669 (1996).
G. M. Jenkins and D. G. Watts, [*Spectral Analysis and Its Applications*]{}, (Holden-Day, San Francisco), (1968).
S. Wolfram, [*Mathematica: User manual*]{}, Wolfram Research, Inc., (2002).
T. Becker, [*“Gröbner Bases: A computational approach to commutative algebra”*]{}, Volker W
M. Kreuzer and L. Robbiano, [*“Computational commutative algebra 1”*]{}, Springer Verlag(2000).
D. Summers, “Algorithm tradeoffs,” oral presentation, 3rd progress meeting of the ESA funded LISA PMS Project. ESTEC, NL, February 2003.
W.M. Folkner, F. Hechler, T.H. Sweetser, M.A. Vincent, and P.L. Bender, [*Clas. Quantum Grav.*]{}, [**14**]{}, 1543, (1997).
M. Tinto, J.W. Armstrong & F.B. Estabrook, [*Phys. Rev. D*]{}, [**63**]{}, 021101 (2001)
R. Nayak and J-Y Vinet, (under preparation), (2004).
B. Noble, [*Applied Linear Algebra*]{}, Prentice/Hall International, p. 378, (1969).
S. Selby, [*Standard of Mathematical Tables*]{}, The Chemical Bubber Co., p. 131, (1964).
L.S. Finn, [*Phys. Rev. D*]{}, [**63**]{}, 102001, (2001).
M. Tinto, F.B. Estabrook, & J.W. Armstrong, [*LISA Pre-Project Publication*]{}, $(http://www.srl.caltech.edu/lisa/tdi\_wp/LISA\_Whitepaper.pdf)$, (2002).
[^1]: A module is an abelian group over a [*ring*]{} as contrasted with a vector space which is an abelian group over a field. The scalars form a ring and just like in a vector space, scalar multiplication is defined. However, in a ring the multiplicative inverses do not exist in general for the elements, which makes all the difference!
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Cataclysmic cosmic events can be plausible sources of both gravitational waves (GW) and high-energy neutrinos (HEN). Both GW and HEN are alternative cosmic messengers that may escape very dense media and travel unaffected over cosmological distances, carrying information from the innermost regions of the astrophysical engines. For the same reasons, such messengers could also reveal new, hidden sources that were not observed by conventional photon astronomy.
Requiring the consistency between GW and HEN detection channels shall enable new searches as one has significant additional information about the common source. A neutrino telescope such as ANTARES can determine accurately the time and direction of high energy neutrino events, while a network of gravitational wave detectors such as LIGO and VIRGO can also provide timing/directional information for gravitational wave bursts. By combining the information from these totally independent detectors, one can search for cosmic events that may arrive from common astrophysical sources.
author:
- '\'
title: 'Searching for high-energy neutrinos in coincidence with gravitational waves with the ANTARES and VIRGO/LIGO detectors'
---
neutrinos; gravitational waves; multi-messenger astronomy.
Introduction
============
Astroparticle physics has entered an exciting period with the recent development of experimental techniques that have opened new windows of observation of the cosmic radiation in all its components. In this context, it has been recognized that not only a multi-wavelength but also a multi-messenger approach was best suited for studying the high-energy astrophysical sources.
In this context, and despite their elusive nature, both high-energy neutrinos (HENs) and gravitational waves (GWs) are now considered seriously as candidate cosmic messengers. Contrarily to high-energy photons (which are absorbed through interactions in the source and with the extragalactic background light) and cosmic rays (which are deflected by ambient magnetic fields, except at the highest energies), both HENs and GWs may indeed escape from dense astrophysical regions and travel over large distances without being absorbed, pointing back to their emitter.
It is expected that many astrophysical sources produce both GWs, originating from the cataclysmic event responsible for the onset of the source, and HENs, as a byproduct of the interactions of accelerated protons (and heavier nuclei) with ambient matter and radiation in the source. Moreover, some classes of astrophysical objects might be completely opaque to hadrons and photons, and observable only through their GW and/or HEN emissions. The detection of coincident signals in both these channels would then be a landmark event and sign the first observational evidence that GW and HEN originate from a common source. GW and HEN astronomies will then provide us with important information on the processes at work in the astrophysical accelerators. A more detailed discussion on the most plausible GW/HEN sources will be presented in Section \[sec:sources\], along with relevant references.
Furthermore, provided that the emission mechanisms are sufficiently well known, a precise measurement of the time delay between HEN and GW signals coming from a very distant source could also allow to probe quantum-gravity effects and possibly to constrain dark energy models[@QG]. Gamma-ray bursts (GRBs) appear as good candidates, [*albeit*]{} quite challenging[@QGlim], for such time-of-flight studies.
Common HEN and GW astronomies are also motivated by the advent of a new generation of dedicated experiments. In this contribution, we describe in more detail the feasibility of joint GW+HEN searches between the ANTARES neutrino telescope[@antares] (and its future, $km^3$-sized, successor KM3NeT[@km3net]) and the GW detectors VIRGO[@virgo] and LIGO[@ligo] (which are now part of the same experimental collaboration). The detection principle and performances of these detectors are briefly described in Section \[sec:det\], along with their respective schedule of operation in the forthcoming years. Section \[sec:ana\] presents an outlook on the analysis strategies that will be set up to optimize coincident GW-HEN detection among the three experiments.
It should be noted that similar studies involving the Ice Cube neutrino telescope[@ice3], currently under deployment at the South Pole and looking for cosmic neutrinos from sources in the Northern Hemisphere, are under way[@aso].
[{width="5in"} \[sub\_fig1\]]{}
Potential common sources of Gravitational waves and high-energy neutrinos {#sec:sources}
=========================================================================
Galactic sources
----------------
Several classes of galactic sources could feature both GW and HEN emission mechanisms potentially accessible to the present generation of GW interferometers and HEN telescopes, among which :
- [**Microquasars**]{} are associated with X-ray binaries involving a compact object (neutron star or black hole) that accretes matter from a companion star and re-emits it in relativistic jets associated with intense radio (and IR) flares. Such objects could emit GWs during both accretion and ejection phases[@pradierVLVNT]; and the latter phase could be correlated with a HEN signal if the jet has a hadronic component[@mq].
- [**Soft Gamma-ray Repeaters (SGRs)**]{} are X-ray pulsars with a soft $\gamma$-ray bursting activity. The [*magnetar*]{} model described them as highly magnetized neutron stars whose outbursts are caused by global star-quakes associated with rearragements of the magnetic field[@magnetar]. The deformation of the star during the outburst could produce GWs, while HENs could emerge from hadron-loaded flares[@ioka]. The fact that both the AMANDA neutrino telescope and the LIGO interferometer have issued relevant limits on the respective GW and HEN emissions from the 2004 giant flare of SGR 1806-20[@sgrlim] indicates that a coincident GW+HEN search might further constrain the parameter space of emission models.
Extragalactic sources
---------------------
Gamma-Ray Bursts (GRBs) are probably the most promising class of extragalactic sources. In the prompt and afterglow phases, HENs ($10^5 - 10^{10}$ GeV) are expected to be produced by accelerated protons in relativistic shocks and several models predict detectable fluxes in km$^3$-scale detectors[@GRBnus].
- [**Short-hard GRBs**]{} are thought to originate from coalescing binaries involving black holes and/or neutron stars; such mergers could emit GWs detectable from relatively large distances[@nakar], with significant associated HEN fluxes.
- [**Long-soft GRBs**]{}, as described by the collapsar model, are compatible with the emission of a strong burst of GWs during the gravitational collapse of the (rapidly rotating) progenitor star and in the pre-GRB phase; however this population is distributed over cosmological distances so that the associated HEN signal is expected to be faint[@kotake].
- [**Low-luminosity GRBs**]{}, with $\gamma$-ray luminosities a few orders of magnitude smaller, are believed to originate from a particularly energetic, possibly rapidly-rotating and jet-driven population of core-collapse supernovae. They could produce stronger GW signals together with significant high- and low-energy neutrino emission; moreover they are often discovered at shorter distances[@gupta].
- [**Failed GRBs**]{} are thought to be associated with supernovae driven by mildly relativistic, baryon-rich and optically thick jets, so that no $\gamma$-rays escape. Such “hidden sources” could be among the most promising emitters of GWs and HENs, as current estimations predict a relatively high occurrence rate in the volume probed by current GW and HEN detectors[@ando].
The detectors {#sec:det}
=============
ANTARES
-------
The ANTARES detector is the first undersea neutrino telescope; its deployment at a depth of 2475m in the Mediterranean Sea near Toulon was completed in May 2008. It consists in a three-dimensional array of 884 photomultiplier tubes (PMTs) distributed on 12 lines anchored to the sea bed and connected to the shore through an electro-optical cable. Before reaching this final (12L) setup, ANTARES has been operating in various configurations with increasing number of lines, from one to five (5L) and ten (10L); the respective periods are indicated on the time chart of Fig. \[fig:tab\].
ANTARES detects the Cherenkov radiation emitted by charged leptons (mainly muons, but also electrons and taus) induced by cosmic neutrino interactions with matter inside or near the instrumented volume. The knowledge of the timing and amplitude of the light pulses recorded by the PMTs allows to reconstruct the trajectory of the muon and to infer the arrival direction of the incident neutrino and to estimate its energy.
Since the Earth acts as a shield against all particles but neutrinos, the design of a neutrino telescope is optimized for the detection of up-going muons produced by neutrinos which have traversed the Earth and interacted near the detector. The field of view of ANTARES is therefore $\sim\, 2 \pi\, \mathrm{sr}$ for neutrino energies $100\ \mathrm{GeV} \lesssim E_\nu \lesssim 100\ \mathrm{TeV}$. Above this energy, the sky coverage is reduced because of neutrino absorption in the Earth; but it can be partially recovered by looking for horizontal and downward-going neutrinos, which can be more easily identified as the background of atmospheric muons is much fainter at these energies. With an effective area $\sim 0,1\ \mathrm{km^2}$), ANTARES is expected to achieve an unprecedented angular resolution (about $0.3^\circ$ for neutrinos above 10 TeV) as a result of the good optical properties of sea water[@opt].
The data acquisition system of ANTARES is based on the “all-data-to-shore” concept, which allows to operate different physics triggers to the same data in parallel, each optimized for a specific (astro)physics signal[@miekedata]. In particular, satellites looking for GRBs can trigger the detector in real time via the GCN (Gamma-Ray Burst Coordinate Network) alert system. About 60s of buffered raw data (sometimes even including data recorded before the alert time) are then written on disk and kept for offline analysis, allowing for a significant gain in detection efficiency[@miekegrb].
Another interesting feature recently implemented in ANTARES is the possibility to trigger an optical telescope network on the basis of “golden” neutrino events (such as neutrino doublets coincident in time and space or single neutrinos of very high energy) selected by a fast, online reconstruction procedure[@damien].
All these characteristics make the ANTARES detector especially suited for the search of astrophysical point sources[@simona] - and transients in particular.
VIRGO and LIGO
--------------
The GW detectors VIRGO[@virgo] (one site in Italy) and LIGO[@ligo] (two sites in the United States) are Michelson-type laser interferometers that consist of two light storage arms enclosed in vacuum tubes oriented at $90^\circ$ from each other. Suspended, highly reflective mirrors play the role of test masses. Current detectors are sensitive to relative displacements (hence GW amplitude) of the order of $10^{20}$ to $10^{22}$ Hz$^{-1/2}$. Their current detection horizon is about 15 Mpc for standard binary sources.
Their sensitivity is essentially limited at high frequencies by laser shot noise and at low frequencies by seismic noise ($<\ \sim O(50\ \mathrm{Hz})$) and by the thermal noise of the atoms in the mirrors (up to few 10 Hz). To reduce the sources of noise, GW interferometers rely on high-power, ultrastable lasers and on sophisticated techniques of position and alignment control and of stabilization of the mirrors.
Both detectors had a data-taking phase during 2007 (Virgo Science Run 1 and Ligo S5), which partially coincided with the ANTARES 5L configuration. They are currently upgrading to VIRGO+ and eLIGO, to improve their sensitivity by a factor of 2 - and hence the probed volume by a factor of 8. The collaborations have merged and are preparing for a common science run starting mid-2009 (Virgo Science Run 2 and LIGO S6), i.e. in coincidence with the operation of ANTARES 12L (see Fig. \[fig:tab\]). Another major upgrade for both classes of detectors is scheduled for the upcoming decade: the Advanced VIRGO/Advanced LIGO and KM3NeT projects should gain a factor of 10 in sensitivities respect to the presently operating instruments.
The VIRGO/LIGO network monitors a good fraction of the sky in common with HEN telescopes: as can be seen from Figure \[fig:map\] the overlap of visibility maps with each telescope is about 4 sr ($\sim 30\%$ of the sky), and the same value holds for Ice Cube.
![Instantaneous common sky coverage for VIRGO + LIGO + ANTARES in geocentric coordinates. This map shows the combined antenna pattern for the gravitational wave detector network (above half-maximum), with the simplifying assumption that ANTARES has 100% visibility in its antipodal hemisphere and 0% elsewhere. The colour scale is from $0\%$ (left, blue) to $100\%$ (right, red)[@ecm].[]{data-label="fig:map"}](VanElewyck_icrc09_fig2){width="3in"}
Outlook on the analysis strategies {#sec:ana}
==================================
GW interferometers and HEN telescopes share the challenge to look for faint and rare signals on top of abundant noise or background events. The GW+HEN search methodology involves the combination of independent GW/HEN candidate event lists, with the advantage of significantly lowering the rate of accidental coincidences.
The information required about any GW/HEN event consists of its timing, arrival direction and associated angular uncertainties (possibly under the form of a sigificance sky map). Each event list is obtained by the combination of reconstruction algorithms specific to each experiment, and quality cuts used to optimize the signal-to-background ratio.
GW+HEN event pairs within a predefined, astrophysically motivated (and possibly source- or model-dependent), time interval can be selected as time-coincident events. Then, the spatial overlap between GW and HEN events is statistically evaluated, e. g. by an unbinned maximum likelihood method, and the significance of the coincident event is obtained by comparing to the distribution of accidental events obtained with Monte-Carlo simulations using data streams scrambled in time (or simulated background events).
Preliminary investigations of the feasibility of such searches have already been performed and indicate that, even if the constituent observatories provide several triggers a day, the false alarm rate for the combined detector network can be maintained at a very low level ($\sim (600\ \mathrm{yr})^{-1}$)[@aso; @pradierVLVNT].
Conclusions and perspectives {#sec:concl}
============================
A joint GW+HEN analysis program could significantly expand the scientific reach of both GW interferometers and HEN telescopes. The robust background rejection arising from the combination of two totally independent sets of data results in an increased sensitivity and the possible recovery of cosmic signals. The observation of coincident triggers would provide strong evidence for the detection of a GW burst and a cosmic neutrino event, and for the existence of common sources. Beyond the benefit of a high-confidence discovery, coincident GW/HEN (non-)observation shall play a critical role in our understanding of the most energetic sources of cosmic radiation and in constraining existing models. They could also reveal new, “hidden” sources unobserved so far by conventional photon astronomy. A new period of concurrent observations with upgraded experiments is expected to start mid-2009. Future schedules involving next-generation detectors with a significantly increased sensitivity (such as KM3NeT and the Advanced LIGO/Advanced VIRGO projects) are likely to coincide as well, opening the way towards an even more efficient GW+HEN astronomy.
Acknowledgements
================
V. V. E. thanks E. Chassande-Mottin for many fruitful discussions and for his help in preparing this manuscript. She acknowledges financial support from the European Community 7th Framework Program (Marie Curie Reintegration Grant) and from the French Agence Nationale de la Recherche (ANR-08-JCJC-0061-01).
[99]{} see e.g. G. Amelino-Camelia, [*Int. J. Mod. Phys.*]{} [**D12**]{} (2003) 1633; S. Choubey and S.F. King, [*Phys. Rev.*]{} [**D67**]{} (2003) 073005; U. Jacob and T. Piran, [*Nature Phys.*]{} [**3**]{} (2007) 87. M. C. Gonzalez-Garcia and F. Halzen, [*JCAP*]{} [**0702**]{} (2007) 008 see e.g. J. Carr \[ANTARES Collaboration\], J. Phys. Conf. Ser. [**136**]{} (2008) 022047; G. Carminati \[ANTARES Collaboration\], arXiv:0905.1373 \[astro-ph.IM\]; P. Coyle \[ANTARES Collaboration\], these Proceedings. http://www.km3net.org/ ; A. Kappes \[KM3NeT Consortium\], arXiv:0711.0563 \[astro-ph\]. http://www.virgo.infn.it/; F. Acernese \[VIRGO Collaboration\], [*Class. Quant. Grav.*]{} [**25**]{} (2008) 184001. http://www.ligo.caltech.edu/; D. Sigg \[LIGO Collaboration\], [*Class. Quant. Grav.*]{} [**25**]{} (2008) 114041. http://icecube.wisc.edu/; A. Karle \[IceCube Collaboration\], arXiv:0812.3981 \[astro-ph\] and these Proceedings. Y. Aso , [*Class. Quant. Grav.* ]{} [**25**]{} (2008) 114039. T. Pradier, [*Nucl. Instrum. Meth. A*]{} [**602**]{} (2009) 268. M. M. Reynoso, G. E. Romero and H. R. Christiansen, [*MNRAS*]{} [**387**]{} (2008) 1745; F. Aharonian , [*J. Phys. Conf. Ser.*]{} [**39**]{} (2006) 408; C. Distefano , [*Astrophys. J.*]{} [**575**]{} (2002) 378. C. Thompson and R. C. Duncan, [*MNRAS*]{} [**275**]{}, 255 (1995). K. Ioka, [*MNRAS*]{} [**327**]{} (2001) 639; K. Ioka , [*Astrophys. J.*]{} [**633**]{} (2005) 1013. A. Achterberg [*et al.*]{} \[IceCube Collaboration\], Phys. Rev. Lett. [**97**]{} (2006) 221101; see e.g. E. Waxman and J. N. Bahcall, Phys. Rev. Lett. [**78**]{} (1997) 2292; J. P. Rachen and P. Meszaros, Phys. Rev. D [**58**]{} (1998) 123005; J. Alvarez-Muniz, F. Halzen and D. W. Hooper, [*Phys. Rev. *]{} [**D62**]{} (2000) 093015. see e.g. E. Nakar, [*Physics Reports*]{} [**442**]{} (2007) 166. see e.g. K. Kotake , [*Repts. Prog. Phys.*]{} [**69**]{} (2006) 971; C. D. Ott, arXiv:0809.0695\[astro-ph\]; M. Vietri, [*Phys. Rev. Lett.*]{} [**80**]{} (2008) 3690. see e.g. K. Murase , [*Astrophys. J.*]{} [**651**]{}(2006):L5; Gupta and Zhang, [*Astropart. Phys.*]{} [**27**]{} (2007) 386-391; X.-Y. Wang , [*Phys. Rev. D*]{} [**76(8)**]{} (2007) 083009; L. Dessart , [*Astrophys. J. Lett.*]{} [**673**]{} (2008) L43. S. Ando and J. Beacom, [*Phys. Rev. Lett.*]{} [**95(6)**]{} (2005) 061103. J. A. Aguilar \[ANTARES Collaboration\], [*Astropart. Phys*]{} [**23**]{} (2005) 131. M. Bouwhuis \[ANTARES Collaboration\], [*Concepts and performance of the Antares data acquisition system*]{}, these Proceedings. M. Bouwhuis \[ANTARES Collaboration\], [*Search for gamma-ray bursts with the Antares neutrino telescope*]{}, these Proceedings. D. Dornic \[ANTARES Collaboration\], these Proceedings. S. Toscano \[ANTARES Collaboration\], these Proceedings. E. Chassande-Mottin, [*talk at the ILIAS Workshop on supernovae, neutrinos and gravitational Waves, Cascina (Italy), November 26th, 2008*]{} (see http://agenda.infn.it/conferenceDisplay.py?confId=811).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We explore here the idea, reminiscent in some respect of Von Weizsäcker’s (1944) and Alfvèn’s (1976) outmoded cosmogonies, that long-lived vortices in a turbulent protoplanetary nebula can capture large amount of solid particles and initiate the formation of planets. Some puzzling features of the solar system appear as natural consequences of our simple model:
- The captured mass presents a maximum near Jupiter’s orbit.
- Outside this optimal orbit, the collected material, mainly composed of low density particles, sinks deeply into the vortices and rapidly collapses into massive bodies at the origin of the solid core of the giant planets.
- Inside this orbit, by contrast, the high density particles are preferentially selected by the vortices and assembled by local gravitational instabilities into planetesimals, massive enough to be released by the vortices and to grow later, in successive collisions, to form the terrestrial planets.
author:
- 'P. Barge'
- 'J. Sommeria'
date: 'Received ??, Accepted ??, 1994 '
title: 'Did planet formation begin inside persistent gaseous vortices?'
---
——————-
Planets are thought to be formed from the dust grains embedded in a gaseous disk, probably like observed around most young low-mass stars: the Protoplanetary Nebula. It is likely that this by-product of the sun formation was also, for a while, a turbulent accretion-disk. During this stage the star completes its accretion and the disk spreads outward with the angular momentum (justifying the repartition of mass and momentum between sun and planets). In the meantime the dust grains are submitted to a turbulent diffusion (due to the gas motions) which speeds up their growth and enables to explain the chemical composition of some meteorites (Morfill, 1983). Due to collisional fragmentation, this turbulent coagulation stalls for centimeter-sized particles in a highly turbulent nebula (Weidenschilling, 1984) and for meter-sized particles in a weakly turbulent nebula (Weidenschilling and Cuzzi, 1993), while gravitational binding becomes effective only in the kilometer range. So, it is commonly thought that the solid material decouples from the gas only after some turbulence decayed, in a two stage process:
\(i) settling of the dust grains toward the mid-plane of the gaseous disk;
\(ii) gravitational collapse of the resulting layer of sediment (when dense enough) into numerous kilometer-sized bodies, the so called “planetesimals”.
Then, as suggested by the cratering of the present planets, gravitationaly bounded bodies grow by the accumulation of planetesimals in successive collisions; this stage of the planet growth is, indeed, reproduced by a number of dynamical models (Safronov 1969; Barge and Pellat 1991, 1993).
However the above scenario faces two major difficulties.
\(1) The solid cores of the giant planets must be formed in less than some $10^6$ years, in order for the gas to be captured before being swept away (Safronov 1969; Strom [*[et al.]{}*]{} 1993) during the sun’s T-Tauri phase; with a reasonable density of solid material, this is difficult to achieve by planetesimal accumulation (Safronov 1969; Wetherill 1988), especially for the outermost planets.
\(2) The formation of the planetesimals themselves is not clearly understood. Indeed the gas, which is supported by a radial pressure gradient, rotates at slightly less than the local Keplerian speed. The resulting velocity difference $\Delta V$ between the sediment layer and the overlying gas induces shear turbulence that prevents the layer from settling to the density required for gravitational instability (Weidenschilling and Cuzzi 1993; Cuzzi [*[et al.]{}*]{} 1993).
Both difficulties are solved by the present model, in which the particles, once settled in the nebula mid-plane, are captured and concentrated into long-lived vortices.
Such vortices may be maintained by specific instability mechanism (Dubrulle 1993), but more generally emerge from random turbulence in rotating shear flows. While three-dimensional eddies are quickly damped by energy cascade toward small scales, two-dimensional turbulence persists without energy dissipation, forming instead larger and larger vortices until a steady solitary vortex is formed. Striking examples are the persistent atmospheric vortices in the giant planets (Ingersoll 1990), like Jupiter’s Great Red Spot. This phenomenon can be reproduced in laboratory experiments (Antipov [*[et al.]{}*]{} 1986; Sommeria [*[et al.]{}*]{} 1988; Nezlin and Snezkhin 1993), and explained in terms of statistical mechanics of two-dimensional turbulence (Sommeria [*[et al.]{}*]{} 1991; Miller [*[et al.]{}*]{} 1992; Michel and Roberts 1994). Observations of accretion-disks around black-holes (Abramowicz [*[et al.]{}*]{} 1992) or T-tauri stars could also indicate the presence of such organized vortices.
In this letter, as we focus on particle trajectories, fluid dynamics will be discussed at an heuristic level, sufficient to justify a simple vortex model. Particle motions are referred to a cartesian frame, in which $x$ and $y$ stand for the azimuthal position and the radial displacement, respectively, rotating around the sun at the Keplerian angular velocity $\Omega=\Omega_0 r^{-3/2}$ ($\Omega_0$ is the Earth’s velocity for $r=1$, in astronomical unit, AU). The $y$ axis is directed outward, and the $x$ axis along the orbital motion, which has then a clockwise (negative) rotation.
Under the standard assumption of hydrostatic balance in the thickness $H$ of the nebula ($H\simeq
C_S/\Omega$, where $C_S$ is the sound speed), the dynamical problem is two-dimensional. Further, neglecting pressure forces, the simplest flow is a set of circular orbits with Keplerian azimuthal velocity ($V_x = - 3\Omega y/2$ and $V_y =0$). The small vortices, possibly rising in this flow with scale $R < H$ and typical vorticity $\Omega$, have a velocity of the order of $\Omega R$ which is less than the sound speed; so, they can be considered as incompressible. Then, vortices spinning like the shear flow are robust and merge one another. (while those with opposite sign are laminated by the shear) (Marcus 1990; Dowling and Ingersoll 1989). The process of vortex growth ends up when the Mach number reaches unity (i.e) $R \simeq H$, beyond which energy losses by sound waves become prohibitive. Finally, the vortex structure should evolve as to minimize the pressure effects responsible for these losses, the streamlines fitting at best with the free-particle trajectories.
An obvious solution is a set of Keplerian ellipses with the same semi-major axis but different eccentricities, corresponding in our rotating frame, to concentric epicycles ($V_x = - 2\Omega y$, $V_y =\Omega
x/2$); this is a steady solution of the fluid equations with uniform pressure. A correspondence between epicyclic motion and vortex flow was used first by Von Weizsäcker (1944); it appears also in the dynamics of non-axisymmetric planetary rings where fluid streamlines can coincide with particle trajectories and describe the ring shapes (Borderies [*[et al.]{}*]{} 1982). We assume a simple matching of this “epicyclic flow” with the azimuthal Keplerian flow at large distances (see Fig.1):
$$\cases{
V_x = - {3\over 2}\Omega y - {1\over 2}\Omega
y~e^{-{{x^2 + y^2}\over {2
R^2}}}
\cr
V_y = {1\over 2}\Omega x ~e^{-{{x^2 + y^2}\over {2
R^2}}} }~~.$$
The characteristic size $R$ (or “radius”) of this vortex is limited to the thickness $H$ of the nebula as discussed above. Its decay time under persistent three-dimensional turbulence can be estimated, using the classical turbulent viscosity for accretion-disks $\nu_t = \alpha
C_S^2/\Omega$ with $\alpha =
10^{-3}$ (following current nebula models $10^{-4}<\alpha<10^{-2}$). The corresponding energy decay time $\tau_D$ is then estimated by dividing the rate of viscous energy dissipation by the vortex kinetic energy. This yields $\Omega
\tau_D = 6.5/\alpha$, so that $\tau_D$ is about $500$ rotation periods $2\pi/\Omega$. For the Great Red Spot of Jupiter or Dark Oval of Neptune, the ratio between the estimated friction time ($\sim$ 10 years) and rotation time (a few days) is similar. We then similarly expect that, at a given distance from the sun, successive mergings have time to produce a unique vortex or at most very few of them. By contrast, vortices with sufficient radial separation (a few times their size $R\sim H$) cannot merge, so that we finally expect a set of independent isolated vortices with radial interval scaling like the nebula thickness. Since $H$ scales with a power law of the distance to the sun ($r^{5/4}$ in the standard model considered below), this is consistent with an approximate geometric progression of the planetary positions.
The particles embedded in the gas of the nebula are submitted to a friction drag whose expression depends on the mean-free-path of the gas molecules relative to the particle size. For decimetric particles (and beyond $2$AU from the sun), mean-free-path exceeds size and the drag reaches the Epstein regime. The motion equations of the particles, submitted to the sun attraction, Coriolis force and friction drag, then read:
$$\cases{
{{dv_x}\over {dt}} = - 2\Omega v_y - {1\over {t_S}} (v_x
- V_x)
\cr
{{dv_y}\over {dt}} = 3\Omega^2 y + 2\Omega v_x - {1\over
{t_S}} (v_y - V_y)
}~~~,$$ where $ t_S = \rho_d s /(\rho_{gas}C_S)$ is the stopping-time for a spherical particle with radius $s$ and density $\rho_d$ in a gas with density $\rho_{gas}$. The dynamical evolution depends on the single non-dimensional friction parameter $\tau_S =\Omega t_S$, that is on the particle mass/area ratio: (i) the ligthest particles ($\tau_S << 1$) come at rest rapidly with the gas and travel with the local flow; (ii) the heaviest particles ($\tau_S >> 1$) cross the vortex with a keplerian motion nearly unaffected by the friction drag.
In the intermediate range of $\tau_s$, a numerical integration of the equations shows that a particle can be captured by the vortex if its impact parameter (initial distance to the $x$ axis) is sufficiently small (Fig.1); otherwise it is dragged by the flow. The corresponding critical impact parameter $\eta_c$ can be fitted by the function (see Fig.2)
$$f(\tau_S) = {{\eta_c}\over R} =
{}~{{A~\tau_S^{1/2}}\over {\tau_S^{3/2} +
B}}
~~,$$ where $A\simeq 2.4$ and $B\simeq 2.2$. This function reaches a maximum when $\tau_S\simeq 1$, and reduces to the power laws $\tau_S^{1/2}$ and $\tau_S^{-1}$, in the limits of the light and heavy particles, respectively. As the approach velocity ($3\eta\Omega /2$) only depends on the impact parameter, the mass capture rate is straightforwardly:
$${{dM_{capt}}\over {dt}} = {3\over 2} \sigma R^2
\Omega f^2({\tau_S})$$
where $\sigma$ is the mean surface density of nebular solid material.
We now evaluate how the capture rate depends on the distance from the sun, choosing a standard model of nebula (Cuzzi [*[et al.]{}*]{} 1993) in which the surface densities (both for gas and for particles) and the temperature are the decreasing power-laws $r^{-3/2}$ and $r^{-1/2}$, respectively; at $1$ AU the densities are set to $1 700 gcm^{-2}$ for the gas and to $20
gcm^{-2}$ for the particles, whereas the temperature is assumed to be $280~K$. Consequently, the thickness of the nebula $H$ ($\simeq R$), approximately $0.04~AU$ near Earth’s orbit, increases as $r^{5/4}$.
Further, as to get the essence of our capture mechanism, it is sufficient to assume that all the particles have the same density $\rho_d=2 gcm^{-3}$ (the density of a composite rock-ice material) and the same size $s_* = 40cm$ (a typical prediction in a gravitationally stable layer of sediment (Weidenschilling and Cuzzi 1993), inside which fragmentation is less effective than in a fully turbulent accretion-disk). The friction parameter, which writes $\tau_S = 2~{\rho_d
s/\sigma_{gas}}$, increases as $r^{3/2}$. As a result the capture rate, proportional to $f^2(\tau_S)$, is optimum at the distance $r_*$ from the sun for which the function $rf^2$ is maximum. With our numerical values this optimum is reached when $r_* \simeq 7.5~AU$, that is in between the present Jupiter’s and Saturn’s orbits, explaining the predominance of these two planets. The mass collected (at constant rate) reaches typical planetary values after a time $\Delta t$ corresponding to $500$ revolutions of the vortex (see table I).
$r~(A.U.)$ $\Delta t~(yrs)$ $\tau_S$ $M_{capt}(M_\oplus)$ $M_{core}(M_\oplus)~^a$
------------ ------------------ ---------- ---------------------- -------------------------
$~1$ $5.00~10^2$ $~0.09$ $~0.6$ $ - $
$~2$ $1.41~10^3$ $~0.27$ $~3.2$ $ - $
$~5$ $6.00~10^3$ $~1.05$ $16.0$ $15-30$
$10$ $1.45~10^4$ $~2.97$ $18.0$ $16-23$
$20$ $4.20~10^4$ $~8.42$ $~7.8$ $11-13$
$30$ $8.25~10^4$ $15.46$ $~3.8$ $14-16$
: Amount of captured mass
[Note: ]{} (a) Classical estimation of the amount of high-Z material contained in the giant planets (Pollack 1985); notice that these values are still under debate and, as recently proposed (Guillot [*[et al.]{}*]{} 1994), could be significantly smaller.
The decrease of the predicted masses at large distance seems a bit too strong, when compared to the estimated masses (Pollack 1985) of heavy elements contained in the giant planets; in fact, it would be reduced by accounting for a dispersion in particle size and density. On the other hand, comparison with the masses of the terrestrial planets has been discarded as requiring the further modelling of collisional accumulation.
The above calculations implicitly assume that the particles are continuously renewed near the vortex orbit. This occurs due to the inward drift under the systematic drag associated with the velocity difference $\Delta V$ between gas and particles. This drift, indeed, which reaches its optimum value $\Delta V$ for $\tau_S \sim 1$ (that is near Jupiter’s orbit), results in a mass flux exceeding easily the capture rate (We have also introduced this drift in the expression for the gas velocity $V_x$ and check that it has no influence on the capture cross section).
This two dimensional capture mechanism adds to the vertical settling toward the nebula midplane, which is known to form a particle sublayer whose typical thickness is $H_p\simeq
10^{-3} H$ (Cuzzi [*[et al.]{}*]{} 1993). It results, inside the vortices, in an increasing surface density $\sigma_{vort}$ and in a stronger volume density $\sigma_{vort}/H_p$ which reaches much more easily the Roche threshold for gravitational instability. In terms of the particle velocity dispersions $C_p\sim \Omega H_p$ the criterion for instability to occur reads:
$$C_p \le {{\pi G \sigma_{vort}}\over {\Omega} }~~~,$$ where $G$ is the gravitational constant. In the absence of any surface density enhancement ($\sigma_{vort} = \sigma$), this velocity threshold is very low, $20 cms^{-1}$ in our nebula model, and is easily exceeded by any residual turbulence. Indeed, according to the classical model of turbulent accretion-disk, the velocity dispersion $C_p = \alpha
C_S$ (with $\alpha =
10^{-3}$) is of the order of $2 ms^{-1}$ at 1 AU. Even with an initially laminar nebula, a minimal velocity dispersion (Weidenschilling and Cuzzi 1993; Cuzzi [*[et al.]{}*]{} 1993) $C_p = 2\Delta V/Re^*$ (where $R_* \sim 100$) would be generated by the turbulent shear between the particle sublayer (considered as heavy fluid) and the overlying gas. This velocity difference $\Delta V$ between gas and particles is approximately $60ms^{-1}$ at any distance from the sun and the resulting velocity dispersion (several $ms^{-1}$) is sufficient to inhibit gravitational instability.
By contrast, inside the vortices, the surface density is increased by several order of magnitude in some ten rotation periods, so that gravitational instabilities become much easier and rapidly gather the material into planetesimals.
The fate of these planetesimals (nearly insensitive to the gas friction) will, in fact, strongly depends on the friction parameter $\tau_S$ of the particles they formed from, that is on the distance from the sun. Indeed preliminary computations indicate that:
\(1) Inside Jupiter’s orbit ($\tau < 1$), particle concentration occurs in an annular region at the vortex periphery, as clearly seen from the dotted trajectory of Fig.1; the resulting planetesimals have wide epicyclic oscillations and are quickly released from the vortex. Afterward the growth of the terrestrial planets would proceed following the “standard” collisional history (Safronov 1969; Barge and Pellat 1991, 1993).
\(2) Outside Jupiter’s orbit ($\tau > 1$), the trapped particles deeply sink into the central vortex region (solid trajectory of Fig.1) and reach an epicyclic, nearly ballistic motion; once formed, planetesimals remain along similar slowly-evolving orbits, until they collapse into a single body, massive enough to form a Giant Planet after the capture of the surrounding gas (Safronov 1969; Pollack 1985).
The direction of planetary rotation depends on the subsequent phases of dust contraction and gas accretion, and the result is far from obvious (Dones and Tremaine, 1993; Coradini et al. 1989). However, it is straigthforward to show that the angular momentum of a swarm of particles contracting under self-gravity and inelastic collisions, when refered to its center of mass, is conserved, in the standard inertial frame of reference; then, simple calculations indicate that this angular momentum is prograde like with Keplerian circular orbits (whereas vorticity remains retrograde, even in the inertial frame).
In summary, after decay of the initial three-dimensional turbulence and settling of the solid particles toward the nebula mid-plane, two-dimensional turbulence could persist for a long time and organizes into long-lived vortices able to strongly concentrate the solid material. This allows us to suggest new solutions to some major problems in the modelling of planetary formation:
\(i) the gravitational instability at the origin of the planetesimals is made easier,
\(ii) the cores of the four giant planets form in less than $10^5~yrs$, while the terrestrial planets result from longer planetesimal accumulation.
Another important consequence of our model is simply related to the fact that, in a given vortex, particles with $\tau_S\sim 1$ are preferentially captured. This corresponds to dense particles inside the optimal radius $r^*$ and to light ones outside. An efficient mechanism of chemical segregation is therefore provided by the mass/area dependance of the friction parameter; it could help explaining some of the strong disparities observed in the compositions of meteorites and planets.
Of course the existence and structure of our long-lived vortices would require further dynamical justification. However it must be stressed that the capture mechanism we describe is unsensitive to the choice of the starting assumptions (i.e) nebula model and constant particle size. Indeed, all the conclusions of the paper hold as long as the friction time $\tau_S$ increases with the sun distance, reaching unity at Jupiter’s orbit. This is due to the fact that $\tau_S$, which depends on the ratio $\Omega/(\rho_{gas} C_S)$, strongly increases with the distance from the sun, a property which holds for various nebula structures and a wide range of particle sizes, and vortex shapes (provided its core fits with epicyclic motion). Our model has therefore a strong predictive potentiality, and it is reasonable to consider it as a possible and fruitful alternative to the classical scenario of planetesimal formation.
Abramowicz, M. A., Lanza, A., Spiegel, E.A., E. Szuszkiewicz, E.: 1992, Nature 356, 41 Alfvèn, H., Arrhenius, G.: 1976, Evolution of the Solar System (N.A.S.A. SP-345) Antipov, S.V., Nezlin M.V., Snezhkin E.N., Trubnikov A.S.: 1986, Nature 323, 238 Barge, P., Pellat, R.: 1991, Icarus 93, 270 Barge, P., Pellat, R.: 1993, Icarus 104, 79 Borderies, N., Goldreich, P., Tremaine, S.: 1982, Nature 299, 209 Coradini, A., Cerroni, P., Magni, G., Federico, C., 1989. In: Atreya, S.K., Pollack, J.B., Matthews, M.S. (eds.) Origin and Evolution of Planetary and Satellite Atmospheres. Univ. Arizona Press, Tucson, p.881 Cuzzi, J.N., Dobrovolskis, A.R., Champney, J.M.: 1993, Icarus 106, 102 Dones, L., Tremaine, S.: 1993, Icarus 103, 67 Dowling, T.E., Ingersoll A.P.: 1989, J. Atmos. Sci. 46 (21), 3256 Dubrulle, B.: 1993, Icarus 106, 59 Greenberg, R., Wacker, J.F., Hartmann, W.K., Chapman, C.R.: 1978, Icarus 35, 1 Guillot, T., Chabrier, G., Morel, P., Gautier, D.: 1994, Icarus (in press) Ingersoll, A. P.: 1990, Science 248, 308 Marcus, P.S.: 1990, J. Fluid Mech. 215, 393 Michel, J., Robert, R.: 1994, J. Stat. Phys., (in press) Miller, J., Weichman, P.B., Cross, M.C.: 1992, Phys. Rev. A 45, 2328 Morfill, G.: 1983, Icarus 53, 41 Nezlin, M.V., Snezhkin, E.N.: 1993, Rossby vortices, spiral structures, solitons (Springer-Verlag) Pollack, J.B., 1985. In: Black, D.C., Matthews, M.S. (eds.) Protostars and Planets II. Univ. Arizona Press, Tucson, p. 791 Safronov, V.S.: 1969, Evolution of the Protoplanetary Cloud and Formation of the Earth and the Planets (Nauka, Moscow) Sommeria, J., Meyers, S.D., Swinney, H.L.: 1988, Nature 331, p 689 ,P 227 Sommeria, J., Nore , C., Dumont T., Robert, R.: 1991, C.R. Acad. Sci. (Paris) 312, II, 999 Strom, S.E., Edwards, S., Skrutskie, N.F., 1993. In: Levy, E.H., Lunine, J.L., Matthews, M.S. (eds.) Protostars and Planets III. Univ. Arizona Press, Tucson, p. 1030 Weidenschilling, S. J.: 1984, Icarus 60, 553 Weidenschilling, S.J., Cuzzi, J.N., 1993. In: Levy, E.H., Lunine, J.L., Matthews, M.S. (eds.) Protostars and Planets III. Univ. Arizona Press, Tucson, p. 1030 Weizsäcker, C.F. von: 1944, Zeit. für Astrophys. 22, 319 Wetherill, G. W., 1988. In: Weaver, H.A., Danly, L. (eds.) The Formation and Evolution of Planetary Systems. Cambridge Univ. Press, Cambridge, p. 344
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
At the first part of our communicate we show how specific umbral extensions of the Stirling numbers of the second kind result in new type of Dobinski-like formulas. In the second part among others one recovers how and why Morgan Ward solution of uncountable family of $\psi$- [*difference calculus*]{} nonhomogeneous equations $\Delta_{\psi}f=\varphi$ in the form $$f(x)=\sum_{n \geq
1}\frac{B_{n}}{n_{\psi}!}\varphi^{(n-1)}(x)+\int_{\psi}\varphi(x)
+p(x)$$ extends to $\psi$- Appell polynomials case automatically. Illustrative specifications to $q$-calculus case and Fibonomial calculus case are made explicit due to the usage of upside down notation for objects of Extended Finite Operator Calculus .
author:
- |
A.K.Kwaśniewski\
\
Higher School of Mathematics and Applied Informatics\
PL - 15-021 Bialystok , ul.Kamienna 17, Poland\
e-mail: [email protected]
title: Information on some recent applications of umbral extensions to discrete mathematics
---
*to appear in Review Bulletin of Calcutta Mathematical Society Vol. 13 (2005)*
MCS numbers: 05A40, 11B73, 81S99
Key words: extended umbral calculus,Dobinski type formulas,difference equations
to be presented at the ISRAMA Congress, Calcuta - India , December 2004
INTRODUCTION {#introduction .unnumbered}
------------
The upside down convenient notation for objects of Extended Finite Operator Calculus **(EFOC)** recently being developed and promoted by the present author [@1; @2; @3; @4] is to be used throughout the whole exposition of both parts.\
The **EFOC** is an implementation of operator approach of Gian Carlo Rota to various by now formulations of extended of umbral difference calculi. For abundant references - thousands of them - see [@5] and [@1; @2; @3; @4].
In [@1; @2; @3; @4; @5] and here also $\psi$ denotes a number or functions‘ sequence - sequence of functions of a parameter $q$. $\psi$ constitutes the Morgan Ward \[6\] extension (see: after then Viskov \[7\]) of $\langle\frac{1}{n!}\rangle_{n\geq 0}$ sequence to quite arbitrary one (the so called - “admissible” in \[8\] and after then in \[1,2,3,4,5\]).The specific choices are for example : Fibonomialy-extended sequence $\langle\frac{1}{F_n!}\rangle_{n\geq
0}$ ($\langle F_n \rangle$ - Fibonacci sequence ) or just “the usual” $\psi$-sequence $\langle\frac{1}{n!}\rangle_{n\geq 0}$ or the famous Gauss $q$-extended $\langle\frac{1}{n_q!}\rangle_{n\geq
0}$ admissible sequence of extended umbral operator calculus, where $ n_q=\frac{1-q^n}{1-q}$ and $n_q!=n_q(n-1)_q! , 0_q!=1$ - see more below. With such type extension we frequently may in a “$\psi$-mnemonic” way repeat reasoning and this what was done by Rota (see at first **(FOC)** of Rota in \[9\]).
The simplicity of the first steps to be done while identifying general properties of $\psi$-extended objects consists in writing objects of these extensions in mnemonic convenient **upside down notation** [@1; @2; @3; @4] , [@10] which is here down introduced as follows:
$$\label{eq1}
\frac {\psi_{(n-1)}}{\psi_n}\equiv n_\psi,
n_\psi!=n_\psi(n-1)_\psi!, n>0 , x_{\psi}\equiv \frac
{\psi{(x-1)}}{\psi(x)} ,$$
$$\label{eq2}
x_{\psi}^{\underline{k}}=x_{\psi}(x-1)_\psi(x-2)_{\psi}...(x-k+1)_{\psi}$$
$$\label{eq3}
x_{\psi}(x-1)_{\psi}...(x-k+1)_{\psi}=
\frac{\psi(x-1)\psi(x-2)...\psi(x-k)} {\psi(x)
\psi(x-1)...\psi(x-k +1)} .$$
If one writes the above in the form $x_{\psi} \equiv \frac
{\psi{(x-1)}}{\psi(x)}\equiv \Phi(x)\equiv\Phi_x\equiv x_{\Phi}$ , one sees that the name upside down notation is legitimate.
You may consult [@1; @2; @3; @4; @5] and [@10] for further development and use of this notation .
In the **first** part of our communicate umbral extensions of the Stirling numbers of the both kinds are considered and the resulting new type of Dobinski-like formulas are discovered \[10\]. These extensions naturally encompass the well known $q$-extensions . The fact that $q$-extended Stirling numbers giving rise to the umbral $q$-extended Dobinski formula interpreted as the average of powers of random variable $X_q$ with the $q$-Poisson distribution and are equivalent by re-scaling with the other Comtet -like $q$-extended Stirling numbers \[10\] - singles out the $q$-extensions itself which appear to be a kind of bifurcation point in the domain of umbral extensions . The further consecutive umbral extensions of Carlitz-Gould $q$-Stirling numbers are therefore realized in \[10\] in a two-fold way.
In the **second** part of our communicate one displays [@11] how and why Morgan Ward solution [@6] of $\psi$- [*difference calculus*]{} nonhomogeneous equation $\Delta_{\psi}f=\varphi$ in the form $$f(x)=\sum_{n \geq
1}\frac{B_{n}}{n_{\psi}!}\varphi^{(n-1)}(x)+\int_{\psi}\varphi(x)
+p(x)$$ recently proposed by the present author [@12]- extends here now to $\psi$- Appell polynomials case - almost [*automatically*]{}. Illustrative specifications to $q$-calculus case and Fibonomial calculus case [@2; @13; @14] were already made explicit in [@12] exactly due to the of upside down notation for objects of the **EFOC** .
The First Part - Stirling numbers extensions and Dobinski-like formulae {#the-first-part---stirling-numbers-extensions-and-dobinski-like-formulae .unnumbered}
-----------------------------------------------------------------------
In this part of our communicate we follow \[10\] and we refer Reader for further details to consult \[10\]. At the start let us recall that the **two** standard [@15; @16; @17] $q$-extensions Stirling numbers of the second kind might be defined as follows:
$$\label{eq4}
x_q^n=\sum_{k=0}^{n}\Big\{{n \atop k}\Big\}_q x_q^{\underline k},$$
where $ x_q=\frac{1-q^x}{1-q}$ and $x_q^{\underline
k}=x_q(x-1)_q...(x-k+1)_q $ , which corresponds to the $\psi$ sequence choice in the $q$-Gauss form $\langle\frac{1}{n_q!}\rangle_{n\geq 0}$
and $q^\sim$-Stirling numbers
$$\label{eq5}
x^n=\sum_{k=0}^{n}\Big\{{n \atop k}\Big\}^\sim_q \chi_{\underline
k}(x)$$
where $\chi_{\underline k}(x)= x(x-1_q)(x-2_q)...(x-[k-1]_q)$
For these two classical by now $q$-extensions of Stirling numbers of the second kind - the “$q$-standard” recurrences hold respectively:
$$\Big\{{{n+1}\atop k}\Big\}_q =
\sum_{l=0}^{n}\binom{n}{l}_q q^l\Big\{{l\atop {k-1}}\Big\}_q ;
n\geq 0 , k\geq 1,$$
$$\Big\{{{n+1}\atop k}\Big\}^\sim_q =
\sum_{l=0}^{n}\binom{n}{l}_q q^{l-k+1}\Big\{{l\atop
{k-1}}\Big\}^\sim_q ; n\geq 0 , k\geq 1.$$
From the above it follows immediately that corresponding $q$-extensions of $B_n$ Bell numbers satisfy respective recurrences:
$$B_q(n+1) =
\sum_{l=0}^{n}\binom{n}{l}_q q^l B_q(l) ; n\geq 0 ,$$
$$B^\sim_q(n+1) =
\sum_{l=0}^{n}\binom{n}{l}_q q^{l-k+1} \overline {B} ^\sim_q(l) ;
n\geq 0$$ where $$\overline {B}^\sim_q(l)=
\sum_{k=0}^{l} q^k\Big\{{l\atop k}\Big\}^\sim_q .$$
**Note** that both definitions via (4) and (5) equations consequently correspond to different $q$-counting [@16] .\
For applications to coherent state phisics see \[17\] and references therein.\
With any other choice out of countless choices of the $\psi$ sequence the equation (5) becomes the definition of $\psi^\sim$- Stirling (vide “Fibonomial-Stirling”) numbers of the second kind$\Big\{{n \atop k}\Big\}^\sim_\psi$ and then $\psi^\sim$-Bell numbers $B^\sim_n(\psi)$ are defined as sums as usual - where now $\chi_{\underline k}(x)$ in (5) is to be replaced by $\psi_{\underline k}(x)=
x(x-1_\psi)(x-2_\psi)...(x-[k-1]_\psi)$. These $\psi^\sim$- Stirling numbers of the second kind recognized for $q$ case properly as Comtet numbers in Wagner‘s terminology [@16] satisfy familiar recursion and are given by familiar formulas to be presented soon.\
The extension of definition (4) of the $q$- Stirling numbers of the second kind beyond this $q$-case i.e. beyond the $\psi=\langle\frac{1}{n_q!}\rangle_{n\geq 0}$ choice is not that mnemonic and the “ behavior ” of the naturally expected recursion under extension comprises quite a surprise (see: appendix 2.2 in \[10\]).\
Therefore further consecutive umbral extension of Carlitz-Gould $q$-Stirling numbers $\Big\{{n\atop k}\Big\}_q $ and $\Big\{{n\atop k}\Big\}^\sim_q $ is realized two-fold way \[10\].
The first “easy way” consists in almost mnemonic sometimes replacement of $q$ subscript by $\psi$ because we learn from [@16] that the equation (5) defines as a matter of fact the specific case of Comtet numbers [@16] i.e. we define $\psi$-extended Comtet-Stirling numbers of the second kind as follows:
$$\label{eq6}
x^n=\sum_{k=0}^{n}\Big\{{n \atop k}\Big\}^\sim_{\psi}
\psi_{\underline k}(x)$$
where $\psi_{\underline k}(x)=
x(x-1_{\psi})(x-2_{\psi})...(x-[k-1]_{\psi}).$
As a consequence we have “for granted” :
$$\label{eq7}
\Big\{{{n+1}\atop k}\Big\}^\sim_{\psi} = \Big\{{n\atop
{k-1}}\Big\}^\sim_{\psi} + k_{\psi}\Big\{{n\atop
k}\Big\}^\sim_{\psi} ;\quad n\geq 0 , k\geq 1 ;$$
where $\Big\{{n\atop 0}\Big\}^\sim_{\psi}=
\delta_{n,0},\quad\Big\{{n\atop k}\Big\}^\sim_{\psi}=0 ,\quad k>n
;\quad $and the easy derivable recurrence relations for ordinary generating function now read
$$\label{eq8}
G^\sim_{k_{\psi}}(x)=\frac{x}{1-k_{\psi}}G^\sim_{k_{\psi}-1}(x) ,
\quad k\geq 1$$
where $$G^\sim_{k_{\psi}}(x)= \sum_{n\geq 0}\Big\{{n\atop
k}\Big\}^\sim_{\psi}x^n ,\quad k\geq 1$$
from where as a consequence we arrive to what follows:
$$\label{eq9}
G^\sim_{k_{\psi}}(x)=\frac{x^k}{(1-1_{\psi}x)(1-2_{\psi}x)...(1-k_{\psi}x)}
\quad, \quad k\geq 0$$
and adapting reasoning from [@18] we derive the following explicit formula
$$\label{eq10}
\Big\{{n\atop k}\Big\}^\sim_{\psi} = \frac
{r_{\psi}^n}{k_{\psi}!}\sum_{r=1}^{k}(-1)^{k-r}\binom{k_{\psi}}{r_{\psi}};
\quad n,k\geq 0 .$$
Expanding the right hand side of (9) results in another explicit formula for these $\psi$-case Comtet numbers i.e. we have
$$\label{eq11}
\Big\{{n\atop k}\Big\}^\sim_{\psi} = \sum_{1\leq i_1 \leq
i_2\leq...\leq i_{n-k}\leq
k}(i_1)_{\psi}(i_2)_{\psi}...(i_{n-k})_{\psi}; \quad n,k\geq 0 .$$
or equivalently (compare with \[16\])
$$\label{eq12}
\Big\{{n\atop k}\Big\}^\sim_{\psi}=\sum_{d_1+ d_2+...+d_k =
n-k,\quad d_i\geq
0}1_{\psi}^{d_1}2_{\psi}^{d_2}...k_{\psi}^{d_k};\quad n,k\geq 0 .$$
With help of $\psi^\sim$-**Stirling numbers** of the second kind being defined equivalently by (6) , (7), (11) or (12) we define now $\psi^\sim$-**Bell numbers** in a standard way $$B^\sim_n(\psi)=\sum_{k=0}^{n} \Big\{{n\atop
k}\Big\}^\sim_{\psi} ,\qquad n\geq 0 .$$
The recurrence for $B^\sim_n(\psi)$ is already quite involved and complicated for the $q$-extension case (see: the first section in \[10\])- and no acceptable readable form of recurrence for the $\psi$-extension case is known to us.
Nevertheless after adapting the the corresponding Wilf‘s reasoning from [@18] we derive for two variable ordinary generating function for $\Big\{{n\atop k}\Big\}^\sim_{\psi}$ Stirling numbers of the second kind and the $\psi$-exponential generating function for $ B^\sim_n(\psi)$ Bell numbers the following formulae
$$\label{eq13}
C^\sim_{\psi}(x,y) = \sum_{n\geq 0} A^\sim_n (\psi,y)x^n ,$$
where the $\psi$- exponential-like polynomials $ A^\sim_n
(\psi,y)$ $$A^\sim_n (\psi,y)=\sum_{k=0}^{n} \Big\{{n\atop
k}\Big\}^\sim_{\psi}y^k$$ do satisfy the recurrence
$$A^\sim_n (\psi,y)= [y(1+\partial_{\psi}]A^\sim_{n-1}(\psi,y) \qquad n\geq 1 ,$$
hence $$A^\sim_n (\psi,y)= [y(1+\partial_{\psi}]^n 1,\quad \qquad n\geq 0 ,$$ where the linear operator $\partial_{\psi}$ acting on the algebra of formal power series is being called (see: \[1,2,3,4,5\] and references therein) the “$\psi$-derivative” and $\partial_{\psi}
y^n = n_{\psi}y^{n-1}.$
The $\psi$-exponential generating function $ F^{\sim}_{\psi}(x)=
\sum_{n\geq 0}B^\sim_n(\psi)\frac{x^n}{n_{\psi}!} $
for $ F^{\sim}_n(\psi)$ Bell numbers - again - after cautious adaptation of the method from the Wilf‘s generatingfunctionology book \[18\] we get only a little bit involved formula
$$\label{eq14}
B^{\sim}_{\psi}(x)= \sum_{r\geq 0}\epsilon(\psi,r)
\frac{e_{\psi}[r_{\psi}x]}{r_{\psi}!}$$
where (see: \[6,7,1,2,3,4\]) $$e_{\psi}(x) =
\sum_{n\geq 0}\frac{x^n}{n_{\psi}!}$$
while $$\label{eq15}
\epsilon(\psi,r)=\sum_{k=r}^{\infty} \frac{(-1)^{k-r}}{(k_{\psi}-
r_{\psi})!}$$
and for $\psi$-extension the Dobinski like formula here now reads
$$\label{eq16}
B^{\sim}_n (\psi)= \sum_{r\geq 0}\epsilon(\psi,r)
\frac{r_{\psi}^n}{r_{\psi}!}.$$
In the case of Gauss $q$-extended choice of $\langle\frac{1}{n_q!}\rangle_{n\geq 0}$ admissible sequence of extended umbral operator calculus we have then
$$\label{eq17}
\epsilon(q,r)=\sum_{k=r}^{\infty}
\frac{(-1)^{k-r}}{(k-r)_q!}q^{-\binom {r}{2}}$$
and the new $q^\sim$-Dobinski formula is given by
$$\label{eq18}
B^{\sim}_n (\psi)= \sum_{r\geq 0}\epsilon(\psi,r)
\frac{r_{\psi}^n}{r_{\psi}!}.$$
which for $ q=1$ becomes the Dobinski formula from 1887 \[19\].\
As for the problem of how eventually one might interpret the $\psi^\sim$-Dobinski formulae (16) and (18) in the Rota-like way see: \[10\].\
In \[10\] you find also details in support of our conviction that $q$-extensions seem appear as a kind of bifurcation point in the domain of umbral extensions.\
For the discussion of the other way to arrive eventually to another type of Dobinski formula - via an attempt to $\psi$-extend the equation (4) - we again refer the Reader to \[10\] as this is quite a longer story with a surprise.
The parallel treatment of the Comtet $ \left[ {{\begin{array}{*{20}c} {n}\\
{k}\end{array}} } \right]^\sim_\psi$ Stirling numbers of the first kind is now not difficult.
Namely we define as in \[10\] the $\psi^\sim$ Stirling numbers of the first kind as the coefficients in the following expansion
$$\label{eq19}
\psi_{\underline k}(x)=\sum_{r=0}^{k}\left[ {{\begin{array}{*{20}c} {k}\\
{r}\end{array}} } \right]^\sim_\psi x^r$$
where - recall $\psi_{\underline k}(x)=
x(x-1_{\psi})(x-2_{\psi})...(x-[k-1]_{\psi});\quad$ From the above we then get
$$\label{eq20}
\sum_{r=0}^{k}\left[ {{\begin{array}{*{20}c} {k}\\
{r}\end{array}} } \right]^\sim_\psi \Big\{{r\atop
l}\Big\}^\sim_{\psi}= \delta_{k,l}.$$
Another (expected Whithney numbers of the first kind) $\psi^c$- Stirling numbers of the first kind \[10\] are defined as follows:
$$\label{eq21}
\psi_{\overline k}(x)=\sum_{r=0}^{k}\left[ {{\begin{array}{*{20}c} {k}\\
{r}\end{array}} } \right]^c_\psi x^r$$
where - now $\psi_{\overline k}(x)=
x(x+1_{\psi})(x+2_{\psi})...(x+[k-1]_{\psi});\quad$ More on that - see \[10\].
The Second Part -$\psi$- Appell polynomials‘ solutions of the $Q(\partial_{\psi})$- difference nonhomogeneous equation {#the-second-part--psi--appell-polynomials-solutions-of-the-qpartial_psi--difference-nonhomogeneous-equation .unnumbered}
----------------------------------------------------------------------------------------------------------------------
The second part is based on \[11,12\] to which we refer for more details.
At first recall [@3; @1] the simple fact to be used in what follows.
$Q(\partial_{\psi})$ is a $\psi$- delta operator iff there exists invertible $S\in \Sigma _{\psi}$ such that $Q(\partial_{\psi})=\partial_{\psi}S$.
Formally: “$S=Q/\partial_{\psi}$” or “$S^{-1}=\partial_{\psi}/Q$”. In the sequel we use this abbreviation $Q(\partial_{\psi})\equiv
Q$.
$\psi$- [*Appell*]{} or generalized [*Appell polynomials*]{} $\left\{A_{n}(x)\right\}_{n \geq 0}$ are defined according to $$\label{psi-A-1}
\partial_{\psi}A_{n}(x)=n_{\psi}A_{n-1}(x)$$ and they naturally do satisfy the $\psi$ - Sheffer-Appell identity [@3; @1] $$\label{psi-A-2}
A_{n}(x+_{\psi}y)=\sum_{s=0}^{n}\binom{n}{s}
_{\psi}A_{s}(y)x^{n-s}.$$ $\psi$-[*Appell*]{} or generalized [*Appell polynomials*]{} $\left\{A_{n}(x)\right\}_{n \geq 0}$ are equivalently characterized via their $\psi$- exponential generating function $$\label{psi-A-3}
\sum_{n \geq
0}z^{n}\frac{A_{n}(x)}{n_{\psi}!}=A(z)\exp_{\psi}\left\{xz\right\},$$ where $A(z)$ is a formal series with constant term different from zero - here normalized to one.
The $\psi$- exponential function of $\psi$-Appell-Ward numbers $A_{n}=A_{n}(0)$ is $$\label{psi-A-4}
\sum_{n \geq 0}z^{n}\frac{A_{n}}{n_{\psi}!}=A(z).$$ Naturally $\psi$- [*Appell*]{} $\left\{A_{n}(x)\right\}_{n \geq
0}$ satisfy the $\psi$- [*difference*]{} equation $$\label{psi-A-5}
QA_{n}(x)=n_{\psi}x^{n-1};\;\;\;'n\geq 0,$$ because $QA_{n}(x)=QS^{-1}x^{n}=Q(\partial_{\psi}/Q)x^{n}=\partial_{\psi}x^{n}=n_{\psi}x^{n-1}\;
;n \geq 0$. Therefore they play the same role in $Q(\partial_{\psi})$- [*difference*]{} calculus as Bernoulli polynomials do in standard difference calculus or $\psi$-[*Bernoulli-Ward polynomials*]{} (see Theorem 16.1 in [@1] and consult also [@12]) in $\psi$-[*difference*]{} calculus due to the following: The central problem of the $Q(\partial_{\psi})$ - [*difference calculus*]{} is: $$Q(\partial_{\psi})f=\varphi\;\;\;\;\;\;\;\;\;\;\varphi=?,$$ where $f, \varphi$ - are for example formal series or polynomials.
The idea of finding solutions is the $\psi$-[*Finite Operator Calculus*]{} [@1; @3; @4; @2] standard. As one knows [@1; @3]) any $\psi$- delta operator $Q$ is of the form
$$Q(\partial_{\psi})=\partial_{\psi}S$$
where$ S\in \Sigma_{\psi}.$ Let $Q(\partial_{\psi})=\sum_{k \geq
1}\frac{q_{k}}{k_{\psi}!}\partial_{\psi}^{k},\;\;q_{1}\neq 0$. Consider then $Q(\partial_{\psi}=\partial_{\psi}S)$ with $S=\sum_{k \geq
0}\frac{q_{k+1}}{(k+1)_{\psi}!}\partial_{\psi}^{k}\equiv \sum_{k
\geq
0}\frac{s_{k}}{k_{\psi}!}\partial_{\psi}^{k};\;\;\;s_{0}=q_{1}\neq
0$.
We thus have for $S^{-1}\equiv\hat{A}$ - call it: $\psi$- [*Appell operator*]{} - the obvious expression
$$\hat{A}\equiv S^{-1}=\frac{\partial_{\psi}}{Q_{\psi}}=\sum_{n\geq 0}\frac{A_{n}}{n_{\psi}!}\partial_{\psi}^{n}.$$
Now multiply the equation $Q(\partial_{\psi})f=\varphi
$by $\hat{A}\equiv \sum_{n\geq
0}\frac{A_{n}}{n_{\psi}!}\partial_{\psi}^{n}$ thus getting
$$\label{psi_A_6}
\partial_{\psi}f=\sum_{n\geq
0}\frac{A_{n}}{n_{\psi}!}\varphi^{(n)},\;\;\;\varphi^{(n)}=\partial_{\psi}\varphi^{(n-1)}.$$
The solution then reads:
$$\label{psi-A-7}
f(x)=\sum_{n \geq
1}\frac{A_{n}}{n_{\psi}!}\varphi^{(n-1)}(x)+\int_{\psi}\varphi(x)+p(x),$$
where $p$ is “$Q(\partial_{\psi})$- periodic” i.e. $Q(\partial_{\psi})p=0$. Compare with [@12]
for “$+_{\psi}1$- periodic” i.e. $p(x+_{\psi}1)=p(x)$ i.e. $\Delta_{\psi}p=0$. Here the relevant $\psi$ - integration $\int_{\psi}\varphi(x)$ is defined as in [@1]. We recall it in brief. Let us introduce the following representation for $\partial_{\psi}$ “difference-ization”
$$\partial_{\psi}=\hat{n}_{\psi}\partial_{0}\; ;\;\;\;\hat{n}_{\psi}x^{n-1}=n_{\psi}x^{n-1};\;\;n\geq 1,$$
where $\partial_{0}x^{n}=x^{n-1}$ i.e. $\partial_{0}$ is the $q=0$ Jackson derivative. $\partial_{0}$ is identical with divided difference operator. Then we define the linear mapping $\int_{\psi}$ accordingly:
$$\int_{\psi}x^{n}=\left( \hat{x}\frac{1}{\hat{n}_{\psi}}\right)x^{n}=\frac{1}{(n+1)_{\psi}}x^{n+1};\;\;\;n\geq 0$$ where of course $\partial_{\psi}\circ \int_{\psi}=id$.
Examples
========
1. The case of $\psi$- [*Bernoulli-Ward*]{} polynomials and $\Delta_{\psi}$- [*difference calculus*]{} was considered in detail in [@12] following [@6].
2. Specification of (a) to the Gauss and Heine originating $q$-umbral calculus case [@6; @3; @4; @5; @2] was already presented in [@12].
3. Specification of (a) to the Lucas originating FFOC - case was also presented in [@12] (here: FFOC=[**F**]{}ibonomial [**F**]{}inite [**O**]{}perator [**C**]{}alculus), see example 2.1 in [@2]). Recall: the [*Fibonomial coefficients*]{} -already known in 19-th century to Lucas \[20\]- are defined “binomially” as
$$\binom{n}{k}_{F}=\frac{F_{n}!}{F_{k}!F_{n-k}!}=\binom{n}{n-k}_{F},$$
where ($F_{n}$- [*Fibonacci numbers*]{} in up-side down notation: $n_{F}\equiv F_{n}\neq
0$,\
$n_{F}!=n_{F}(n-1)_{F}(n-2)_{F}(n-3)_{F}\ldots
2_{F}1_{F};\;\;0_{F}!=1$;\
$n_{F}^{\underline{k}}=n_{F}(n-1)_{F}\ldots
(n-k+1)_{F};\;\;\;\binom{n}{k}_{F}\equiv
\frac{n^{\underline{k}}_{F}}{k_{F}!}$.
We shall call the corresponding linear difference operator $\partial_{F};\;\;\partial_{F}x^{n}=n_{F}x^{n-1};\;\;n\geq 0$ the $F$-derivative. Then in conformity with [@6] and with notation as in [@1]-[@13; @14] one has:
$$E^{a}(\partial_{F})=\sum_{n\geq 0}\frac{a^{n}}{n_{F}!}\partial_{F}^{n}$$
for the corresponding \[3,1,4\] generalized translation operator $E^{a}(\partial_{F})$. The $\psi$- integration for the moment is still not explored $F$- integration and we arrive at the **$F$- Bernoulli** polynomials unknown till now.\
[**Note:**]{} recently a combinatorial interpretation of Fibonomial coefficient has been found [@14] by the present author.
4. The other examples of $Q(\partial_{\psi})$- [*difference calculus*]{} - expected naturally to be of primary importance in applications are provided by the possible use of such $\psi$- Appell polynomials as:
- $\psi$-Hermite polynomials $\left\{H_{n,\psi}\right\}_{n\geq 0}$:
$$H_{n,\psi}(x)=\left[\sum_{k\geq 0}\left(-\frac{1}{2}\right)^{k}
\frac{\partial_{\psi}^{2k}}{k_{\psi}!}\right]x^{n}\;\;\;n\geq 0;$$
- $\psi$ - Laguerre polynomials $\left\{ L_{n,\psi}\right\}_{n\geq
0}$ [@3]:
$$\begin{gathered}
L_{n,q}(x) = \frac{{n_{q}} }{{n}}\hat {x}_{\psi}
\left[{\frac{{1}}{{\partial _{\psi} - 1}}}\right]^{-n}x^{n-1} =
\frac{{n_{\psi}} }{{n}}\hat {x}_{\psi} \left( {\partial _{\psi} -
1} \right)^{n}x^{n-1} = \\ = \frac{{n_{\psi}} }{{n}}\hat
{x}_{\psi} \sum\limits_{k = 1}^{n} \left( { - 1} \right)^{k}\left(
{{\begin{array}{*{20}c}
{n} \hfill \\
{k} \hfill \\
\end{array}} } \right) \partial _{\psi} ^{n - k}x^{n - 1} =\\
=\frac{{n_{\psi}} }{{n}}\sum\limits_{k = 1}^{n} {} \left( { - 1}
\right)^{k}\left( {{\begin{array}{*{20}c}
{n} \hfill \\
{k} \hfill \\
\end{array}} } \right) \left( {n - 1} \right)_{\psi} ^{\underline {n - k}
}\frac{{k}}{{k_{\psi}} }x^{k} .\end{gathered}$$
For $q=1$ in $q$-extended case one recovers the known formula :
$L_{n,q=1}(x) = \sum\limits_{k = 1}^{n} {} \left( { - 1}
\right)^{k}\frac{{n_{q} !}}{{k_{q} !}}\left(
{{\begin{array}{*{20}c}
{n - 1} \hfill \\
{k - 1} \hfill \\
\end{array}} } \right)x^{k}$.
[99]{}
A.K.Kwaśniewski [*Main theorems of extended finite operator calculus*]{} Integral Transforms and Special Functions, [**14**]{} No 6 (2003): 499-516
A.K.Kwaśniewski [*On Simple Characterizations of Sheffer $psi$-polynomials and Related Propositions of the Calculus of Sequences*]{} ,Bulletin de la Soc. des Sciences et de Lettres de Lodz, [**52**]{} Ser. Rech. Deform. 36 (2002):45-65, ArXiv: math.CO/0312397 .
A.K.Kwaśniewski [*Towards $\psi$-extension of Finite Operator Calculus of Rota*]{} Rep. Math. Phys. [**48**]{} No3 (2001):305-342
A.K.Kwaśniewski[*On extended finite operator calculus of Rota and quantum groups*]{} Integral Transforms and Special Functions [**2**]{} No4 (2001):333-340
A.K.Kwaśniewski [*First contact remarks on umbra difference calculus references streams*]{} Bulletin de la Soc. des Sciences et de Lettres de Lodz, in press; ArXiv: math.CO/0403139 v1 8 March 2004
M. Ward: [*A calculus of sequences*]{}, Amer.J.Math. Vol.58, 1936,pp.255-266
O.V. Viskov: [*Operator characterization of generalized Appel polynomials* ]{} Soviet Math. Dokl. [**16**]{}, 1521 (1975).
G. Markowsky, [*Differential operator and theory of binomial enumeration*]{}, Math. Anal. Appl. [**63**]{} (1978), 145–155.
G.-C. Rota: [*Finite Operator Calculus*]{}, Academic Press, New York 1975.
A.K.Kwaśniewski [*On umbral extensions of Stirling numbers and Dobinski-like formulas*]{} submitted for publication ArXiv:math.CO/0411002 30 October 2004
A.K.Kwaśniewski [*$\psi$-Appell polynomials‘ solutions of an umbral difference nonhomogeneous equation*]{} Bulletin de la Societe des Sciences et des Lettres de Lodz (54) Serie: Recherches sur les Deformations [**45**]{} (2004), 11-15. ArXiv: math.CO/0405578 30 May 2004
A. K. Kwaśniewski: [*On basic Bernoulli-Ward polynomials* ]{}, Bulletin de la Societe des Sciences et des Lettres de Lodz (54) Serie: Recherches sur les Deformations [**45**]{} (2004), 5-10. ArXiv:math.CO/0405577 30 May 2004
A.K.Kwaśniewski Fibonomial cumulative connection constants Bulletin of the ICA [**44**]{} (2005), 81-92. ArXiv: math.CO/0406006 1 Jun 2004
A.K.Kwaśniewski: More on combinatorial interpretation of the fibonomial coefficients Bulletin de la Soc. des Sciences et de Lettres de Lodz (54) Serie: Recherches sur les Deformations [**44**]{} (2004),23-38. ArXiv: math.CO/0402344 v1 22 Feb 2004
S.C. Milne [*A $q$-analog of restricted growth functions, Dobinski’s equality, and Charlier polynomials*]{} ,Trans. Amer. Math. Soc. [**245**]{} (1978) 89-118
Carl G. Wagner [*Partition Statistics and q-Bell Numbers (q = -1)*]{}, Journal of Integer Sequences [**7**]{}(2004) Article 04.1.1
J.Katriel, M. Kibler [*Normal ordering for deformed boson operators and operator-valued deformed Stirling numbers*]{} J. Phys. A: Math. Gen. [**25**]{} (1992): 2683-26-91
H.S.Wilf [*Generatingfunctionology*]{} Boston: Academic Press, 1990
G. Dobinski [*Summierung der Reihe S .... für m = 1, 2, 3, 4, 5, ....*]{} Grunert Archiv (Arch. Math. Phys) [**61**]{}, 333-336, (1877)
E. Lucas, [*Théorie des fonctions numériques simplement périodiques*]{}, American Journal of Mathematics [**1**]{} (1878): 184–240
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present atmospheric parameters for 51 nearby F and G dwarf and subgiant stars uniformly distributed over the $-2.60 < {\rm [Fe/H]} < +0.20$ metallicity range that is suitable for the Galactic chemical evolution research. Lines of iron in the two ionization stages, Fe I and Fe II, were used to derive a homogeneous set of effective temperatures, surface gravities, iron abundances, and microturbulence velocities. Our spectroscopic analyses took advantage of employing high-resolution ($R \ge$ 60000) Shane/Hamilton and CFHT/ESPaDOnS observed spectra and non-local thermodynamic equilibrium (NLTE) line formation for Fe I and Fe II in the classical 1D model atmospheres. The spectroscopic method was tested in advance with the 20 benchmark stars, for which there are multiple measurements of the infrared flux method (IRFM) effective temperature and their Hipparcos parallax error is less than 10%. We found NLTE abundances from lines of Fe I and Fe II to be consistent within 0.06 dex for every benchmark star, when applying a scaling factor of [$S_{\rm H}$]{} = 0.5 to the Drawinian rates of inelastic Fe+H collisions. The obtained atmospheric parameters were checked for each program star by comparing its position in the log g – [$T_{\rm eff}$]{} plane with the theoretical evolutionary track of given metallicity and $\alpha$-enhancement in the @Yi2004 grid. Our final effective temperatures lie exactly in between the $T_{\rm IRFM}$ scales of @Alonso1996irfm and @Casagrande2011, with a mean difference of +46 K and $-51$ K, respectively. NLTE leads to higher surface gravity compared with that for LTE. The shift in log g is smaller than 0.1 dex for stars with either \[Fe/H\] $\ge -0.75$, or [$T_{\rm eff}$]{} $\le$ 5750 K, or log g $\ge$ 4.20. NLTE analysis is crucial for the VMP turn-off and subgiant stars, for which the shift in log g between NLTE and LTE can be up to 0.5 dex. The obtained accurate atmospheric parameters will be used in the forthcoming papers to determine NLTE abundances of important astrophysical elements from lithium to europium and to improve observational constraints on the chemo-dynamical models of the Galaxy evolution.'
author:
- 'T. Sitnova'
- 'G. Zhao'
- 'L. Mashonkina'
- 'Y. Chen'
- 'F. Liu'
- 'Yu. Pakhomov'
- 'K. Tan'
- 'M. Bolte'
- 'S. Alexeeva'
- 'F. Grupp'
- 'J.-R. Shi'
- 'H.-W. Zhang'
title: |
**Systematic NLTE study of the $-2.6 \le$ \[Fe/H\] $\le 0.2$ F and G dwarfs\
in the solar neighbourhood.\
I. Stellar atmosphere parameters [^1]**
---
Introduction
============
In recent decades the construction of large telescopes, the developement of efficient spectrometers, and progress in stellar atmosphere and synthetic spectrum numerical modelling provided a considerable improvement in the quality and quantity of stellar abundance determinations. Elemental abundances for the FGK-type stars provide important clues to understand the main processes at play in formation and evolution of the Milky Way (MW). Studies of stellar samples with metallicity \[Fe/H\][^2] $\ge -1$ showed how trends in various chemical elements can be used to learn the chemodynamical evolution of the Galactic disk and to resolve the thick disk and thin disk [@Gratton1996ASPC...92..307G; @Fuhrmann1998; @Fuhrmann2004; @mash_eu; @2003MNRAS.340..304R; @Mishenina2004; @Zhang2006; @Adibekyan2013; @Bensby2014 and references therein]. Studies of very metal-poor (VMP, \[Fe/H\] $\le -2$) stars unraveled the major processes of chemical enrichment of the Milky Way .
Elemental abundances of stars with different metallicities are the main constraint for Galactic chemodynamical models ( e.g. ). For the Galactic chemical evolution research it would be useful to deal with a homogeneous set of accurate stellar abundances in wide metallicity range, from supersolar down to extremely low iron abundances. Large high-resolution spectroscopic surveys were proposed to increase the statistics of observations and to improve a homogeneity of derived stellar chemical abundances over wide metallicity range. The Apache Point Observatory Galactic Evolution Experiment (APOGEE) [@APOGEE2013AJ....146...81Z], with its $~10^5$ red giants observed in the near-infrared H-band with a spectral resolving power of $R \simeq$ 22500, enables users to address numerous Galactic structure and stellar populations issues. The Gaia-ESO Survey consortium [@2012Msngr.147...25G; @2013Msngr.154...47R; @Smiljanic2014arXiv1409.0568S] is obtaining high-quality spectroscopic data for about 10$^5$ stars using FLAMES at the VLT. Spectra for a million stars will be acquired by the coming Galactic Archaeology with HERMES (GALAH) survey [@2015arXiv150204767D].
We initiate a new project of deriving a homogeneous set of stellar atmosphere parameters and chemical abundances for the Galactic field FGK-type stars in the metallicity range $-3 \le$ \[Fe/H\] $\le +0.3$ that is suitable for a systematic research of Galactic chemical evolution. By employing high-resolution spectral observations, deriving accurate stellar atmosphere parameters, and treating the non-local thermodynamic equilibrium (NLTE) line formation for the key chemical species, we attempt to push the accuracy of the abundance analysis to the points where the trends with metallicity could be meaningfully discussed. The selected stars are uniformly distributed in metallicity and they can serve as a calibration sample for the existing and coming large-scale stellar surveys, such as the LAMOST Experiment for Galactic Understanding and Exploration [LEGUE @2012RAA....12..735D], designed to determine from low-resolution spectra elemental abundances for hundreds of thousands to millions of stars, spread over much larger distances than ever before.
The goal of the present paper is the determination of precise atmospheric parameters, i.e. the effective temperature, ${T_{\rm eff}}$, the surface gravity, log g, the iron abundance, \[Fe/H\], and the microturbulence velocity, ${\xi_t}$, for the selected sample of, presumably, dwarf stars. The next papers in the series will concern with calculations of the NLTE abundances for many chemical elements from Li to Eu and analysis of the obtained abundance trends. Taking advantage of the Shane/Hamilton and CFHT/ESPaDOnS high-resolution observational material with sufficient spectral coverage and the NLTE line formation for Fe I-Fe II, we determined stellar atmosphere parameters spectroscopically from lines of iron in the two ionization stages, Fe I and Fe II.
The spectroscopic methods are, in particular, useful for evaluating ${T_{\rm eff}}$ and log g of the VMP stars that are mostly distant, and, therefore, their temperatures cannot be reliably derived from photometric colours due to the uncertainty in the interstellar reddening data and, at present, surface gravities cannot be calculated from the trigonometric parallax because it is either rather uncertain or non-measurable. We look forward to seeing soon accurate parallaxes from Gaia[^3]. In the literature there is no consensus on a validity of the excitation temperatures, $T_{\rm exc}$, derived through forcing no dependence of the Fe I line abundance on the excitation energy of the lower level, ${E_{\rm exc}}$. For their sample of VMP cool giants @Cayrel2004 reached in LTE good agreement between the photometric and spectroscopic temperatures after they excluded strong lines with ${E_{\rm exc}}$ = 0 eV. A similar approach (LTE, but not cutting ${E_{\rm exc}}$ = 0 eV) yielded the same outcome for the VMP near-main-sequence stars in @2013ApJ...769...57F, while up to several hundred degrees cooler excitation temperatures than the photometric ones were found for the cool giants.
@Ruchti2013 determined surface gravities of the stellar sample selected from the RAVE survey, using a grid of the NLTE abundance corrections from @lind2012. For common ${T_{\rm eff}}$ in the NLTE and LTE analysis, they found systematic biases in log g of up to 0.2 dex in the metallicity range $-2 <$ \[Fe/H\] $<-0.5$ and up to 0.3 dex for the more metal-poor stars.
The NLTE effects for lines of Fe I were accounted for by @Bensby2014 in their study of 714 F and G dwarf stars, by applying the NLTE abundance corrections interpolated in the grid of @lind2012. They found minor shifts between NLTE and LTE, with $\Delta{T_{\rm eff}}$ = $-12\pm28$ K, $\Delta$log g = +0.012 $\pm$ 0.059, and $\Delta{\log\varepsilon}({\rm Fe}) = -0.013\pm0.016$, on average. This is because the majority of their stars are close-to-solar metallicity and mildly metal-deficient ones, with \[Fe/H\] $> -1.2$, and the NLTE abundance corrections are small, at the level of few units of hundredth, as predicted by @mash_fe and @lind2012 for this metallicity range. Furthermore, the NLTE calculations of @lind2012 resulted in small departures from LTE for Fe I until the extremely low metallicities because they were performed assuming a high efficiency of the Fe+H collisions and using the Drawinian rates [@Drawin1968; @Drawin1969]. For nearby stars with very good [Hipparcos]{} parallaxes ($\Delta\pi/\pi < 0.05$), @Bensby2014 found that essentially all stars with log g $>$ 4.2 and ${T_{\rm eff}}<$ 5650 K do not show ionization equilibrium between Fe I and Fe II, when determining the surface gravity from [Hipparcos]{} parallaxes. @Bensby2014 suggested that classical plane-parallel (1D) models have limitations and cannot properly handle excitation balance and/or ionization balance, and they applied, therefore, empirical corrections to the atmospheric parameters from ionization balance. In this study we check with our stellar sample, whether spectroscopic methods of ${T_{\rm eff}}$ and log g determination from lines of Fe I and Fe II have any limitations and what they are.
The paper is structured as follows. In Sect. \[Sect:observations\] we describe the stellar sample, observations, and their reduction. Kinematical properties of the selected stars are calculated in Sect. \[Sect:Kinematics\]. They are used to identify a membership of individual stars to the galactic stellar populations. Section \[Sect:method\] concerns with the methodical issues. Stellar effective temperatures, surface gravities, iron abundances, microturbulence velocities, and masses are derived in Sect. \[Sect:parameters\]. The obtained results are discussed in Sect. \[Sect:comparisons\]. Our conclusions are presented in Sect. \[Sect:Conclusions\].
Stellar sample, observations, spectra reduction {#Sect:observations}
===============================================
The stars were selected from the \[Fe/H\] catalogue of @cayrel2001 based on the following criteria.
1. The stars have a declination of $\delta > -20^\circ$ to be observed at the nothern sky.
2. The selected stars should cover as broad as possible metallicity range and be uniformly distributed, with 2-3 stars in each 0.1 dex metallicity interval.
3. The stellar sample should be homogeneous in temperature and luminosity. We selected the F-G-K dwarfs and subgiants based on the literature data.
4. Binaries, variables, and stars with any chemical peculiarity (carbon-enhanced stars, low \[$\alpha$/Fe\] stars, etc.) were excluded.
5. For testing purposes the stellar sample should include a dozen of well-studied stars of various metallicities, for which their stellar atmosphere parameters ${T_{\rm eff}}$ and log g were determined in the literature, presumably, from the non-spectroscopic methods.
Thus, 50 stars, in total, were selected, and they cover the $-2.6 <$ \[Fe/H\] $< +0.3$ metallicity range.
Spectra of 48 stars were obtained using the Hamilton Echelle Spectrograph mounted on the Shane 3-m telescope of the Lick observatory during the two observation runs March 15-16, 2011 and January 5-11, 2012. Most stars were observed, at least, twice. The resolving power is $R = \lambda/\Delta\lambda \simeq$ 60000, and the spectral coverage is 3700 Å to 9300 Å. The signal-to-noise ratio ($S/N$) at 5500 Å is higher than 100 for most stars. The exception is BD$-04^\circ$ 3208, where $S/N$($\lambda$ = 5500Å) $\simeq$ 70. It is worth noting that the spectra suffer from the fringing effect in the infrared band.
For two stars, we used the spectra observed with the ESPaDOnS echelle spectrograph at the 3.6-m telescope of the Canada-France-Hawaii (CFH) observatory in queued service observing mode during several nights in 2011 and 2012. The spectra cover 3700 Å to 10450 Å. They have $R \simeq$ 81000 and $S/N \simeq 150$ at 5030 Å and higher value of about 200 at 8090 Å.
The Shane/Hamilton spectra were reduced with the IRAF package following standard procedures that include the aperture definition, flat-field correction, background subtraction, one-dimensional spectra extraction, wavelength calibration, and continuum normalization. The wavelength calibration was performed using the lamp with the titanium cathode in argon environment. Such a lamp is rare in the astronomical observations, and an appropriate list of the Ti and Ar lines was compiled by @pakhomov2013. The uncertainty in wavelength calibration is estimated to be 0.006 Å.
The CFHT/ESPaDOnS spectra were reduced with the special reduction package Libre-ESpRIT[^4], which includes two steps. The first one performs a geometrical analysis from a sequence of calibration exposures, and the second step achieves spectrum optimal extraction in itself, using the geometrical information derived in the first step. The reduction procedure returns the normalised spectra.
To improve the statistics of the VMP stars, we include the well-studied halo star HD140283 in our sample. Its high-quality observed spectrum was taken from the [ESO UVESPOP]{} survey [@2003Msngr.114...10B]. For seven stars our observational material was complemented with the data from different sources. We employed the spectra obtained by Klaus Fuhrmann with the fiber optics Cassegrain echelle spectrograph FOCES at the 2.2-m telescope of the Calar Alto Observatory in 1996 to 2000 [@Fuhrmann1998; @Fuhrmann2004]. The star BD$-04^\circ$3208 was observed using UVES/VLT in April 2001 within our project 67.D-0086A (PI: T. Gehren). Details of spectra reduction were described by @Mashonkina2003. For HD49933 its high-quality spectrum was observed with the HARPS spectrograph at the 3.6-m ESO La Silla telescope. Details of spectra reduction were described by @Ryabchik2009.
This paper deals with 51 stars, in total. The investigated stars together with characteristics of the observed spectra are listed in Table\[Tab:observations\].
Kinematic properties of the selected stars {#Sect:Kinematics}
==========================================
The stellar kinematics is closely related to stellar populations in the Galaxy, and it is commonly applied to identify a membership of given star to the galactic stellar populations, namely the thin disk, the thick disk, and the halo.
The galactic space velocity components U, V, W were calculated using the equations and formalism of @1987AJ.....93..864J. They were defined with respect to the local standard of rest (LSR), adopting the standard solar motion (U, V, W) = (10.00, 5.25, 7.17)[kms$^{-1}$]{} of @1998MNRAS.298..387D. In computations of the X, Y, Z-coordinates, we used the best current estimate of the Galactocentric distance of the Sun $R_G$ = 8.0kpc, which was inferred from a comparison of different statistical techniques [@2013ARep...57..128M]. The parallaxes and proper motions were taken from the updated version of the [Hipparcos]{} catalogue and the radial velocities from the [Hipparcos]{} Input catalogue [@1993BICDS..43....5T]. The obtained results are presented in Table\[kin\]. The star BD$-13^\circ$3442 is missing in the [Hipparcos]{} catalogue, and it was assigned to the galactic halo based on its very low Fe abundance, with \[Fe/H\] = $-2.62$ (Sect.\[Sect:parameters\]), and high proper motion as indicated by the Simbad database[^5].
Following @Fuhrmann2000 and later studies , we identify the stars with peculiar space velocities with respect to the LSR in the range 85 [kms$^{-1}$]{} $< v_{pec} <$ 180 [kms$^{-1}$]{} as belonging to the thick disk stellar population, the stars with $v_{pec} <$ 85 [kms$^{-1}$]{} to the thin disk, and all the halo stars of our sample have $v_{pec} >$ 200 [kms$^{-1}$]{} (Fig.\[Fig:toomre\]). The star G090-003 is not shown in Fig.\[Fig:toomre\], because its peculiar space velocity, $v_{pec}$ = 847 $\pm$ 1580 [kms$^{-1}$]{}, is rather uncertain due to large error of the [Hipparcos]{} parallax, $\pi$ = 1.12 $\pm$ 2.14 mas. For six stars with peculiar space velocities close to $v_{pec}$ = 85 [kms$^{-1}$]{}, their membership to either the thin disk or the thick disk cannot be decided using only the kinematic properties. Following @Fuhrmann1998, we also took into account their chemical properties, as determined in Sect.\[Sect:parameters\]. For example, HD30562, with $v_{pec}$ = 93 [kms$^{-1}$]{}, was assigned to the thin disk because it has a supersolar Fe abundance, with \[Fe/H\] = 0.17, and it does not reveal Mg enhancement relative to Fe.
Method of calculations {#Sect:method}
======================
Stellar parameters of the selected sample were determined in this study applying different methods, including the spectroscopic one based on the NLTE analysis of lines of Fe I and Fe II. This section describes calculations of the theoretical spectra.
Codes and model atmospheres
---------------------------
The statistical equilibrium (SE) of Fe I-Fe II was calculated with a comprehensive model atom treated by @mash_fe. The main source of the uncertainty in the NLTE results for Fe I is poorly known inelastic collisions with hydrogen atoms. We employed the formula of @Drawin1968 [@Drawin1969], as implemented by @Steenbock1984, for allowed $b-b$ and $b-f$ transitions and a simple relation between hydrogen and electron collisional rates, $C_H = C_e \sqrt{(m_e/m_H)} N_H/N_e$, for forbidden transitions, following @1994PASJ...46...53T. @mash_fe recommended to scale the Drawin rates by a factor of [$S_{\rm H}$]{} = 0.1 based on their analysis of element abundances derived from the two ionization stages, Fe I and Fe II, in the selected VMP stars. @Bergemann_fe_nlte estimated a larger value of [$S_{\rm H}$]{} = 1. In this study, [$S_{\rm H}$]{} was further constrained using the benchmark stars.
The coupled radiative transfer and SE equations were solved with a revised version of the DETAIL code [@detail]. The update was presented by @mash_fe. The obtained departure coefficients were then used by the codes binmag3 [@oleg], synthV-NLTE [Vadim Tsymbal, private communication, based on SynthV code @synthv] and SIU [@Reetz] to calculate the synthetic line profiles.
We used the MARCS model structures [@Gustafssonetal:2008][^6], which were interpolated for given [$T_{\rm eff}$]{}, log g, and \[Fe/H\] using a FORTRAN-based routine written by Thomas Masseron[^7].
Selection of iron lines
-----------------------
The investigated lines were selected according to the following criteria.
\(i) The lines should be almost free of visible/known blends in the solar spectrum.
\(ii) For Fe I the linelist should cover a range of excitation energies of the lower level, ${E_{\rm exc}}$, as large as possible to investigate the excitation equilibrium of neutral iron. On the other hand, the low-excitation (${E_{\rm exc}}<$ 1.2 eV) lines in metal-poor stars often appear to have higher abundances than lines with higher excitation energy [see, for example @Cayrel2004; @Lai2008]. The calculations with ab initio 3D, time-dependent, hydrodynamical model atmospheres of cool stars predict that the 3D effects for lines of Fe I are strongly ${E_{\rm exc}}$ dependent . For example, the 3D abundance corrections amount to $-0.8$ dex and $-0.25$ dex for the $E_{\rm exc} = 0$ lines of Fe I in the models [$T_{\rm eff}$]{} / log g / \[M/H\] = 5860/4/$-2$ and 5850/4/$-1$, respectively, while they are $-0.2$ dex and $-0.07$ dex for the $E_{\rm exc} = 2$ eV lines and close to 0 for the $E_{\rm exc} = 4$ eV lines . To minimise a possible influence of the 3D effects on [$T_{\rm eff}$]{} determinations, we did not consider lines of Fe I with $E_{\rm exc} < 2$ eV.
\(iii) For both close-to-solar metallicity and VMP stars, lines of various strength have to be present in the spectrum to evaluate stellar microturbulence velocity ${\xi_t}$.
For individual stars, we also avoided using of saturated lines with the equivalent width $EW >$ 180mÅ to minimise an influence of possible uncertainty in the van der Waals damping constants on final abundances.
The selected lines are listed in Table\[lines\] along with their atomic parameters. For neutral iron we employed experimental $gf$-values from @1979MNRAS.186..633B [@1982MNRAS.199...43B; @1982MNRAS.201..595B; @1991JOSAB...8.1185O], and . Two sets of oscillator strengths from @mb2009 and @ru1998 were inspected for Fe II. Van der Waals broadening of the selected lines is accounted for using the most accurate data as provided by @BPM for Fe I and for Fe II.
Sun as a reference star
-----------------------
To minimise the effect of the uncertainty in $gf$-values on the final results, we applied a line-by-line differential NLTE and LTE approach, in the sense that stellar line abundances were compared with individual abundances of their solar counterparts. The solar flux observations were taken from the Kitt Peak Solar Flux Atlas [@Atlas]. The calculations were performed with the MARCS model atmosphere $5777/ 4.44/ 0$. A microturbulence velocity of 0.9[kms$^{-1}$]{} was adopted.
The solar LTE and NLTE abundances from individual lines are presented in Table\[lines\]. The NLTE calculatoins were performed using [$S_{\rm H}$]{} = 0.5 (see sect.\[Sect:parameters\]). For the Fe I lines, the solar mean LTE and NLTE abundances are log $A_{\rm LTE} = -4.57 \pm 0.09$ and log $A_{\rm NLTE} = -4.56 \pm 0.09$. Hereafter, $A_X = N_X/N_{tot}$ is the element abundance taken relative to the total number of atoms. For lines of Fe II, the NLTE abundance corrections do not exceed 0.01 dex, in absolute value, and the solar mean abundance amounts to ${\log A_{\rm }} = -4.56 \pm 0.05$ and ${\log A_{\rm }} = -4.50 \pm 0.05$, when using $gf$-values from @mb2009 and @ru1998, respectively. Hereafter, the statistical abundance error is the dispersion in the single line measurements about the mean: $\sigma = \sqrt{\Sigma (x - x_i )^2 /(N - 1)}$, where N is the total number of lines used, x is their mean abundance, $x_i$ the individual abundance of each line. It is worth noting that an uncertainty of 0.06 dex in abundance from the Fe II lines leads to an 0.12 dex uncertainty in log g for solar-type stars. This is crucial for surface gravity determinations from the Fe I/Fe II ionization equilibrium.
Determination of stellar parameters {#Sect:parameters}
===================================
To derive atmospheric parameters of our stellar sample, the following strategy was applied. First, we selected the stars, for which their effective temperatures and surface gravities were determined reliably using the non-spectroscopic methods, namely ${T_{\rm eff}}$ from the infrared flux method (IRFM) and log g from the well-known relation between log g, stellar mass $M$, ${T_{\rm eff}}$, and absolute bolometric magnitude ${\rm M_{bol}}$:
$$[g] = [M] + 4[{T_{\rm eff}}] + 0.4 ({\rm M_{bol}}-{\rm M_{bol\odot}}). \quad \label{formula1}$$
Here, square brackets denote the logarithmic ratio with respect to the solar value. The ${\rm M_{bol}}$ magnitudes were obtained using the $V$ magnitudes from , the revised [Hipparcos]{} parallaxes of , and the bolometric corrections (BC) from @Alonso1995. For the Sun the absolute visual magnitude ${\rm M_{V\odot}} = 4.87$ and BC$_\odot = -0.12$ were adopted. The star’s mass was estimated from its position in the ${\rm M_{bol}}$ - $\log{T_{\rm eff}}$ diagram by interpolating in the Dartmouth isochrones [@Dotter2008]. “Reliably” means that multiple measurements of the IRFM temperature are available and the uncertainty in the [Hipparcos]{} parallax based surface gravity, log g$_{Hip}$, is less than 0.1 dex. This implies a parallax error of smaller than 10%. The 20 stars in our sample meet these requirements. They are listed in Table\[params\] as the benchmark stars.
Atmospheric parameters of the remaining stars were derived or improved spectroscopically from the NLTE analysis of lines of Fe I and Fe II.
In the stellar parameter range, with which we concern, lines of Fe I are weakened towards higher [$T_{\rm eff}$]{}, resulting in higher derived element abundance, while they are nearly insensitive to variation in log g. In contrast, lines of Fe II are only weakly sensitive to variation in [$T_{\rm eff}$]{}, while they are weakened with log g increasing. If [$T_{\rm eff}$]{} is fixed, the surface gravity is obtained from the requirement that abundances from lines of Fe I and Fe II must be equal. One needs to be cautious, when deriving both [$T_{\rm eff}$]{} and log g from Fe I and Fe II. Any shift in [$T_{\rm eff}$]{} leads to shifting log g in the sense: an increase in [$T_{\rm eff}$]{} leads to an increase in log g. Therefore, in this study we used non-spectroscopic (IRFM) determinations of [$T_{\rm eff}$]{}, where available, and for every star its final parameters, [$T_{\rm eff}$]{}/log g/\[Fe/H\], were checked with the corresponding evolutionary track. For each benchmark star, abundances from lines of Fe I and Fe II were derived under various line-formation assumptions, i.e., NLTE with [$S_{\rm H}$]{} = 0.1, 0.5, 1 and LTE, and we investigated which of them leads to consistent element abundances from both ionization stages. Our calculations showed that the departures from LTE are small for the \[Fe/H\] $> -1$ stars. Indeed, in this metallicity range the NLTE abundance correction, $\Delta_{\rm NLTE} = \log A_{\rm NLTE} - \log A_{\rm LTE}$, does not exceed 0.04 dex for [$S_{\rm H}$]{} = 0.5, and the difference in NLTE abundances between [$S_{\rm H}$]{} = 0.1 and 0.5 is smaller than 0.04 dex. The three of six \[Fe/H\] $< -1$ stars are cool, with ${T_{\rm eff}}\le$ 5400 K, and they are not useful for constraining [$S_{\rm H}$]{} because of small $\Delta_{\rm NLTE} \le$ 0.02 dex. Table\[shtab\] lists the abundance differences between Fe I and Fe II from the calculations with different [$S_{\rm H}$]{} for the three benchmark stars with the largest NLTE effects. For every star, the acceptable abundance difference is achieved with [$S_{\rm H}$]{} = 0.5 and 1. Hereafter, we choose [$S_{\rm H}$]{} = 0.5 to perform the NLTE calculations for Fe I-FeII.
[rcccc]{} 84937 & -0.06 & -0.02 & 0.00 & 0.11 \
94028 & -0.06 & -0.05 & -0.04 & -0.01 \
140283 & -0.06 & -0.04 & -0.02 & 0.10
Spectroscopic stellar parameters were obtained in the iterative procedure. For each star, our initial guess was ${T_{\rm eff}}$ = $T_{\rm IRFM}$, if available, or ${T_{\rm eff}}$ = $T_{b-y}$ and log g = log g$_{Hip}$. Effective temperature and surface gravity were corrected until the ionisation equilibrium between Fe I and Fe II and the Fe I excitation equilibrium were fulfilled.
Effective temperatures
----------------------
Although there are multiple measurements of the IRFM effective temperature for individual stars of our sample , we found no study that provides the data for every investigated star. Figure\[Fig:teff\_comp\] (top panel) displays the differences between different sources for the stars in common with this work. It can be seen that the ${T_{\rm eff}}$ scale of @Casagrande2011 [C11] is hotter than that of @Alonso1996irfm [A96], with $\Delta T_{\rm IRFM}$(C11 - A96) = 114 $\pm$ 62 K for 25 stars. The difference between C11 and is smaller, $\Delta T_{\rm IRFM}$(C11 - GB09) = 32 $\pm$ 83 K for 26 stars. Here, we did not count the two outliers HD34411 and HD142373, with large uncertainties in $T_{\rm IRFM}$ of 479 K and 342 K, respectively, that were caused by using saturated 2MASS photometry.
We calculated four sets of the photometric temperatures from the $b-y$ and $V-K$ colour indices using two different calibrations of @Casagrande2010 [C10] and @Alonso1996. The $V$ magnitudes and $b-y$ colours were taken from for 45 stars, in total, and the $K$ magnitudes from the 2MASS catalogue [@2MASS2006AJ....131.1163S] for 24 stars, in total. Figure\[Fig:teff\_comp\] display the differences between $T_{\rm IRFM}$ and $T_{b-y}$ for both effective temperature scales. We found that $T_{b-y}$ and $T_{V-K}$ agree well with those from the IRFM, when using the calibrations of C10, with $T_{\rm IRFM}$(C11) – $T_{b-y}$(C10) = 10 $\pm$ 46 K (42 stars) and $T_{\rm IRFM}$(C11) – $T_{V-K}$(C10) = 22 $\pm$ 68 K (21 stars). For the @Alonso1996 calibration, the differences are larger, with $T_{\rm IRFM}$(A96) – $T_{b-y}$ = 41 $\pm$ 78 K (25 stars) and $T_{\rm IRFM}$(A96) – $T_{V-K}$ = 72 $\pm$ 58 K (12 stars), although the statistics is small for the $V-K$ photometry.
For the [*benchmark stars*]{}, we aimed to have a homogeneous set of ${T_{\rm eff}}$ based on a single source of the IRFM temperatures. The only study that provides $T_{\rm IRFM}$ for every benchmark star is GB09. Using their $T_{\rm IRFM}$, we could not achieve the Fe I/Fe II ionization equilibrium and the Fe I excitation equilibrium for half benchmark stars, independent of applying either LTE or NLTE approach. No preference was, therefore, given to any source of $T_{\rm IRFM}$. Instead, for each star its temperature from A96, GB09, C10, and C11, where available, and the corresponding log g$_{Hip}$ value were checked with the difference in NLTE abundances between Fe I and Fe II and a slope of the \[Fe/H\] versus ${E_{\rm exc}}$ plot for Fe I. For six stars, corrections up to 50 K were applied to the most appropiate $T_{\rm IRFM}$ to obtain the final effective temperature (Table\[params\]).
Surface gravities, metallicities, and microturbulence velocities
----------------------------------------------------------------
For each benchmark star, the NLTE abundances from the two ionization stages, Fe I and Fe II, were found to be consistent within 0.06 dex. Hereafter, 0.06 dex was considered as an admissible difference between Fe I and Fe II, when deriving spectroscopic stellar parameters of the remaining, non-benchmark, stars. For given star, we started from checking the photometric temperature, $T_{\rm IRFM}$ or $T_{b-y}$, and the [Hipparcos]{} parallax based gravity, log g$_{Hip}$. Effective temperature and surface gravity were allowed to vary within the corresponding error bars. The procedure was iterated until the NLTE abundance difference Fe I – Fe II is getting smaller than 0.06 dex, the excitation trend for Fe I disappears, and lines of different equivalent width give consistent iron abundances.
We found that the differential approach largely removes a line-to-line scatter for the \[Fe/H\] $\ge -1.5$ stars. For example, when moving from the absolute to the differential NLTE abundances, the dispersion reduces from $\sigma$ = 0.10 dex to $\sigma$ = 0.04 dex for lines of Fe I and from $\sigma$ = 0.08 dex to $\sigma$ = 0.05 dex for lines of Fe II in HD22484 (6000/4.07/0.01, Fig.\[abs\]). For more metal-poor stars, the differential approach is less efficient in removing a line-to-line scatter. For example, for HD 84937 (6350/4.09/$-2.12$) the scatter of data reduces for Fe II, but not Fe I lines (Fig.\[abs1\]). When using the ${E_{\rm exc}}>$ 2 eV lines of Fe I, we obtained $\sigma = 0.035$ dex and 0.078 dex for the absolute and differential abundances, respectively. This is, probably, due to the uncertainties in the van der Waals damping constant, $C_{6}$. The lines, which can be measured in the VMP stars, have strong van der Waals broadened wings in the solar spectrum (for example Fe I 5383Å, Fig.\[kb5383\]). For such a line, the uncertainty in the derived solar abundance is contributed from the uncertainties in both $gf$- and $C_{6}$-values. When dealing with the VMP stars, the differential analysis is able to cancel the uncertainty in $gf$, but not $C_{6}$.
Although our final results are based on using the ${E_{\rm exc}}\ge$ 2 eV lines of Fe I, for most stars we also checked abundances determined from the lower excitation lines. We found no abundance difference between the low and high excitation lines for close-to-solar metallicity stars. For the most metal-poor stars, such as HD 84937, the absolute abundances reveal an excitation trend, but it disappears for the differential abundances (Fig.\[abs1\]).
We did not determine stellar parameters under the LTE assumption and cannot evaluate the differences in log g between using LTE and NLTE. However, the NLTE effects can be estimated based on our LTE and NLTE calculations for Fe I-Fe II and the sensitivity of iron lines to variation in log g. Since the departures from LTE lead to weakened lines of Fe I, but they do not affect lines of Fe II until the extremely low metallicities, the ionization equilibrium Fe I/Fe II is achieved for higher gravities in NLTE than in LTE. The shifts in log g increase towards lower metallicity and surface gravity. For our program stars they may be up to 0.1 dex at \[Fe/H\] $> -1.5$, up to 0.2 dex for $-2.2 <$ \[Fe/H\] $< -1.8$ and reach 0.45 dex for the most MP star of our sample BD$-13^\circ$3442, with \[Fe/H\] = $-2.62$.
Checking atmospheric parameters with evolutionary tracks
--------------------------------------------------------
For each star the obtained effective temperature and surface gravity were checked by comparing its position in the log g – [$T_{\rm eff}$]{} plane with the theoretical evolutionary track of given metallicity and $\alpha$-enhancement in the @Yi2004 grid (Fig.\[halo\_thick\]). An $\alpha$-enhancement was assumed to be equal to \[Mg/Fe\] as derived in this study and presented in Table\[kin\]. Stellar masses corresponding to the best fit evolutionary tracks are indicated in Table\[params\]. They range between 0.60 and 0.85 solar mass for the halo and thick disk stars that does not contradict with their evolutionary status and old age. For the thin disk stars, their masses range between 0.85$M_\odot$ and 1.5$M_\odot$ and ages between 1 Gyr and 9 Gyr.
For two halo cool dwarfs we had to revise [$T_{\rm eff}$]{} / log g, to enforce the stars to sit on the main sequence of the appropriate evolutionary track. We stress that the changes in atmospheric parameters were not allowed to destroy the ionization equilibrium between Fe I and Fe II and the Fe I excitation equilibrium. For HD64090, we adopted [$T_{\rm eff}$]{} = 5400 K, which is close to $T_{\rm IRFM}$ = 5440 K (A96), and log g = 4.70, which is 1.5$\sigma_{\rm log g}$ higher than log g$_{Hip}$. For BD+66$^\circ$0268 our final temperature, [$T_{\rm eff}$]{} = 5300 K, is close to $T_{\rm IRFM}$ = 5280 K (A96) and log g = 4.72 is 2$\sigma_{\rm log g}$ higher than log g$_{Hip}$.
Final atmospheric parameters
----------------------------
The final atmospheric parameters together with the NLTE and LTE abundance differences between Fe I and Fe II are presented in Table\[params\]. For the star’s metallicity we adopted \[Fe/H\] determined from lines of Fe II, although the difference in NLTE abundances between Fe I and Fe II nowhere exceeds 0.06 dex. The Fe I – Fe II abundance differences are also shown in Fig. \[tgf50\]. In NLTE, they reveal no trends with any atmospheric parameter, either \[Fe/H\], or ${T_{\rm eff}}$, or log g.
Figure \[slopes\] displays the excitation temperature slopes, d\[Fe/H\]/d${E_{\rm exc}}$, calculated from lines of Fe I in the individual stars. The data reveal no trend with effective temperature, and $d$\[Fe/H\]/$d{E_{\rm exc}}$ = 0.0044 dex/eV, on average. The excitation slope seems to depend slightly on metallicity at \[Fe/H\] $\ge -1$, although nowhere its magnitude exceeds the error bars. The scatter of data is larger for the \[Fe/H\] $< -1.8$ than less metal-poor stars due to smaller number of observed lines of Fe I.
In Fig. \[sp\_hipp\] we compare surface gravities from the Fe I/Fe II NLTE ionization equilibrium and Hipparcos parallax methods for the 20 stars, which do not belong to the benchmark stellar sample, but have a relative parallax uncertainty less than 10%. The differences are minor, with $\Delta$log g(Sp – Hip) = $0.008\pm0.037$ dex, on average, and they do not show any trends with surface gravity or effective temperature. Having in mind that for each benchmark star the Fe I/Fe II ionization equilibrium was achieved with log g from the Hipparcos parallax method, we infer that the spectroscopic method of gravity determination is working in the 5130 K $\le {T_{\rm eff}}\le$ 6600 K and 3.12 $\le$ log g $\le$ 4.66 range. An exception is HD64090, for which we determined log g$_{Sp}$ = 4.70 that is 0.13 dex higher than log g$_{Hip}$. We, thus, do not confirm a temperature trend of $\Delta$log g(Sp – Hip) for cool (${T_{\rm eff}}<$ 5300 K) dwarf (log g $> 4.2$) stars obtained by @Bensby2014.
Final slopes of the \[Fe/H\] vs. $EW$ plots amount to $-0.0002$dex/mÅ, on average. Using the derived microturbulence velocities ${\xi_t}$ and basic atmosheric parameters ${T_{\rm eff}}$, log g, and \[Fe/H\], we built up the approximation formula:
$${\xi_t}= -0.21 + 0.06 \times {\rm [Fe/H]} + 5.6 \times ({T_{\rm eff}}/10^4) - 0.43 \times {\rm log~g}.\label{formula2}$$
Microturbulence velocity grows towards higher [$T_{\rm eff}$]{} and lower surface gravity. The metallicity dependence is weak. Similar relations were previously found by other authors, for example , and @Ramirez2013. In the first three studies theirs formulae did not include the \[Fe/H\] term because these studies dealt with the limited metallicity range. Using the formula (\[formula2\]) results in ${\xi_t}$ = 1.1[kms$^{-1}$]{} and 1.4[kms$^{-1}$]{} for the solar atmospheric parameters and 6000/4.0/$-1.0$, respectively. For comparison, the corresponding values are 1.1[kms$^{-1}$]{} and 1.5[kms$^{-1}$]{} from the @Ramirez2013 formula.
Notes on individual stars {#Sect:notes}
-------------------------
We provide additional comments on stars that turned out to be interesting for one or the other reason in the course of the analysis.
[*HD59984 and HD105755*]{}, with \[Fe/H\] = $-0.69$ and $-0.73$, respectively, may be the most metal-poor thin-disk stars. Both kinematics, $v_{pec}$ = 25[kms$^{-1}$]{} and 28[kms$^{-1}$]{}, and an age of 8 Gyr for both stars support their thin-disk status.
[*HD106516*]{}: We failed to obtain self-consistent stellar parameters for this star. On the one hand, the star exhibits a typical thick-disk kinematics, $v_{pec}$ = 108[kms$^{-1}$]{}, and chemical signatures, \[Mg/Fe\] = 0.38 and \[Fe/H\] = $-0.73$. On the other hand, its age was estimated to be 6 Gyr and unlikely older than 9 Gyr, identifying the star as a thin-disk star. According to @2001AJ....122.3419C, it is a single-lined spectroscopic binary, with the period $P$ = 843.9 days.
[*HD134169*]{}: kinematics, $v_{pec}$ = 5[kms$^{-1}$]{} identifies it as a thin-disk star. However, HD134169 exhibits a typical thick-disk chemical signatures, \[Mg/Fe\] = 0.32 and \[Fe/H\] = $-0.78$, and old age, $\tau \simeq$ 11 Gyr. Enhancement of Al relative to Fe, with \[Al/Fe\] = 0.54, found by also suggests a thick-disk origin.
[*BD+37$^\circ$ 1458*]{}: this is a halo star. With spectroscopic [$T_{\rm eff}$]{} = 5500 K, log g = 3.70, and \[Fe/H\] = $-1.95$, the star sits on the evolutionary track of $M = 0.67 M_\odot$ at $\tau$ = 24 Gyr. The younger age can only be obtained for higher [$T_{\rm eff}$]{} and log g. However, there is no ground for a substantial revision of atmospheric parameters. The star’s temperature is well fixed by several studies. @Alonso1996irfm [@2005ApJ...626..446R], and derived $T_{\rm IRFM}$ = 5510 K, 5516 K, and 5582 K, respectively. determined ${T_{\rm eff}}$ = 5450 K from the Balmer line wing fits. With ${T_{\rm eff}}$ = 5500 K, log g$_{Hip}$ = $3.41\pm0.18$ is even lower than the spectroscopically derived value. For this star we choose the spectroscopic stellar parameters as the final ones.
Uncertainties in derived atmospheric parameters {#Sect:uncertain}
-----------------------------------------------
The following approaches were applied to evaluate the uncertainties in the derived atmospheric parameters.
For the benchmark stars we adopted the $T_{\rm IRFM}$ errors as indicated by the original sources. Statistical error of the Hipparcos parallax based surface gravity was computed as the quadratic sum of errors of the star’s parallax, effective temperature, mass, visual magnitude, and bolometric correction: $$\sigma_{\rm log g}^2 = (2 \sigma_{\log\pi})^2 + (4 \sigma_{\log T})^2 + \sigma_{\log M}^2 + (0.4 \sigma_V)^2 + (0.4 \sigma_{\rm BC})^2 .$$
The uncertainties in $V$ magnitude and bolometric correction together contribute less than 0.01 dex to the total log g error. For [$T_{\rm eff}$]{} the contribution nowhere exceeds 0.03 dex. Stellar masses were well constrained using the evolutionary tracks, with an uncertainty of no more than 0.1 $M_\odot$, resulting in 0.04 dex contribution to $\sigma_{\rm log g}$.
In Fig.\[this\_irfm\] our final ${T_{\rm eff}}$s are compared with the IRFM temperatures from C11 and A96. For 47 stars in common with C11, our values are 51 K, on average, lower. In contrast, this study determined 46 K higher temperatures compared with that of A96 for 29 common stars. We note very similar statistical errors of 63 K and 69 K for the temperature differences (this work – C11) and (this work – A96), respectively. Based on these comparisons, we estimate the systematic and statistical errors of ${T_{\rm eff}}$ for the 31 stars with the spectroscopically derived atmospheric parameters to be 50 K and 70 K.
For 20 stars with the spectroscopic surface gravities, log g$_{Sp}$, their [Hipparcos]{} parallaxes were measured with a relative parallax error less than 10%, and we calculated reliable log g$_{Hip}$ values. As shown in Fig.\[sp\_hipp\], log g$_{Sp}$ – log g$_{Hip}$ = $0.008\pm0.037$ dex. This led us to estimate the uncertainty in log g$_{Sp}$ to be 0.04 dex.
Statistical error of \[Fe/H\] was defined by the dispersion, $\sigma$, for lines of Fe II in given star.
Statistical error of the microturbulence velocity was adopted to be common for the whole stellar sample, and it was defined by the dispersion in the single ${\xi_t}$ measurements about the relation (\[formula2\]). It amounts to 0.14 [kms$^{-1}$]{}.
It is worth noting that the uncertainty in spectroscopically derived atmospheric parameters can be larger than that quoted above. For example, for the halo star HD74000 we found the two sets of stellar parameters that reproduce well the Fe I/Fe II ionization and Fe I excitaion equilibrium (Fig.\[hd74000\]) and fit the appropriate evolutionary tracks. These are ${T_{\rm eff}}$ = 6225 K, log g = 4.13, and \[Fe/H\] = $-1.97$ suggesting the star’s age of 15 Gyr and ${T_{\rm eff}}$ = 6360 K, log g = 4.30, and \[Fe/H\] = $-1.93$ resulting in $\tau$ = 13.5 Gyr. The first set relied, in fact, on $T_{\rm IRFM}$(A96) and the second one on $T_{\rm IRFM}$(C11). A shift of +135 K in [$T_{\rm eff}$]{} leads to +0.17 dex shift in log g, and both sets of parameters provide an acceptable star’s age. No of two sets can be preferred from a comparison of log g$_{Sp}$ with the corresponding [Hipparcos]{} parallax based surface gravity, log g$_{Hip} = 4.16\pm0.16$ for ${T_{\rm eff}}$ = 6225 K and log g$_{Hip} = 4.20\pm0.16$ for ${T_{\rm eff}}$ = 6360 K. From analysis of H$_\alpha$ and H$_\beta$ in HD74000 @Mashonkina2003 inferred ${T_{\rm eff}}$ = 6225 K. Two pairs of ${T_{\rm eff}}$ / log g, which provide the Fe I/Fe II ionization equilibrium and reasonable star’s mass and age, were also obtained for HD108177 (6100/4.22/$-1.67$ and 6250/4.40/$-1.62$), and G090-003 (6100/3.90/$-2.04$ and 5930/3.80/$-2.10$). We, therefore, recommend to apply various spectroscopic and non-spectroscopic methods to given star to find an unique solution for the star’s [$T_{\rm eff}$]{} / log g.
Comparison with other studies {#Sect:comparisons}
=============================
For ten stars of our sample, their ${T_{\rm eff}}$s were determined in the literature based on measurements of the angular diameters, trigonometric parallaxes, and bolometric fluxes. @Boyajian2013 [B13] published the most numerous list of the interferometric temperatures, $T_{int}$, based on angular diameters measured with the CHARA array (330 m maximum baseline). Figure\[this\_irfm\] displays the temperature differences between this study and interferometric measurements of B13, @North2009 [@Creevey2012; @Creevey2014], and @vonBraun2014. Our values are, on average, higher, with $\Delta{T_{\rm eff}}$ = 135 $\pm$ 126 K for ten stars and 78 $\pm$ 81 K, when excluding the two outliers, HD 103095 and HD 140283. We selected three stars for a detailed comparison.
[*HD 102870.*]{} Two successive determinations of @Boyajian2012 and B13 resulted in $T_{int}$ = 6132 K and 6054 K, with small temperature errors of 36 K and 13 K, respectively. A downward revision by 78 K was due to employing the different bolometric fluxes. The most recent temperature of B13 is in line with the earlier data of @North2009, $T_{int}$ = 6059 $\pm$ 49 K. Surface gravity based on the asteroseismic measurements amounts to log g$_{seis}$ = 4.11 $\pm$ 0.02 [@creevey2013]. With [$T_{\rm eff}$]{} = 6060 K and log g = 4.11, we failed to achieve the ionization equilibrium between Fe I and Fe II and obtained Fe I – Fe II = $-0.12$ dex in NLTE. It is worth noting, log g$_{Hip}$ = 4.14 adopted in this study agrees well with log g$_{seis}$. For solar-type stars the asteroseismology method depends only weakly on [$T_{\rm eff}$]{}. Indeed, a 200 K temperature difference produces a difference of 0.007 dex in log g. To obtain consistent NLTE abundances from lines of Fe I and Fe II using log g = 4.14, one needs to have 110 K higher temperature of HD 102870 (Table\[params\]) compared with its $T_{int}$.
[*HD 103095.*]{} Using the CHARA array measurements, B13 and @Creevey2012 derived $T_{int}$ = 4771 $\pm$ 18 K and 4818 $\pm$ 54 K, respectively. We checked [$T_{\rm eff}$]{} = 4820 K and log g$_{Hip}$ = 4.60 with various spectroscopic temperature and gravity indicators.
\(i) The Fe I/Fe II ionization equilibrium is not fulfilled, and the NLTE abundance difference amounts to Fe I – Fe II = $-0.27$ dex.
\(ii) We compared the NLTE abundances from C I 9094 Å, 9111 Å with the carbon abundance from a number of molecular CH bands, which are known to be sensitive to [$T_{\rm eff}$]{} variation. The NLTE method for C I and atomic and molecular data were taken from @Alexeeva_carbon. With the 4820/4.60/$-1.3$ model, the abundance difference, C I – CH = 0.56 dex, is very large, while it amounts to $-0.06$ dex for our final parameters, [$T_{\rm eff}$]{} = 5130 K and log g = 4.66.
\(iii) Effective temperature can also be constrained from analysis of the H$_\alpha$ line wings (Fig.\[Halpha\_HD103095\]). The theoretical NLTE profiles of H$_\alpha$ were computed following @Mashonkina2008. It can be seen that [$T_{\rm eff}$]{} = 4820 K leads to shallower wings of H$_\alpha$ compared with the observations, but [$T_{\rm eff}$]{} = 5130 K to deeper core-to-wing transition region. The best fit was achieved for [$T_{\rm eff}$]{} = 5030 K. It is worth noting, @Cayrel2011 obtained 100 K, on average, lower temperatures from the H$_\alpha$ line wings compared with accurate effective temperatures from the apparent angular diameter for the eleven FGK-type stars in the $-0.7 \le$ \[Fe/H\] $\le 0.2$ range. In contrast, $T_{int}$ = 4771 K and 4818 K of HD103095, as determined by B13 and @Creevey2012, respectively, are more than 200 K lower than the temperature from H$_\alpha$.
Thus, no spectroscopic temperature indicator supports the literature data on $T_{int}$ for HD103095.
[*HD140283.*]{} @Creevey2014 derived $T_{int}$ = 5534 $\pm$ 103 K and 5647 $\pm$ 105 K assuming zero-reddening and A$_V$ = 0.1$^m$, respectively. It is worth noting, HD140283 is a nearby star, at a distance of 58 pc from the Sun, and an interstellar absorption of A$_V$ = 0.1$^m$ is unlikely produced. Of the two temperatures $T_{int}$ = 5534 K should be preferred, nevertheless we checked both using the same spectroscopic indicators as for HD103095. The NLTE abundance differences Fe I – Fe II were found to be $-0.18$ dex and $-0.09$ dex, respectively. For both temperatures an abundance difference between carbon atomic and molecular lines is large, with (C I – CH) = 0.69 dex and 0.32 dex, while consistent abundances, (C I – CH) = 0.00 dex, were obtained for our final parameters, [$T_{\rm eff}$]{} = 5780 K and log g = 3.70. Figure\[Halpha\_HD140283\] shows theoretical profiles of H$_\alpha$ in the three different model atmospheres compared with the observed spectrum of HD140283. It is evident that $T_{int}$ = 5534 K is too low and it does not fit any spectroscopic indicator of [$T_{\rm eff}$]{}.
It is worth noting, both HD 103095 and HD 140283 have rather small angular diameters, $\theta_{int}$ = 0.679 $\pm$ 0.007 mas [@Creevey2012] and 0.353 $\pm$ 0.013 mas [@Creevey2014], respectively, which can be overestimated resulting in underestimated temperatures. Indeed, @2014MNRAS.439.2060C suspected systematic trends in the @Boyajian2012 [B12] dataset that was also based on the CHARA array interferometric measurements. They obtained the differences ($T_{\rm IRFM}$(C11) - $T_{int}$(B12)) growing towards smaller angular diameter and reaching +200 K, on average, at $\theta_{int}$ = 0.8 mas.
Asteroseismic measurements of surface gravity were made for one more star of our sample, HD 49933 [@Kallinger2010]. They resulted in log g$_{seis}$ = 4.22, which is only 1.4$\sigma_g$ higher, than our value.
Spectroscopic stellar parameters were determined in a number of recent studies. We selected three of them for comparison. @Bensby2014 used lines of Fe I and Fe II to derive temperatures and gravities of the extended stellar sample. The LTE abundances from individual lines were corrected using the grid of the NLTE abundance corrections from @lind2012. For nearby stars with very good [Hipparcos]{} parallaxes, @Bensby2014 calculated also the log g$_{Hip}$ values and found systematic discrepancies between the surface gravities from the two methods for the stars with log g $>$ 4.2 and ${T_{\rm eff}}<$ 5650 K. Therefore, they applied empirical corrections to the atmospheric parameters from ionization balance. It should be reminded, we obtained no temperature trend of $\Delta$log g(Sp – Hip) for our program stars (Sect.\[Sect:parameters\]). Figure\[this\_spectr\] (top panel) shows the differences in ${T_{\rm eff}}$ and log g for 17 stars in common with @Bensby2014. On average, $\Delta{T_{\rm eff}}$ = $-34\pm$87 K, $\Delta$log g = $-0.02\pm$0.11, and $\Delta$\[Fe/H\] = $-0.02\pm$0.11 in the sense “this work minus Bensby”. The two independent research provide consistent stellar parameters, with the small systematic shifts and the statistical errors being typical of stellar parameter determinations. However, for some individual stars the discrepancies are uncomfortably large. For example, $\Delta{T_{\rm eff}}$ exceeds 100 K for five stars and $\Delta$log g $>$ 0.2 for HD22879. We note, in particular, the differences of different sign for two well-studied very metal-poor stars HD 84937 and HD140283, with $\Delta{T_{\rm eff}}= -190$ K and +120 K and $\Delta$log g $= -0.14$ and +0.13, respectively. This can be due to large uncertainty in evaluating a slope of the log A(Fe I) versus ${E_{\rm exc}}$ trend, where the total number of investigated lines is limited. Since the Fe I/Fe II ionization equilibruim depends on not only surface gravity, but also [$T_{\rm eff}$]{}, an overestimation (underestimation) of [$T_{\rm eff}$]{} produces an upward (downward) shift in log g.
@Takeda2005 derived stellar [$T_{\rm eff}$]{} and log g from lines of Fe I and Fe II under the LTE assumption. However, all the 18 stars in common with our work lie in the $-1.3 \leq$ \[Fe/H\] $\leq$ 0.2 metallicity range, where the departures from LTE for Fe I and Fe II are small. Therefore, the obtained differences in ${T_{\rm eff}}$ and log g (Fig.\[this\_spectr\], middle panel) cannot be caused by applying different line-formation treatments. The data in Fig.\[this\_spectr\] suggest, most probably, that the uncertainty in the derived effective temperature was directly translated to the uncertainty in surface gravity.
@Fuhrmann1998 [@Fuhrmann2004] applied different spectroscopic approaches to derive [$T_{\rm eff}$]{} from the Balmer line wing fits and log g from the Mg Ib line profile fits under the LTE assumption. For 16 stars in common with this work, the mean differences in effective temperature and iron abundance, $\Delta{T_{\rm eff}}$ = 29 $\pm$ 71 K (Fig.\[this\_spectr\], bottom panel) and $\Delta$\[Fe/H\] = 0.00 $\pm$ 0.07, do not exceed the error bars. The only star, HD30743, was found with $\Delta{T_{\rm eff}}>$ 100 K. However, this study obtains higher surface gravities, with $\Delta$log g = 0.08 $\pm$ 0.07. The discrepancy can be explained, in part, by using the LTE approach in @Fuhrmann1998 [@Fuhrmann2004]. In the stellar parameter range, with which we concern, Mg I is subject to overionization resulting in weakened Mg Ib line. Ignoring the departures from LTE leads to an underestimation of the derived gravity.
Conclusions {#Sect:Conclusions}
===========
The sample of 51 nearby FG dwarf stars uniformly distributed over the $-2.60 < {\rm [Fe/H]} < +0.20$ metallicity range was selected for a systematic NLTE investigation of the Galaxy chemical evolution. A membership of individual stars to the particular galactic stellar population, namely the thin disk, the thick disk, and the halo, was mainly identified using the star’s kinematics and for few stars using also the chemical signatures, \[Fe/H\] and \[Mg/Fe\]. A combination of the photometric and spectroscopic methods was applied to derive a homogeneous set of the stellar atmosphere parameters: [$T_{\rm eff}$]{}, log g, \[Fe/H\], and ${\xi_t}$. Our spectroscopic analyses took advantage of employing the high-resolution ($R \ge$ 60000) observed spectra and NLTE line formation for Fe I and Fe II in the classical 1D model atmospheres. Spectroscopic method of [$T_{\rm eff}$]{}/log g determination was tested using the 20 benchmark stars, for which there are multiple measurements of the IRFM effective temperature and their Hipparcos parallax error is less than 10%. An efficiency of poorly known inelastic Fe+H collisions in the SE of Fe I-Fe II was estimated empirically from analysis of their different influence on lines of Fe I and Fe II in the MP benchmark stars. We found abundances from the two ionisation stages to be consistent within 0.06 dex for every star, when applying [$S_{\rm H}$]{} = 0.5.
Stellar parameters of the remaining 31 stars were obtained spectroscopically from the NLTE analysis of the iron lines. For lines of Fe II in our program stars the NLTE abundance corrections do not exceed 0.01 dex in absolute value. The deviations from LTE for Fe I grow towards higher ${T_{\rm eff}}$ and lower \[Fe/H\] and log g. For stars with either \[Fe/H\] $\ge -0.75$, or [$T_{\rm eff}$]{} $\le$ 5750 K, or log g $\ge$ 4.20, the difference in average abundance between NLTE and LTE was found to be less than 0.06 dex, which translates to a shift of less than 0.1 dex in log g. Since NLTE leads to weakened lines of Fe I, but it does not affect lines of Fe II until the extremely low metallicities, the ionization equilibrium Fe I/Fe II is achieved for higher gravities in NLTE than in LTE. The NLTE analysis is crucial for the VMP turn-off and subgiant stars. Indeed, the shift in log g between NLTE and LTE reaches +0.45 dex for BD$-13^\circ$ 3442, with [$T_{\rm eff}$]{} = 6400 K, log g(NLTE) = 3.95, and \[Fe/H\] = $-2.62$.
The obtained effective temperatures and surface gravities were checked by comparing the star’s position in the log g – [$T_{\rm eff}$]{} plane with the theoretical evolutionary track of given metallicity and $\alpha$-enhancement in the @Yi2004 grid. Most stars reveal self-consistent data on star’s mass, age, and kinematics.
Our final effective temperatures lie exactly in between the $T_{\rm IRFM}$ scales of A96 and C11, with a systematic shift of +46 K and $-51$ K, respectively. We estimate the [$T_{\rm eff}$]{} statistical error to be about 70 K. Surface gravities obtained from the Fe I/Fe II and Hipparcos parallax methods were found to agree well. We do not confirm a temperature trend of $\Delta$log g(Sp – Hip) reported by @Bensby2014 for the log g $>$ 4.2 and ${T_{\rm eff}}<$ 5650 K stars.
We recommend to apply a line-by-line differential approach relative to the Sun, to minimise the effect of the uncertainty in $gf$-values on the final results. It is quite efficient for the \[Fe/H\]$ > -1.5$ stars. However, in the more metal-poor stars, the line-to-line scatter for Fe I has a similar magnitude for the differential and absolute abundances. This is, probably, due to uncertainties in $C_{6}$-values for the high-excitation lines. We note, in particular, Fe I 5367Å (${E_{\rm exc}}$ = 4.41 eV) and Fe I 5383 Å (${E_{\rm exc}}$ = 4.31 eV).
It was found that none of checked spectroscopic temperature indicators, such as Fe I versus Fe II, C I versus CH, and the H$_\alpha$ line wings, do support interferometric effective temperatures of HD 103095, $T_{int}$ = 4818 K $\pm$ 54 K [@Creevey2012], and HD 140283, $T_{int}$ = 5534 $\pm$ 103 K [@Creevey2014]. Both stars have rather small angular diameters, and their measurements can be affected by some systematic trends, as suspected by @2014MNRAS.439.2060C for the @Boyajian2012 dataset.
We conclude that the NLTE analysis of lines of iron in the two ionisation stages, Fe I and Fe II, is efficient in deriving reliable atmospheric parameters for the FG-type dwarf stars in the broad metallicity range, down to \[Fe/H\] = $-2.6$. The stellar parameter determinations would benefit from using the complementary data on interferometric and IRFM temperatures, trigonometric parallaxes, and asteroseismology measurements. We also recommend to combine different spectroscopic temperature and gravity indicators and check the obtained [$T_{\rm eff}$]{}/log g values with the theoretical evolutionary tracks / isochrones.
The obtained accurate atmospheric parameters will be used in the forthcoming papers to determine NLTE abundances of important astrophysical elements from lithium to europium and improve observational constraints on the chemo-dynamical models of the Galaxy evolution. We also plan to extend our sample toward lower metallicity and surface gravity.
The authors thank Prof. Debra Fisher for carrying out some of the observations, Oleg Kochukhov and Vadim Tsymbal for providing the codes binmag3 and synthV-NLTE at our disposal and Klaus Fuhrmann for providing the FOCES spectra. Y.Q. Chen acknowledges the observation of ESPaDOnS/CFHT telescope for two stars obtained through the Telescope Access Program (TAP, 2011B), which is funded by the National Astronomical Observatories. This study was supported by the Russian Foundation for Basic Research (grant 14-02-91153 and 14-02-31780), the National Natural Science Foundation of China (grants 11390371, 11233004, 11222326, 11103034, 11473033, 11473001), and by the National Basic Research Program of China (grant 2014CB845701/02/03). We made use the Simbad, MARCS, and VALD databases.
We thank the anonymous referee for comments that helped to improve the manuscript.
[rrrllll]{} & & &\
19373 & 4.05 & 0.10 & 234, 228 & 30.00, 30.00 & 2 & 2012-01-09, 2012-01-09\
22484 & 4.28 & 0.01 & 171, 160 & 30.00, 30.00 & 2 & 2012-01-09, 2012-01-09\
22879 & 6.74 & -0.84 & 170, 195 & 360.00, 360.00 & 2 & 2012-01-09, 2012-01-09\
24289 & 9.96 & -1.94 & 112, 128 & 2700.00, 2700.00 & 2 & 2012-01-09, 2012-01-09\
30562 & 5.77 & 0.17 & 180, 190 & 120.00, 120.00 & 2 & 2012-01-09, 2012-01-09\
30743 & 6.26 & -0.44 & 188, 184 & 240.00, 240.00 & 2 & 2012-01-09, 2012-01-09\
34411 & 4.70 & 0.01 & 220, 218, 209F$^1$ & 45.00, 45.00, 84.19 & 3 & 2012-01-11, 2012-01-11, 2012-01-07\
43318 & 5.65 & -0.19 & 210, 212, 183F & 100.00, 100.00, 215.14 & 3 & 2012-01-11, 2012-01-11, 2012-01-07\
45067 & 5.90 & -0.16 & 230, 235 & 180.00, 180.00 & 2 & 2012-01-09, 2012-01-09\
45205 & 8.00 & -0.87 & 140, 177 & 1200.00, 1800.00 & 2 & 2012-01-09, 2012-01-09\
49933 & 5.78 & -0.47 & 216, 203 & 659.71, 135.0 & 2 & 2012-01-08, 2012-01-10\
52711 & 5.93 & -0.21 & 256, 163F & 160.00, 119.18 & 2 & 2012-01-10,2012-01-06\
58855 & 5.36 & -0.29 & 224, 167F & 150.00, 75.13 & 2 & 2012-01-09, 2012-01-06\
59374 & 8.50 & -0.88 & 122F, 131F & 1174.47, 812.31 & 2 & 2012-01-07, 2012-01-06\
59984 & 5.93 & -0.69 & 182F, 165F & 269.17, 229.17 & 2 & 2012-01-07, 2012-01-06\
62301 & 6.75 & -0.70 & 197, 158F & 450.00, 238.20 & 2 & 2012-01-09, 2012-01-06\
64090 & 8.30 & -1.73 & 280, 136F & 2500.00, 687.35 & 2 & 2012-12-26, 2012-01-06\
69897 & 5.10 & -0.25 & 186, 180F & 240.00, 66.18 & 2 & 2012-01-08, 2012-01-06\
74000 & 9.67 & -1.97 & 144, 142, 74F & 2400.00, 2400.00, 1454.51 & 3 & 2012-01-10, 2012-01-10, 2012-01-06\
76932 & 5.86 & -0.98 & 170, 181F, 210F & 612.32, 258.23, 489.24 & 3 & 2012-01-08, 2012-01-06, 2012-01-07\
82943 & 6.54 & 0.19 & 214, 170F & 270.00, 342.25 & 2 & 2012-01-10, 2012-01-06\
84937 & 8.28 & -2.16 & 122, 153, 95 & 1475.81, 1800.0, 846.41 & 3 & 2012-01-08, 2012-01-09, 2012-01-06\
89744 & 5.74 & 0.13 & 220, 165F & 135.00, 121.16 & 2 & 2012-01-10, 2012-01-06\
90839 & 4.83 & -0.18 & 190, 169F & 60.00, 137.20 & 2 & 2012-01-10, 2012-01-06\
92855 & 7.28 & -0.12 & 171, 163F & 600.00, 309.20 & 2 & 2012-01-10, 2012-01-06\
94028 & 8.23 & -1.47 & 118, 172 & 1800.00, 1317.69 & 2 & 2011-03-15, 2011-03-15\
99984 & 5.95 & -0.38 & 181F, 183F & 517.22, 491.24 & 2 & 2012-01-07, 2012-01-07\
100563 & 5.70 & 0.06 & 171F, 173F & 413.25, 383.21 & 2 & 2012-01-07, 2012-01-07\
102870 & 3.61 & 0.11 & 208F, 211F & 77.12, 91.13 & 2 & 2012-01-07, 2012-01-07\
103095 & 6.45 & -1.26 & 188 & 300.00 & 1 & 2012-01-10\
105755 & 8.59 & -0.73 & 175, 103F & 2400.0, 1052.44 & 2 & 2012-01-10, 2012-01-06\
106516 & 6.11 & -0.73 & 151F, 155F & 434.25, 546.25 & 2 & 2012-01-07, 2012-01-07\
108177 & 9.66 & -1.67 & 60, 111 & 3000.00, 3600.00 & 2 & 2011-03-15, 2011-03-15\
110897 & 6.00 & -0.57 & 260, 170F, 172F & 600.00, 568.30, 508.23 & 3 & 2012-01-08, 2012-01-07, 2012-01-07\
114710 & 4.26 & 0.06 & 184F, 186F & 112.13, 129.15 & 2 & 2012-01-07, 2012-01-07\
115617 & 4.74 & -0.10 & 177, 170 & 160.00, 160.00 & 2 & 2012-01-08, 2012-01-08\
134088 & 8.00 & -0.80 & 130 & 1800.00 & 1 & 2011-03-15\
134169 & 7.67 & -0.78 & 214, 212 & 750.00, 750.00 & 2 & 2012-01-10, 2012-01-10\
138776 & 8.72 & 0.24 & 211 & 1800.00 & 1 & 2012-01-11\
142091 & 4.82 & -0.07 & 277, 275 & 60.00, 60.00 & 2 & 2012-01-11, 2012-01-11\
142373 & 4.62 & -0.54 & 140, 147 & 240.00, 240.00 & 2 & 2012-01-08, 2012-01-08\
$-04^\circ$3208 & 9.99 & -2.20 & 60, 66 & 1500.00, 2400.00 & 2 & 2012-01-08, 2012-01-08\
$-13^\circ$3442 & 10.37 & -2.62 & 103, 102, 104, 102 & 2700.00, 2700.00, & 4 & 2012-01-09, 2012-01-09,\
& & & & 2700.00, 2700.00 & &2012-01-09, 2012-01-09\
+24$^\circ$1676 & 10.80 & -2.44 & 86, 90, 92, 92, 57F & 2400.00, 2400.00, 2400.00, & 5 & 2012-01-11, 2012-01-11, 2012-01-11,\
& & & & 2400.00, 1581.60 & & 2012-01-11, 2012-01-06\
+29$^\circ$2091 & 10.22 & -1.91 & 82, 121, 115, 120, 83F & 2400.00, 2700.00, 2700.00, & 5 & 2012-01-08, 2012-01-10, 2012-01-10,\
& & & & 2700.00, 1287.46 & & 2012-01-10, 2012-01-06\
+37$^\circ$1458 & 8.92 & -1.95 & 235, 105F & 2500.00, 897.38 & 2 & 2012-01-10, 2012-01-06\
+66$^\circ$0268 & 9.88 & -2.06 & 112, 110, 115 & 2400.00, 2400.00, 2400.00 & 3 & 2012-01-10, 2012-01-10, 2012-01-10\
G090-003 & 10.50 & -2.04 & 106, 120, 122, 66F & 2700.00, 2700.00, & 4 & 2012-01-11, 2012-01-11,\
& & & & 2700.00, 1181.44 & & 2012-01-11, 2012-01-06\
& & &\
+07$^\circ$4841 & 10.38 & -1.46 & 152 & 1470 & 2\
+09$^\circ$0352 & 10.17 & -2.09 & 160 & 2400 & 2\
& & &\
22879 & 6.74 & -0.84 & 200 & 900.00, 1800.00, 1800.00 & 3 & 1996-09-02, 1999-01-02, 1996-10-31\
34411 & 4.70 & 0.01 & 200 & 1500.00, 1500.00 & 2 & 1999-08-21, 1998-12-27\
84937 & 8.28 & -2.16 & 200 & 3600.00, 3300.00 & 2 & 1999-03-01, 1999-03-01\
94028 & 8.23 & -1.47 & 200 & 2400.00, 2400.00 & 2 & 2000-01-17, 2000-01-17\
103095 & 6.45 & -1.26 & 200 & 900.00 & 1 & 2000-05-19\
142373 & 4.62 & -0.54 & 200 & 600.00, 600.00 & 2 & 1998-06-09, 1998-06-09\
& & &\
49933$^2$ & 4.15 & -0.47 & 500 & & &\
140283$^3$ & 7.21 & -2.46 & 200 & & &\
$-04^\circ$3208$^3$ & 9.99 & -2.20 & 200 & & &\
[rrrrrrrrc]{} 19373& 10& -10& -1& -75.2& -16.1& 21.5& 0.00 0.03 & Thin disk\
22484& 10& 0& -9& 1.7& -15.5& -42.3& -0.05 0.02 & Thin disk\
22879& 20& 0& -17& -109.9& -88.2& -41.3& 0.30 0.02 & Thick disk\
24289&160& 30& -139& -116.7& -172.6& 173.9& 0.21 0.04 & Halo\
30562& 20& 10& -13& -54.9& -72.0& -21.6& 0.04 0.05 & Thin disk\
30743& 20& 20& -18& 24.8& -5.2& -22.5& 0.08 0.06 & Thin disk\
34411& 10& 0& 0& -74.5& -35.4& 4.4& -0.03 0.03 & Thin disk\
43318& 30& 20& -5& 47.0& 1.2& -37.0& -0.05 0.05 & Thin disk\
45067& 30& 20& -3& -16.5& -64.4& 13.7& 0.01 0.04 & Thin disk\
45205& 60& -30& 26& -35.9& -89.5& 17.0& 0.39 0.03 & Thick disk\
49933& 30& 20& 0& 25.7& -13.1& -9.2& 0.16 0.10 & Thin disk\
52711& 20& 0& 5& -18.6& -77.8& -9.0& 0.01 0.05 & Thin disk\
58855& 20& 0& 9& 25.3& -15.4& -3.9& 0.07 0.03 & Thin disk\
59374& 40& 20& 14& -48.5& -120.3& -9.8& 0.30 0.02 & Thick disk\
59984& 20& 20& 3& 8.0& -11.4& -21.2& 0.10 0.02 & Thin disk\
62301& 30& 0& 16& -7.4& -110.0& -22.1& 0.26 0.03 & Thick disk\
64090& 30& 0& 13& 266.9& -227.6& -89.6& 0.24 0.04 & Halo\
69897& 20& 0& 9& -23.9& -38.7& 7.1& -0.01 0.06 & Thin disk\
74000& 60& 110& 34& 216.8& -348.0& 62.2& 0.37 0.03 & Halo\
76932& 10& 20& 7& -49.1& -91.3& 70.0& 0.32 0.05 & Thick disk\
82943& 10& 20& 13& 13.3& -13.5& -12.6& -0.05 0.05 & Thin disk\
84937& 40& 30& 52& 205.0& -214.7& -7.9& 0.21 0.08 & Halo\
89744& 20& 0& 33& -10.6& -29.9& -14.3& -0.02 0.04 & Thin disk\
90839& 10& 0& 10& -13.9& -1.8& 2.4& 0.02 0.03 & Thin disk\
92855& 20& 0& 32& -42.1& -22.4& -13.9& -0.02 0.04 & Thin disk\
94028& 20& 10& 42& -33.7& -129.3& 13.5& 0.35 0.05 & Thick disk\
99984& 20& -10& 49& -6.4& 11.0& -36.1& 0.02 0.03 & Thin disk\
100563& 0& 10& 24& -14.5& -20.9& -9.8& -0.03 0.03 & Thin disk\
102870& 0& 10& 10& 40.4& 3.2& 7.3& -0.02 0.03 & Thin disk\
103095& 0& 0& 9& 278.3& -157.2& -14.8& 0.24 0.06 & Halo\
105755& 30& -30& 78& 26.7& 5.3& 8.3& 0.34 0.08 & Thin disk\
106516& 0& 10& 18& 53.4& -73.4& -57.7& 0.38 0.04 & Thick disk\
108177& -20& 50& 99& 126.7& -262.9& 32.0& 0.23 0.04 & Halo\
110897& 0& 0& 17& -41.7& 7.0& 75.3& 0.18 0.03 & Thin disk\
114710& 0& 0& 9& -49.9& 11.5& 8.4& -0.07 0.04 & Thin disk\
115617& 0& 0& 6& -23.3& -47.8& -31.3& -0.10 0.02 & Thin disk\
134088& -30& 0& 26& -25.6& -72.4& -68.7& 0.36 0.03 & Thick disk\
134169& -40& 0& 44& 2.5& -3.3& -1.9& 0.32 0.02 & Thin disk\
138776& -50& 0& 45& 6.4& -56.4& -5.0& 0.02 0.04 & Thin disk\
140283& -50& 0& 32& -249.0& -253.3& 41.1& 0.23 0.03 & Halo\
142091& -10& -20& 24& 33.6& -40.7& -18.1& 0.07 0.05 & Thin disk\
142373& 0& -10& 12& -41.6& 10.8& -68.3& 0.18 0.01 & Thin disk\
$-04^\circ$ 3208& -30& 110& 169& -123.5& -309.9& -120.2& 0.21 0.06 & Halo\
$-13^\circ$ 3442& & & & & & & 0.23 0.01 & Halo\
+07$^\circ$ 4841& -50& -150& -123& -224.4& -283.9& -55.3& 0.43 0.05 & Halo\
+09$^\circ$ 0352& 120& -40& -120& -156.5& -176.9& 97.7& 0.45 0.05 & Halo\
+24$^\circ$ 1676& 360& 90& 126& 166.8& -483.7& 106.4& 0.21 0.06 & Halo\
+29$^\circ$ 2091& 40& 20& 85& 157.3& -345.9& 88.4& 0.34 0.04 & Halo\
+37$^\circ$ 1458& 150& -10& 26& -280.3& -233.0& -26.0& 0.38 0.05 & Halo\
+66$^\circ$ 0268 & 40& -40& 10& -181.0& -426.4& -73.1& 0.15 0.07 & Halo\
G090-003 & 820& 90& 332& 10.4& -824.6& -192.4& 0.25 0.05 & Halo\
[rcrrccr]{} Fe I & & & & & &\
4920.50 & 2.83 & 0.07 & -30.51 & -4.61 & -4.61 & 466.0\
5198.72 & 2.22 & -2.14 & -31.32 & -4.52 & -4.53 & 102.0\
5217.40 & 3.21 & -1.07 & -30.37 & -4.57 & -4.57 & 130.3\
5232.94 & 2.94 & -0.06 & -30.54 & -4.68 & -4.68 & 371.1\
5236.20 & 4.19 & -1.50 & -31.32 & -4.68 & -4.69 & 33.7\
5242.50 & 3.63 & -0.97 & -31.56 & -4.47 & -4.47 & 93.4\
5281.79 & 3.04 & -0.83 & -30.53 & -4.68 & -4.69 & 163.6\
5324.18 & 3.21 & -0.10 & -30.42 & -4.54 & -4.54 & 332.5\
5367.47 & 4.41 & 0.44 & -30.20 & -4.78 & -4.78 & 170.1\
5379.58 & 3.69 & -1.51 & -31.56 & -4.46 & -4.47 & 64.1\
5383.37 & 4.31 & 0.64 & -30.37 & -4.72 & -4.72 & 220.5\
5393.17 & 3.24 & -0.72 & -30.42 & -4.60 & -4.60 & 171.2\
5491.83 & 4.19 & -2.19 & -31.33 & -4.53 & -4.54 & 14.5\
5586.76 & 3.37 & -0.10 & -30.38 & -4.58 & -4.58 & 289.3\
5662.52 & 4.18 & -0.57 & -30.52 & -4.41 & -4.42 & 105.7\
5696.09 & 4.55 & -1.72 & -30.21 & -4.68 & -4.69 & 14.6\
5705.46 & 4.30 & -1.36 & -30.47 & -4.61 & -4.62 & 41.2\
5778.45 & 2.59 & -3.44 & -31.37 & -4.60 & -4.61 & 22.7\
5855.08 & 4.61 & -1.48 & -30.21 & -4.58 & -4.59 & 24.4\
5916.25 & 2.45 & -2.99 & -31.45 & -4.43 & -4.44 & 56.7\
6065.49 & 2.61 & -1.53 & -31.41 & -4.51 & -4.52 & 132.1\
6082.71 & 2.22 & -3.57 & -31.74 & -4.55 & -4.56 & 34.7\
6151.62 & 2.18 & -3.30 & -31.58 & -4.52 & -4.53 & 51.1\
6200.32 & 2.61 & -2.44 & -31.43 & -4.47 & -4.48 & 74.6\
6213.43 & 2.22 & -2.48 & -31.58 & -4.58 & -4.59 & 86.0\
6229.23 & 2.85 & -2.80 & -31.32 & -4.65 & -4.66 & 37.8\
6252.55 & 2.40 & -1.69 & -31.52 & -4.52 & -4.53 & 135.4\
6393.61 & 2.43 & -1.43 & -31.53 & -4.63 & -4.64 & 150.8\
6411.65 & 3.65 & -0.60 & -30.38 & -4.57 & -4.57 & 154.7\
6421.35 & 2.28 & -2.03 & -31.80 & -4.56 & -4.57 & 112.7\
6518.37 & 2.83 & -2.46 & -31.37 & -4.59 & -4.60 & 64.0\
Fe II & & & & & &\
4233.17 & 2.58 & -1.97 & -32.01 & -4.64 & -4.64 & 132.1\
4508.29 & 2.84 & -2.44 & -32.00 & -4.48 & -4.48 & 93.0\
4582.83 & 2.83 & -3.18 & -32.03 & -4.54 & -4.54 & 57.7\
4620.52 & 2.82 & -3.21 & -32.02 & -4.63 & -4.63 & 53.9\
4923.93 & 2.88 & -1.26 & -32.03 & -4.69 & -4.69 & 196.1\
5018.44 & 2.88 & -1.10 & -32.04 & -4.72 & -4.72 & 219.1\
5197.58 & 3.22 & -2.22 & -32.02 & -4.58 & -4.58 & 85.4\
5264.81 & 3.22 & -3.13 & -32.01 & -4.48 & -4.48 & 49.2\
5284.11 & 2.88 & -3.11 & -32.04 & -4.57 & -4.57 & 60.3\
5325.55 & 3.21 & -3.16 & -32.03 & -4.55 & -4.55 & 45.2\
5414.07 & 3.21 & -3.58 & -32.02 & -4.52 & -4.52 & 29.8\
5425.26 & 3.20 & -3.22 & -32.04 & -4.56 & -4.56 & 43.4\
5991.38 & 3.15 & -3.55 & -32.05 & -4.58 & -4.58 & 32.1\
6239.95 & 3.89 & -3.46 & -32.00 & -4.55 & -4.55 & 13.2\
6247.56 & 3.89 & -2.30 & -32.00 & -4.58 & -4.58 & 55.0\
6369.46 & 2.89 & -4.11 & -32.06 & -4.59 & -4.59 & 20.7\
6432.68 & 2.89 & -3.57 & -32.07 & -4.56 & -4.56 & 43.2\
6456.38 & 3.90 & -2.05 & -32.00 & -4.59 & -4.59 & 66.9\
[rllrrrrrrr]{}\
19373 & 6045 & 4.24 & 0.10 & 1.2 & 1.11 & 0.06 & 0.06 & 26 & 15\
22484 & 6000 & 4.07 & 0.01 & 1.1 & 1.18 & -0.02 & 0.00 & 27 & 16\
22879 & 5800 & 4.29 & -0.84 & 1.0 & 0.75 & -0.06 & -0.03 & 23 & 14\
30562 & 5900 & 4.08 & 0.17 & 1.3 & 1.12 & 0.02 & 0.02 & 26 & 16\
34411 & 5850 & 4.23 & 0.01 & 1.2 & 1.10 & -0.01 & -0.02 & 26 & 15\
49933 & 6600 & 4.15 & -0.47 & 1.7 & 1.11 & -0.05 & -0.04 & 26 & 15\
59374 & 5850 & 4.38 & -0.88 & 1.2 & 0.75 & -0.06 & -0.02 & 24 & 15\
59984 & 5930 & 4.02 & -0.69 & 1.4 & 0.89 & -0.06 & -0.06 & 23 & 18\
64090 & 5400 & 4.70 & -1.73 & 0.7 & 0.62 & -0.02 & -0.02 & 22 & 11\
69897 & 6240 & 4.24 & -0.25 & 1.4 & 1.15 & -0.06 & -0.04 & 28 & 15\
84937 & 6350 & 4.09 & -2.12 & 1.7 & 0.76 & -0.06 & 0.00 & 12 & 7\
94028 & 5970 & 4.33 & -1.47 & 1.3 & 0.70 & -0.06 & -0.04 & 20 & 15\
102870 & 6170 & 4.14 & 0.11 & 1.5 & 1.35 & -0.03 & -0.01 & 23 & 14\
103095 & 5130 & 4.66 & -1.26 & 0.9 & 0.60 & 0.01 & 0.01 & 22 & 8\
105755 & 5800 & 4.05 & -0.73 & 1.2 & 0.85 & -0.01 & 0.00 & 23 & 15\
114710 & 6090 & 4.47 & 0.06 & 1.1 & 1.17 & 0.03 & 0.04 & 26 & 15\
134169 & 5890 & 4.02 & -0.78 & 1.2 & 0.87 & 0.01 & 0.06 & 26 & 18\
140283 & 5780 & 3.70 & -2.46 & 1.6 & 0.80 & -0.06 & -0.02 & 20 & 14\
142091 & 4810 & 3.12 & -0.07 & 1.2 & 1.16 & 0.05 & 0.05 & 20 & 13\
+66$^\circ$ 0268 & 5300 & 4.72 & -2.06 & 0.6 & 0.60 & -0.03 & -0.03 & 21 & 9\
\
24289 & 5980 & 3.71 & -1.94 & 1.1 & 0.83 & -0.10 & -0.03 & 16 & 10\
30743 & 6450 & 4.20 & -0.44 & 1.8 & 1.08 & -0.03 & 0.01 & 26 & 18\
43318 & 6250 & 3.92 & -0.19 & 1.7 & 1.26 & -0.05 & -0.01 & 28 & 15\
45067 & 5960 & 3.94 & -0.16 & 1.5 & 1.18 & -0.03 & -0.01 & 24 & 16\
45205 & 5790 & 4.08 & -0.87 & 1.1 & 0.82 & -0.08 & -0.05 & 23 & 16\
52711 & 5900 & 4.33 & -0.21 & 1.2 & 0.97 & 0.03 & 0.05 & 26 & 16\
58855 & 6410 & 4.32 & -0.29 & 1.6 & 1.15 & -0.01 & 0.01 & 26 & 15\
62301 & 5840 & 4.09 & -0.70 & 1.3 & 0.85 & -0.07 & -0.04 & 29 & 16\
74000 & 6225 & 4.13 & -1.97 & 1.3 & 0.76 & -0.08 & -0.02 & 15 & 7\
76932 & 5870 & 4.10 & -0.98 & 1.3 & & -0.01 & 0.03 & 27 & 16\
82943 & 5970 & 4.37 & 0.19 & 1.2 & 1.17 & 0.01 & 0.01 & 25 & 13\
89744 & 6280 & 3.97 & 0.13 & 1.7 & 1.50 & 0.01 & 0.02 & 26 & 13\
90839 & 6195 & 4.38 & -0.18 & 1.4 & 1.10 & 0.04 & 0.06 & 28 & 17\
92855 & 6020 & 4.36 & -0.12 & 1.3 & 1.08 & 0.00 & 0.02 & 24 & 11\
99984 & 6190 & 3.72 & -0.38 & 1.8 & 1.33 & -0.01 & 0.03 & 24 & 15\
100563 & 6460 & 4.32 & 0.06 & 1.6 & 1.30 & 0.00 & 0.02 & 23 & 10\
106516 & 6300 & 4.44 & -0.73 & 1.5 & 0.95 & -0.03 & 0.00 & 22 & 15\
108177 & 6100 & 4.22 & -1.67 & 1.1 & 0.72 & -0.08 & -0.06 & 16 & 5\
110897 & 5920 & 4.41 & -0.57 & 1.2 & 0.85 & 0.03 & 0.05 & 29 & 18\
115617 & 5490 & 4.40 & -0.10 & 1.1 & 0.93 & -0.04 & -0.05 & 26 & 13\
134088 & 5730 & 4.46 & -0.80 & 1.1 & 0.75 & -0.02 & 0.00 & 25 & 15\
138776 & 5650 & 4.30 & 0.24 & 1.3 & 0.93 & 0.02 & 0.00 & 21 & 11\
142373 & 5830 & 3.96 & -0.54 & 1.4 & 0.95 & -0.02 & -0.02 & 28 & 17\
–4$^\circ$ 3208 & 6390 & 4.08 & -2.20 & 1.4 & 0.77 & -0.08 & 0.00 & 16 & 11\
–13$^\circ$ 3442 & 6400 & 3.95 & -2.62 & 1.4 & 0.79 & -0.14 & 0.00 & 8 & 5\
+7$^\circ$ 4841 & 6130 & 4.15 & -1.46 & 1.3 & 0.79 & 0.00 & 0.02 & 21 & 17\
+9$^\circ$ 0352 & 6150 & 4.25 & -2.09 & 1.3 & 0.70 & -0.01 & 0.03 & 15 & 6\
+24$^\circ$ 1676 & 6210 & 3.90 & -2.44 & 1.5 & 0.78 & -0.06 & 0.04 & 12 & 5\
+29$^\circ$ 2091 & 5860 & 4.67 & -1.91 & 0.8 & 0.70 & -0.05 & -0.05 & 18 & 10\
+37$^\circ$ 1458 & 5500 & 3.70 & -1.95 & 1.0 & & -0.07 & -0.05 & 21 & 14\
G090-003 & 6010 & 3.90 & -2.04 & 1.3 & 0.78 & -0.06 & -0.02 & 17 & 12\
[95]{} natexlab\#1[\#1]{}
, V. Z., [et al.]{} 2013, , 554, A44
, S., & [Mashonkina]{}, L. 2015, MNRAS, 00, 0, in preparation
, C., [Barklem]{}, P. S., [Lambert]{}, D. L., & [Cunha]{}, K. 2004, , 420, 183
, A., [Arribas]{}, S., & [Martinez-Roger]{}, C. 1995, , 297, 197
—. 1996, , 117, 227
—. 1996, , 313, 873
, M., [Fuhrmann]{}, K., & [Gehren]{}, T. 1994, , 291, 895
, S., [Jehin]{}, E., [Ledoux]{}, C., [Cabanac]{}, R., [Melo]{}, C., [Gilmozzi]{}, R., & [ESO Paranal Science Operations Team]{}. 2003, The Messenger, 114, 10
, A., [Kock]{}, A., & [Kock]{}, M. 1991, , 248, 315
, P. S., & [Aspelund-Johansson]{}, J. 2005, , 435, 373
, P. S., [Piskunov]{}, N., & [O’Mara]{}, B. J. 2000, Astron. and Astrophys. Suppl. Ser., 142, 467
, T., [Alves-Brito]{}, A., [Oey]{}, M. S., [Yong]{}, D., & [Mel[é]{}ndez]{}, J. 2010, , 516, L13
, T., [Feltzing]{}, S., & [Oey]{}, M. S. 2014, , 562, A71
, M., [Lind]{}, K., [Collet]{}, R., [Magic]{}, Z., & [Asplund]{}, M. 2012, , 427, 27
, D. E., [Ibbetson]{}, P. A., [Petford]{}, A. D., & [Shallis]{}, M. J. 1979, , 186, 633
, D. E., [Petford]{}, A. D., [Shallis]{}, M. J., & [Simmons]{}, G. J. 1982, , 199, 43
, D. E., [Petford]{}, A. D., & [Simmons]{}, G. J. 1982, , 201, 595
, P., [et al.]{} 2009, , 501, 519
, T. S., [et al.]{} 2012, , 746, 101
—. 2013, , 771, 40
, K., & [Giddings]{}, J. 1985, Newsletter on the analysis of astronomical spectra, No. 9, University of London
, B. W., [Latham]{}, D. W., [Laird]{}, J. B., [Grant]{}, C. E., & [Morse]{}, J. A. 2001, , 122, 3419
, L., [Ram[í]{}rez]{}, I., [Mel[é]{}ndez]{}, J., [Bessell]{}, M., & [Asplund]{}, M. 2010, , 512, A54
, L., [Sch[ö]{}nrich]{}, R., [Asplund]{}, M., [Cassisi]{}, S., [Ram[í]{}rez]{}, I., [Mel[é]{}ndez]{}, J., [Bensby]{}, T., & [Feltzing]{}, S. 2011, , 530, A138
, L., [et al.]{} 2014, , 439, 2060
, R., [van’t Veer-Menneret]{}, C., [Allard]{}, N. F., & [Stehl[é]{}]{}, C. 2011, , 531, A83
, R., [et al.]{} 2004, , 416, 1117
, G., [Soubiran]{}, C., & [Ralite]{}, N. 2001, , 373, 159
, Y. Q., [Nissen]{}, P. E., & [Zhao]{}, G. 2004, , 425, 697
, Y. Q., [Nissen]{}, P. E., [Zhao]{}, G., [Zhang]{}, H. W., & [Benoni]{}, T. 2000, , 141, 491
, C., [Matteucci]{}, F., & [Romano]{}, D. 2001, , 554, 1044
, R., [Asplund]{}, M., & [Trampedach]{}, R. 2007, , 469, 687
, O. L., [et al.]{} 2012, , 545, A17
—. 2013, , 431, 2419
—. 2015, , 575, A26
, G. M., [et al.]{} 2015, ArXiv e-prints
, W., & [Binney]{}, J. J. 1998, , 298, 387
, L.-C., [et al.]{} 2012, Research in Astronomy and Astrophysics, 12, 735
, V., [Ku[č]{}inskas]{}, A., [Steffen]{}, M., [Ludwig]{}, H.-G., [Prakapavi[č]{}ius]{}, D., [Klevas]{}, J., [Caffau]{}, E., & [Bonifacio]{}, P. 2013, , 559, A102
, A., [Chaboyer]{}, B., [Jevremovi[ć]{}]{}, D., [Kostov]{}, V., [Baron]{}, E., & [Ferguson]{}, J. W. 2008, , 178, 89
, H.-W. 1968, Zeitschrift fur Physik, 211, 404
, H. W. 1969, Zeitschrift fur Physik, 225, 483
, B., [Andersen]{}, J., [Gustafsson]{}, B., [Lambert]{}, D. L., [Nissen]{}, P. E., & [Tomkin]{}, J. 1993, , 275, 101
, A., [Casey]{}, A. R., [Jacobson]{}, H. R., & [Yu]{}, Q. 2013, , 769, 57
, K. 1998, , 338, 161
—. 2000, http://www.ing.iac.es/ klaus/ (Paper II)
—. 2004, Astronomische Nachrichten, 325, 3
, T., [Shi]{}, J. R., [Zhang]{}, H. W., [Zhao]{}, G., & [Korn]{}, A. J. 2006, , 451, 1065
, G., [et al.]{} 2012, The Messenger, 147, 25
, J. I., & [Bonifacio]{}, P. 2009, , 497, 497
, R., [Carretta]{}, E., [Matteucci]{}, F., & [Sneden]{}, C. 1996, in Astronomical Society of the Pacific Conference Series, Vol. 92, Formation of the Galactic Halo...Inside and Out, ed. H. L. [Morrison]{} & A. [Sarajedini]{}, 307
Gustafsson, B., Edvardsson, B., Eriksson, K., Jorgensen, U. G., Nordlund, [Å]{}., & Plez, B. 2008, A&A, 486, 951
, D. R. H., & [Soderblom]{}, D. R. 1987, , 93, 864
, T., [Gruberbauer]{}, M., [Guenther]{}, D. B., [Fossati]{}, L., & [Weiss]{}, W. W. 2010, , 510, A106
, C., [Karakas]{}, A. I., & [Umeda]{}, H. 2011, , 414, 3231
, O. 2010, http://www.astro.uu.se/ oleg/binmag.html
, R. L. 2007, Robert L. Kurucz on-line database of observed and predicted atomic transitions
, R. L., [Furenlid]{}, I., [Brault]{}, J., & [Testerman]{}, L. 1984, [Solar flux atlas from 296 to 1300 nm]{} (New Mexico: National Solar Observatory)
, D. K., [Bolte]{}, M., [Johnson]{}, J. A., [Lucatello]{}, S., [Heger]{}, A., & [Woosley]{}, S. E. 2008, , 681, 1524
, K., [Bergemann]{}, M., & [Asplund]{}, M. 2012, , 427, 50
, Z. M. 2013, Astronomy Reports, 57, 128
, L., & [Gehren]{}, T. 2000, , 364, 249
, L., [Gehren]{}, T., [Shi]{}, J.-R., [Korn]{}, A. J., & [Grupp]{}, F. 2011, , 528, A87
, L., [Gehren]{}, T., [Travaglio]{}, C., & [Borkova]{}, T. 2003, , 397, 275
, L., [et al.]{} 2008, , 478, 529
, A., [Preston]{}, G. W., [Sneden]{}, C., & [Searle]{}, L. 1995, , 109, 2757
, J., & [Barbuy]{}, B. 2009, , 497, 611
, T. V., [Soubiran]{}, C., [Kovtyukh]{}, V. V., & [Korotin]{}, S. A. 2004, , 418, 551
, J. R., [et al.]{} 2009, , 393, 245
, P. E., 1981, , 97, 145
, T. R., [Wickliffe]{}, M. E., [Lawler]{}, J. E., [Whaling]{}, W., & [Brault]{}, J. W. 1991, Journal of the Optical Society of America B Optical Physics, 8, 1185
, E. H. 1983, , 54, 55
—. 1993, , 102, 89
, Y. V., & [Zhao]{}, G. 2013, , 146, 97
, A. J. J., & [Uylings]{}, P. H. M. 1998, VizieR Online Data Catalog, 334, 300
, I., [Allende Prieto]{}, C., & [Lambert]{}, D. L. 2013, , 764, 78
, I., & [Mel[é]{}ndez]{}, J. 2005, , 626, 446
, S., [Gilmore]{}, G., & [Gaia-ESO Consortium]{}. 2013, The Messenger, 154, 47
, B. E., [Tomkin]{}, J., [Lambert]{}, D. L., & [Allende Prieto]{}, C. 2003, , 340, 304
Reetz, J. K. 1991, Diploma Thesis (Universität München)
, D., [Karakas]{}, A. I., [Tosi]{}, M., & [Matteucci]{}, F. 2010, , 522, A32
, G. R., [Bergemann]{}, M., [Serenelli]{}, A., [Casagrande]{}, L., & [Lind]{}, K. 2013, , 429, 126
, T., [Fossati]{}, L., & [Shulyak]{}, D. 2009, , 506, 203
, M. F., [et al.]{} 2006, , 131, 1163
, R., [et al.]{} 2014, , 570, A122
, W., & [Holweger]{}, H. 1984, , 130, 319
, Y. 1994, , 46, 53
, Y., [Ohkubo]{}, M., [Sato]{}, B., [Kambe]{}, E., & [Sadakane]{}, K. 2005, , 57, 27
, V. 1996, in Astronomical Society of the Pacific Conference Series, Vol. 108, M.A.S.S., Model Atmospheres and Spectrum Synthesis, ed. S. J. [Adelman]{}, F. [Kupka]{}, & W. W. [Weiss]{}, 198
, C., [et al.]{} 1993, Bulletin d’Information du Centre de Donnees Stellaires, 43, 5
, F. 2007, , 474, 653
, K., [et al.]{} 2014, , 438, 2413
, S. K., [Demarque]{}, P., & [Kim]{}, Y.-C. 2004, , 291, 261
, G., [et al.]{} 2013, , 146, 81
, H. W., & [Zhao]{}, G. 2005, , 364, 712
—. 2006, , 449, 127
[^1]: Based on observations collected at the UCO/Lick observatory, USA and Canada-France-Hawaii Telescope
[^2]: In the classical notation, where \[X/H\] = $\log(N_{\rm X}/N_{\rm H})_{star} - \log(N_{\rm X}/N_{\rm H})_{Sun}$.
[^3]: [http://sci.esa.int/gaia/]{}
[^4]: [http://www.cfht.hawaii.edu/Instruments/Spectroscopy/Espadons]{}
[^5]: [http://simbad.u-strasbg.fr/simbad/]{}
[^6]: http://marcs.astro.uu.se
[^7]: [http://marcs.astro.uu.se/software.php]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have detected strong dichroism in the Ni $L_{2,3}$ x-ray absorption spectra of monolayer NiO films. The dichroic signal appears to be very similar to the magnetic linear dichroism observed for thicker antiferromagnetic NiO films. A detailed experimental and theoretical analysis reveals, however, that the dichroism is caused by crystal field effects in the monolayer films, which is a non trivial effect because the high spin Ni $3d^{8}$ ground state is not split by low symmetry crystal fields. We present a practical experimental method for identifying the independent magnetic and crystal field contributions to the linear dichroic signal in spectra of NiO films with arbitrary thicknesses and lattice strains. Our findings are also directly relevant for high spin $3d^{5}$ and $3d^{3}$ systems such as LaFeO$_{3}$, Fe$_{2}$O$_{3}$, VO, LaCrO$_{3}$, Cr$_{2}$O$_{3}$, and Mn$^{4+}$ manganate thin films.'
author:
- 'M. W. Haverkort'
- 'S. I. Csiszar'
- 'Z. Hu'
- 'S. Altieri'
- 'A. Tanaka'
- 'H. H. Hsieh'
- 'H.-J. Lin'
- 'C. T. Chen'
- 'T. Hibma'
- 'L. H. Tjeng'
title: Magnetic versus crystal field linear dichroism in NiO thin films
---
Magnetic linear dichroism (MLD) in soft-x-ray absorption spectroscopy (XAS) has recently developed into one of the most powerful tools to study the magnetic properties of antiferromagnetic thin films [@Thole85; @Sinkovic90; @Kuiper93; @Alders95; @Alders98]. The contrast that one can obtain as a result of differences in the magnitude and orientation of local moments is essential to determine the spin anisotropy and important parameters like the Néel temperature $(T_{N})$, as well as to map out spatially the different magnetic domains that are present in antiferromagnetic films [@Stohr98; @Spanke98; @Stohr99; @Scholl00; @Nolting00; @Ohldag01a; @Ohldag01b; @Zhu01; @Hille01]. Such information is extremely valuable for the research and application of magnetic devices that make use of exchange bias.
![Energy level diagram for Ni$^{2+}$ ($3d^{8}$) in: (a) $O_{h}$ symmetry with $pd\sigma$=-1.29 and $10Dq$=0.85 eV; (b) $O_{h}$ symmetry with additional exchange field of 0.16 eV; (c) $D_{4h}$ symmetry with $pd\sigma$=-1.29, $10Dq$=0.85, and $Ds$=0.12 eV; (d) $D_{4h}$ symmetry with additional exchange. The $3d$ spin-orbit interaction is included, but the states are labelled as if the spin-orbit interaction was not present[]{data-label="fig1"}](fig1.eps){width="45.00000%"}
Much of the modern MLD work has been focussed on NiO and LaFeO$_{3}$ thin films, and the observed dichroism has been attributed entirely to magnetic effects [@Stohr98; @Spanke98; @Stohr99; @Scholl00; @Nolting00; @Ohldag01a; @Ohldag01b; @Zhu01; @Hille01]. Other sources that could contribute to linear dichroism, however, such as crystal fields of lower than octahedral symmetry, have been neglected or not considered. Indeed, one would expect that such low symmetry crystal fields are negligible for bulk-like NiO and LaFeO$_{3}$ films, and, more fundamentally, that such crystal fields will not split the high-spin Ni $3d^{8}$ or Fe $3d^{5}$ ground state. We have illustrated this insensitivity in Fig. 1 for the Ni$^{2+}$ case, where the energy level diagram in an $O_{h}$ environment is compared to that in a $D_{4h}$ point group symmetry [@SOremark]. In contrast, an exchange field will split the Ni$^{2+}$ ground state into three levels with $M_{S}$=-1,0,1 with an energy separation given by the exchange coupling $J$, see Fig. 1. The basis for obtaining strong dichroism in the Ni $L_{2,3}$ ($2p$$\rightarrow$$3d$) absorption spectra is that dipole selection rules dictate which of the quite different final states can be reach and with what probability for each of the initial states. The isotropic spectrum of each of these three states will be the same, but each state with a different $|M_{S}|$ value will have a different polarization dependence [@Thole85; @Sinkovic90; @Kuiper93; @Alders95; @Alders98]. A completely analogous argumentation can be given for the orbitally highly symmetric high spin $3d^{5}$ and $3d^{3}$ cases, e.g. Mn$^{2+}$, Fe$^{3+}$, V$^{2+}$, Cr$^{3+}$, Mn$^{4+}$.
In this paper we report on XAS measuments on single monolayer (ML) NiO films which are grown on a Ag(100) substrate and capped by a 10 ML MgO(100) film. We have observed strong linear dichroism in the Ni $L_{2,3}$ spectra, very similar to that measured for thicker NiO films. From a detailed theoretical and experimental analysis, however, we discovered that the dichroism can not be attributed to the presence of some form of magnetic order, but entirely to crystal field effects. The analysis provides us also with a practical guide of how to disentangle quantitatively the individual contributions to the linear dichroic signal, i.e. the contribution from magnetic interactions versus that from low symmetry crystal fields. This is important for a reliable determination of, for instance, the spin moment orientation in NiO as well as LaFeO$_{3}$, Fe$_{2}$O$_{3}$, VO, LaCrO$_{3}$, Cr$_{2}$O$_{3}$, and Mn$^{4+}$ manganate ultra thin films, surfaces and strained films, where the low symmetry crystal field splittings may not be negligible as compared to the exchange field energies.
The polarization dependent XAS measurements were performed at the Dragon beamline of the National Synchrotron Radiation Research Center in Taiwan. The spectra were recorded using the total electron yield method in an XAS chamber with a base pressure of 3x10$^{-10}$ mbar. The photon energy resolution at the Ni $L_{2,3}$ edges ($h\nu \approx 850-880$ eV) was set at 0.3 eV, and the degree of linear polarization was $\approx 98 \%$. A NiO single crystal is measured *simultaneously* in a separate chamber upstream of the XAS chamber in order to obtain a relative energy reference with an accuracy of better than 0.02 eV. The 1 ML NiO film on Ag(100) was prepared in Groningen, by using NO$_{2}$ assisted molecular beam epitaxy. Immediately after the NiO growth, the sample was capped *in-situ* with an epitaxial 10 ML MgO(100) film. Reflection high energy electron diffraction (RHEED) intensity oscillations recorded during growth of thicker films demonstrated the layer-by-layer growth mode and provided an accurate thickness calibration [@Altierithesis; @Altieri99].
![Experimental polarization dependent Ni $L_{2,3}$ XAS of 1 ML NiO(100) on Ag(100) covered with MgO(100). $\theta$ is the angle between the light polarization vector and the (001) surface normal ($\theta$=$90^{\circ}$ means normal light incidence). []{data-label="fig2"}](fig2.eps){width="45.00000%"}
Fig. 2 shows the polarization dependent Ni $L_{2,3}$ XAS spectra of the 1 ML NiO film, taken at room temperature. The angle between the light polarization vector and the (001) surface normal is given by $\theta$ ($\theta = 90^{\circ}$ means normal light incidence). The general lineshape of the spectra is very similar to that of thicker NiO films and bulk NiO [@Alders98].
{width="45.00000%"}
Fig. 3 presents a close-up of the $L_{2}$ edge, the region most often used to measure the magnitude of the magnetic linear dichroic effect in antiferromagnetic NiO films [@Stohr98; @Spanke98; @Stohr99; @Ohldag01a; @Ohldag01b; @Zhu01; @Hille01]. The spectra of the 1 ML NiO film show a very clear polarization dependence. This linear dichroic effect is as strong as that for a 20 ML NiO film grown on MgO(100) (taken from Alders *et al.* [@Alders98]), albeit with an opposite sign, as can be seen from Fig. 3. Relying on the analysis by Alders *et al.* [@Alders98] for the antiferromagnetic 20 ML film, one may be tempted to conclude directly that the spin orientation in the 1 ML film is quite different to that of the 20 ML film, i.e. that the spins for the 1 ML would be lying more parallel to the interface while those of the thicker films are pointing more along the interface normal. However, Alders *et al.* [@Alders98] have also shown that the magnetic ordering temperature of NiO films decreases strongly if the film is made thinner. In fact, for a 5 ML NiO film on MgO(100), it was found that $T_{N}$ is around or below room temperature, i.e. that no linear dichroism can be observed at room temperature. A simple extrapolation will therefore suggest that 1 ML NiO will not be magnetically ordered at room temperature. This is in fact supported by the 80 K data of the 1 ML NiO on Ag(100) as shown in Figs. 2 and 3: the spectra and the dichroism therein are identical to those at 300 K, indicating that $T_{N}$ must be at least lower than 80 K.
![Theoretical and experimental polarization dependence of the Ni $L_{3}$-XAS of a 1 ML NiO(100) on Ag(100) covered with MgO(100). The theoretical spectra are calculated in $D_{4h}$ symmetry without exchange.[]{data-label="fig4"}](fig4.eps){width="45.00000%"}
In order to resolve the origin of the linear dichroism in the 1 ML NiO system, we now resort to the Ni $L_{3}$ part of the spectrum. A close-up of this region is given in Fig. 4. We can easily observe that the strong polarization dependence of the spectra is accompanied by an energy shift $\Delta E$ of 0.35 eV in the main peak of the $L_{3}$ white line. This shift seems small compared to the 852 eV photon energy being used, but it is very reproducible and well detectable since the photon energy calibration is done with an accuracy of better than 0.02 eV thanks to the simultaneous measurement of a NiO single crystal reference. We now take this energy shift as an indicator for the presence and strength of local crystal fields with a symmetry lower than $O_{h}$, i.e. crystal fields that do not split the ground state but do alter the energies of the XAS final states, and, via second order processes, also causes spectral weight to be transferred between the various peaks as we will show below.
To understand the Ni $L_{2,3}$ spectra quantitatively, we perform calculations for the atomic $2p^{6}3d^{8} \rightarrow 2p^{5}3d^{9}$ transitions using the same method as described earlier by Alders *et al.* [@Alders98], but now in a $D_{4h}$ point group symmetry. The method uses the full atomic multiplet theory and includes the effects of the solid. It accounts for the intra-atomic $3d$-$3d$ and $2p$-$3d$ Coulomb and exchange interactions, the atomic $2p$ and $3d$ spin-orbit couplings, the O $2p$ - Ni $3d$ hybridization with $pd\sigma$ = -1.29 eV, and an $O_{h}$ crystal field splitting of $10Dq$ = 0.85 eV. The local symmetry for the Ni ion sandwiched between the Ag(100) substrate and the MgO(100) film is in principle $C_{4v}$, but for $d$ electrons one can ignore the odd part of the crystal field, so that effectively one can use the tetragonal $D_{4h}$ point group symmetry. As we will explain below, the $D_{4h}$ parameters $Ds$ and $Dt$ [@Ballhausen62] are set to 0.12 and 0.00 eV, respectively, and the exchange field (the molecular field acting on the spins) to zero. The calculations have been carried out using the XTLS 8.0 programm[@Tanaka94].
The right panel of Fig. 4 shows the calculated $L_{3}$ spectrum for the light polarization vector perpendicular and parallel to the $C_{4}$ axis ($\theta =
90^{\circ}$ and $\theta = 0^{\circ}$, respectively). One can clearly see that the major experimental features are well reproduced, including the 0.35 eV energy shift between the two polarizations. This shift can be understood in a single electron picture. The ground state has the $\underline{3d}_{x^{2}-y^{2}}\underline{3d}_{z^{2}}$ configuration, where the underline denotes a hole. The final state has a $2p^{5}\underline{3d}_{x^{2}-y^{2}}$ or $2p^{5}\underline{3d}_{z^{2}}$ configuration. For $z$ polarized light the $3d_{z^{2}}$ state can be reached, but the $3d_{x^{2}-y^{2}}$ can not, and the final state will be of the form $2p^{5}\underline{3d}_{x^{2}-y^{2}}$. For $x$ polarized light the final state will be of the form $2p^{5}\underline{3d}_{z^{2}-y^{2}}$$=
$$\sqrt{3/4}(2p^{5}\underline{3d}_{z^{2}})$$+
$$\sqrt{1/4}(2p^{5}\underline{3d}_{x^{2}-y^{2}})$. In a pure ionic picture, the $2p^{5}\underline{3d}_{x^{2}-y^{2}}$ state will be 4$Ds$+5$Dt$ lower in energy than the $2p^{5}\underline{3d}_{z^{2}}$ state. In the presence of the O $2p$ - Ni $3d$ hybridization, we find that $Ds$=0.12 and $Dt$=0.00 eV reproduce the observed 0.35 eV shift.
Going back to the $L_{2}$ edge, we can see in Fig. 3 that the calculations can also reproduce very well the observed linear dichroism in the 1 ML NiO spectra. In fact, one now could also see the same 0.35 eV shift at this edge, although it is not as clear as in the $L_{3}$ edge. We would like to stress here that the good agreement has been achieved without the inclusion of an exchange splitting, i.e. the dichroism is solely due to the low symmetry crystal field splitting. It is a final state effect and the change in the ratio between the two peaks of the $L_{2}$ edge as a function of polarization can be understood as follows. In $O_{h}$ symmetry the first peak is due to two final states, one of $T_{2}'$ and one of $E_{1}'$ symmetry. The second peak is due to a final state of $T_{1}'$ symmetry. All three states have the $2p^{5}\underline{3d}_{{e}_{g}}$ configuration. If one reduces the crystal field to $D_{4h}$ symmetry the peaks will split. The $T_{2}'$ state will split into two states of $B_{2}'$ and $E_{1}'$ symmetry, the $E_{1}'$ into $A_{1}'$ and $B_{1}'$, and the $T_{1}'$ into $A_{2}'$ and $E_{1}'$. The energy splitting can be measured, but this is much easier done using the $L_{3}$ edge. We note that each of the two peaks in the L$_{2}$ edge will have a state of $E_{1}'$ symmetry, so that these two will mix and transfer spectral weight. This can be seen with isotropic light, but will show up more pronounced as a linear dichroic effect if polarized light is used.
In contrast to the 1 ML NiO case, the good agreement between theory and experiment for the polarization dependent spectra of a 20 ML NiO film [@Alders98] have been achieved by assuming the presence of an antiferromagnetic order with an exchange field of about 0.16 eV in a pure local $O_h$ symmetry. It is surprising and also disturbing that a low symmetry crystal field could induce a spectral weight transfer between the two peaks of the $L_{2}$ white line such that the resulting linear dichroism appears to be very similar as a dichroism of magnetic origin. It is obvious that the ratio between the two peaks can not be taken as a direct measure of the spin orientation or magnitude of the exchange field in NiO films [@Alders98] if one has not first established what the crystal field contribution could be.
![Calculated ratio of the two peaks in the $L_{2}$ edge as a function of $\Delta E$, which is the shift in energy of the $L_{3}$ main peak in going from normal ($\theta = 90^{\circ}$) to grazing ($\theta = 0^{\circ}$) incidence of the linearly polarized light. The $L_{2}$ ratio is calculated for $\theta = 90^{\circ}$, $\theta = 0^{\circ}$, and for the isotropic spectrum. []{data-label="fig5"}](fig5.eps){width="45.00000%"}
We now can identify two strategies for finding out which part of the linear dichroism is due to low symmetry crystal field effects. The first one is to study the temperature dependence as we have done above. Here we have made use of the fact that those crystal fields do not split the high spin ground state, so that there are no additional states to be occupied with different temperatures other than those already created by the presence of exchange fields. Thus there should not be any temperature dependence in the crystal field dichroism. The linear dichroism due to magnetism, however, is temperature dependent and scales with $<$$M^2$$>$. By going to temperatures high enough such that there is no longer any temperature dependence in the linear dichroism, i.e. when all magnetic ordering has been destroyed, one will find the pure crystal field induced dichroism.
The second strategy to determine the low symmetry crystal field contribution is to measure carefully the energy shift $\Delta E$ in the main peak of the Ni $L_{3}$ white line for $\theta$=$0^{\circ}$ vs. $\theta$=$90^{\circ}$. We now calculate the ratio between the two peaks of the Ni $L_{2}$ edge as a function of $\Delta E$, and the results are plotted in Fig. 5 for $\theta =90^{\circ}$, $\theta =0^{\circ}$, and $\theta =54.7^{\circ}$ (isotropic spectrum). Since $\Delta E$ is a function of $Ds$ and $Dt$ combined, we have carried out the calculations with $Ds$ as a running variable for several fixed values of $Dt$, and plotted the resulting $L_{2}$ ratios vs. $\Delta E$. We now can use Fig. 5 as a road map to determine how much of the linear dichroism in the Ni $L_{2}$ edge is due to crystal field effects and how much due to magnetism. We can see directly that the 1 ML data lie on curves with the same $Dt$, meaning that the measured $L_{2}$ ratios are entirely due to crystal fields. The same can also be said for the 20 ML NiO on MgO at 528 K, which is not surprising since this temperature is above $T_{N}$. However, for the 20 ML NiO at 195 K, one can see that the data points do not lie on one of the $Dt$ curves (one may look for larger $Dt$ curves, but this results in lineshapes very different from experiment) or let alone on curves with the same $Dt$, indicating that one need magnetism to explain the $L_{2}$ ratios. In other words, knowing the $L_{2}$ ratio and $\Delta E$ together allows us to determine the magnitude of the exchange interaction and the orientation of the spin moments. It is best to use the $\theta = 0^{\circ}$ spectra since here the $L_{2}$ ratio is determined almost by $\Delta E$ alone and is not too sensitive to the individual values of $Ds$ and $Dt$.
To conclude, we have observed strong linear dichroism in the 1 ML NiO on Ag, very similar to the well known magnetic linear dichroism found for bulk like antiferromagnetic NiO films. The dichroism in the 1 ML, however, can not be attributed to the presence of some form of magnetic order, but entirely to crystal field effects. We provide a detailed analysis and a practical guide of how to disentangle quantitatively the magnetic from the crystal field contributions to the dichroic signal. This is important for a reliable determination of, for instance, the spin moment orientation in NiO as well as LaFeO$_{3}$, Fe$_{2}$O$_{3}$, VO, LaCrO$_{3}$, Cr$_{2}$O$_{3}$, and Mn$^{4+}$ manganate ultra thin films, surfaces and strained films, where the low symmetry crystal field splittings may not be negligible as compared to the exchange field energies.
We acknowledge the NSRRC staff for providing us with an extremely stable beam. We would like to thank Lucie Hamdan for her skillful technical and organizational assistance in preparing the experiment. The research in Cologne is supported by the Deutsche Forschungsgemeinschaft through SFB 608.
[99]{} B. T. Thole *et al.*, Phys. Rev. Lett. **55**, 2086 (1985) B. Sinkovic *et al.*, NSLS Annual Report 1989, p. 146 (unpublished) P. Kuiper *et al.*, Phys. Rev. Lett. **70**, 1549 (1993) D. Alders *et al.*, Europhys. Lett. **32**, 259 (1995) D. Alders *et al.*, Phys. Rev. B **57**, 11623 (1998) J. Stöhr *et al.*, Surf. Rev. Lett. **5**, 1297 (1998) D. Spanke *et al.*, Phys. Rev. B **58**, 5201 (1998) J. Stöhr *et al.*, Phys. Rev. Lett. **83**, 1862 (1999) A. Scholl *et al.*, Science **287**, 1014 (2000) F. Nolting *et al.*, Nature **405**, 767 (2000) H. Ohldag *et al.*, Phys. Rev. Lett. **86**, 2878 (2001) H. Ohldag *et al.*, Phys. Rev. Lett. **87**, 247201 (2001) W. Zhu *et al.*, Phys. Rev. Lett. **86**, 5389 (2001) F. U. Hillebrecht *et al.*, Phys. Rev. Lett. **86**, 3419 (2001) Spin-orbit interaction in the $3d$ shell will not split the $^{3}A_{2}$ ground state of a Ni $3d^{8}$ ion in $O_{H}$ symmetry. Spin-orbit coupling does split the $^{3}$B$_{1}$ state for a Ni ion in $D_{4h}$, but this splitting is negligibly small, of order $10^{-4}$ eV. S.Altieri, PhD. Thesis, Groningen (1999) S. Altieri *et al.*, Phys. Rev. B **59**, 2517 (1999) C.J. Ballhausen *Introduction to Ligand Field Theory, Chap. 5* (McGraw-Hill, New York, 1962.) A. Tanaka, and T. Jo, J. Phys. Soc. Jpn. **63**, (1994) 2788
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'U. Geppert , M. Rheinhardt'
- 'J. Gil'
date: 'Received date ; accepted date'
title: 'Spot–like Structures of Neutron Star Surface Magnetic Fields'
---
Observational evidence {#obs}
======================
There exists both observational and theoretical evidence of the existence of strong small–scale magnetic field structures in the surface layers of isolated neutron stars. More and more, it becomes clear that neutron stars of nearly all ages must posses magnetic field structures, which are much more complicated than the simple assumption of a star centered magnetic dipole suggests. The polar surface magnetic field strength of such a dipole is conventionally estimated by $B_d \sim 6.4 \cdot 10^{19} \sqrt{P\dot{P}}$ G, where standard neutron star quantities are assumed and the rotational period $P$ (in seconds) and its temporal derivative $\dot{P}$ are taken from radio and/or X–ray timing observations.
Recently, @BSP03 reported the (marginal) detection of an electron cyclotron line at about $3.3$ in the Chandra–spectra of the millisecond pulsar . The energetic location of that line corresponds to a surface magnetic field strength of $3\cdot 10^{11}$ G, while $P$ and $\dot P$ yield a dipolar surface field strength of about $4.5\cdot 10^9$ G, almost two orders of magnitude lower.
@HSH03 analysed the XMM–spectra of the isolated neutron star and found evidence for a proton cyclotron absorption line in the energy range of $0.1\ldots0.3$ , which is consistent with a magnetic surface field strength of $(2\ldots6)\cdot 10^{13}$ G. While $P=10.31$ s seems to be a well settled value for the rotational period, the determination of $\dot{P}$ is a much more delicate task in this case. Recent evaluations of XMM data indicate, however, that the global dipolar field attributed to $P$ and $\dot{P}$ is at least one order of magnitude smaller than the field corresponding to the proton cyclotron line (Schwope 2003, private communication). Likewise, for the pulsar a surface field of $\sim 1.5\cdot 10^{14}$ G has been estimated [@S02], whereas the estimate for its dipolar field is $(2\ldots4)\cdot10^{12}$ G [@P02].
These X–ray observations indicate that, apart from a large–scale dipolar magnetic field which determines the spin–down behaviour of the pulsar, much stronger but short–ranged field components close to the neutron star surface exist. They do not affect significantly the braking of the star’s rotation, but can affect the magnetospheric processes in the vicinity of the polar cap (defined by the totality of all open dipolar field lines).
It is commonly accepted that pulsar radio emission is generated within a dense electron–positron plasma, the creation of which requires an ultra–strong potential drop that accelerates charged particles along curved magnetic field lines. The observed phenomenon of drifting subpulses strongly suggests that this potential drop results from the deficiency in the actual charge density with respect to the so–called co–rotational charge density [@GJ69] just above the polar cap surface [see @GMG03]. The formation of such a charge–depleted region (called ‘polar gap’ after @RS75, who proposed it for the first time) requires a strong and highly non–dipolar surface magnetic field, with radii of curvature much smaller than $10^6$ cm and magnitudes close to $10^{13}$ G, irrespective of the dipolar field strength inferable from the pulsar spin–down [@GM01; @GM02; @GMG03]. Moreover, several periodicities observed in the phenomenon of drifting subpulses strongly suggest that the subpulse–associated plasma filaments [called ‘sparks’ after @RS75] circulate around a local magnetic pole[^1] [@DR99; @DR01; @GS03]. Accordingly, the small–scale surface magnetic field anomalies are supposed to show spot–like structures allowing a persistent arrangement of drifting sparks (due to the well known ${\vec{E}}\times {\vec{B}}$ plasma drift mechanism) in the form of quasi–annular patterns [@GS00; @FCM01].
The Hall–drift induced instability
==================================
It is well known that the Hall–effect, probably via a cascade [@GR92], causes the generation of smaller scaled magnetic field components out of an existing large scale field. This process has been discussed for neutron star magnetic fields by a number of authors [see, e.g., @SU97; @VCO00; @HR02]. Here, however, we want to refer to the fact that small–scale poloidal fields close to the neutron star surface can be generated from a subsurface toroidal magnetic field by a Hall–drift induced instability (HDI). The basic prerequisite for this instability is that a sufficiently strong [*and*]{} inhomogeneous background field exists (@RG02 [@GR02] and @RKG03\[RKG03\]). We demonstrate, that strength and spatial structure of the HDI modes resulting from a realistic NS crust model are consistent with quantities derived from observations.
The decay of a magnetic field in the almost crystallised crust of a NS is governed by $$\begin{aligned}
\phantom{.}&\itdot{{\vec{B}}} = -\curl\big(\,\eta\curl{\vec{B}}+\alpha(\curl{\vec{B}}\times{\vec{B}})
\,\big)\\
\phantom{.}&\dv {\vec{B}}= 0\; ,
\end{aligned}\label{indeq}$$ where the diffusivity $\eta$ and the Hall–parameter $\alpha$ determine the intensity of Ohmic decay and Hall–drift, respectively, in their spatial distribution. Linearization of Eq. (\[indeq\]) with respect to a background field $\vec B_0$ yields $$\begin{aligned}
\phantom{.}&\begin{aligned}
\dot{{\vec{b}}} = &-\curl\,(\eta\,\curl{\vec{b}}) \\
&-\curl\big(\alpha\,(\,\curl{\vec{B}}_0
\times{\vec{b}}+\curl{\vec{b}}\times{\vec{B}}_0\,)\big)
\end{aligned}\\
\phantom{.}&\dv {\vec{b}}= 0
\end{aligned}\label{indeqlin}$$ describing the evolution of a small magnetic perturbation $\vec b$. For conditions as realised in the crusts of cooling isolated neutron stars, unstably growing perturbations $\vec b = \hat{\vec b}(\vec r) \exp{\gamma t}\,,\;\gamma > 0\,$, can exist, with typical growth times $\gamma^{-1} \sim 10^{4...5}$ years.
The existence of this instability has been shown in a slab geometry, approximating the neutron star crust geometry locally. Typical crustal density profiles were adopted. Boundary conditions are defined by assuming a perfect conductor being adjacent to the slab at the bottom and vacuum being adjacent at the top. That is, corresponding to the superconductivity of the NS core, the magnetic and electric fields do not penetrate under the bottom of the slab whereas the magnetic field is continued above the slab as a potential field approximating the conditions outside the NS.
The background fields ${\vec{B}}_0$ we used are parallel to the slab; they vary strongly with depth and vanish at the bottom of the slab. Fields of that kind, especially toroidal ones, may exist as relics of a short, but very efficient convective dynamo phase acting in the proto–neutron star [@TD93; @UG03]. However, it is perhaps more likely that such background fields are due to the action of a thermoelectric instability, which amplifies toroidal seed fields very effectively and sets on compulsory when $T_{\mathrm{s6}
}^4/g_{\mathrm{s14}} \ga 100$ (@WG96, $T_{\mathrm{s6}}$ – surface temperature in $10^6$ K, $g_{\mathrm{s14}}$ – surface gravity in $10^{14}$ cm s$^{-2}$). Such conditions are met in almost all newly born neutron stars and are maintained up to an age of $\sim\!1000$ years. Moreover, when a strong temperature gradient is established again by any external or internal process, the thermoelectric instability will be switched on again. It can thus, at later stages too, produce a toroidal background field which is capable of providing the conditions for the onset of the HDI. Hence, the initial strength of the background field is determined essentially either by the vigour of the dynamo or by the strength of the temperature gradient and may well exceed $10^{14}$ G locally inside the crust.
Of course, in any real situation the background field is not exactly parallel to the slab: Within the region we are interested in, the polar cap, a normal background field arises quite naturally. Its influence on the properties of the HDI can at the moment only be extrapolated from the special case of a [*homogeneous*]{} normal background field component in a homogeneous slab. Then, the tangential part of the background field rotates about a normal axis with an angular velocity proportional to the normal field component. The HDI continues to exist if the ratio of normal to tangential background field components is not too large. If the normal field is small, what we want to assume here, the background configuration rotates slowly and we expect no major differences in the perturbation modes in comparison with those obtained for completely tangential background fields to appear.
The HDI is described in detail for conditions realised in the crusts of isolated cooling neutron stars in RKG03 and we refer the reader to that extended presentation. Its main result is that the HDI occurs for a variety of realistic crust conditions and that the obtained growth times are short enough to cause observable consequences. The easiest way to get an idea on how the HDI acts is considering its analogy to field generation by hydrodynamic dynamo action. One major ingredient of dynamo models, e.g., those explaining the solar magnetic field, is shear motion (in axisymmetry typically occurring as differential rotation) which is capable of generating strong toroidal (in axisymmetry: azimuthal) fields from weak poloidal (in axisymmetry: meridional) fields by ‘winding up’. In our context, the motion of the electron fluid, in which the field is partially frozen in, is well able to provide shear (see the term $\alpha\,\curl{\vec{B}}_0 \times{\vec{b}}$ in ). To get a successful dynamo, shear has to be completed by another effect generating in turn poloidal fields from toroidal ones. Mean–field dynamo theory has identified various possibilities for that, amongst which one (the so–called $\vec{\omega}\times{\vec{j}}$ effect) has exactly the same mathematical structure like the second Hall term $\alpha\curl{\vec{b}}\times{\vec{B}}_0$ in [see, e.g., @R69]. Thus, a cycle of mutual amplification of poloidal and toroidal fields can establish resulting in a growing total field.
The small–scale poloidal field
==============================
The structure of a typical[^2] unstable (i.e., exponentially growing) perturbation field (or eigenmode) is shown in Figs. \[eigfunc1\] and \[eigfunc2\] (for detailed explanations see RKG03). While the background fields used in that paper are derived from the equatorial regions of global dipolar fields, we nevertheless apply here the corresponding results in the vicinity of the polar cap region. There, a global poloidal field can no longer provide a suitable background field structure – a toroidal field is needed. Applicability of the results of RKG03 then requires that the toroidal background field near the pole and the poloidal one near the equator may have more or less similar radial profiles. (Note, that the properties of the HDI modes are only moderately sensitive with respect to the radial profile.) This seems to be at least possible in the case of convective–dynamo generated fields and even a reasonable assumption for thermoelectrically generated fields with not too low multipolarity. For a maximum strength of the toroidal background field in the order of about $10^{14}$ G the perturbation rises with a characteristic growth time of about $5\cdot10^4$ years. We assume, that at saturation it may reach a significant fraction of the background field’s strength. Then it is possible that its surface strength reaches $5 \cdot 10^{12}$ G and we scaled the perturbation field (the amplitude of which is not determined by the linear Eq. (\[indeqlin\])) simply to this value.
The lifetime of such a perturbation is bounded from above by its ohmic decay time. Very likely, it is shorter because the perturbation is subject to the Hall–effect which is supposed to accelerate the decay in general. With the conditions at the depth in which the perturbation currents circulate and their scales, we estimate the ohmic decay time of the field presented in Figs. \[eigfunc1\] and \[eigfunc2\] as $\gtrsim 10^7$ yrs.
Both the major radial and tangential scales of the unstable perturbations are given approximately by the radial extent of the background field. According to the assumptions about the latter they are significant portions of the crust thickness (say $\lesssim 50$%) and scale with it. The crust thickness, in turn, is model dependent: Stiff equations of state (EOS) result in smaller compactnesses and hence larger star radii and crust thicknesses in comparison with softer ones. For example, the specific EOSs considered in RKG03, the stiff Pandharipande–Smith and medium soft Friedman–Pandharipande ones yield a crust thickness of $\sim$ 3.8 km and 700 m, respectively. Generalising from these two cases we suggest that the scales of the eigenmodes are not very sensitive with respect to the star model.
Let us stress that the spot–like structure presented in Figs. \[eigfunc1\] and \[eigfunc2\] satisfies all conditions to create the vacuum gap and to generate the electron–positron plasma within it. In fact, @GM01 and @GM02 argued that the vacuum gap can only be formed under the so–called near–threshold conditions, when the surface magnetic field $B_s \ga 0.1 B_q
\sim 4.4 \cdot 10^{12}$ G. In addition, it requires small radii of curvature $R_c \ll 10^6$ cm. As can be seen from Fig. \[eigfunc1\], in our case $R_c < 10^5$ cm. In such a strong and curved surface magnetic field the magnetic pair creation via curvature and/or ICS (inverse Compton scattering) photons is very efficient.
Moreover, the magnetic field lines of the perturbation field converge at local poles (spot centers) as can be seen in Fig. \[eigfunc2\]. As already mentioned at the end of , this property of the surface magnetic field is crucial for the subpulse drift phenomenon. Of course, as an implicit assumption one has to keep in mind that just one spot structure plays the role of a local magnetic pole requiring that the canonical polar cap (i.e., the one formed by the dipolar field alone) coincides at least partially with this spot. The actual polar cap is defined by the open field lines of the superimposed global star–centered–dipole and local perturbation fields, along which an ultra–high accelerating potential drop can exist, or, in other words, by those perturbation field lines which merge with the open dipolar field lines before entering the radio emission region [see @GM02a for details].
When assuming that close to the surface the global dipolar field is at least five times smaller than the perturbation field the superposition of both fields would yield a field very similar to the one shown where only the symmetry between the regions with positive and negative normal flux is slightly disturbed.
Note, that the spot–like structures of the surface field must not be confused with the Deshpande–Rankin ‘hot spots’ or ‘sparks’ [@DR99]. The former can be considered stationary within all the time scales relevant for observations and provide just the conditions for the development of sparks. By virtue of the ${\vec{E}}\times{\vec{B}}$ plasma drift the latter move around the local magnetic pole. As the appearance of the pure ${\vec{E}}\times{\vec{B}}$ drift is confined to homogeneous fields only, additional drift components and/or acceleration prevent the charge carriers forming the sparks from moving exactly periodically along closed paths. So it remains to be examined whether our field spots are suited to reproduce the periodicities of the subpulse phenomenon in detail.
\
\
This paper is supported in part by Polish grant 2P03D00819. J.G. acknowledges the renewal of the Alexander von Humboldt fellowship. U.G. and M.R. are grateful to the Arbeitsamt Berlin for financial support and to the AIP for hospitality.
SmSc6.bbl
[^1]: A surface region with a magnetic flux prevailingly normal to the surface and with a certain degree of axisymmetry.
[^2]: Pandharipande–Smith equation of state, initial penetration density $\rho_0=10^{13}$ g cm$^{-3}$, cubic background field ${\vec{B}}_0$ with initial polar field strength $10^{14}$ G, model age $3 \cdot 10^5$ years.
| {
"pile_set_name": "ArXiv"
} |
[**OBSERVABLE CONSEQUENCES\
OF CROSSOVER-TYPE DECONFINEMENT PHASE TRANSITION**]{}
V.D. Toneev$^{\dag}$
[\
$\dag$ [*E-mail: [email protected]* ]{}]{}
**Abstract**
Equation of State (EoS) for hot and dense nuclear matter with a quark-hadron phase transition is constructed within a statistical mixed-phase model assuming coexistence of unbound quarks in nuclear surrounding. This model predicts the deconfinement phase transition of the crossover type. The so-called “softest point” effect of EoS is analyzed and confronted to that for other equations of state which exhibit the first order phase transition (two-phase bag model) or no transition at all (hadron resonance gas). The collective motion of nucleons from high-energy heavy-ion collisions is considered within a relativistic two-fluid hydrodynamics for different EoS. It is demonstrated that the beam energy dependence of the directed flow is a smooth function in the whole range from SIS till SPS energies and allows to disentangle different EoS, being in good agreement with experimental data for the statistical mixed-phase model. In contrast, excitation functions for relative strangeness abundance turn out to be insensitive to the order of phase transition.\
[**Key-words:** ]{} QCD phase transition, heavy-ion collisions, hydrodynamics, directed flow, strange particle production.
Introduction
============
The predicted phase transition from confined hadrons to a deconfined phase of their constituents ([*i.e.*]{} the asymptotically-free quarks and gluons or the so-called Quark-Gluon Plasma, QGP) is a challenge to the theory of strong interaction. Over the past two decades a lot of efforts has been spent to both the theoretical study of deconfinement phase transition and the search for its possible manifestation in relativistic heavy ion collisions, properties of neutron stars and Universe evolution. A unique opportunity provided by relativistic heavy-ion collisions allowed to reach a state with temperature and energy density exceeding the critical values, $T_c\sim 170 \ MeV$ and $\varepsilon_c \sim 1 \ GeV/fm^3$, specific for the deconfinement phase transition. A rather long list of various signals for the QGP formation in hot and dense nuclear matter is available now and it has been probed in experiments with heavy ions. Unfortunately, there is no crucial signal for unambiguous identification of the deconfinement phase and, for a particular reaction at the given bombarding energy, practically every proposed signal can be simulated to some extent by hadronic interactions.
In this paper we turn to the study of excitation functions for observables to be sensitive to the expected QCD deconfinement phase transition. Its manifestation has been considered already some time ago by [@SZ89; @V83]. Since a phase transition slows down the time evolution of the system due to [*softening*]{} of the EoS, and one can expect a remarkable loss of correlations around some critical incident energy resulting in definite observable effects.
Equation of state in mixed phase model
======================================
Following the common strategy of the two-phase (2P) bag model [@Cleym], one can determine the deconfinement phase transition by means of the Gibbs conditions matching the EoS of a relativistic gas of hadrons and resonances, whose interactions are simulated by the Van der Waals excluded volume correction, to that of an ideal gas of quarks and gluons, where the change in vacuum energy in a QGP state is parameterized by the bag constant $B$. Thermodynamics of the hadron gas is described in the grand canonical ensemble. All hadrons with the mass $m_j < 1.6 \ GeV$ i are taken into consideration. One should emphasize that the phase transition in the 2P model is right along of the first order by constructing.
To reproduce the variety of phase transitions predicted by QCD lattice calculations we represent a phenomenological Mixed Phase (MP) model [@NST98; @TNS98]. The underlying assumption of the MP model is that unbound quarks and gluons [*may coexist*]{} with hadrons forming a [*homogeneous*]{} quark/gluon–hadron phase. Since the mean distance between hadrons and quarks/gluons in this mixed phase may be of the same order as that between hadrons, the interaction between all these constituents (unbound quarks/gluons and hadrons) plays an important role and defines the order of the phase transition.
Within the MP model [@NST98; @TNS98] the effective Hamiltonian is expressed in the quasiparticle approximation with density-dependent mean-field interactions. Under quite general requirements of confinement for color charges, the mean-field potential of quarks and gluons is approximated by $$U_q(\rho)=U_g(\rho)={A\over\rho^{\gamma}}~;
\ \ \ \gamma >0
\label{eq6}$$ with [*the total density of quarks and gluons*]{} $$\rho=\rho_q + \rho_g +\sum\limits_{j}\;\nu_j\rho_{j}~,$$ where $\rho_q$ and $\rho_g$ are the densities of unbound quarks and gluons outside of hadrons, while $\rho_{j}$ is the density of hadron type $j$ and $\nu_j$ is the number of valence quarks inside. The presence of the total density $\rho$ in (\[eq6\]) implies interactions between all components of the mixed phase. The approximation (\[eq6\]) mirrors two important limits of the QCD interaction. For $\rho \to 0$, the interaction potential approaches infinity, [*i.e.*]{} an infinite energy is necessary to create an isolated quark or gluon, which simulates the confinement of color objects. In the other extreme case of large energy density corresponding to $\rho \to \infty$, we have $U_q=U_g=0$ which is consistent with asymptotic freedom.
The use of the density-dependent potential (\[eq6\]) for quarks and the hadronic potential, described by a modified non-linear mean-field model [@Zim], requires certain constraints to be fulfilled, which are related to thermodynamic consistency [@NST98; @TNS98]. For the chosen form of the Hamiltonian these conditions require that $U_g(\rho)$ and $U_q(\rho)$ do not depend on temperature. From these conditions one also obtains a form for the quark–hadron potential [@NST98].
A detailed study of the pure gluonic $SU(3)$ case with a first-order phase transition allows one to fix the values of the parameters as $\gamma =0.92$ and $\displaystyle A^{1/(3\gamma+1)}
= 250$ MeV. These values are then used for the $SU(3)$ system including quarks. As is shown in Fig.1 for the case of quarks of two light flavors at zero baryon density ($n_B=0$), the MP model is consistent with lattice QCD data providing a continuous phase transition if the crossover type with a deconfinement temperature $T_{dec}=153$ MeV. For a two-phase approach based on the bag model a first-order deconfinement phase transition occurs with a sharp jump in energy density $\varepsilon$ at $T_{dec}$ close to the value obtained from lattice QCD.
=10.cm
Though at a glimpse the temperature dependencies of the energy density $\varepsilon$ and pressure $p$ for the different approaches presented in Fig.1 look quite similar, there is a large difference revealed when $p/\varepsilon$ is plotted versus $\varepsilon$ (cf. Fig.2, left panel). The lattice QCD data differ at low $\varepsilon$, which is due to difficulties within the Kogut–Susskind scheme [@berna] in treating the hadronic sector. A particular feature in the MP model is that, for $n_B=0$, the [*softest point*]{} of the EoS, defined as a minimum of the function $p(\varepsilon)/\varepsilon$ [@HS95], is not
=10.cm
very pronounced and located at comparatively low values of the energy density: $\varepsilon_{SP} \approx 0.45$ GeV/fm$^3$, which roughly agrees with the lattice QCD value [@RS86]. This value of $\varepsilon $ is close to the energy density inside a nucleon, and hence, reaching this value indicates that we are dealing with a single [*big hadron*]{} consisting of deconfined matter. In contradistinction, the bag-model EoS exhibits a very pronounced softest point at large energy density $\varepsilon_{SP} \approx
1.5$ GeV/fm$^3$ [@HS95; @R96].
The MP model can be extended to baryon-rich systems in a parameter-free way [@NST98; @TNS98]. As demonstrated in Fig.2 (right panel), the softest point for baryonic matter is gradually washed out with increasing baryon density and vanishes for $n_B
\gsim 0.3 \ n_0$ ($n_0$ is normal nuclear matter density). This behavior differs drastically from that of the two-phase bag-model EoS, where $\varepsilon_{SP}$ is only weakly dependent on $n_B$ [@HS95; @R96]. It is of interest to note that the interacting hadron gas model has no softest point at all and, in this respect, its thermodynamic behavior is close to that of the MP model at high energy densities [@TNS98].
These differences between the various models of EoS should manifest themselves in dynamics discussed below.
Directed flow of baryons
========================
The EoS described above is applied to a two-fluid (2F) hydrodynamic model [@MRS91], which takes into account finite stopping power of colliding heavy ions. In this dynamical model, the total baryonic current and energy-momentum tensor are written as $$\begin{aligned}
J^{\mu} &=& J^{\mu}_p + J^{\mu}_t~~,
\label{eq7.6}
\\
T^{\mu\nu} &=& T^{\mu\nu}_p + T^{\mu\nu}_t~~,
\label{eq7}
\end{aligned}$$ where the baryonic current $J^{\mu}_{\alpha}=n_{\alpha}u_{\alpha}^{\mu}$ and energy-momentum tensor $T^{\mu\nu}_{\alpha}$ of the fluid $\alpha$ are initially associated with either target ($\alpha=t$) or projectile ($\alpha=p$) nucleons. Later on these fluids contain all hadronic and quark–gluon species, depending on the model used for describing the fluids. The twelve independent quantities (the baryon densities $n_{\alpha}$, 4-velocities $u_{\alpha}^{\mu}\;$ normalized as $u_{\alpha\mu}u_{\alpha}^{\mu}=1$, as well as temperatures $T$ and pressures $p$ of the fluids) are obtained by solving the following set of equations of two-fluid hydrodynamics [@MRS91] $$\begin{aligned}
\label{eq8}
\partial_{\mu} J_{\alpha}^{\mu} &=& 0~~, \\
\partial_{\mu} T^{\mu\nu}_{\alpha} &=& F_{\alpha}^\nu~~,
\label{eq8a}
\end{aligned}$$ where the coupling term $$\begin{aligned}
F_{\alpha}^\nu=n^s_p n^s_t
\left < V_{rel} \int d\sigma_{NN\to NX}(s)\; (p - p_{\alpha})^\nu
\right >
\label{eq9}\end{aligned}$$ characterizes friction between the counter-streaming fluids. The cross sections $d\sigma_{NN\to
NX}$ take into account all elastic and inelastic interactions between the constituents of different fluids at the invariant collision energy $s^{1/2}$ with the local relative velocity $V_{rel} =[s(s-4m_N^7)]^{1/2}/2m_N^2~.$ The average in (\[eq9\]) is taken over all particles in the two fluids which are assumed to be in local equilibrium intrinsically [@MRS91]. The set of Eqs. (\[eq8\]) and (\[eq8a\]) is closed by EoS, which is naturally the same for both colliding fluids.
Following the original paper [@MRS91], it is assumed that a fluid element decouples from the hydrodynamic regime, when its baryon density $n_B$ and densities in the eight surrounding cells become smaller than a fixed value $n_f$. A value $n_f = 0.8 n_0$ is used for this local freeze-out density which corresponds to the actual density of the freeze-out fluid element of about $0.6-0.7\ n_0$.
The directed flow characterizes the deflection of emitted hadrons away from the beam axis within the reaction $x-z$ plane. In particular, one defines the differential directed flow by the mean in-plane component $\left< p_x(y)\right>$ of the transverse momentum at a given rapidity $y$. This deflection is believed to be quite sensitive to the [*elasticity*]{} or [*softness*]{} of the EoS and can be quantified in two ways: In terms of the derivative (a slope parameter) at mid-rapidity $$\label{eq12} F_y = \left. \frac{d \; \left<p_x(y)\right>}{dy}
\right|_{y=y_{cm}}~~,$$ which is quite suitable for analyzing the flow excitation function, and by another integral quantity to be less sensitive to possible rapidity fluctuations of the in-plane momentum: $$\left< P_x\right> = \frac{\displaystyle\int dp_xdp_ydy \ p_x \
\left( E{\displaystyle
\frac{d^3N}{dp^3}}\right)}{\displaystyle\int dp_xdp_ydy \ \left(
E{\displaystyle
\frac{d^3N}{dp^3}}\right) }~~,
\label{eq13}$$ where the integration in the c.m.system runs over the rapidity region $\displaystyle [0,y_{cm}]$. Excitation functions in the SIS-AGS-SPS energy range are plotted in Fig.3 for both characteristics [@INNST02].
\[fig3\]
Our first 2F hydrodynamic calculations of $F_y (E_{lab})$ are in a good agreement with experiment in the whole energy range considered. In the left lower panel of Fig.3 our results are compared with transport calculations. The ARC and ART are cascade models, while the RQMD takes also into account mean-field effects. Though all these models agree with experimental data at $E_{lab}
\approx 10$ A$\cdot$GeV (considered as a reference point), values of $F_y$ at lower energies are clearly underestimated, as is evident from comparison with results of the E895 Collaboration [@E895] (see empty squares in Fig.3). Recently, a good description of experimental points (including the E895 data) was reported within a relativistic BUU (RBUU) model [@cas99]. The good agreement with experiment was achieved here by a special fine tuning of the mean fields involved in the particle propagation.
The calculated excitation functions of $\left< P_x\right>$ for baryons within different hydrodynamic models are shown in the right panel of Fig.3. Conventional 1F hydrodynamics for pure hadronic matter [@R96] results in a very large directed flow due to the inherent instantaneous stopping of the colliding matter. This instantaneous stopping is unrealistic at high beam energies. If the deconfinement phase transition, based on the bag-model EoS [@R96], is included into this model, the excitation function of $\left<
P_x\right>$ exhibits a deep minimum near $E_{lab}\approx 6$ A$\cdot$GeV, which manifests the softest-point effect of the bag-model EoS as shown in the right panel of Fig.2.
The result of 2F hydrodynamics with the MP EoS noticeably differs from the 1F calculations. After a maximum around 1 A$\cdot$GeV, the average directed flow decreased slowly and smoothly. This difference is caused by two reasons. First, as follows from Fig.2, the softest point of the MP EoS is washed out for $n_B \gsim 0.4$. The second reason is dynamical: the finite stopping power and direct pion emission change the evolution pattern. The latter point is confirmed by comparison to three-fluid calculations with the bag EoS [@3DF] plotted in the right lower panel of Fig.3. The third pionic fluid in this model is assumed to interact only with itself neglecting the interaction with baryonic fluids. Therefore, with regard to the baryonic component, this three-fluid hydrodynamics [@3DF; @3DF97] is completely equivalent to our two-fluid model and the main difference is due to the different EoS. As seen in Fig.3, the minimum of the directed flow excitation function, predicted by the one-fluid hydrodynamics with the bag-model EoS, survives in the three-fluid (nonunified) regime but its value decreases and its position shifts to higher energies. If one applies the unification procedure of [@3DF], which favors fusion of two fluids into a single one, and thus making stopping larger, three-fluid hydrodynamics practically reproduces the one-fluid result and predicts in addition a bump at $E_{lab}\approx 40$ A$\cdot$GeV.
Strangeness production
======================
Enhanced strangeness production as compared to proton-proton or proton-nucleus collisions is one of the QGP signals proposed a long time ago. In the hydrodynamic model described above only baryon charge rather than strangeness exchange is included. So, to see experimental consequences of EoS with different phase transition order, we consider an expanding homogeneous blob of the compressed and heated QCD matter (a fireball) formed in heavy-ion collisions. The initial state ($\varepsilon_0$ and $n_B$) for this fireball is estimated from results of the QGSM transport calculations in the center-of-mass frame inside a cylinder of the volume $V_0$ with radius $R=5 \ fm$ and Lorentz-contracted length $L=2R/\gamma_{c.m.}$. Isoentropic expansion is treated in an approximate manner assuming $V \sim V_0
t$ until the freeze-out point defined by $\varepsilon_f = 0.15\
GeV/fm^3 \approx m_N n_0$ (for more detail see [@FNNRT02]).
One should note that till this point the grand canonical ensemble was used where complete chemical equilibrium is assumed and the strangeness conservation is controlled on average by the strange chemical potential $\mu_S$. In the thermodynamical limit, fluctuations in a number of strange particles are small and coincide with those for the canonical ensemble. However, it is not the case for finite systems at relatively small $T$ where the strangeness canonical ensemble should be applied, taking into account the associative nature of strange particle creation by exact and local conservation of strangeness. Using the general formalism for the canonical strangeness conservation proposed in [@HR85; @CRS91], the partition function of a gas of hadrons with strangeness $s_i =0, \pm 1, \pm 2, \pm3$ and total strangeness $S=0$ can be written as follows $$Z_S=\frac{1}{2\pi} \int_{-\pi}^{\pi}d\phi \ \exp(\sum_{s=-3}^{3}
{\cal S}_s \ e^{is\phi}) \label{eq26}$$ where ${\cal S}_s=V\sum_i Z_i~. $ Here $Z_i$ is the one-particle partition function for species $i$ and the sum is taken over all particles and resonances carrying strangeness $s_i$. A number of strange particles can be found by the appropriate differentiation of the partition function $Z_S$, Eq.(\[eq26\]). It is easy to see that canonical result can be obtain in the Boltzmann approximation from the grand canonical one by replacing the strange fugacity in the following way : $$\exp (\mu_s/T) \to \left(\frac{{\cal S}_1}{\sqrt{{\cal S}_1 {\cal
S}_{-1}}}\right)^s \ \frac{I_s(x)}{I_0(x)}~,
\label{eq31}$$ where the argument of the Bessel functions $x\equiv 2\sqrt {{\cal
S}_1{\cal S}_{-1}} \sim V$. This receipt was applied to our treatment of particle abundance at the freeze-out point. Generally, the correlation volume in the suppression factor $I_s(x)/I_0(x)$ of (\[eq31\]) does not coincide with the system volume $V$. In our model, the initial Lorentz-contracted volume $V_0$ is considered as a strangeness correlation volume.
![ Particle ratios for strange hadrons in full $4\pi$ angle interval for central $Au+Au$ collision as a function of bombarding energy. The compilation of available experimental points is taken from [@Redlich01; @Stock02]. The calculated excitation functions represent four modeling EoS with the canonical suppression factor $I_s(x)/I_0(x)$. For the case of the MP model, the grand canonical results (dashed lines) are given, as well.[]{data-label="fig4"}](ratios.eps){width="10.3cm"}
Inspection of Fig.\[fig4\] shows that the inclusion of the canonical strangeness suppression factor allows one to decrease noticeably strange particle abundance. However, comparing canonical and grand canonical results for the MP model, one can see that they do not coincide at high energies as it would be expected. This is explained by the beam energy dependence of the strangeness correlation volume in contrast with usual canonical description [@HR85; @CRS91]. The most striking result followed from Fig.\[fig4\] is that all the models considered in the strangeness canonical ensemble predict practically the same relative abundance of strange hadrons in the whole energy range studied. The measured excitation function for $K^+/\pi^+$ are reproduced reasonably well excluding maybe the SIS energy. In the case of $K^-/\pi^-$ the general form of excitation functions also agrees with experimental one but the relative abundance is overestimated what mainly originates from neglecting the electric charge (isospin) conservation. A simple estimate shows that taking into account the isospin conservation the $K^-/\pi^-$ ratio decreases by about $18\%$ and $12 \%$ at $E_{lab}=10$ and $150 \
AGeV$, respectively, without any essential influence on the $K^+/\pi^+$ ratio. The relative yield of hyperons, in particular $\Lambda/\pi^+$’s, seems to be overshot in the energy range $E_{lab} \lsim \ 10 \ AGeV$ what can result from the simplified dynamical treatment: The Bjorken-like longitudinal expansion can be applied at the SPS energies, but at the SIS energies the transverse expansion is not negligible. It is worthy to note that all calculations have been done with the same shock-like freeze-out condition for every EoS without any special tuning.
Conclusions
===========
It has been shown that the directed flow excitation functions $F_y$ and $\left< P_x\right>$ for baryons are sensitive to the EoS, but this sensitivity is significantly masked by nonequilibrium dynamics of nuclear collisions. Nevertheless, the results indicate that the widely used two-phase EoS, based on the bag model [@HS95; @R96] and giving rise to a first-order phase transition, seems to be inappropriate. The neglect of interactions near the deconfinement temperature results in an unrealistically strong softest-point effect within this two-phase EoS. Smooth experimental excitation function of the directed flow is reasonably reproduced within the MP model.
The dynamical trajectories for a fireball state in the $T-\mu_B$ plane are quite different for different EoS [@FNNRT02; @TCN01]. However, after using the shock-like freeze-out, the global strangeness production is turned out to be completely insensitive to the particular EoS as illustrated by the calculation of excitation functions for $K^+, K^-$ and hyperons. To get agreement with experiment the canonical suppression factor should be taken into account. The only trace of dynamics is the beam-energy dependence of the strangeness correlation volume. This effect results in a negative slope of the $K^+/\pi^+$ excitation function at high energies which is not reproduced by the equilibrium statistical model with canonical account of strangeness [@CRS91; @Redlich01].
Among other signals of the QCD deconfinement phase transition, the dilepton and hard photon production is the most promising. Being sensitive to the whole evolution time of a system, these signals can become experimentally observable to disentangle the crossover phase transition, predicted by the MP model, from the first order one which is peculiar for the bag-model EoS.
Useful and numerous discussions with B. Friman, Yu.B. Ivanov, E.G. Nikonov, W. Nörenberg and K. Redlich are acknowledged. This work was supported in part by DFG (project 436 RUS 113/558/0) and RFBR (grant 00-02-04012).
[99]{}
E. Shuryak and O.V. Zhirov, [*Phys. Lett.*]{} B [**89**]{}, 253 (1979).
L. van Hove, [*Z. Phys.*]{} C [**21**]{}, 93 (1983).
J. Cleymans, R.V. Gavai, and E. Suhonen, [*Phys. Rep.*]{} [**130**]{}, 217 (1986).
E.G. Nikonov, A.A. Shanenko, and V.D. Toneev, [*Heavy Ion Physics*]{} [**8**]{} (1998) 89.
V.D. Toneev, E.G. Nikonov, and A.A. Shanenko, in [*Nuclear Matter in Different Phases and Transitions*]{}, eds. J.-P. Blaizot, X. Campi, and M. Ploszajczak, Kluwer Academic Publishers (1999), p.309.
J. Zimanyi [*et al*]{}., [*Nucl. Phys.*]{} [**A484**]{}, 147 (1988).
K. Redlich and H. Satz, [*Phys. Rev.*]{} D [**33**]{}, 3747 (1986).
C. Bernard [*et al*]{}., [*Nucl. Phys. (Proc. Suppl.)*]{} [**B47**]{}, 499 (1996); [*ibid*]{} 503.
C.M. Hung and E.V. Shuryak, [*Phys. Rev. Lett.*]{} [**75**]{}, 4063 (1995).
D.H. Rischke [*et al.*]{}, [*Heavy Ion Physics*]{} [**9**]{}, 309 (1996);\
D.H. Rischke, [*Nucl. Phys.*]{} [**A610**]{}, 28c (1996).
I.N. Mishustin, V.N. Russkikh, and L.M. Satarov, [*Nucl. Phys.*]{} [**A494**]{}, 595 (1989);\
[*Yad. Fiz.*]{} [**54**]{}, 479 (1991) (translated as [*Sov. J. Nucl. Phys.*]{} [**54**]{}, 260 (1991).
Yu.B. Ivanov, E.G. Nikonov, W. Nörenberg, A.A. Shanenko and V.D. Toneev, [*Heavy Ion Physics*]{} [**15**]{} (2002) 117.
N.N. Ajitanand [*et al.*]{}, [*Nucl. Phys.*]{} [**A638**]{}, 451c (1998).
P.K. Sahu, W. Cassing, U. Mosel, and A. Ohnishi, [*Nucl. Phys.*]{} [**A672**]{}, 376 (2000).
J. Brachmann [*et al.*]{}, [*Phys. Rev.*]{} C [**61**]{}, 024909 (2000).
H. Liu for the E895 Collaboration, [*Nucl. Phys.*]{} [**A638**]{}, 451c (1998).
J. Brachmann [*et al.*]{}, [*Nucl. Phys.*]{} [**A619**]{}, 391 (1997).
B. Friman, E.G. Nikonov, W. Nörenberg, K. Redlich and V.D. Toneev, [*Strangeness Production in Nuclear Matter and Expansion Dynamics*]{} (in preparation). R. Hagedorn and K. Redlich, [*Z. Phys.*]{} A [**27**]{}, 541 (1985).
J. Cleymans, K. Redlich, and E. Suhonen, [*Z. Phys.*]{} C [**51**]{}, 137 (1991).
K. Redlich, [*Nucl. Phys.*]{} [**A698**]{}, 94 (2002).
The NA49 Collaboration, [*nucl-ex/0205002*]{}.
V.D. Toneev, J. Cleymans, E.G. Nikonov, K. Redlich, and A.A. Shanenko, [*J. Phys.*]{} G [**27**]{}, 827 (2001).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper studies the utility maximization problem on the terminal wealth with both random endowments and proportional transaction costs. To deal with unbounded random payoffs from some illiquid claims, we propose to work with the acceptable portfolios defined via the consistent price system (CPS) such that the liquidation value processes stay above some stochastic thresholds. In the market consisting of one riskless bond and one risky asset, we obtain a type of the super-hedging result. Based on this characterization of the primal space, the existence and uniqueness of the optimal solution for the utility maximization problem are established using the convex duality analysis. As an important application of the duality theory, we provide some sufficient conditions for the existence of a shadow price process with random endowments in a generalized form similar to [@Chris22] as well as in the usual sense using acceptable portfolios.'
address:
- 'Erhan Bayraktar, Department of Mathematics, University of Michigan, 530 Church Street, Ann Arbor, MI 48109, USA'
- 'Xiang Yu, Department of Applied Mathematics, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong'
author:
- Erhan Bayraktar
- Xiang Yu
bibliography:
- 'ReferenceEndowment.bib'
date: '.'
title: 'Optimal Investment with Random Endowments and Transaction Costs: Duality Theory and Shadow Prices'
---
[^1]
Introduction
============
The optimal investment problem via expected utility maximization is a classical research topic in modern quantitative finance. In the frictionless markets, the problem with both liquid assets and illiquid contingent claims has recently received much attention and has been significantly developed. It is assumed in the model that the investor receives random payoffs from some contingent claims at the terminal time $T$. In complete markets, random endowments can be perfectly hedged using a dynamic trading portfolio with the liquid assets. As a consequence, the optimal investment problem with some payoffs reduces to the one without random endowments, but with some augmented initial wealth. In the case when the market is incomplete, the problem becomes more delicate. In particular, to build the convex duality theory, the unhedgeable random endowments demand new techniques, especially if they are unbounded, see for example [@Cv], [@kram04] and also extensions by [@KarZit03], [@Zit05] and [@Mos] when the intermediate consumption is allowed.
In the presence of market frictions, the utility maximization problem relies heavily on the definition of working portfolio processes. In virtue of the generalized market models, the conventional semimartingale properties and stochastic integrals are no longer inherited in the setting with transaction costs. New approaches are therefore required. In the multi-asset foreign exchange markets, the self-financing admissible portfolios are defined carefully using the convex solvency cones and strictly consistent price systems (SCPS). The superhedging theorems are developed under different market assumptions, see among [@MAFI:MAFI004], [@Kab], [@MAFI:MAFI180] and [@Campi06]. In a more concrete setting with one bond and one risky asset, the admissible portfolios are defined by requiring that the corresponding liquidation value processes stay above some constant lower bounds, see [@Sch2] and [@Sch33]. Recently in [@Sch33], an easy-to-apply version of the superhedging theorem is established under the assumption that the stock price process admits CPS for arbitrarily small transaction costs.
As an important add-on to the existing literature, this paper aims to study the utility maximization problem under transaction costs together with unbounded random endowments. We note that the optimal investment problem defined on multi-asset accounts with random endowments has been studied by [@campi1]. In order to apply the superhedging theorem in [@Campi06], however, [@campi1] still works with admissible portfolios and their random endowments are assumed to satisfy $\mathcal{E}_T\in\mathbb{L}^{\infty}$ in order to guarantee the existence of the optimal solution. Due to the unboundedness assumption on random endowments in our framework, the definition of admissible portfolios is no longer suitable and needs to be modified since the constant lower bound will become an unnatural constraint. In the frictionless market, a definition of acceptable portfolios is introduced by [@DS97] and [@kram04] in which some maximal elements in the set of wealth processes can serve as the stochastic thresholds. However, with transaction costs, the definition of acceptable portfolios is not clear. Indeed, the naive choice of the maximal element from the admissible portfolio processes with transaction costs can not be applied as lower bounds, see [@jacka2007] for example.
Recently, in the Kabanov’s multi-asset framework where the transaction costs are modeled by matrix-valued processes, a new definition of acceptable portfolios is proposed in [@Yu] using convex solvency cones and SCPS. One of our main contributions of the current work is to give a definition of acceptable portfolios in a simple setting consisting of one stock and one bond as in [@Sch2] and [@Sch33]. In this paper, the self-financing portfolio is called acceptable if liquidation value processes are bounded below by some processes related to all CPS $(\mathbb{Q}, \tilde{S})$, see Definition $\ref{accptpot}$. The main idea behind our definition is to choose some maximal elements as stochastic thresholds from the set of wealth processes without transaction costs where the CPS $\tilde{S}$ is taken as the underlying asset. We require in the definition that the stochastic lower bounds hold for arbitrary CPS $\tilde{S}$ and for all time $t\in[0,T]$. Comparing with the Definition $2.3$ in [@Yu] (see also Definition $2.7$ in [@Campi06]), it is worth noting that the condition of self-financing portfolio in our framework is more explicit with nice financial interpretations. However, we need to pay the price that it is generally more difficult to prove the super-hedging theorem as we need to verify a certain limit of a sequence of self-financing processes is still self-financing. With the assistance of convex solvency cones, this convergence result is taken as granted in [@Campi06] and [@Yu]. Meanwhile, unlike Definition $2.3$ in [@Yu] where the stochastic lower bounds are mandated on each portfolio process, we focus on each liquidation value process instead. Several new mathematical challenges arise since we are lack of the supermartingale property as in Lemma $2.8$ of [@Campi06] to prove the closedness property of the set of acceptable portfolios. In particular, the backward implication in Proposition $\ref{auxlem}$ of this paper is crucial for us to obtain the closedness result which is clearly not required in [@Campi06] and [@Yu]. Given the assumption that the stock price process admits CPS for all small transaction costs, we eventually are able to verify Proposition $\ref{auxlem}$ and thereby establish a super-hedging result using acceptable portfolios. The existence and uniqueness of the optimal solution are consequent on the Bipolar relationship implied by the super-hedging theorem and the standard duality approach.
Based on the duality theory, this paper also contributes to the existence of a shadow price process. Roughly speaking, a process $\hat{S}$ is called a shadow price process if it evolves inside the bid-ask spread and the optimal frictionless trading in $\hat{S}$ leads to the same utility value function as in the original market under transaction costs and two optimal portfolio processes coincide. As stated in [@Chris22], a candidate shadow price process is defined by $\hat{S}\triangleq \frac{Y^{1,\ast}}{Y^{0,\ast}}$ where $(Y^{0,\ast}, Y^{1,\ast})$ is the minimizer in the duality theory. If the stock price process $S$ is [càdlàg ]{}, $\hat{S}$ may not be a semimartingale since it may fail to be [càdlàg ]{}. To overcome this difficulty, [@Chris22] considers a shadow price process $\hat{\mathbf{S}}=(\hat{S}^p, \hat{S})$ defined in a general *sandwiched sense* such that $\hat{S}^p=\frac{Y^{1,\ast,p}}{Y^{0,\ast,p}}$ and $\hat{S}=\frac{Y^{1,\ast}}{Y^{0,\ast}}$ where $((Y^{0,\ast,p}, Y^{1,\ast,p}), (Y^{0,\ast}, Y^{1,\ast}))$ is a *sandwiched strong supermartingale deflator*, see Definition $\ref{sandsuper}$ and Definition $\ref{sanddefla}$. Despite that the shadow price process fails to be [càdlàg ]{}, the stochastic integrals are still well defined using predictable processes of finite variation as integrands. In [@Chris22], the modified self-financing and admissible portfolio processes can therefore be defined and the verification of the shadow price process can be worked out. With unbounded random endowments, the definition of sandwiched shadow price process given [@Chris22] can be extended in our setting using the modified acceptable portfolios. To the best of our knowledge, the study of a shadow price process in observing random endowments is new to the literature and we hope to add some interesting perspectives to this research direction. However, in contrast with [@Chris22], the existence of a sandwiched shadow price process can only be checked under some sufficient conditions which involve the marginal utility-based prices of the given contingent claims. It is not surprising that the unhedgeable random endowments will increase the complexity of verifications of the candidate shadow price process. But it is an interesting byproduct for us to reveal the relationship between the sandwiched shadow price and the marginal utility-based prices. In addition, we also give a formal definition of a shadow price process $\hat{S}$ in the usual sense, see Definition $\ref{shadowprice}$. The existence of a classic shadow price process can also be obtained under some stronger assumptions.
The rest of the paper is organized as follows: Section $\ref{section2}$ introduces the market model with transaction costs and the definition of acceptable portfolio processes. The utility maximization problem with unbounded random endowments is formulated in Section $\ref{section3}$. The dual space and the corresponding dual optimization problem are introduced afterwards. The main result of the duality theory is presented at the end. Section $\ref{section4}$ provides some sufficient conditions and establishes the existence of a sandwiched shadow price process consisting of a predictable and an optional strong supermartingales. The existence of a shadow price process in the classic sense is also discussed. Section $\ref{section5}$ contains the proofs of main theorems and some auxiliary results.
Market Model {#section2}
============
We consider the market which consists with one riskless bond and one risky asset. The riskless bond $B$ is assumed to be constant $1$ which amounts to serve as the numéraire. The stock price is modeled by a strictly positive adapted [càdlàg ]{}process $(S_t)_{0\leq t\leq T}$ on some filtered probability space $(\Omega, \mathcal{F}, (\mathcal{F}_t)_{0\leq t\leq T},\mathbb{P})$ satisfying the usual assumptions of right continuity and completeness. The time horizon is given by $T>0$. Moreover, we assume that $\mathcal{F}_0$ is trivial, $\mathcal{F}_T=\mathcal{F}_{T-}$ and $S_T=S_{T-}$. Trading the risky asset incurs transaction costs, that is to say, we can buy the stock at the price $S$ but can only sell it at the price $(1-\lambda)S$. Here, $S$ denotes the ask price, $(1-\lambda)S$ denotes the bid price and $[(1-\lambda)S, S]$ is called the bid-ask spread.
For a given price process $S=(S_t)_{0\leq t\leq T}$ and transaction costs $0<\lambda<1$. A $\lambda$-consistent price system ($\lambda$-CPS) is a pair $(\mathbb{Q}, \tilde{S})$ such that $\mathbb{Q}$ is a probability measure equivalent to $\mathbb{P}$, $\tilde{S}=(\tilde{S}_t)_{0\leq t\leq T}$ takes its values in the bid-ask spread $[(1-\lambda)S, S]=([(1-\lambda)S_t, S_t])_{0\leq t\leq T}$ and $\tilde{S}$ is a $\mathbb{Q}$-local martingale.
Denote $\mathcal{S}(\lambda, S)$ ( short as $\mathcal{S}$) as the set of all $\tilde{S}$ such that $(\mathbb{Q},\tilde{S})$ is a CPS with transaction costs $\lambda$. For each $\tilde{S}\in\mathcal{S}$, also denote set $\mathcal{M}(\tilde{S})$ as the set of all probability measures $\mathbb{Q}$ such that $(\mathbb{Q}, \tilde{S})$ is a $\lambda$-CPS. Define the set $\mathcal{M}(\lambda, S)$ (short as $\mathcal{M}$) by $\mathcal{M}\triangleq \bigcup_{\tilde{S}\in\mathcal{S}}\mathcal{M}(\tilde{S})$. Notice that each $\tilde{S}$ is a semimartingale under the physical probability measure $\mathbb{P}$. Given the initial wealth $a>0$, denote $\mathcal{X}(\tilde{S}, a)$ as the set of all nonnegative wealth processes in the $\tilde{S}$-market, $\tilde{S}\in\mathcal{S}$. That is $$\mathcal{X}(\tilde{S}, a)\triangleq \Big\{X\geq 0: X_t=a+(H\cdot\tilde{S})_t,\ \ \text{where $H$ is predictable and $\tilde{S}$-integrable,\ $t\in[0,T]$} \Big\}.$$ A wealth process in $\mathcal{X}(\tilde{S}, a)$ is called *maximal*, denoted by $X^{\max, \tilde{S}}$, if its terminal value $X_T^{\max,\tilde{S}}$ can not be dominated by any other processes in $\mathcal{X}(\tilde{S}, a)$.
\[Assum\] For each $0<\lambda'<1$, the price process $S$ admits $\lambda'$-CPS.
The trading strategy $\phi=(\phi^0, \phi^1)_{0\leq t\leq T}$ represents the holdings in units of the riskless and the risky asset, respectively, after rebalancing the portfolios at time $t$. $(\phi^0,\phi^1)$ is called *self-financing with transaction costs $\lambda$* (see [@Sch2] and [@Chris22]) if
- $\phi=(\phi^0,\phi^1)_{0\leq t\leq T}$ is a pair of predictable processes of finite variation.
- For any process $\phi$ of finite variation, $\phi=x+\phi^{\uparrow}-\phi^{\downarrow}$ represents its Jordan-Hahn decomposition into two non-decreasing processes $\phi^{\uparrow}$ and $\phi^{\downarrow}$ both null at zero. $(\phi^0, \phi^1)$ satisfies the condition $$\label{self}
\int_{s}^{t}d\phi_u^0 \leq -\int_s^t S_ud\phi_u^{1,\uparrow}+\int_s^t(1-\lambda)S_ud\phi_u^{1,\downarrow}$$ a.s. for all $0\leq s<t\leq T$, where $$\int_s^t S_ud\phi_u^{1,\uparrow}\triangleq \int_s^t S_u d\phi_u^{1,\uparrow, c}+\sum_{s<u\leq t}S_{u-}\triangle \phi_{u}^{1,\uparrow}+\sum_{s\leq u<t}S_u\triangle_{+}\phi_u^{1,\uparrow},\nonumber$$ and $$\int_s^t(1-\lambda)S_ud\phi_u^{1,\downarrow}\triangleq \int_s^t(1-\lambda)S_ud\phi_u^{1,\downarrow,c}+\sum_{s<u\leq t}(1-\lambda)S_{u-}\triangle \phi_u^{1,\downarrow}+\sum_{s\leq u<t}(1-\lambda)S_u\triangle_+\phi_u^{1,\downarrow}\nonumber$$ can be defined as Riemann-Stieltjes integrals since $S$ is [càdlàg ]{}. Here we define $\triangle \phi_t\triangleq \phi_t-\phi_{t-}$ and $\triangle_+ \phi_t\triangleq \phi_{t+}-\phi_t$.
It is worth noting that since $S$ is [càdlàg ]{}, we need to take care of both left and right jumps of the portfolio process $\phi$. In general, three values $\phi_{\tau-}$, $\phi_{\tau}$ and $\phi_{\tau+}$ may very well be different. If the stopping time $\tau$ is totally inaccessible, the predictability of $\phi$ implies that $\triangle \phi_{\tau}=0$ almost surely. But if the stopping time $\tau$ is predictable, it may happen that both $\triangle \phi_{\tau}\neq 0$ and $\triangle_+\phi_{\tau}\neq 0$.
Given the initial position $(\phi^0_0,\phi^1_0)=(x,0)$ in the bond and risky asset separately, where $x\in\mathbb{R}$. We define the *liquidation value* at time $t$ by $$V(\phi)_t\triangleq \phi_t^0+(\phi_t^1)^+(1-\lambda)S_t-(\phi_t^1)^-S_t.\nonumber$$
The conventional definition of working portfolios in the existing literature assumes constant thresholds for the liquidation value processes, see [@Sch2]:
\[admissiblephi\] For an $\mathbb{R}_+$-valued adapted [càdlàg ]{}process $S=(S_t)_{0\leq t\leq T}$ with transaction costs $0<\lambda<1$, a self-financing trading strategy $\phi$ is called admissible if there exists a constant $a\geq 0$ and for every $[0,T]$-valued stopping time $\tau$, $$V(\phi)_{\tau}=\phi_{\tau}^0+(\phi_{\tau}^1)^+(1-\lambda)S_{\tau}-(\phi_{\tau}^1)^-S_{\tau}\geq -a, \ \ \text{a.s.}\nonumber$$
From now on, the market is enlarged by allowing trading $N$ European contingent claims at time $t=0$ with final payoff $\mathcal{E}_T=(\mathcal{E}_T^i)_{1\leq i\leq N}$. We denote $q=(q^i)_{1\leq i\leq N}$ as static holdings in contingent claims $\mathcal{E}_T$. Without loss of generality, it is assumed that $\mathcal{E}_T^i\geq 0$ for all $1\leq i\leq N$. Each $\mathcal{E}_T^i$ may be unbounded, but we assume that $\sum_{i=1}^{N}\mathcal{E}_T^i$ is integrable uniformly with respect to the set $\mathcal{M}$ in the following sense:
\[assE\] $$\label{unic}
\lim_{m\rightarrow\infty}\sup_{\mathbb{Q}\in\mathcal{M}}\mathbb{E}^{\mathbb{Q}}\bigg[ \bigg(\sum_{i=1}^{N}\mathcal{E}_T^i\bigg)\mathbf{1}_{\{\sum_{i=1}^{N}\mathcal{E}_T^i>m\}}\bigg]=0.$$
It is clear that $(\ref{unic})$ implies that $\sum_{i=1}^{N}\mathcal{E}^i_T$ has a finite superhedging price under all $\lambda$-CPS, i.e., $\sup_{\mathbb{Q}\in\mathcal{M}}\mathbb{E}^{\mathbb{Q}}[\sum_{i=1}^{N}\mathcal{E}^i_T]<\infty$. If we require that $\mathcal{E}_T\in\mathbb{L}^{\infty}$, Assumption $\ref{assE}$ holds true. Assumption $\ref{assE}$ is a technical condition that we need for the later proof of the super-hedging theorem with all CPS.
Under Assumption $\ref{assE}$, the following holds (see the proof of Lemma $2.1$ in [@Yu]).
\[boundE\] Under Assumption $\ref{assE}$, there exists a constant $a>0$ such that for each $\tilde{S}\in\mathcal{S}$, there exits a maximal element $X^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S}, a)$ and $\sum_{i=1}^{N}\mathcal{E}_T^i\leq X_T^{\max,\tilde{S}}$.
\[assNo\] For any $q\in\mathbb{R}^N$ such that $q\neq 0$, the random variable $q\cdot\mathcal{E}_T$ is not replicable in the market under $\lambda$-CPS.
To deal with unbounded random endowments, the above definition of admissible portfolios is not appropriate. The constant lower bounds needs to be relaxed as stochastic thresholds. Following the idea of [@Yu], we shall propose the modified definition of our working portfolios as below.
\[accptpot\] For an $\mathbb{R}_+$-valued adapted [càdlàg ]{}process $S=(S_t)_{0\leq t\leq T}$ with transaction costs $0<\lambda<1$, a self-financing trading strategy $\phi$ is called **acceptable** if there exists a constant $a\geq 0$ and for each $\tilde{S}\in\mathcal{S}$, there exists a maximal element $X^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S}, a)$ such that for every $[0,T]$-valued stopping time $\tau$, $$V(\phi)_{\tau}=\phi_{\tau}^0+(\phi_{\tau}^1)^+(1-\lambda)S_{\tau}-(\phi_{\tau}^1)^-S_{\tau} \geq -X^{\max,\tilde{S}}_{\tau},\ \ \text{a.s.}\nonumber$$
Each admissible portfolio process is acceptable since any given constant $a>0$ is a maximal element in $\mathcal{X}(\tilde{S},a)$. Indeed, for each $\tilde{S}\in\mathcal{S}$, there exists $\mathbb{Q}\sim\mathbb{P}$ such that $\tilde{S}$ is a $\mathbb{Q}$-local martingale. It follows that each $\tilde{S}$ is a semimartingale and satisfies the No Free Lunch with Vanishing Risk condition, see [@Sch94] for details. Therefore a contradiction arises if there exists a maximal element in $\mathcal{X}(\tilde{S},a)$ which dominates the constant $a$.
Denote $\mathcal{A}_x(\lambda, S)$ (short as $\mathcal{A}_x$) as the set of all pairs $(\phi^0, \phi^1)\in\mathbb{L}^0(\mathbb{R}^2)$ of acceptable portfolios with transaction costs $\lambda$ starting at $\phi_{0}=(\phi_{0}^0,\phi_{0}^{1})=(x,0)$. We call $\mathcal{U}_x(\lambda, S)$ (short as $\mathcal{U}_x$) the set of all terminal values of the pair $(\phi^0,\phi^1)\in\mathcal{A}_x$, i.e., $$\mathcal{U}_x=\{(\phi^0_T,\phi^1_T): (\phi^0,\phi^1)\in\mathcal{A}_x\}.\nonumber$$ Let us also denote $\mathcal{V}_x(\lambda, S)$ (short as $\mathcal{V}_x$) the set of all terminal values of these liquidation value processes such that the position in the stock is liquidated at time $T$, i.e., $$\mathcal{V}_x=\{V_T: V_T=\phi_T^0,\ \phi_T^1=0,\ \ (\phi^0,\phi^1)\in\mathcal{A}_x\}.\nonumber$$
Contrary to the admissible portfolios, the definition of acceptable portfolio in our setting is more difficult to check since it involves all $\tilde{S}\in\mathcal{S}$ and all $0\leq t\leq T$. The following result asserts that to check the self-financing portfolio is acceptable or not, it is enough to check the terminal tim $T$.
\[auxlem\] Fix the [càdlàg ]{}, adapted process $S$ and transaction costs $0<\lambda<1$ as above and let Assumption $\ref{Assum}$ hold. Fix $\hat{a}>0$ and for each $\tilde{S}\in\mathcal{S}$, pick and fix one $\hat{X}^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S},\hat{a})$. For any $(\phi^0,\phi^1)\in\mathcal{A}_x$ and for each $\tilde{S}\in\mathcal{S}$, if we have $$\label{later}
V(\phi^0,\phi^1)_T=\phi_T^0+(\phi_T^1)^+(1-\lambda)S_T-(\phi_T^1)^-S_T\geq -\hat{X}_T^{\max,\tilde{S}},$$ then for every $[0,T]$-valued stopping time $\tau$, we also have $$\label{bef}
V(\phi^0,\phi^1)_{\tau}=\phi_{\tau}^0+(\phi_{\tau}^1)^+(1-\lambda)S_{\tau}-(\phi_{\tau}^{1})^-S_{\tau}\geq -\hat{X}_{\tau}^{\max,\tilde{S}}.$$
Proposition $\ref{auxlem}$ provides us a convenient way to check the definition of acceptable portfolios. For example, if there exists a random variable $B$ which satisfies $\sup_{\mathbb{Q}\in\mathcal{M}}\mathbb{E}^{\mathbb{Q}}[B]<\infty$ and $V(\phi^0, \phi^1)_T\geq -B$ holds, Proposition $\ref{auxlem}$ together with Lemma $\ref{boundE}$ imply that the self-financing portfolio $(\phi^0,\phi^1)$ is acceptable. More importantly, the backward implication in Proposition $\ref{auxlem}$ can replace the super-martingale property in the later proof of the super-hedging theorem. To verify a certain type of closedness, the condition $(\ref{bef})$ is crucial which demonstrates the mathematical difference between our framework and multi-asset model in [@Yu].
Utility Maximization with Unbounded Random Endowments {#section3}
=====================================================
We first introduce the set of acceptable portfolio processes with the initial wealth $x\in\mathbb{R}$ whose terminal liquidation value dominates the payoff $-q\cdot\mathcal{E}_T$ by $$\label{primeH}
\mathcal{H}(x,q)\triangleq \{ V_T: V_T+q\cdot\mathcal{E}_T \geq 0,\ \ V\in\mathcal{V}_x\},$$ where the effective domain is defined by $$\mathcal{K}\triangleq \text{int}\{ (x,q)\in\mathbb{R}^{1+N}:\mathcal{H}(x,q)\neq \emptyset\}.\nonumber$$ The set $\mathcal{H}(x,q)$ serves as the primal space that we will work on.
The agent’s preference is presented by a utility function $U:(0,\infty)\rightarrow\mathbb{R}$, which is assumed to be strictly increasing, strictly concave and continuously differentiable. It is assumed that the utility function satisfies the Inada conditions $$U'(0)\triangleq\lim_{x\rightarrow 0}U'(x)=\infty,\ \ \ \ U'(\infty)\triangleq \lim_{x\rightarrow\infty}U'(x)=0.\nonumber$$ Moreover, we make the assumption on the asymptotic elasticity of the utility function $$\label{asym}
AE(U)\triangleq \underset{x\rightarrow\infty}{\lim\sup}\frac{xU'(x)}{U(x)}<1.$$ The convex conjugate of $U(x)$ is defined by $$\tilde{U}(y)\triangleq \sup_{x>0}\Big(U(x)-xy\Big),\ \ y>0.\nonumber$$
Given the initial position $x$ and the initial static holding $q$ such that $(x,q)\in\mathcal{K}$, the agent is to maximize the expected utility defined on the terminal wealth consisting of the terminal liquidation value and the final payoff from the contingent claims. The **primal utility optimization problem** is defined by $$\label{primeu}
u(x,q)\triangleq \sup_{V_T\in\mathcal{H}(x,q)}\mathbb{E}[U(V_T+q\cdot\mathcal{E}_T)],\ \ \ (x,q)\in\mathcal{K}.$$
Let $\mathcal{C}(x,q)$ be the solid hull of the primal space $\mathcal{H}(x,q)$ $$\label{primsp}
\mathcal{C}(x,q)\triangleq \{g\in\mathbb{L}_+^0: g\leq V_T+q\cdot\mathcal{E}_T,\ \ V_T\in\mathcal{H}(x,q)\},\ \ (x,q)\in\mathcal{K}.$$
The monotonicity of $U(x)$ implies that $$u(x,q)=\sup_{g\in\mathcal{C}(x,q)}\mathbb{E}[U(g)],\ \ \ (x,q)\in\mathcal{K}.\nonumber$$
Following [@kram04], we consider the *relative interior* of the polar cone of $-\mathcal{K}$ defined by $$\mathcal{L}\triangleq \text{ri}\{(y,r)\in\mathbb{R}^{1+N}: xy+q\cdot r\geq 0\ \ \text{for all}\ (x,q)\in\mathcal{K}\}.\nonumber$$
Denote $\mathcal{B}$ as the set of density processes of $\lambda$-CPS in the sense that $$\begin{split}
\mathcal{B}\triangleq \bigg\{(Z^0, Z^1)\geq 0: &Z_t^0=\mathbb{E}\Big[\frac{d\mathbb{Q}}{d\mathbb{P}}\Big|\mathcal{F}_t\Big],\ \ \text{and}\ \ Z_t^1=\tilde{S}_tZ_t^0,\\
&\text{where}\ \mathbb{Q}\in\mathcal{M}(\tilde{S}),\ \text{for each}\ \tilde{S}\in\mathcal{S} \bigg\}.\nonumber
\end{split}$$
In general, the set $\mathcal{B}$ lacks the closedness property and a proper enlargement is needed to serve as a dual space of $\mathcal{C}(x,q)$. Let us first consider the non-negative liquidation value processes using admissible portfolios.
Starting with a strictly positive initial position $(\phi^0_0,\phi^1_0)=(x,0)$ where $ x>0$, the admissible portfolio $(\phi^0,\phi^1)$ is called $0$-admissible if for every $[0,T]$-valued stopping time $\tau$, we have $$V(\phi)_{\tau}=\phi_{\tau}^0+(\phi_{\tau}^1)^+(1-\lambda)S_{\tau}-(\phi_{\tau}^1)^-S_{\tau}\geq 0, \ \ \text{a.s.}\nonumber$$ Given $x>0$, we shall denote the set of all $0$-admissible portfolio by $\mathcal{A}_x^{\text{adm}}$ and the set of all terminal values of the $0$-admissible portfolio by $\mathcal{U}_x^{\text{adm}}$, i.e., $$\label{0admU}
\mathcal{U}_x^{\text{adm}}=\{(\phi_T^0,\phi^1_T)\in\mathbb{L}^0(\mathbb{R}^2): (\phi^0,\phi^1)\in\mathcal{A}_x^{\text{adm}}\},\ \ x>0.$$ Also denote the set $\mathcal{V}_x^{\text{adm}}$ the set of the terminal value of all $0$-admissible liquidation value processes with initial position $(x,0)$ such that the position in the stock is liquidated at $t=T$, i.e., $$\label{0admV}
\mathcal{V}_x^{\text{adm}}=\{V_T\in\mathbb{L}^0_+(\mathbb{R}): \exists (\phi_T^0,\phi_T^1)\in\mathcal{U}_x^{\text{adm}}\ \text{such that}\ \phi_T^0=V_T,\ \phi_T^1=0\}.$$
In our framework, the stock price process $S=(S_t)_{0\leq t\leq T}$ is [càdlàg ]{}. All self-financing portfolio processes $(\phi^0_t, \phi_t^1)_{0\leq t\leq T}$ need to be predictable of finite variation which can have both left and right jumps in order to obtain that $\mathcal{U}_x^{\text{adm}}$ is closed under convergence in probability, see [@Campi06] and [@Sch33] for details. To retain supermartingale properties, a new limit is required instead of Fatou’s limit; see [@Chris] and [@Chris22]. The convergence in probability at all finite stopping times and the concept of *optional strong supermartingales* seem to be tailor-made for analyzing problems with transaction costs. The following definition introduced by [@Chris] plays an important role in the proper definition of the dual space.
An optional process $X=(X_t)_{0\leq t\leq T}$ is called an optional strong supermartingale if, for all stopping times $0\leq \sigma\leq \tau\leq T$, we have $$\mathbb{E}[X_{\tau}|\mathcal{F}_{\sigma}]\leq X_{\sigma},\nonumber$$ where we impose that $X_{\tau}$ is integrable for any $[0,T]$-valued stopping time $\tau$.
We shall extend the dual space using the optional strong supermartingales. For $y>0$, $$\label{deflator}
\begin{split}
\mathcal{Z}(y)\triangleq \bigg\{&(Y^0, Y^1)\ \text{are nonnegative optional strong supermartingales}: Y_0^0=y, \frac{Y^1}{Y^0}\in[(1-\lambda)S, S], \\
&\phi^0Y^0+\phi^1Y^1\ \text{is a non-negative optional strong supermartingale},\ \forall(\phi^0,\phi^1)\in\mathcal{A}_1^{\text{adm}}\bigg\},
\end{split}$$ and $$\label{dualY}
\mathcal{Y}(y)\triangleq \{Y_T\in\mathbb{L}_+^0(\mathbb{R}): \exists (Y^0, Y^1)\in\mathcal{Z}(y)\ \text{with}\ Y_T=Y_T^0\},\ \ y>0.$$ Due to Proposition $1.6$ in [@Sch2], we have that $y\mathcal{B}\subset\mathcal{Z}(y)$.
Given $(y,r)\in\mathcal{L}$, we are interested in the subset $$\label{dualsp}
\mathcal{Y}(y,r)\triangleq \{Y_T\in\mathcal{Y}(y): \mathbb{E}[Y_T(V_T+q\cdot\mathcal{E}_T)]\leq xy+q\cdot r,\ \ V_T\in\mathcal{H}(x,q),\ \ (x,q)\in\mathcal{K}\},$$ which is the correct dual space to work on since the random endowments can be hidden by its definition.
Define the abstract set $\mathcal{D}(y,r)$ as the solid hull of $\mathcal{Y}(y,r)$, $$\mathcal{D}(y,r)=\{h\in\mathbb{L}_+^0(\mathbb{R}^2): h\leq Y_T,\ \ Y_T\in\mathcal{Y}(y,r)\},\ \ (y,r)\in\mathcal{L}.\nonumber$$
We are now ready to define the corresponding **dual optimization problem** to problem $(\ref{primeu})$ by $$\label{dualv}
v(y,r)\triangleq \inf_{Y_T\in\mathcal{Y}(y,r)}\mathbb{E}[\tilde{U}(Y_T)]=\inf_{h\in\mathcal{D}(y,r)}\mathbb{E}[\tilde{U}(h)],\ \ (y,r)\in\mathcal{L}.$$
The following theorem constitutes the duality theory on the existence and uniqueness of the optimal solution to the utility maximization problem $(\ref{primeu})$.
\[mainthm\] Let Assumption $\ref{Assum}$, $\ref{assE}$ and $\ref{assNo}$ and condition $(\ref{asym})$ hold and we assume that $$u(x,q)<\infty\ \ \text{for some}\ \ (x,q)\in\mathcal{K}.\nonumber$$ Then we have
- The function $u$ is finitely valued on $\mathcal{K}$ and the function $v$ is finitely valued on $\mathcal{L}$. The value functions $u$ and $v$ are conjugate $$\begin{split}
u(x,q)=\inf_{(y,r)\in\mathcal{L}}\Big( v(y,r)+xy+q\cdot r\Big),\ \ \ (x,q)\in\mathcal{K},\\
v(y,r)=\sup_{(x,q)\in\mathcal{K}}\Big(u(x,q)-xy-q\cdot r \Big),\ \ \ (y,r)\in\mathcal{L}.\nonumber
\end{split}$$
- The optimal solution $Y_T^{\ast}(y,r)$ to $(\ref{dualv})$ exists and is unique for all $(y,r)\in\mathcal{L}$.
- The optimal solution $V_T^{\ast}(x,q)$ to $(\ref{primeu})$ exists and is unique for all $(x,q)\in\mathcal{K}$.
- There are $(\phi^{0,\ast}, \phi^{1,\ast})\in\mathcal{A}_x$ and $(Y^{0,\ast}, Y^{1,\ast})\in\mathcal{Z}(y)$ such that $$V(\phi^{0,\ast}, \phi^{1,\ast})_T=V_T^{\ast}(x,q),\ \ \text{and}\ \ \ Y_T^{0,\ast}=Y_T^{\ast}(y,r).\nonumber$$
- The sub-differential of $u$ maps $\mathcal{K}$ into $\mathcal{L}$, i.e., $$\partial u(x,q)\subset \mathcal{L},\ \ (x,q)\in\mathcal{K}.\nonumber$$
- If $(y,r)\in\partial u(x,q)$, the optimal solutions are related by $$\label{ok1}
\begin{split}
Y^{\ast}_T(y,r)&=U'(V^{\ast}_T(x,q)+q\cdot \mathcal{E}_T),\\
\mathbb{E}[Y^{\ast}_T(y,r)(V_T^{\ast}(x,q)+q\cdot\mathcal{E}_T)]&=xy+q\cdot r.
\end{split}$$
Denote $\mathcal{P}(x,q;U)$ the set of all marginal utility-based prices at $(x,q)\in\mathcal{K}$ $$\label{marprice}
\mathcal{P}(x,q;U)\triangleq\{p\in\mathbb{R}: u(x-q'p, q+q')\leq u(x,q)\ \text{for all}\ q'\in\mathbb{R}\}.$$ The definition asserts that the agent’s holdings $q$ in $\mathcal{E}_T$ is optimal in the model where the contingent claims can be traded at the marginal utility-based price $p$ at time zero. Equivalently, see [@kram04] and [@Hugo], we have $$\label{marprice2}
\mathcal{P}(x,q;U)= \Big\{\frac{r}{y}: (y,r)\in\partial u(x,q)\Big\}.$$ The duality theory above is the first step to perform the sensitivity analysis and first order expansion of the marginal utility-based prices as in [@Sirbu2006] and [@Kramkov20071606] but in market models with proportional transaction costs. It is possible for us to discuss the sensitivity analysis of the marginal utility-based price also on the transaction costs $0<\lambda<1$. These will be left as some future research projects.
Connections to the Shadow Prices {#section4}
================================
In this section, we analyze the connections between the duality theory and the existence of the shadow price process in the frictionless market with random endowments. However, as we can expect, the existence of a shadow price may not be guaranteed in general and we are interested in addressing this issue by providing some reasonable sufficient conditions. To simplify the notation, we shall take $N=1$ and hence $q\in\mathbb{R}$.
First, we introduce the concept of a classic shadow price in the usual sense. To this end, we need some preparations of definitions. For a fixed $\lambda$-CPS $(\mathbb{Q}, \hat{S})$, i.e., $\hat{S}\in\mathcal{S}$ and the positive initial wealth $x>0$, we define the set of self-financing and $0$-admissible trading strategies in the market without transaction costs by $$\begin{split}
\mathcal{A}_x^{\text{adm}}(\hat{S})\triangleq \bigg\{(\phi^0, \phi^1): \ \ &x+\int_0^t\phi_u^1d\hat{S}_u\geq 0,\ \forall t\in[0,T],\ \ \text{$\phi^1$ is predictable and $\hat{S}$-integrable},\\
&\phi_t^0=x+\int_0^t\phi_u^1d\hat{S}_u-\phi_t^1\hat{S}_t,\ \forall t\in[0,T] \bigg\}.\nonumber
\end{split}$$
The set of wealth process under $0$-admissible strategies in the $\hat{S}$-market without transaction costs is define by $$\mathcal{X}(\hat{S},x)\triangleq \bigg\{X: X_t=x+\int_0^t\phi_u^1d\hat{S}_u\geq 0,\ \ \forall t\in[0,T],\ \ (\phi^0,\phi^1)\in\mathcal{A}_x^{\text{adm}}(\hat{S}) \bigg\},\ \ x>0.\nonumber$$ We denote $X^{\max}$ the maximal element in the set $\mathcal{X}(\hat{S}, x)$ for some $x>0$ and $\mathcal{X}(\hat{S})\triangleq \bigcup_{x>0}\mathcal{X}(\hat{S}, x)$.
\[notransactions\] For a fixed $\hat{S}\in\mathcal{S}$, the self-financing portfolio is called acceptable in the $\hat{S}$-market without transaction costs if the wealth process $X$ admits a representation $X=X'-X^{\max}$, where $X'$ is a wealth process under some $0$-admissible portfolios and $X^{\max}$ is a maximal element in $\mathcal{X}(\hat{S})$. That is to say, the set of all acceptable portfolios $\mathcal{A}_x(\hat{S})$ can be written as $$\label{shadowaccpt}
\begin{split}
\mathcal{A}_x(\hat{S})\triangleq \bigg\{(\phi^0, \phi^1):\ \ &x+\int_0^t\phi_u^1d\hat{S}_u=X'_t-X_t^{\max},\ \phi^1\ \text{is predictable and $\hat{S}$-integrable},\\
&\text{where}\ X', X^{\max}\in\mathcal{X}(\hat{S})\ \text{and}\ \phi_t^0=x+\int_0^t\phi_u^1d\hat{S}_u-\phi_t^1\hat{S}_t,\ \forall t\in[0,T] \bigg\}.\nonumber
\end{split}$$
The set of all terminal wealth processes in the $\hat{S}$-market is denoted by $$\mathcal{V}_x(\hat{S})\triangleq \bigg\{X_T\in\mathbb{L}^0(\mathbb{R}): X_T=x+\int_0^T\phi_u^1d\hat{S}_u=\phi_T^0+\phi_T^1\hat{S}_T,\ \ (\phi^0,\phi^1)\in\mathcal{A}_x(\hat{S}) \bigg\},\nonumber$$ and the set of terminal wealth values under acceptable portfolios dominating the payoff $-q\mathcal{E}_T$ is defined by $$\mathcal{H}(x,q;\hat{S})\triangleq \{X_T: X_T+q\mathcal{E}_T\geq 0,\ X_T\in\mathcal{V}_x(\hat{S}) \}.\nonumber$$ The corresponding effective domain is given by $$\mathcal{K}(\hat{S})\triangleq \text{int} \{(x,q)\in\mathbb{R}^{2}: \mathcal{H}(x,q;\hat{S})\neq \emptyset \}.\nonumber$$
\[shadowprice\] A process $\hat{S}\in\mathcal{S}$, i.e., $(\mathbb{Q}, \hat{S})$ is a $\lambda$-CPS for some $\mathbb{Q}\in\mathcal{M}(\hat{S})$, is called a *classic shadow price process*, if the optimal solution $(\hat{\phi}^0, \hat{\phi}^1)$ with the terminal wealth $X(\hat{\phi}^0,\hat{\phi}^1)_T\in\mathcal{H}(x,q;\hat{S})$ to the frictionless utility maximization problem $$\label{frictionless}
u(x,q;\hat{S})\triangleq \sup_{X_T\in\mathcal{H}(x,q;\hat{S})}\mathbb{E}[U(X_T+q\mathcal{E}_T)],$$ exists for $(x,q)\in\mathcal{K}\cap\mathcal{K}(\hat{S})$ and coincides with the optimal solution $\phi^{\ast}=(\phi^{0,\ast}, \phi^{1, \ast})$ to the problem $(\ref{primeu})$ under transaction costs $\lambda$. In particular, we have $u(x,q)=u(x,q;\hat{S})$.
We shall make the two important observations.
Our definition of the classic shadow price process $\hat{S}$ is more restrictive than [@Chris22] and [@Chris33] and a classic shadow price process $\hat{S}$ satisfies NFLVR condition by its definition. The acceptable portfolio process differs from the admissible portfolio process and the existence of equivalent local martingale measures for $\hat{S}$ are required to build the duality theory in the shadow price market without transaction costs. Therefore, unlike [@Chris33], even when the stock price process $S$ is continuous, the existence of consistent local martingale system $(Z^0,Z^1)$ (see [@BayYu] for its definition and the equivalent characterization) is no longer the sufficient condition for the existence of a classic shadow price process. The existence of a CPS becomes crucial in our framework with random endowments.
Comparing Definition $\ref{accptpot}$ and Definition $\ref{notransactions}$, it is easy to see that $\mathcal{A}_x\subseteq\mathcal{A}_x(\hat{S})$ since we require $\hat{S}\in\mathcal{S}$. Therefore, it follows that $\mathcal{H}(x,q)\neq\emptyset$ implies that $\mathcal{H}(x,q;\hat{S})\neq\emptyset$ and hence $\mathcal{K}\subset\mathcal{K}(\hat{S})$. In Definition $\ref{shadowprice}$, it is then enough to require $(x,q)\in\mathcal{K}$ for the well-posedness of both $u(x,q)$ and $u(x,q;\hat{S})$.
If a classic shadow price $\hat{S}$ exists, an optimal strategy $(\hat{\phi})=(\hat{\phi}^0, \hat{\phi}^1)$ for the utility maximization problem $(\ref{frictionless})$ in the frictionless market can be realized in the market with transaction costs. In particular, we wish to see that the optimal strategy $(\phi^{0,\ast}, \phi^{1,\ast})$ to the problem $(\ref{primeu})$ under transaction costs only trades if $\hat{S}$ is at the bid or ask price, i.e., $$\{d\phi^{1,\ast}>0\}\subseteq\{\hat{S}=S\},\ \ \text{and}\ \ \ \{d\phi^{1,\ast}<0\}\subseteq\{\hat{S}=(1-\lambda)S\}\nonumber$$ in the sense that $$\label{stratno}
\begin{split}
\{d\phi^{1,\ast,c}>0\} &\subseteq\{\hat{S}=S\},\ \ \ \ \{d\phi^{1,\ast,c}<0\}\subseteq\{ \hat{S}=(1-\lambda)S\},\\
\{\triangle \phi^{1,\ast}>0 \}&\subseteq \{\hat{S}_{-}=S_{-}\},\ \ \ \ \{ \triangle \phi^{1,\ast}<0\}\subseteq \{\hat{S}_{-}=(1-\lambda)S_{-}\},\\
\{\triangle_{+} \phi^{1,\ast}>0 \}&\subseteq \{\hat{S}=S\},\ \ \ \ \{ \triangle_{+} \phi^{1,\ast}<0\}\subseteq \{\hat{S}=(1-\lambda)S\}.
\end{split}$$
Define the dual space for the shadow price $\hat{S}$ by $$\begin{split}
\mathcal{Y}(y; \hat{S})\triangleq \bigg\{Y\geq 0:& Y_0=y\ \text{and}\ Y_t(\phi_t^0+\phi_t^1\hat{S}_t)=Y_t\bigg(1+\int_0^t\phi^1_ud\hat{S}_u\bigg),\ 0\leq t\leq T,\\
&\text{is a {c\`{a}dl\`{a}g }supermartingale for all $(\phi^0, \phi^1)\in\mathcal{A}_1^{\text{adm}}(\hat{S})$} \bigg\}.\nonumber
\end{split}$$
The of the polar cone of $-\mathcal{K}(\hat{S})$ is denoted by $$\mathcal{L}(\hat{S})\triangleq \text{ri}\{(y,r)\in\mathbb{R}^{2}: xy+qr\geq 0\ \text{for all}\ (x,q)\in\mathcal{K}(\hat{S})\}.\nonumber$$
Define $\mathcal{Y}(y,r;\hat{S})$ as the subset of $\mathcal{Y}(y;\hat{S})$ by $$\mathcal{Y}(y,r;\hat{S})\triangleq \{Y_T\in\mathcal{Y}(y;\hat{S}):\ \mathbb{E}[Y_T(X_T+q\mathcal{E}_T)]\leq xy+qr,\ \ X_T\in\mathcal{H}(x,q;\hat{S}),\ (x,q)\in\mathcal{K}(\hat{S})\}.\nonumber$$
The dual optimization problem to $(\ref{frictionless})$ is then formulated as $$\label{shadowdualv}
v(y,r;\hat{S})=\inf_{Y_T\in\mathcal{Y}(y,r;\hat{S})}\mathbb{E}[\tilde{U}(Y_T)].$$
Example $4.1$ of [@Chris22] shows that if $S$ is [càdlàg ]{}, the dual optimizer $(Y^{0,\ast}, Y^{1,\ast})$ to problem $(\ref{dualv})$ and the candidate of the shadow price process $\hat{S}\triangleq \frac{Y^{1,\ast}}{Y^{0,\ast}}$ may not be [càdlàg ]{}and therefore may not be semimartingales. The existence of a classic shadow price may fail in general. However, the stochastic integral $\int_0^t\hat{\phi}^1_ud\hat{S}_u$ can still be well-defined as long as $\hat{\phi}^1$ is a predictable process of finite variation and $\hat{S}$ is [làdlàg ]{}(see [@Chris] and [@Chris22]) and $$\label{wrongsha}
\int_0^t\hat{\phi}^1_ud\hat{S}_u=\int_0^t\hat{\phi}_u^{1,c}d\hat{S}_u+\sum_{0<u\leq t}\triangle\hat{\phi}_u^{1}(\hat{S}_t-\hat{S}_{u-})+\sum_{0\leq u<t}\triangle_+\hat{\phi}_u^1(\hat{S}_t-\hat{S}_u),\ \ 0\leq t\leq T.$$ The integral above can still be interpreted as the gains from trading of the self-financing portfolio $(\hat{\phi}^1_t)_{0\leq t\leq T}$ without transaction costs under the price process $\hat{S}=(\hat{S}_t)_{0\leq t\leq T}$, although $\hat{S}$ is not a semimartingale. Therefore, the natural question is that whether or not can we choose the quotient $\hat{S}=\frac{Y^{1,\ast}}{Y^{0,\ast}}$ as the underlying asset and define the wealth process in this general shadow price market by the stochastic integral $(\ref{wrongsha})$ ? Unfortunately, the answer is negative in general. Example $4.2$ in [@Chris22] points out that we may not be able to verify properties $(\ref{stratno})$ using the wealth process defined by $(\ref{wrongsha})$. In particular, it is hard to guarantee that $$\{\triangle \phi^{1,\ast}>0 \}\subseteq \{\hat{S}_{-}=S_{-}\},\ \ \ \ \{ \triangle \phi^{1,\ast}<0\}\subseteq \{\hat{S}_{-}=(1-\lambda)S_{-}\},\nonumber$$ where $\phi^{1,\ast}$ is the optimal portfolio process in Theorem $\ref{mainthm}$. As a consequence, we are not able to verify that $(\phi^{0,\ast}, \phi^{1,\ast})$ is the optimal solution in the shadow price market driven by $\hat{S}=\frac{Y^{1,\ast}}{Y^{0,\ast}}$. It requires us to modify either the definition of the general shadow price $\hat{S}$ or the wealth process given by $(\ref{wrongsha})$.
To examine the shadow price process in a correct generalized form, Example $4.2$ in [@Chris22] shows that it is important to introduce the following concepts.
A predictable process $X=(X_t)_{0\leq t\leq T}$ is called a predictable strong supermartingale if, for all predictable stopping times $0\leq \sigma\leq \tau\leq T$, we have $$\mathbb{E}[X_{\tau}|\mathcal{F}_{\sigma}]\leq X_{\sigma},\nonumber$$ where we impose that $X_{\tau}$ is integrable for any $[0,T]$-valued predictable stopping time $\tau$.
\[sandsuper\] A sandwiched strong supermartingale is a pair $\mathbf{X}=(X^p,X)$ such that $X^p$ (resp. $X$) is a predictable (resp. optional) strong supermartingale and such that $$\label{sand}
X_{\tau-}\geq X_{\tau}^p\geq \mathbb{E}[X_{\tau}|\mathcal{F}_{\tau-}],$$ for all predictable stopping times $\tau$.
For a sandwiched strong supermartingale $\mathbf{X}=(X^p, X)$ and a predictable process $\phi$ of finite variation, as in [@Chris], the stochastic integral is defined in *a sandwiched sense* by $$\int_0^t\phi_ud\mathbf{X}_u\triangleq \int_0^t\phi_u^cdX_u+\sum_{0\leq u<t}\triangle\phi_u(X_t-X_u^p)+\sum_{0<u\leq t}\triangle_+\phi_u(X_t-X_u),\ \ 0\leq t\leq T. \nonumber$$
\[sanddefla\] We call $\mathbf{Y}=(Y^p, Y)=((Y^{0,p}, Y^{1,p}), (Y^0, Y^1))$ a sandwiched strong supermartingale deflator if $Y=(Y^0, Y^1)\in\mathcal{Z}(y)$ (see $(\ref{deflator})$) and $(Y^{0,p}, Y^0)$ and $(Y^{1,p}, Y^1)$ are sandwiched strong supermartingales and the process $\hat{S}^p$ stays inside the bid-ask spread, $$\hat{S}^p_t=\frac{Y^{1,p}}{Y^{0,p}}\in\ [(1-\lambda)S_{t-}, S_{t-}],\ \ t\in[0,T]. \nonumber$$
Following the proof of Lemma $A.1$ of [@Chris22], by passing to the forward convex combinations if necessary, the following convergence results hold.
\[seplemma\] Fix $(x,q)\in\mathcal{K}$. For any $(y,r)\in\partial u(x,q)$, there exists a minimizing sequence $Z^n(y,r)=(Z_t^{0,n}(y,r), Z_t^{1,n}(y,r))_{0\leq t\leq T}$ in $\mathcal{B}(1)$ to the dual problem $(\ref{dualv})$, i.e., $$\mathbb{E}[\tilde{U}(yZ_T^{0,n}(y,r))]\searrow v(y,r),\ \ \text{as}\ n\rightarrow\infty,\nonumber$$ and a sandwiched strong supermartingale deflator $\mathbf{Y}^{\ast}(y,r)=(Y^{\ast, p}(y,r), Y^{\ast}(y,r))$ such that $$\label{con111}
(yZ^{0,n}_{\tau-}(y,r), yZ^{1,n}_{\tau-}(y,r))\xrightarrow{\mathbb{P}}(Y^{0,\ast,p}_{\tau}(y,r), Y^{1,\ast,p}_{\tau}(y,r)),$$ and $$\label{con222}
(yZ^{0,n}_{\tau}(y,r), yZ^{1,n}_{\tau}(y,r))\xrightarrow{\mathbb{P}}(Y^{0,\ast}_{\tau}(y,r), Y^{1,\ast}_{\tau}(y,r)),$$ as $n\rightarrow\infty$ for all $[0,T]$-valued stopping time $\tau$, where $Y^{0,\ast}(y,r)$ is the dual optimizer to the problem $(\ref{dualv})$.
\[importantnew\] Fix $(x,q)\in\mathcal{K}$. Assume that there exists some $(y,r)\in\partial u(x,q)$ such that the minimizing sequence $Z^n(y,r)=(Z_t^{0,n}(y,r), Z_t^{1,n}(y,r))_{0\leq t\leq T}$ in $\mathcal{B}(1)$ satisfies $$\label{marginalprice}
\underset{n\rightarrow\infty}{\lim\inf}\ \mathbb{E}[Z_T^{0,n}(y, r)\mathcal{E}_T]=\frac{r}{y}.$$
Denote $\mathcal{P}$ the set of all arbitrage-free prices. For any $Z\in\mathcal{B}(1)$, we have $\mathbb{E}[Z_T^{0}\mathcal{E}_T]\in\mathcal{P}$ and $\mathcal{P}(x,q;U)\subset\mathcal{P}$, where $\mathcal{P}(x,q;U)$ is the set of all marginal utility-based prices, see $(\ref{marprice})$ and $(\ref{marprice2})$ for its definition. The condition $(\ref{marginalprice})$ states that there exists a marginal utility based price $\frac{r}{y}\in\mathcal{P}(x,q;U)$ such that the minimizing sequence $Z^{0,n}(y,r)$ satisfies $\underset{n\rightarrow\infty}{\lim\inf}\ \mathbb{E}[Z_T^{0,n}(y, r)\mathcal{E}_T]=\frac{r}{y}$. Here, we reveal a sufficient condition for the existence of a sandwiched shadow price process related to the property of some marginal utility-based prices. The following two examples provide some concrete market models satisfying the condition $(\ref{marginalprice})$ for separate cases when $q>0$ and $q<0$.
Consider $(x,q)\in\mathcal{K}$ and $q>0$ and we assume that $\mathcal{E}_T\leq a(1-\lambda)S_T$ for some $a>0$. By rescaling the values of $S_0$ and $Y_0^{1,\ast}(y,r)$, we can assume without loss of generality that there exists some $(y,r)\in\partial u(x,q)$ such that $$\begin{split}
&(1-\lambda)S_0\leq \frac{r}{ay}\leq S_0,\\
&\text{and}\ \ \ Y_0^{1,\ast}(y,r)=\frac{r}{a}.\nonumber
\end{split}$$ Consider the minimizing sequence $(Z^{0,n}(y,r), Z^{1,n}(y,r))\in\mathcal{B}(1)$, we have $\tilde{S}^n\triangleq \frac{Z^{1,n}(y,r)}{Z^{0,n}(y,r)}\in[(1-\lambda)S, S]$. Therefore, it is easy to see that $$\begin{split}
\mathbb{E}[Z_T^{0,n}(y, r)\mathcal{E}_T]&\leq a\mathbb{E}[Z_T^{0,n}(y,r)(1-\lambda)S_T]\\
&\leq a\mathbb{E}[Z_T^{0,n}(y,r)\tilde{S}_T^n]\\
&=a\mathbb{E}[Z_T^{1,n}(y,r)]\leq aZ_0^{1,n}(y,r).\nonumber
\end{split}$$ Since $yZ_0^{1,n}(y,r)$ converges to $Y_0^{1,\ast}(y,r)=\frac{r}{a}$, it follows that $$\label{exampleinq}
\underset{n\rightarrow\infty}{\lim\inf}\mathbb{E}[Z_T^{0,n}(y, r)\mathcal{E}_T]\leq \underset{n\rightarrow\infty}{\lim\inf}aZ_0^{1,n}(y,r)\leq \frac{r}{y}.$$
On the other hand, for the same pair $(y,r)\in\partial u(x,q)$, we have that $$\label{exampleinq2}
\begin{split}
xy+qr=&\mathbb{E}[Y_T^{0,\ast}(y,r)(V_T^{\ast}(x,q)+q\mathcal{E}_T)]\\
\leq &\underset{n\rightarrow\infty}{\lim\inf}\mathbb{E}[yZ_T^{0,n}(y,r)(V_T^{\ast}(x,q)+q\mathcal{E}_T)]\\
\leq & xy+qy\underset{n\rightarrow\infty}{\lim\inf}\mathbb{E}[Z_T^{0,n}(y,r)\mathcal{E}_T].\nonumber
\end{split}$$ Since $q>0$, it follows that $\underset{n\rightarrow\infty}{\lim\inf}\mathbb{E}[Z_T^{0,n}(y,r)\mathcal{E}_T]\geq \frac{r}{y}$. The last inequality together with $(\ref{exampleinq})$ implies $(\ref{marginalprice})$.
Consider $(x,q)\in\mathcal{K}$ and $q<0$. Assume that $\mathcal{E}_T\geq aS_T$ for some constant $a>0$. Without loss of generality, we can assume that there exists some $(y,r)\in\partial u(x,q)$ such that $$\label{exampleinq3}
\mathbb{E}[Y_T^{1,\ast}(y,r)]\geq \frac{r^{\ast}}{a},$$ where $r^{\ast}$ is defined as the smallest value of $r$ such that $(y,r)\in\partial u(x,q)$. This can be achieved either by assuming that $a$ is large enough or by rescaling the process $S$ and the process $Y^{1,\ast}$.
Since for any $(y,r)$ and $(\bar{y},\bar{r})\in\partial u(x,q)$, we always have $Y_T^{0,\ast}(y,r)=Y_T^{0,\ast}(\bar{y},\bar{r})$. We shall pick the pair $(y^{\ast}, r^{\ast})\in\partial u(x,q)$. For the minimizing sequence $(Z^{0,n}(y^{\ast},r^{\ast}), Z^{1,n}(y^{\ast},r^{\ast}))\in\mathcal{B}(1)$, Fatou’s lemma together with $(\ref{exampleinq3})$ and the fact that $\frac{Y^{1,\ast}(y^{\ast}, r^{\ast})}{Y^{0,\ast}(y^{\ast}, r^{\ast})}\in [(1-\lambda)S, S]$ imply that $$\begin{split}
\underset{n\rightarrow\infty}{\lim\inf}\mathbb{E}[Z_T^{0,n}(y^{\ast},r^{\ast})\mathcal{E}_T]\geq &\mathbb{E}\Big[\frac{1}{y^{\ast}}Y_T^{0,\ast}(y^{\ast}, r^{\ast})\mathcal{E}_T\Big]\geq \mathbb{E}\Big[\frac{1}{y^{\ast}}Y_T^{0,\ast}(y^{\ast}, r^{\ast})aS_T\Big]\\
\geq &\mathbb{E}\bigg[\frac{1}{y^{\ast}}Y_T^{0,\ast}(y^{\ast}, r^{\ast})a \frac{Y_T^{1,\ast}(y^{\ast}, r^{\ast})}{Y^{0,\ast}_T(y^{\ast}, r^{\ast})}\bigg]\\
=& a\mathbb{E}\Big[\frac{1}{y^{\ast}}Y_T^{1,\ast}(y^{\ast}, r^{\ast})\Big]\geq \frac{r^{\ast}}{y^{\ast}}.\nonumber
\end{split}$$ For the same pair $(y^{\ast},r^{\ast})\in\partial u(x,q)$, following $(\ref{exampleinq2})$ and the fact $q<0$, we will have $$\underset{n\rightarrow\infty}{\lim\inf}\mathbb{E}[Z_T^{0,n}(y^{\ast},r^{\ast})\mathcal{E}_T]\leq\frac{r^{\ast}}{y^{\ast}},\nonumber$$ which verifies $(\ref{marginalprice})$ with the choice of $r=r^{\ast}$ and $y=y^{\ast}$.
The next theorem provides the existence of a sandwiched supermartingale deflator related to the dual minimizer of the problem $(\ref{dualv})$, and hence the candidate sandwiched shadow price process is well-defined.
\[theorem1\] Fix $(x,q)\in\mathcal{K}$. Under all assumptions of Theorem $\ref{mainthm}$ and under Assumption $\ref{importantnew}$, there exists at least a pair of $(y,r)\in\partial u(x,q)$ and for the optimizer $\phi^{\ast}(x,q)=(\phi^{0,\ast}(x,q), \phi^{1,\ast}(x,q))$ to the primal utility maximization problem $(\ref{primeu})$, we have $$\label{equiphi}
Y^{0,\ast}(y,r)\phi^{0,\ast}(x,q)+Y^{1,\ast}(y,r)\phi^{1,\ast}(x,q)=Y^{0,\ast}(y,r)(x+\phi^{1,\ast}(x,q)\cdot\hat{\mathbf{S}}),$$ where $$\hat{\mathbf{S}}=(\hat{S}^p, \hat{S})=\Big(\frac{Y^{1,\ast,p}(y, r)}{Y^{0,\ast,p}(y, r)}, \frac{Y^{1,\ast}(y, r)}{Y^{0,\ast}(y, r)}\Big)\nonumber$$ and $$\label{selfphixq}
(\phi^{1,\ast}(x,q)\cdot\hat{\mathbf{S}})_t\triangleq \int_0^t\phi^{1,\ast,c}_u(x,q)d\hat{S}_u+\sum_{0\leq u<t}\triangle \phi^{1,\ast}_u(x,q)(\hat{S}_t-\hat{S}^p_u)+\sum_{0<u\leq t}\triangle_+\phi_u^{1,\ast}(x,q)(\hat{S}_t-\hat{S}_u).$$ It follows that $$\label{behavphi}
\begin{split}
\{d\phi^{1,\ast,c}(x,q)>0\}&\subseteq \{\hat{S}=S\},\ \ \ \ \{d\phi^{1,\ast,c}(x,q)<0\}\subseteq\{\hat{S}=(1-\lambda)S\},\\
\{\triangle\phi^{1,\ast}(x,q)>0\}&\subset\{\hat{S}^p=S_{-}\},\ \ \ \ \{\triangle\phi^{1,\ast}(x,q)<0\}\subseteq \{\hat{S}^p=(1-\lambda)S_{-}\},\\
\{\triangle_+\phi^{1,\ast}(x,q)>0\}&\subset\{\hat{S}=S\},\ \ \ \ \ \{\triangle_+\phi^{1,\ast}(x,q)<0\}\subseteq \{\hat{S}=(1-\lambda)S\}.
\end{split}$$
For any sandwiched supermartingale deflator $\mathbf{Y}=(Y^p,Y)$ with the associated price process $\hat{\mathbf{S}}=(\hat{S}^p,\hat{S})=\Big(\frac{Y^{1,p}}{Y^{0,p}}, \frac{Y^{1}}{Y^0}\Big)$, and any acceptable trading strategy $\phi\in\mathcal{A}_x$, it is easy to verify that the liquidation value $V(\phi^0, \phi^1)$ satisfies $$\begin{split}
V(\phi^0,\phi^1)_t&=\phi_t^0+(\phi_t^1)^+(1-\lambda)S_t-(\phi_t^1)^{-}S_t\\
&\leq x+\int_0^t\phi_u^{1,c}d\hat{S}_u+\sum_{0\leq u<t}\triangle \phi_u^1(\hat{S}_t-\hat{S}^p_u)+\sum_{0<u\leq t}\triangle_+\phi_u^1(\hat{S}_t-\hat{S}_u)=x+(\phi^1\cdot\hat{\mathbf{S}})_t. \nonumber
\end{split}$$
Thanks to $(\ref{equiphi})$ and $(\ref{behavphi})$, we are able to verify that the optimal strategy $(\phi^{0,\ast}, \phi^{1,\ast})$ only trades when the sandwiched shadow price process $\hat{\mathbf{S}}=(\hat{S}^p, \hat{S})$ assumes the least favorable position in the bid-ask spread.
In order to verify the existence of the sandwiched shadow price process, it is important to give a new definition of the acceptable portfolios for the underlying price process $\hat{\mathbf{S}}$. Clearly, the definition of $\mathcal{A}_x(\hat{S})$ in $(\ref{shadowaccpt})$ is too wide in general since we can only work with integrand processes of finite variation. The equality $(\ref{equiphi})$ gives us a hint of the definition of self-financing portfolios for the sandwiched shadow prices. Let us also recall that the important property behind the concept of a sandwiched shadow price $\hat{\mathbf{S}}$ is that any self-financing and acceptable portfolio trading with $\hat{\mathbf{S}}$ can not do better than the optimizer $(\phi^{0,\ast},\phi^{1,\ast})$ given in Theorem $\ref{mainthm}$ for the price process $S$ with transaction costs $\lambda$. Moreover, the strategy $(\phi^{0,\ast}, \phi^{1,\ast})$ trading in $\hat{\mathbf{S}}$ without transaction costs brings the same expected utility value as the case of trading in $S$ under transaction costs $\lambda$. Similar to the definition of admissible portfolios in [@Chris22], we can now give the following modified definition of self-financing and acceptable portfolios for the sandwiched shadow price process such that it is comparable with respect to the definition of acceptable portfolios for $S$ with transaction costs $\lambda$.
\[sandSaccpt\] The portfolio process $(\phi^0_t, \phi^1_t)_{0\leq t\leq T}$ is called **acceptable** for the sandwiched shadow price process $\hat{\mathbf{S}}$ if
- $(\phi^0, \phi^1)$ is predictable process of finite variation.
- $(\phi^0, \phi^1)$ is self-financing for $\hat{\mathbf{S}}$ without transaction costs in the sense that $$\phi_t^0=x+\int_0^t\phi_u^1d\hat{\mathbf{S}}_u-\phi_t^1\hat{S}_t,\ \ 0\leq t\leq T.\nonumber$$
- Define the auxiliary liquidation value process by $$V(\phi^0,\phi^1)_t\triangleq \phi_t^0+(\phi_t^1)^+(1-\lambda)S_t-(\phi_t^1)^-S_t.\nonumber$$ There exists a constant $a>0$ such that for each $\tilde{S}\in\mathcal{S}$, there exists a maximal element $X^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S}, a)$ and $$V(\phi^0, \phi^1)_{\tau}\geq -X^{\max,\tilde{S}}_{\tau},\nonumber$$ for all $[0,T]$-valued stopping time $\tau$.
Denote $\mathcal{A}_x(\hat{\mathbf{S}})$ the set of all acceptable portfolio processes for the sandwiched shadow price process $\hat{\mathbf{S}}$ starting with initial position $(\phi^0_0, \phi^1_0)=(x,0)$. Also, denote $\mathcal{V}_x(\hat{\mathbf{S}})$ the set of terminal value of all wealth processes generated by the acceptable portfolios $$\mathcal{V}_x(\hat{\mathbf{S}})\triangleq \bigg\{X_T: X_T=\phi_T^0+\phi_T^1\hat{S}_T=x+\int_0^T\phi_u^1d\hat{\mathbf{S}}_u,\ \ (\phi^0, \phi^1)\in\mathcal{A}_x(\hat{\mathbf{S}})\bigg\}.\nonumber$$
Similar to the case of classic shadow price process, given the same random endowment $\mathcal{E}_T$ and initial static position $q\in\mathbb{R}$, let us consider the primal set $$\mathcal{H}(x,q;\hat{\mathbf{S}})\triangleq \{X_T:\ X_T+q\mathcal{E}_T\geq 0,\ X_T\in\mathcal{V}_x(\hat{\mathbf{S}})\},\ \ (x,q)\in\mathcal{K}(\hat{\mathbf{S}})\nonumber$$ where $\mathcal{K}(\hat{\mathbf{S}})$ is defined by $$\mathcal{K}(\hat{\mathbf{S}})=\text{int}\{(x,q)\in\mathbb{R}^{2}: \mathcal{H}(x,q;\hat{\mathbf{S}})\neq\emptyset\}.\nonumber$$
The next theorem is one of our main results concerning the existence of a sandwiched shadow price process.
\[theosandsha\] Fix $(x,q)\in\mathcal{K}$. Under all assumptions in Theorem $\ref{theorem1}$, let $(y,r)\in\partial u(x,q)$ satisfy Assumption $\ref{importantnew}$ and let $(Y^{0,\ast}(y,r), Y^{1,\ast}(y,r))$ be the dual optimizer of problem $(\ref{dualv})$. Consider any $X(\phi^0,\phi^1)_T\in\mathcal{H}(x,q;\hat{\mathbf{S}})$ for the sandwiched shadow price process defined by $\hat{\mathbf{S}}=(\hat{S}^p,\hat{S})=\Big(\frac{Y^{1,\ast,p}(y,r)}{Y^{0,\ast,p}(y,r)}, \frac{Y^{1,\ast}(y,r)}{Y^{0,\ast}(y,r)}\Big)$. We have $$\begin{split}
\mathbb{E}\Big[U(X(\phi^0,\phi^1)_T+q\mathcal{E}_T)\Big]\leq &\mathbb{E}\bigg[U(x+\int_0^T\phi^{1,\ast}_ud\hat{\mathbf{S}}_u+q\mathcal{E}_T)\bigg]\\
=&\mathbb{E}\Big[U(\phi_T^{0,\ast}+\phi_T^{1,\ast}\hat{S}_T +q\mathcal{E}_T)\Big]\\
=&\mathbb{E}\Big[U(V(\phi^{0,\ast}, \phi^{1,\ast})_T+q\mathcal{E}_T)\Big],\nonumber
\end{split}$$ where $(\phi^{0,\ast}(x,q), \phi^{1,\ast}(x,q))$ is the optimal solution to the primal utility maximization problem $(\ref{primeu})$.
Comparing with Proposition $3.7$ in [@Chris22] and Theorem $3.1$ in [@Chris33], the existence of a classic shadow price process under random endowments becomes much more delicate and may fail in general even for continuous price processes. In our framework, it is even not enough to require that the dual optimizer $(Y^{0,\ast}(y,r), Y^{1,\ast}(y,r))$ satisfies the condition that $Y^{0,\ast}(y,r)$ is a martingale and $Y^{1,\ast}(y,r)$ is a local martingale. Actually, first, we need to require that the classic shadow price process admits NFLVR condition so that the duality theory can be obtained in the shadow price market. Second, in order to check that the dual optimizer $Y^{0,\ast}(y,r)$ is in the dual space $\mathcal{Y}(y,r;\hat{S})$ of the shadow price market and to compare utility value functions in two corresponding markets, we have to make the assumption that $Y^{0,\ast}(y,r)\in y\mathcal{M}(\frac{r}{y})\subset \mathcal{Y}(y,r)$ where we define $$\mathcal{M}(p)=\{\mathbb{Q}\in\mathcal{M}: \mathbb{E}^{\mathbb{Q}}[\mathcal{E}_T]=p\},\ \ p\in\mathcal{P}(x,q;U),\nonumber$$ and $\mathcal{P}(x,q;U)$ is the set of all marginal utility-based prices. Therefore, it is assumed that there exists some $(y,r)\in\partial u(x,q)$ such that the arbitrage-free price of $\mathcal{E}_T$ under the measure $\frac{d\mathbb{Q^{\ast}}}{d\mathbb{P}}=\frac{1}{y}Y_T^{0,\ast}(y,r)$ equals the chosen marginal utility-based price, i.e., $\mathbb{E}^{\mathbb{Q}^{\ast}}[\mathcal{E}_T]=\frac{r}{y}$.
The next theorem summarizes the existence of classic shadow price process under some sufficient conditions discussed above.
\[classical1\] Fix $(x,q)\in\mathcal{K}$ and $q>0$ and consider some $(y,r)\in\partial u(x,q)\subset\mathcal{L}$. If the dual minimizer $(Y^{0,\ast}(y,r), Y^{1,\ast}(y,r))$ to the problem $(\ref{dualv})$ satisfies that $(Y^{0,\ast}(y,r), Y^{1,\ast}(y,r))\in y\mathcal{B}$ and $Y^{0,\ast}_T(y,r)\in y\mathcal{M}\Big(\frac{r}{y}\Big)$. The process $\hat{S}(y,r)$ defined by $\hat{S}(y,r)\triangleq \frac{Y^{1,\ast}(y,r)}{Y^{0,\ast}(y,r)}$ is a classic shadow price process given in Definition $\ref{shadowprice}$ to the utility maximization problem $(\ref{primeu})$ with the price process $S$ and the transaction costs $\lambda$.
Proofs of Main Results {#section5}
======================
This section contains proofs of all main theorems and auxiliary results in the previous sections.
Proof of Proposition $\ref{auxlem}$
-----------------------------------
The proof of Theorem $1.7$ in [@Sch2] can be modified to our setting using acceptable portfolios. Assume that $(\ref{bef})$ does not hold for one fixed $\tilde{S}\in\mathcal{S}(\lambda, S)$, we may find $\frac{\lambda}{2}>\alpha>0$ and a stopping time $0\leq \tau\leq T$ such that either $\mathbb{P}(A_+)>0$ or $\mathbb{P}(A_{-})>0$, where we define $$A_{+}\triangleq \Big\{\phi_{\tau}^{1}\geq 0,\ \phi_{\tau}^{0}+\phi_{\tau}^{1}\frac{1-\lambda}{1-\alpha}S_{\tau}<-\hat{X}_{\tau}^{\max,\tilde{S}} \Big\},\nonumber$$ and $$A_{-}\triangleq \Big\{\phi_{\tau}^{1}\leq 0, \phi_{\tau}^{0}+\phi_{\tau}^{1}(1-\alpha^2)S_{\tau}<-\hat{X}_{\tau}^{\max,\tilde{S}} \Big\}.\nonumber$$
For the fixed $\tilde{S}\in\mathcal{S}(\lambda, S)$ and $\hat{X}^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S},a)$, consider any $\mathbb{Q}\in\mathcal{M}(\tilde{S})$, $\hat{X}^{\max,\tilde{S}}$ is a supermartingale under $\mathbb{Q}$. Hence for the stopping time $\tau$ chosen above, we will have $$\mathbb{Q}(\hat{X}_{\tau}^{\max,\tilde{S}} \geq \hat{X}_T^{\max,\tilde{S}})>0.\nonumber$$ Also, since $\mathbb{P}\sim\mathbb{Q}$, we deduce that $\mathbb{P}(\hat{X}_{\tau}^{\max,\tilde{S}} \geq \hat{X}_T^{\max,\tilde{S}})>0$. Let us define two auxiliary sets $$B_+\triangleq \{\hat{X}_{\tau}^{\max,\tilde{S}} \geq \hat{X}_T^{\max,\tilde{S}}\}\cap A_+,\nonumber$$ and $$B_{-}\triangleq \{\hat{X}_{\tau}^{\max,\tilde{S}} \geq \hat{X}_T^{\max,\tilde{S}}\}\cap A_{-}.\nonumber$$ Clearly it follows that $\mathbb{P}(B_+)>0$ or $\mathbb{P}(B_{-})>0$.
Choose $0<\lambda'<\alpha$ and consider a $\lambda'$-CPS with $\bar{S}$ taking values in the spread $[(1-\lambda')S, S]$ and $\mathbb{Q}\in\mathcal{M}(\bar{S}; \lambda')$, where we denote $\mathcal{M}(\bar{S}; \lambda')$ as the set of all $\mathbb{Q}$ such that $(\mathbb{Q}, \bar{S})$ is a $\lambda'$-CPS. It is easy to check that $(1-\alpha)\bar{S}$ and $\frac{1-\lambda}{1-\alpha}\bar{S}$ stays in the spread $[(1-\lambda)S, S]$, and it follows that for any $\mathbb{Q}\in\mathcal{M}(\bar{S}; \lambda')$, $(\mathbb{Q}, (1-\alpha)\bar{S})$ and $(\mathbb{Q}, \frac{1-\lambda}{1-\alpha}\bar{S})$ are both $\lambda$-CPS. Moreover, thanks to Proposition $1.6$ of [@Sch33], we deduce that $\phi_t^0+\phi_t^1(1-\alpha)\bar{S}_t$ and $\phi_t^0+\phi_t^1\frac{1-\lambda}{1-\alpha}\bar{S}_t$, $0\leq t\leq T$ are both local optional strong $\mathbb{Q}$-supermartingales. Since $(\phi^0,\phi^1)$ is an acceptable portfolio, there exists a constant $a>0$ and for $\bar{S}\in\mathcal{S}(\lambda', S)\subset\mathcal{S}$, there exists a $X^{\max,\bar{S}}\in\mathcal{X}(\bar{S},a)$ as the lower bound. It follows that $$0\leq V(\phi^0,\phi^1)_{\tau}+X_{\tau}^{\max,\bar{S}}\leq \phi_{\tau}^{0}+\phi_{\tau}^{1}(1-\alpha)\bar{S}_{\tau}+X_{\tau}^{\max,\bar{S}},\nonumber$$ for any $[0,T]$-valued stopping time $\tau$. Therefore $\phi_{t}^{0}+\phi_{t}^{1}(1-\alpha)\bar{S}_{t}+X_t^{\max,\bar{S}}$ is an optional strong $\mathbb{Q}$-supermartingale for any $\mathbb{Q}\in\mathcal{M}(\bar{S}; \lambda')$. Consider the subset $$\mathcal{M}'(\bar{S}; \lambda')\triangleq \{\mathbb{Q}\in\mathcal{M}(\bar{S};\lambda'): \text{$X^{\max,\bar{S}}$ is a UI martingale under $\mathbb{Q}$}\}.\nonumber$$ Pick any fixed $\mathbb{Q}\in\mathcal{M}'(\bar{S}; \lambda')$. Since $\bar{S}\geq (1-\alpha)S$, we obtain that $$\begin{split}
\mathbb{E}^{\mathbb{Q}}[V(\phi^0,\phi^1)_T| B_{-}]&\leq \mathbb{E}^{\mathbb{Q}}[\phi_T^0+\phi_T^1(1-\alpha)\bar{S}_T+X_T^{\max,\bar{S}}|B_{-}]-\mathbb{E}^{\mathbb{Q}}[X_T^{\max,\bar{S}}|B_{-}]\\
&\leq \mathbb{E}^{\mathbb{Q}}[\phi_{\tau}^0+\phi_{\tau}^1(1-\alpha)\bar{S}_{\tau}+X_{\tau}^{\max,\bar{S}}|B_{-}]-\mathbb{E}^{\mathbb{Q}}[X_{\tau}^{\max,\bar{S}}|B_{-}]\\
&=\mathbb{E}^{\mathbb{Q}}[\phi_{\tau}^0+\phi_{\tau}^1(1-\alpha)\bar{S}_{\tau}|B_{-}]\\
&\leq \mathbb{E}^{\mathbb{Q}}[\phi_{\tau}^{0}+\phi_{\tau}^{1}(1-\alpha)^2 S_{\tau}|B_{-}]\\
&< \mathbb{E}^{\mathbb{Q}}[-\hat{X}_{\tau}^{\max,\tilde{S}}|B_{-}]\leq \mathbb{E}^{\mathbb{Q}}[-\hat{X}_{T}^{\max,\tilde{S}}|B_{-}].\nonumber
\end{split}$$
Similarly, for the same lower bound $X^{\max,\bar{S}}$ chosen above, we have $\phi_t^0+\phi_t^1\frac{1-\lambda}{1-\alpha}\bar{S}_t+X_t^{\max,\bar{S}}$ is an optional strong $\mathbb{Q}$-supermartingale for any $\mathbb{Q}\in\mathcal{M}(\bar{S}; \lambda')$. Again, pick one $\mathbb{Q}\in\mathcal{M}'(\bar{S}; \lambda')$, we have $$\begin{split}
\mathbb{E}^{\mathbb{Q}}[ V(\phi^0,\phi^1)_T| B_+]&\leq \mathbb{E}^{\mathbb{Q}}\Big[\phi_T^0+\phi_T^1\frac{1-\lambda}{1-\alpha}\bar{S}_T+X_T^{\max,\bar{S}} \Big|B_+\Big]-\mathbb{E}^{\mathbb{Q}}[X_T^{\max,\bar{S}}|B_+]\\
&\leq \mathbb{E}^{\mathbb{Q}}\Big[\phi_{\tau}^0+\phi_{\tau}^1\frac{1-\lambda}{1-\alpha}\bar{S}_{\tau}+X_{\tau}^{\max,\bar{S}} \Big|B_+\Big]-\mathbb{E}^{\mathbb{Q}}[X_{\tau}^{\max,\bar{S}}|B_+]\\
&= \mathbb{E}^{\mathbb{Q}}\Big[\phi_{\tau}^0+\phi_{\tau}^1\frac{1-\lambda}{1-\alpha}\bar{S}_{\tau} \Big|B_+\Big]\\
&\leq \mathbb{E}^{\mathbb{Q}}\Big[\phi_{\tau}^0+\phi_{\tau}^{1}\frac{1-\lambda}{1-\alpha}S_{\tau}\Big|B_+\Big]\\
&< \mathbb{E}^{\mathbb{Q}}[-\hat{X}_{\tau}^{\max,\tilde{S}}|B_+]\leq \mathbb{E}^{\mathbb{Q}}[-\hat{X}_{T}^{\max,\tilde{S}}|B_+].\nonumber
\end{split}$$ Since either $\mathbb{P}(B_+)>0$ or $\mathbb{P}(B_{-})>0$, we arrive at a contradiction to $V(\phi^0,\phi^1)_T\geq -\hat{X}_T^{\max,\tilde{S}}$ $\mathbb{P}$-a.s., and our conclusion holds.
Proof of Theorem $\ref{mainthm}$
--------------------------------
The following proposition plays a central role to build a bipolar result required in the proof of Theorem $\ref{mainthm}$.
\[dualprop\] Let Assumption $\ref{Assum}$, $\ref{assE}$ and $\ref{assNo}$ hold. The families $(\mathcal{C}(x,q))_{(x,q)\in\mathcal{K}}$ and $(\mathcal{D}(y,r))_{(y,r)\in\mathcal{L}}$ defined in $(\ref{primsp})$ and $(\ref{dualsp})$ have the following properties:
- For any $(x,q)\in\mathcal{K}$, the set $\mathcal{C}(x,q)$ contains a strictly positive constant. A nonnegative function $g$ belongs to $\mathcal{C}(x,q)$ if and only if $$\label{bipo1}
\mathbb{E}[gh]\leq xy+q\cdot r,\ \ \text{for all}\ (y,r)\in\mathcal{L}\ \text{and}\ h\in\mathcal{D}(y,r).$$
- For any $(y,r)\in\mathcal{L}$, the set $\mathcal{D}(y,r)$ contains a strictly positive random variable. A nonnegative function $h$ belongs to $\mathcal{D}(y,r)$ if and only if $$\label{bipo2}
\mathbb{E}[gh]\leq xy+q\cdot r,\ \ \text{for all}\ (x,q)\in\mathcal{K}\ \text{and}\ g\in\mathcal{C}(x,q).$$
The proof of Proposition $\ref{dualprop}$ is based on a sequel of the following lemmas among which a type of super-hedging theorem for acceptable portfolios is critical. Lemma $\ref{ineqC}$ together with Lemma $\ref{charac}$ below provide us this type of super-hedging result. In particular, the characterization of the set $\mathcal{C}(x,q)$ below gives one side of our super-hedging theorem.
\[ineqC\] If $(x,q)\in\mathcal{K}$, for any $g\in\mathcal{C}(x,q)$, we have $$\label{gineq}
\mathbb{E}^\mathbb{Q}[g]\leq x+\mathbb{E}^{\mathbb{Q}}[q\cdot\mathcal{E}_T],\ \ \forall \mathbb{Q}\in\mathcal{M}.$$
For any $g\in\mathcal{C}(x,q)$, there exits a $V\in\mathcal{H}(x,q)$, and $g\leq V_T+q\cdot\mathcal{E}_T$. It is hence enough to verify that $$\mathbb{E}^{\mathbb{Q}}[V_T+q\cdot\mathcal{E}_T]\leq x+\mathbb{E}^{\mathbb{Q}}[q\cdot\mathcal{E}_T],\ \ \forall \mathbb{Q}\in\mathcal{M},\nonumber$$ which is equivalent to show that $$\label{ineq1}
\mathbb{E}^{\mathbb{Q}}[V_T+q\cdot\mathcal{E}_T]\leq x+\mathbb{E}^{\mathbb{Q}}[q\cdot\mathcal{E}_T],\ \ \forall \mathbb{Q}\in\mathcal{M}(\tilde{S}),\ \ \forall \tilde{S}\in\mathcal{S}.$$ By the definition of acceptable portfolios, there exists a constant $a>0$ such that for each fixed $\tilde{S}\in\mathcal{S}$, there exists a maximal element $X^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S}, a)$ with $V_{\tau}+X_{\tau}^{\max,\tilde{S}}\geq 0$ for all $[0,T]$-valued stopping times. Therefore we can rewrite $$\label{rew}
\mathbb{E}^{\mathbb{Q}}[V_T+q\cdot\mathcal{E}_T]=\mathbb{E}^{\mathbb{Q}}[V_T+X_T^{\max,\tilde{S}}]-\mathbb{E}^{\mathbb{Q}}[X_T^{\max,\tilde{S}}]+\mathbb{E}^{\mathbb{Q}}[q\cdot\mathcal{E}_T].$$ Define the set $$\mathcal{M}'(\tilde{S})\triangleq \{\mathbb{Q}\in\mathcal{M}(\tilde{S}):\ X^{\max,\tilde{S}}\ \text{is a UI martingale under $\mathbb{Q}$}\},\nonumber$$ Theorem $5.2$ of [@DS97] asserts that $\mathcal{M}'(\tilde{S})$ is not empty and dense in $\mathcal{M}(\tilde{S})$ with respect to the norm topology of $\mathbb{L}^1(\Omega, \mathcal{F}, \mathbb{P})$. We shall first verify that the inequality $(\ref{ineq1})$ holds for all $\mathbb{Q}\in\mathcal{M}'(\tilde{S})$. Since $X_T^{\max,\tilde{S}}$ is a UI martingale under $\mathbb{Q}\in\mathcal{M}'(\tilde{S})$, it is sufficient to verify that $$\label{sinq}
\mathbb{E}^{\mathbb{Q}}[V_T+X_T^{\max,\tilde{S}}]\leq x+a,\ \ \forall \mathbb{Q}\in\mathcal{M}'(\tilde{S}).$$ As $\tilde{S}\in[(1-\lambda)S, S]$, it is easy to see that $V_t\leq \tilde{V}_t$ where $\tilde{V}_t\triangleq \phi_t^0+\phi_t^1\tilde{S}_t$ for $t\in[0,T]$. Follow the proof of Proposition $1.6$ of [@Sch2], we get that $\tilde{V}_t$ is a local optional strong supermartingale under each $\mathbb{Q}\in\mathcal{M}'(\tilde{S})$, therefore $\tilde{V}_t+X_t^{\max,\tilde{S}}$ is also a local optional strong supermartingale under $\mathbb{Q}$. Since $\tilde{V}_{t}+X_t^{\max,\tilde{S}}\geq V_t+X_t^{\max,\tilde{S}}\geq 0$, we can deduce that $\tilde{V}_t+X_t^{\max,\tilde{S}}$ is an optional strong supermartingale under $\mathbb{Q}$ by Fatou’s Lemma. And we obtain that $$\mathbb{E}^{\mathbb{Q}}[\tilde{V}_T+X_T^{\max,\tilde{S}}]\leq x+a,\ \ \forall \mathbb{Q}\in\mathcal{M}'(\tilde{S}),\nonumber$$ which implies that $(\ref{sinq})$ holds. Hence, it follows that for each $\tilde{S}\in\mathcal{S}$, $$\label{ineq2}
\mathbb{E}^{\mathbb{Q}}[V_T+q\cdot\mathcal{E}_T]\leq x+\mathbb{E}^{\mathbb{Q}}[q\cdot\mathcal{E}_T],\ \ \forall \mathbb{Q}\in\mathcal{M}'(\tilde{S}).$$ Denote $\gamma_T\triangleq V_T+q\cdot\mathcal{E}_T$. The density property of $\mathcal{M}'(\tilde{S})$ in $\mathcal{\tilde{S}}$ in the norm topology of $\mathbb{L}^1$ implies the existence of a sequence of $\mathbb{Q}^n\in\mathcal{M}'(\tilde{S})$ and by $(\ref{ineq2})$, we have $$\begin{split}
\mathbb{E}^{\mathbb{Q}}[\gamma_T]&=\lim_{m\rightarrow\infty}\mathbb{E}^{\mathbb{Q}}[\gamma_T\mathbf{1}_{\{\gamma_T\leq m\}}]=\lim_{m\rightarrow\infty}\lim_{n\rightarrow\infty}\mathbb{E}^{\mathbb{Q}^n}[\gamma_T\mathbf{1}_{\{\gamma_T\leq m\}}]\\
&\leq \lim_{n\rightarrow\infty}\mathbb{E}^{\mathbb{Q}^n}[\gamma_T]\leq x+\lim_{n\rightarrow\infty}\mathbb{E}^{\mathbb{Q}^n}[q\cdot\mathcal{E}_T].\nonumber
\end{split}$$
Clearly for any $m>0$ and each $1\leq i\leq N$, we have $$\mathcal{E}^i_T\mathbf{1}_{\{\mathcal{E}^i_T>m\}}\leq \sum_{i=1}^{N}\mathcal{E}_T^i\mathbf{1}_{\{\sum_{i=1}^{N}\mathcal{E}_T^i >m\}},\ \text{a.s. under $\mathbb{P}$}.\nonumber$$ The assumption that $\mathcal{E}^i_T\geq 0$ a.s. under $\mathbb{P}$ implies $\mathcal{E}^i_T\geq 0$ a.s. under $\mathbb{Q}\in\mathcal{M}$, it follows that $$\label{woo}
\lim_{m\rightarrow\infty}\sup_{\mathbb{Q}\in\mathcal{M}}\mathbb{E}^{\mathbb{Q}}[\mathcal{E}^i_T\mathbf{1}_{\{\mathcal{E}^i_T>m\}}]=0,\ \ 1\leq i\leq N.$$
Given Assumption $(\ref{assE})$, Moore-Osgood Theorem (see Theorem $5$, p.$102$ of [@13]) and Monotone Convergence Theorem give us that $$\begin{split}
\lim_{n\rightarrow\infty}\mathbb{E}^{\mathbb{Q}^n}[\mathcal{E}^i_T]=&\lim_{n\rightarrow\infty}\lim_{m\rightarrow\infty}\mathbb{E}^{\mathbb{Q}^n}[\mathcal{E}^i_T\mathbf{1}_{\{\mathcal{E}^i_T\leq m\}}]=\lim_{m\rightarrow\infty}\lim_{n\rightarrow\infty}\mathbb{E}^{\mathbb{Q}^n}[\mathcal{E}^i_T\mathbf{1}_{\{\mathcal{E}^i_T\leq m\}}]\\
=&\lim_{m\rightarrow\infty}\mathbb{E}^{\mathbb{Q}}[\mathcal{E}^i_T\mathbf{1}_{\{\mathcal{E}^i_T\leq m\}}]=\mathbb{E}^{\mathbb{Q}}[\mathcal{E}^i_T],\ \ 1\leq i\leq N. \nonumber
\end{split}$$ We thereby obtain that $$\lim_{n\rightarrow\infty}\mathbb{E}^{\mathbb{Q}^n}[q\cdot\mathcal{E}_T]=\mathbb{E}^{\mathbb{Q}}[q\cdot\mathcal{E}_T].\nonumber$$
Therefore, it follows that $(\ref{ineq1})$ holds for any $\mathbb{Q}\in\mathcal{M}(\tilde{S})$ and any $\tilde{S}\in\mathcal{S}$, which completes the proof.
We can see that the Assumption $\ref{assE}$ on random endowments is required so that we can apply the Moore-Osgood Theorem to exchange the order of double limits in the proof above.
For the other side of the super-hedging result, i.e., the proof of Lemma $\ref{charac}$, we need more delicate work. Lemma $\ref{lemm}$ up to Lemma $\ref{lem13}$ below all serve as preparations for this purpose.
Fix a constant $\hat{a}>0$ and define $\mathcal{A}_{0,\hat{a}}$ as the set of all pairs $\phi=(\phi^0, \phi^1)\in\mathcal{A}_0$ and for each $\tilde{S}\in\mathcal{S}$, there exits a $\hat{X}^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S}, \hat{a})$ such that $V(\phi)_T+\hat{X}_T^{\max,\tilde{S}}\geq 0$. We intend to show that elements in the set $\mathcal{A}_{0,\hat{a}}$ are bounded in probability. In fact, any convex combinations of the elements in $\mathcal{A}_{0,\hat{a}}$ are also bounded in probability. This is the first step to obtain the almost surely convergence result for any sequence in $\mathcal{A}_{0,\hat{a}}$ by passing to convex combinations.
\[lemm\] Let $S$ and $0<\lambda<1$ satisfy the previous assumptions and suppose that $(CPS^{\lambda'})$ is satisfied in the local sense for some $0<\lambda'<\lambda$. For $\hat{a}>0$, we can find one probability measure $\mathbb{Q}\sim\mathbb{P}$ and there exist constants $C_0>0$ and $C_1>0$ such that for all $(\phi^0,\phi^1)\in \mathcal{A}_{0,\hat{a}}$, we have $$\label{pro1}
\mathbb{E}^{\mathbb{Q}}\Big[ \|\phi^0\|_T\Big]\leq C_0\hat{a},$$ and $$\label{pro2}
\mathbb{E}^{\mathbb{Q}}\Big[ \|\phi^1\|_T\Big]\leq C_1\hat{a},$$ where $\|\phi\|$ denotes the total variation of $\phi$.
Fix $0<\lambda'<\lambda$ as above. Consider $\tilde{S}\in\mathcal{S}(\lambda',S)$ such that $\tilde{S}_t\in[(1-\lambda')S_t,S_t]$ and $(\tilde{S}_t)_{0\leq t\leq T}$ is a local $\mathbb{Q}$-martingale for all $\mathbb{Q}\in\mathcal{M}(\tilde{S};\lambda')$. Since the assertion of the lemma is of local type, we can assume by choosing stopping, that $\tilde{S}$ is a true martingale. We may also assume that $\phi_T^1=0$ so that the position in stock is liquidated at time $T$.
Assume $(\phi^0,\phi^1)$ is an acceptable portfolio under transaction costs $\lambda$ and $(\phi_0^0, \phi_0^1)=(0,0)$. Define the new process $\phi'=((\phi^0)',(\phi^1)')$ by $$\phi'_t=((\phi^0)_t',(\phi^1)_t')=\Big(\phi_t^0+\frac{\lambda-\lambda'}{1-\lambda}\phi_t^{0,\uparrow}, \phi_t^1 \Big),\ \ 0\leq t\leq T.\nonumber$$
Thanks to the proof of Lemma $3.1$ of [@Sch33], $((\phi^0)',(\phi^1)')$ is a self-financing process under the transaction costs $\lambda'$. Since $(\phi^0,\phi^1)$ is acceptable under transaction costs $\lambda$, and for any $\lambda'<\lambda$, it is clear that $\mathcal{S}(\lambda', S)\subset\mathcal{S}$. It follows that there exists a constant $a>0$ and for each $\tilde{S}\in\mathcal{S}(\lambda', S)$, there exists a $X^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S}, a)$ such that $$V(\phi)_{\tau}\geq -X_{\tau}^{\max,\tilde{S}},\ \text{a.s.}.\nonumber$$ Moreover, it is easy to see that $V_{\tau}(\phi')\geq V_{\tau}(\phi)$ by the definition of $\phi'$. Therefore, we obtain that $\phi'=((\phi^0)',(\phi^1)')$ is an acceptable portfolio under the smaller transaction costs $\lambda'$.
Following the proof of Proposition $1.6$ of [@Sch33], we see that $V(\phi')_t\leq \tilde{V}(\phi')_t$ where $\tilde{V}(\phi')_t\triangleq (\phi^0)'_t+(\phi^1)'_t\tilde{S}$, $\tilde{S}\in\mathcal{S}(\lambda', S)$ and $\tilde{V}(\phi')$ is a local optional strong super-martingale under all $\mathbb{Q}\in\mathcal{M}(\tilde{S};\lambda')$. For each fixed $\tilde{S}\in\mathcal{S}(\lambda',S)$, by the definition of acceptable portfolio, there exists a constant $a$ and a maximal element $X^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S}, a)$ such that $V(\phi')_t\geq -X^{\max,\tilde{S}}_t$. Hence, we get $\tilde{V}(\phi')_t+X_t^{\max,\tilde{S}}\geq 0$ is an optional strong supermartingale. And for this fixed $X^{\max,\tilde{S}}$, consider the set $$\mathcal{M}'(\tilde{S}; \lambda')\triangleq \{\mathbb{Q}\in\mathcal{M}(\tilde{S}; \lambda'):\ X^{\max,\tilde{S}}\ \text{is a UI martingale under $\mathbb{Q}$}\}.\nonumber$$ For each $\mathbb{Q}\in\mathcal{M}'(\tilde{S} ; \lambda')$, we obtain that $$\begin{split}
\mathbb{E}^{\mathbb{Q}}[(\phi^0)'_T+(\phi^1)'_T\tilde{S}_T]=&\mathbb{E}^{\mathbb{Q}}[(\phi^0)'_T+(\phi^1)'_T\tilde{S}_T+X_T^{\max,\tilde{S}}]-\mathbb{E}^{\mathbb{Q}}[X_T^{\max,\tilde{S}}]\\
\leq &0+a-a=0.\nonumber
\end{split}$$ By the definition of $(\phi^0)'$ and $(\phi^1)'$, we deduce that $$\mathbb{E}^{\mathbb{Q}}[\phi_T^0+\phi_T^1\tilde{S}_T]+\frac{\lambda-\lambda'}{1-\lambda}\mathbb{E}^{\mathbb{Q}}[\phi_T^{0,\uparrow}]\leq 0.\nonumber$$ Since $\phi^0_T+\phi_T^1\tilde{S}_T\geq V(\phi)_T$ and $\tilde{S}\in\mathcal{S}$, by definition, there exists a constant $\hat{a}>0$ and $\hat{X}^{\max,\tilde{S}}$ such that $V(\phi)_T\geq -\hat{X}_T^{\max,\tilde{S}}$. Therefore, we obtain that $$\mathbb{E}^{\mathbb{Q}}[\phi_T^{0,\uparrow}]\leq \frac{1-\lambda}{\lambda-\lambda'}\mathbb{E}^{\mathbb{Q}}[\hat{X}_T^{\max,\tilde{S}}]\leq \frac{(1-\lambda)\hat{a}}{\lambda-\lambda'},\ \ \forall \mathbb{Q}\in\mathcal{M}'(\tilde{S};\lambda')\nonumber$$ as $\hat{X}^{\max,\tilde{S}}$ is a supermartingale under $\mathbb{Q}\in\mathcal{M}'(\tilde{S}; \lambda')$. Since the set $\mathcal{M}'(\tilde{S}; \lambda')$ is dense in $\mathcal{M}(\tilde{S}; \lambda')$ with respect to the norm topology of $\mathbb{L}^1$, for any $\mathbb{Q}\in\mathcal{M}(\tilde{S}; \lambda')$, Fatou’s lemma leads to $$\mathbb{E}^{\mathbb{Q}}[\phi_T^{0,\uparrow}]\leq \frac{1-\lambda}{\lambda-\lambda'}\hat{a}.\nonumber$$ For each $\tilde{S}\in\mathcal{S}(\lambda')$, $\phi_T^0=\phi_T^0+\phi_T^1\tilde{S}\geq -\hat{X}^{\max,\tilde{S}}$ by the previous argument and $\phi_T^1=0$. Therefore, it follows that $\phi_T^{0,\downarrow}\leq \phi_T^{0,\uparrow}+\hat{X}_T^{\max,\tilde{S}}$. For each $\mathbb{Q}\in\mathcal{M}(\tilde{S}; \lambda')$, since $X^{\max,\tilde{S}}$ is a supermartingale under $\mathbb{Q}$, we can derive that $$\mathbb{E}^{\mathbb{Q}}[\phi_T^{0,\uparrow}+\phi_T^{0,\downarrow}]\leq 2\frac{1-\lambda}{\lambda-\lambda'}\hat{a}+\hat{a},\nonumber$$ which completes the proof of $(\ref{pro1})$.
As regards $(\ref{pro2})$, we can follow the proof of Lemma $3.1$ of [@Sch33]. First, we have that $$\label{phi1}
d\phi_t^{1,\uparrow}\leq \frac{d\phi_t^{0,\downarrow}}{S_t}.$$ Since $\tilde{S}$ is a $\mathbb{Q}$-local supermartingale and it follows that it is a $\mathbb{Q}$-supermartingale for $\mathbb{Q}\in\mathcal{M}(\tilde{S}; \lambda')$. It follows that $\inf_{0\leq t\leq T}\tilde{S}_t(\omega)$ is $\mathbb{Q}$-a.s. strictly positive since $\tilde{S}_T>0$, $\mathbb{Q}$-a.s.. Therefore, $\inf_{0\leq t\leq T}\tilde{S}_t(\omega)$ is $\mathbb{P}$-a.s. as well. We can obtain that for any $\epsilon>0$, there exists $\delta>0$ such that $$\label{phi12}
\mathbb{P}\Big[ \inf_{0\leq t\leq T}S_t<\delta\Big]<\frac{\epsilon}{2}.$$ Combining $(\ref{pro1})$, $(\ref{phi1})$ and $(\ref{phi12})$, it is easy to derive a control such that $\mathbb{E}^{\mathbb{Q}}[\phi_T^{1,\uparrow}]\leq k\hat{a}$ for some $k>0$. Finally, we recall that $\phi_T^1=0$ which implies that $\phi_T^{1,\uparrow}=\phi_T^{1,\downarrow}$. It follows that $(\ref{pro2})$ holds.
It is now important for us to verify the closedness property of the set $\mathcal{U}_x$ (resp. $\mathcal{V}_x$) for the purpose of the super-hedging result. In particular, it is enough to consider the case $x=0$. For the admissible portfolio processes in Definition $\ref{admissiblephi}$, the Fatou-closedness is an appropriate concept, see Appendix $5.5$ of [@KS09]. As in [@Sch33], a sequence $(\phi_T^n)_{n=1}^{\infty}$ in $\mathbb{L}^0(\mathbb{R}^2)$ *Fatou-converges* to $\phi_T\in\mathbb{L}^0(\mathbb{R}^2)$ if there is $M>0$ such that $V(\phi^{0,n},\phi^{1,n})_T\geq -M$ and $\phi_T^n$ converges a.s. to $\phi_T$. A set is Fatou closed if it is closed under the Fatou convergence. Due to the stochastic lower-bounds for our acceptable portfolios, the above definition of Fatou-closedness has to be modified using the CPS $(\mathbb{Q}, \tilde{S})$ and some corresponding maximal elements $X^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S}, a)$ for some $a>0$. The following alternative definition can replace the role of the usual Fatou-closedness in the literature.
Fix some $\hat{a}>0$, for each $\tilde{S}\in\mathcal{S}$, we pick and fix one maximal element $\hat{X}^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S},\hat{a})$. The set $\mathcal{U}_0$ ( $\mathcal{V}_0$) is said to be **relatively Fatou closed** if for any sequence $(\phi^{0,n}_T,\phi^{1,n}_T)\in\mathcal{U}_0$ which satisfies $V(\phi^n)_T\geq -\hat{X}_T^{\max,\tilde{S}}$ ( $\phi_T^{0,n}\geq -\hat{X}_T^{\max,\tilde{S}}$) and converges to $(\phi^0_T,\phi^1_T)\in\mathbb{L}^0(\mathbb{R}^2)$ ( $\phi_T^0\in\mathbb{L}^0(\mathbb{R})$) almost surely, we have that $(\phi_T^0,\phi_T^1)\in\mathcal{U}_0$ ( $\phi_T^0\in\mathcal{V}_0$).
\[order\] In the two dimensional setting, let us introduce the partial order on $\mathbb{L}^0(\mathbb{R}^2)$ by $(\phi^0,\phi^1)\succeq (\psi^0, \psi^1)$ if $V(\phi^0-\psi^0, \phi^1-\psi^1)_T\geq 0$, a.s.. Therefore, in the above definition, we can also say $(\phi^{0,n}, \phi^{1,n})$ relatively Fatou converges to $(\phi^0,\phi^1)$, if there exists a constant $\hat{a}>0$ such that for each $\tilde{S}\in\mathcal{S}$, we can find a $\hat{X}^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S}, \hat{a})$ such that $(\phi_T^{0,n}, \phi_T^{1,n})\succeq (-X_T^{\max,\tilde{S}}, 0)$ and $(\phi_T^{0,n},\phi_T^{1,n})$ converges to $(\phi^0_T, \phi^1_T)$ almost surely.
\[fatoulem\] Fix $S=(S_t)_{0\leq t\leq T}$ and $0<\lambda<1$ as above and let Assumption $\ref{Assum}$ hold. The sets $\mathcal{U}_0$ and $\mathcal{V}_0$ are both relatively Fatou closed.
Fix $\hat{a}>0$ and for each $\tilde{S}\in\mathcal{S}$, choose and fix $\hat{X}^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S}, \hat{a})$. Consider a sequence $(\phi_T^{0,n},\phi_T^{1,n})\in\mathcal{U}_0$ such that $V(\phi^n)_T\geq -\hat{X}_T^{\max,\tilde{S}}$ and $(\phi_T^{0,n},\phi_T^{1,n})$ converges a.s. to some $(\phi_T^0,\phi_T^1)\in\mathbb{L}^0(\mathbb{R}^2)$. Thanks to Lemma $\ref{auxlem}$, we can deduce that for any $[0,T]$-valued stopping time $\tau$, $V(\phi^n)_{\tau}\geq -\hat{X}_{\tau}^{\max,\tilde{S}}$. Decompose canonically these processes $\phi_T^{0,n}=\phi_T^{0,n,\uparrow}-\phi_T^{0,n,\downarrow}$ and $\phi_T^{1,n}=\phi_T^{1,n,\uparrow}-\phi_T^{1,n,\downarrow}$. Thanks to Lemma $\ref{lemm}$, the proof of Theorem $3.4$ of [@Sch33] can be carried over verbatim in our setting, and we can find a predictable increasing process $\phi^{0,\uparrow}=(\phi^{0,\uparrow}_t)_{0\leq t\leq T}$ such that the sequence $\phi_t^{0,n,\uparrow}$ converges almost surely to $\phi_t^{0,\uparrow}$ for all $0\leq t\leq T$. Similar results hold for $\phi^{0,\downarrow}$, $\phi^{1,\uparrow}$ and $\phi^{1,\downarrow}$. These processes are all predictable, increasing and satisfy condition $(\ref{self})$.
Define the process $(\phi^0_t,\phi^1_t)_{0\leq t\leq T}$ by $\phi_t^0=\phi_t^{0,\uparrow}-\phi_t^{0,\downarrow}$ and $\phi_t^1=\phi_t^{1,\uparrow}-\phi_t^{1,\downarrow}$. The process $\phi$ is a predictable and self-financing portfolio process. Moreover, for each $\tilde{S}\in\mathcal{S}$, since $V(\phi^n)_{\tau}\geq -\hat{X}_{\tau}^{\max,\tilde{S}}$ for all $[0,T]$-valued stopping time $\tau$, we obtain that $V(\phi)_{\tau}\geq -\hat{X}_{\tau}^{\max,\tilde{S}}$ as the convergence of $(\phi^n)_{n=1}^{\infty}$ takes place for all $t\in[0,T]$. We can conclude that $(\phi^0,\phi^1)$ is an acceptable portfolio, i.e., $(\phi^0,\phi^1)\in\mathcal{U}_0$, and therefore $\mathcal{U}_0$ is relatively Fatou closed. The proof for $\mathcal{V}_0$ follows the same arguments.
After the closedness property, we need to proceed to characterize the set $\mathcal{U}_x$ for the purpose of the super-hedging inequality. To this end, we can first try to characterize its polar set. Let $\mathcal{U}_x^{\infty}\triangleq \mathcal{U}_x\cap\mathbb{L}^{\infty}(\mathbb{R}^2)$, we introduce the following auxiliary set $\bar{\mathcal{Z}}$:
\[nameZ\] Denote $\bar{\mathcal{Z}}(\lambda, S)$ (short as $\bar{\mathcal{Z}}$) as the set of all pairs $Z_T=(Z^0_T, Z^1_T)\in\mathbb{L}^1_+(\mathbb{R}^2;\mathcal{F}_T)$ such that $\mathbb{E}[Z_T^0]=1$ and $$\mathbb{E}[\phi_T^0Z_T^0+\phi_T^1Z_T^1]\leq 0,\nonumber$$ for all $\phi_T=(\phi_T^0, \phi_T^1)\in\mathcal{U}_0^{\infty}$.
Each $(Z_T^0, Z_T^1)\in\bar{\mathcal{Z}}$ can be identified with a pair $(\mathbb{Q}, \tilde{S})$ by setting $$Z^i_t=\mathbb{E}[Z^i_T|\mathcal{F}_t],\ \ i=0,1,\ \ \tilde{S}_t=\frac{Z_t^1}{Z_t^0},\ \ \text{and}\ \ \ \frac{d\mathbb{Q}}{d\mathbb{P}}=Z_T^0.\nonumber$$ However, here the measure $\mathbb{Q}$ is only absolutely continuous with respect to $\mathbb{P}$.
The following lemma builds the relationship between the definition of $\bar{\mathcal{Z}}$ and $\lambda$-CPS.
\[lemZ\] Assume that $S_t\leq K$ for some constant $K$ for all $0\leq t\leq T$. For each $Z\in\bar{\mathcal{Z}}$, define the martingale $Z=(Z_t^0, Z_t^1)_{0\leq t\leq T}$ by $$Z^i_t\triangleq \mathbb{E}[Z^i_T|\mathcal{F}_t],\ \ i=0,1,\ \ 0\leq t\leq T.\nonumber$$ We will have that $$\tilde{S}_t\triangleq \frac{Z_t^1}{Z_t^0}\in [(1-\lambda)S_t,S_t],\ \ 0\leq t\leq T,\ \text{a.s.}\nonumber$$ Conversely, suppose that $Z=(Z_t^0,Z_t^1)_{0\leq t\leq T}$ is an $\mathbb{R}_+^2$-valued $\mathbb{P}$-martingale such that $Z_0^0=1$ and $\tilde{S}_t=\frac{Z_t^1}{Z_t^0}$ takes values in $[(1-\lambda)S_t, S_t]$ a.s. on $\{Z_t^0>0\}$. Then we have $Z_T=(Z_T^0,Z_T^1)\in\bar{\mathcal{Z}}$.
Choose any $Z_T\in\bar{\mathcal{Z}}$ and suppose that there exits a $[0,T)$-valued stopping time $\tau$ such that $\mathbb{Q}(\tilde{S}_{\tau} >S_{\tau})>0$. Similar to the proof of Proposition $4.2$ of [@Sch33], we can consider the strategy $$a_t=\Big(-1, \frac{1}{S_{\tau}}\Big)\mathbf{1}_{\{\tilde{S}_{\tau}>S_{\tau}\}}\mathbf{1}_{\rrbracket \tau, T\rrbracket}(t),\ \ 0\leq t\leq T. \nonumber$$ It is a self-financing strategy and $a_T\in\mathbb{L}^{\infty}(\mathbb{R}^2)$, moreover, $V(a)_t$ is uniformly bounded below by $-1$ for $0\leq t\leq T$, therefore, we clearly have $a_T\in\mathcal{U}_0^{\infty}$. Using the fact that $\tilde{S}$ is a martingale, we can follow the proof of Proposition $4.2$ of [@Sch33] to deduce a contradiction. To see the detail, we observe that $$\begin{split}
&\mathbb{E}^{\mathbb{P}}\Big[ \Big( -Z_T^0+\frac{Z_T^1}{S_{\tau}}\Big)\mathbf{1}_{\{\tilde{S}_{\tau}>S_{\tau}\}}\Big]=\mathbb{E}^{\mathbb{P}}\Big[ \mathbb{E}^{\mathbb{P}}\Big[ \Big(-Z_T^0+\frac{Z_T^1}{S_{\tau}}\Big)\mathbf{1}_{\{\tilde{S}_{\tau}>S_{\tau}\}} \Big|\mathcal{F}_{\tau}\Big] \Big]\\
=&\mathbb{E}^{\mathbb{P}}\Big[ Z_{\tau}^0\Big( -1+\frac{\tilde{S}_{\tau}}{S_{\tau}}\Big)\mathbf{1}_{\{\tilde{S}_{\tau}>S_{\tau}\}}\Big]=\mathbb{E}^{\mathbb{Q}}\Big[\Big( -1+\frac{\tilde{S}_{\tau}}{S_{\tau}}\Big)\mathbf{1}_{\{\tilde{S}_{\tau}>S_{\tau}\}}\Big]>0,\nonumber
\end{split}$$ which is a contradiction. If $\mathbb{Q}(\tilde{S}_T>S_T)>0$, we can instead consider the portfolio process $a'_t=\Big(-1, \frac{1}{S_T}\Big)\mathbf{1}_{\{\tilde{S}_T>S_T\}}\mathbf{1}_{\llbracket T\rrbracket}$ and deduce a similar contradiction. Therefore, we obtain that $\tilde{S}_t\leq S_t$ for all $0\leq t\leq T$.
As given in the proof of Proposition $4.2$ of [@Sch33], the strategies $$b_t=((1-\lambda)S_{\tau}, -1)\mathbf{1}_{\{\tilde{S}_{\tau}<(1-\lambda)S_{\tau}\}}\mathbf{1}_{\rrbracket \tau,T\rrbracket}(t),\ \ 0\leq t\leq T, \nonumber$$ and $$b'_t=((1-\lambda)S_{T}, -1)\mathbf{1}_{\{\tilde{S}_{T}<(1-\lambda)S_{T}\}}\mathbf{1}_{\llbracket T\rrbracket}(t),\ \ 0\leq t\leq T \nonumber$$ satisfy $b_T\in\mathbb{L}^{\infty}(\mathbb{R}^2)$ (resp. $b'_T\in\mathbb{L}^{\infty}(\mathbb{R}^2)$) and $V(b)_t\geq -K$ (resp. $V(b')_t\geq -K$) for all $0\leq t\leq T$, and therefore $b_T\in\mathcal{U}_0^{\infty}$ (resp. $b'_T\in\mathcal{U}_0^{\infty}$). Following the proof of Proposition $4.2$ of [@Sch33], we can derive that $\tilde{S}_t\geq (1-\lambda)S_t$ for all $0\leq t\leq T$ using the above constructions of portfolios $(b_t)_{0\leq t\leq T}$ and $(b'_t)_{0\leq t\leq T}$.
For the other direction, we first have $$\phi_T^0Z_T^0+\phi_T^1Z_T^1=(\phi_T^0+\phi_T^1\tilde{S}_T)Z_T^0,\nonumber$$ By the definition of acceptable portfolio and the fact that $(\phi_T^0,\phi_T^1)\in\mathbb{L}^{\infty}(\mathbb{R}^2)$ and $S_T\leq K$, for this $\tilde{S}\in\mathcal{S}$, we can clearly find a constant $a$ and a maximal element $X^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S}, a)$ such that $X_T^{\max, \tilde{S}}\in\mathbb{L}^{\infty}$ with $V(\phi^0, \phi^1)_T\geq -X_T^{\max,\tilde{S}}$. Lemma $\ref{auxlem}$ implies $\phi_t^0+\phi_t^1\tilde{S}_t+X_t^{\max, \tilde{S}}\geq 0$ for all $0\leq t\leq T$ and hence it is an optional strong supermartingale under $\mathbb{Q}$. Since $(\phi_T^0,\phi_T^1)\in\mathbb{L}^{\infty}(\mathbb{R}^2)$, it follows that $\phi_T^0+\phi_T^1\tilde{S}_T$ is $\mathbb{Q}$-integrable. Also, since $X_T^{\max,\tilde{S}}\geq 0$ $\mathbb{P}$-a.s. and hence $\mathbb{Q}$-a.s., it is also $\mathbb{Q}$-integrable. It is easy to see that $$\begin{split}
\mathbb{E}[\phi_T^0Z_T^0+\phi_T^1Z_T^1]&=\mathbb{E}^{\mathbb{Q}}[\phi_T^0+\phi_T^1\tilde{S}_T]+\mathbb{E}^{\mathbb{Q}}[X_T^{\max,\tilde{S}}]-\mathbb{E}^{\mathbb{Q}}[X_T^{\max,\tilde{S}}]\\
&=\mathbb{E}^{\mathbb{Q}}[\phi_T^0+\phi_T^1\tilde{S}_T+X_T^{\max,\tilde{S}}]-\mathbb{E}^{\mathbb{Q}}[X_T^{\max,\tilde{S}}]\\
&\leq a-\mathbb{E}^{\mathbb{Q}}[X_T^{\max,\tilde{S}}].\nonumber
\end{split}$$
For this fixed $\tilde{S}$ and $\mathbb{Q}$, we can find a sequence $\mathbb{Q}^n\in\mathcal{M}'(\tilde{S})$ converging to $\mathbb{Q}$ in the norm topology of $\mathbb{L}^1$, where $$\mathcal{M}'(\tilde{S})\triangleq \{\mathbb{Q}\in\mathcal{M}(\tilde{S}):\ \ \text{$X^{\max,\tilde{S}}$ is a UI martingale under $\mathbb{Q}$}\}.\nonumber$$ Since $X_T^{\max,\tilde{S}}\in\mathbb{L}^{\infty}$, it follows that $$\begin{split}
\mathbb{E}[\phi_T^0Z_T^0+\phi_T^1Z_T^1]&\leq a-\mathbb{E}^{\mathbb{Q}}[X_T^{\max,\tilde{S}}]\leq a-\lim_{n\rightarrow\infty}\mathbb{E}^{\mathbb{Q}^n}[X_T^{\max,\tilde{S}}]\\
&=a-a=0,\nonumber
\end{split}$$ which completes the proof and $(Z_T^0, Z_T^1)\in\bar{\mathcal{Z}}$.
We now pass from $\mathcal{U}_x^{\infty}$ to $\mathcal{U}_x$ and characterize the set $\mathcal{U}_x$ still using the set $\bar{\mathcal{Z}}$.
\[lem13\] Assume that $S_t\leq K$ for some constant $K>0$, $0\leq t\leq T$. We can characterize the set $\mathcal{U}_x(\lambda, S)$ (short as $\mathcal{U}_x$) using the set $\bar{\mathcal{Z}}$ by $$\label{chaUacp}
\begin{split}
\mathcal{U}_x\triangleq\Big\{& (\phi^0_T, \phi^1_T)\in\mathbb{L}^0(\mathbb{R}^2):\ \text{there exists a constant $a>0$ such that for each $\tilde{S}\in\mathcal{S}$}\\
& \exists X^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S}, a)\ \text{with}\ (\phi^0_T, \phi^1_T)\succeq (-X_T^{\max,\tilde{S}}, 0)\ \text{and}\\
&\mathbb{E}[(\phi^0_T+X_T^{\max,\tilde{S}})Z_T^0+\phi_T^1Z_T^1]\leq x+\mathbb{E}[X_T^{\max,\tilde{S}}Z_T^0],\ \ \forall Z_T\in\bar{\mathcal{Z}}\Big\}.
\end{split}$$
In the two dimensional setting with partial order defined in Remark $\ref{order}$, similar to the proof of Lemma $2.5$ of [@Yu], we can obtain that $\mathcal{U}_0^{\infty}$ is $\text{weak}^{\ast}$ closed (i.e., closed in $\sigma(\mathbb{L}^{\infty}, \mathbb{L}^1)$). It is shown in Theorem $5.5.3$ of [@Kab] that we have the following characterization $$\mathcal{U}_0^{\infty}=\{(\phi_T^0, \phi_T^1)\in\mathbb{L}^{\infty}_r(\mathbb{R}^2): \mathbb{E}[\phi_T^0Z_T^0+\phi_T^1Z_T^1]\leq 0,\ \ \forall Z\in(\mathcal{U}_0^{\infty})^{\circ}\},\nonumber$$ where $\mathbb{L}_r^{\infty}(\mathbb{R}^2)$ is the set of all $\mathcal{F}_T$-measurable random vector $(\phi_T^0, \phi_T^1)\in\mathbb{L}^{\infty}(\mathbb{R}^2)$ such that there exists a constant $a>0$ and for each $\tilde{S}$, there exits a $X^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S}, a)$ with $(\phi_T^0, \phi_T^1)\succeq (-X_T^{\max,\tilde{S}}, 0)$. And since $\mathcal{U}_0^{\infty}$ contains the negative orthant $-\mathbb{L}^{\infty}(\mathbb{R}^2)$, its polar $(\mathcal{U}_0^{\infty})^{\circ}$ is therefore defined by $$(\mathcal{U}_0^{\infty})^{\circ}=\{(Z_T^0,Z_T^1)\in\mathbb{L}^1(\mathbb{R}^2): \mathbb{E}[\phi_T^0Z_T^0+\phi_T^1Z_T^1]\leq 0\}.\nonumber$$ By Definition $\ref{nameZ}$ and Lemma $\ref{lemZ}$, it is clear that $\bar{\mathcal{Z}}=(\mathcal{U}_0^{\infty})^{\circ}$, and hence, we get that $$\label{chaU}
\mathcal{U}_0^{\infty}=\{(\phi_T^0, \phi_T^1)\in\mathbb{L}^{\infty}_r(\mathbb{R}^2): \mathbb{E}[\phi_T^0Z_T^0+\phi_T^1Z_T^1]\leq 0,\ \ \forall Z\in\bar{\mathcal{Z}}\}.$$
We then claim that $\mathcal{U}_x^{\infty}$ is relatively Fatou dense in $\mathcal{U}_x$. Consider any $(\phi_T^0,\phi_T^1)\in\mathcal{U}_x$, there exists $\hat{a}>0$ and for each $\tilde{S}\in\mathcal{S}$, there exits a $\hat{X}^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S}, \hat{a})$ such that $V(\phi^0,\phi^1)_T+\hat{X}_T^{\max,\tilde{S}}\geq 0$. We need to show that there exits a sequence $(\phi^{0,n}_T, \phi^{1,n}_T)\in\mathcal{U}_x^{\infty}$ such that $(\phi_T^{0,n}, \phi_T^{1,n})\rightarrow (\phi^0_T, \phi_T^1)$ a.s.. Define the sequence $(\phi^{0,n}_T, \phi^{1,n}_T)$ by $$\phi_T^{0,n}=\phi_T^0\mathbf{1}_{\{ |\phi_T^0|\leq n, |\phi_T^1|\leq n\}},\ \ \ \phi_T^{1,n}=\phi_T^1\mathbf{1}_{\{|\phi_T^0|\leq n, |\phi_T^1|\leq n\}}.\nonumber$$ Clearly, $(\phi_T^{0,n}, \phi_T^{1,n})\rightarrow (\phi^0_T, \phi_T^1)$ a.s.. and $(\phi_T^{0,n}, \phi_T^{1,n})\in\mathbb{L}^{\infty}(\mathbb{R}^2)$. Also, it follows by definition that $(\phi_T^{0,n}, \phi_T^{1,n})\succeq (-\hat{X}_T^{\max,\tilde{S}}, 0)$ since we have $$\begin{split}
V(\phi^{0,n}, \phi^{1,n})_T+\hat{X}_T^{\max,\tilde{S}}&=V(\phi^0,\phi^1)_T\mathbf{1}_{\{ |\phi_T^0|\leq n, |\phi_T^1|\leq n\}}+\hat{X}_T^{\max,\tilde{S}}\\
&\geq V(\phi^0,\phi^1)_T\mathbf{1}_{\{ |\phi_T^0|\leq n, |\phi_T^1|\leq n\}}+\hat{X}_T^{\max,\tilde{S}}\mathbf{1}_{\{ |\phi_T^0|\leq n, |\phi_T^1|\leq n\}}\\
&=\Big(V(\phi^0,\phi^1)_T + \hat{X}_T^{\max,\tilde{S}}\Big)\mathbf{1}_{\{ |\phi_T^0|\leq n, |\phi_T^1|\leq n\}}\geq 0.\nonumber
\end{split}$$ By choosing $$\phi_t^{0,n}=\phi_t^0,\ \ \ \phi_t^{1,n}=\phi_t^1,\nonumber$$ for $0\leq t<T$. The pair $(\phi^{0,n},\phi^{1,n})$ is clearly self-financing. Therefore, $(\phi^{0,n}_T, \phi^{1,n}_T)\in\mathcal{U}_x^{\infty}$ and the claim holds.
Following the proof of Lemma $2.6$ of [@Yu], using $(\ref{chaU})$ and the fact that $\mathcal{U}_x^{\infty}$ is relatively Fatou dense in $\mathcal{U}_x$, we can obtain the characterization $(\ref{chaUacp})$.
Based on all previous results from Lemma $\ref{lemm}$ to Lemma $\ref{lem13}$, we can finally build the other side of the super-hedging theorem for acceptable portfolios. This is an analogue of Lemma $5$ of [@kram04] in the market without transaction costs. The proof relies heavily on the characterization $(\ref{chaUacp})$ in Lemma $\ref{lem13}$. However, we remind the reader that it is assumed that the price process $S$ is uniformly bounded in Lemma $\ref{lem13}$. Therefore, the trick of working with the localizing sequence is necessary here.
\[charac\] Fix some $x\in\mathbb{R}$. Let $g$ be an $\mathbb{R}$-valued, $\mathcal{F}_T$-measurable random variable such that there exits a constant $a>0$ and for each $\tilde{S}\in\mathcal{S}(\lambda)$, there exists a $X^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S}, a)$ with $g+X_T^{\max,\tilde{S}}\geq 0$. If for each $\lambda$-CPS $(\mathbb{Q},\tilde{S})$, i.e., for all $\mathbb{Q}\in\mathcal{M}$, we have $$\label{ineassum}
\mathbb{E}^{\mathbb{Q}}[g]\leq x,$$ then there exits a pair $(\phi^0, \phi^1)\in\mathcal{A}_x$ such that $(\phi^0_0, \phi_0^1)=(x,0)$ and $(\phi_T^0, \phi_T^1)=(g,0)$.
Define the $[0,T]$-valued stopping time $\tau_n$ by $$\tau_n\triangleq \inf\{t:S_t\geq n\}\wedge T.\nonumber$$ Also define $$g^n=\left\{
\begin{array}{rl}
g,&\ \ \ \ \text{on}\ \ \{\tau_n=T\},\\
\underset{\mathbb{Q}\in\mathcal{M}(\lambda, S)}{\text{essinf}}\mathbb{E}^{\mathbb{Q}}\bigg[-\underset{\tilde{S}\in\mathcal{S}(\lambda)}{\text{essinf}}\ \Big(X_T^{\max,\tilde{S}}\Big) \bigg|\mathcal{F}_{\tau_n}\bigg],&\ \ \ \ \text{on}\ \ \{\tau_n<T\}.\nonumber
\end{array}\right.$$ It is clear that $(g^n)_{n=1}^{\infty}$ is $\mathcal{F}_{\tau_n}$-measurable and $g^n$ converges to $g$, $\mathbb{P}$-a.s..
Let $0<\lambda_n <\lambda$ be a sequence of real numbers increasing to $\lambda$. For each fixed $n\in\mathbb{N}$, we consider the stopped process $S^{\tau_n}$ with the transaction costs $\lambda^n$. It is easy to check that for $0<\lambda'<1$, any stopped $\lambda'$-CPS $(\mathbb{Q}, \tilde{S}^{\tau_n})$ for $S$ is also a $\lambda'$-CPS for $S^{\tau_n}$. Moreover, by Proposition $6.1$ of [@Sch33], for any stopped $\lambda'$-CPS $(\mathbb{Q},\tilde{S}^{\tau_n})$, we obtain that $\tilde{S}^{\tau_n}$ is a true $\mathbb{Q}$-martingale instead of a $\mathbb{Q}$-local martingale. Therefore, for any $0<\lambda'<1$, the stopped process $S^{\tau_n}$ admits a $\lambda'$-CPS $(\mathbb{Q}, \tilde{S})$ such that $\tilde{S}$ is a $\mathbb{Q}$-martingale.
Following the proof of Theorem $1.4$ of [@Sch33], for each fixed $n$, we will only consider the $\lambda_n$-CPS $(\mathbb{Q}, \tilde{S})$ such that $\tilde{S}$ takes values in the spread $[(1-\lambda_n)S^{\tau_n}, S^{\tau_n}]$ and $\tilde{S}$ is a true $\mathbb{Q}$-martingale. Let $(Z^0,Z^1)$ denote the associated martingales with respect to the $\lambda_n$-CPS $(\mathbb{Q}, \tilde{S})$ for the stopped price process $S^{\tau_n}$. We will construct a $\lambda$-CPS $(\bar{Z}^0, \bar{Z}^1)$ for the original price process $S$. Fix $0<\lambda' <\frac{\lambda-\lambda_n}{2}$. Assumption $\ref{Assum}$ gives the existence of a $\lambda'$-CPS $(\hat{Z}^0, \hat{Z}^1)$ for $S$ where $\hat{Z}^0$ is a martingale and $\hat{Z}^1$ is a local martingale.
Let us define $$\bar{Z}_t^0=\left\{
\begin{array}{rl}
Z_t^0,&\ \ \ \ 0\leq t\leq \tau_n,\\
\hat{Z}_t^0\frac{Z^0_{\tau_n}}{\hat{Z}^0_{\tau_n}},&\ \ \ \ \tau_n\leq t\leq T,
\end{array}\right.\nonumber$$ and also $$\bar{Z}_t^1=\left\{
\begin{array}{rl}
(1-\lambda')Z_t^1,&\ \ \ \ 0\leq t\leq \tau_n,\\
(1-\lambda')\hat{Z}_t^1\frac{Z^1_{\tau_n}}{\hat{Z}^1_{\tau_n}},&\ \ \ \ \tau_n\leq t\leq T.
\end{array}\right.\nonumber$$
It is clear that $\bar{Z}^0$ (resp. $\bar{Z}^1$) is a positive martingale (resp. local martingale) under $\mathbb{P}$ and $\frac{d\bar{\mathbb{Q}}}{d\mathbb{P}}=\bar{Z}^0_T$ defined a probability measure on $\mathcal{F}$ which is equivalent to $\mathbb{P}$. Moreover, for $0\leq t\leq \tau_n$, we have $\frac{\bar{Z}_t^1}{\bar{Z}_t^0}$ stays in the spread $[(1-\lambda_n)(1-\lambda')S_t, (1-\lambda')S_t]$. On the other hand, for $\tau_n\leq t\leq T$, we can verify that $\frac{\bar{Z}_t^1}{\bar{Z}_t^0}$ lies in $[(1-\lambda_n)(1-\lambda')^2S_t, (1-\lambda')^2S_t]$. It follows that $\frac{\bar{Z}^1}{\bar{Z}^0}$ takes its values in $[(1-\lambda)S, S]$. We first claim that $$\label{addgn}
\mathbb{E}^{\bar{\mathbb{Q}}}[g^n]\leq \mathbb{E}^{\bar{\mathbb{Q}}}[g].$$ To see this, let us denote $f\triangleq -\underset{\tilde{S}\in\mathcal{S}(\lambda)}{\text{essinf}}\ \Big(X_T^{\max,\tilde{S}}\Big)$. Since we have $\bar{\mathbb{Q}}\in\mathcal{M}(\lambda, S)$, it follows that $$\begin{split}
\mathbb{E}^{\bar{\mathbb{Q}}}[g^n]&=\mathbb{E}^{\bar{\mathbb{Q}}}[g\mathbf{1}_{\{\tau_n=T\}}]+\mathbb{E}^{\bar{\mathbb{Q}}}[g^n\mathbf{1}_{\{\tau_n<T\}}]\\
&\leq \mathbb{E}^{\bar{\mathbb{Q}}}[g\mathbf{1}_{\{\tau_n=T\}}]+\mathbb{E}^{\bar{\mathbb{Q}}}[ \mathbb{E}^{\bar{\mathbb{Q}}}[ f|\mathcal{F}_{\tau_n}]\mathbf{1}_{\{\tau_n<T\}}]\\
&=\mathbb{E}^{\bar{\mathbb{Q}}}[g\mathbf{1}_{\{\tau_n=T\}}]+\mathbb{E}^{\bar{\mathbb{Q}}}[f \mathbf{1}_{\{\tau_n<T\}} ].\nonumber
\end{split}$$ Recall that $g\geq f$, $\mathbb{P}$-a.s., and therefore $g\mathbf{1}_{\{\tau_n<T\}}\geq f\mathbf{1}_{\{\tau_n<T\}}$, $\mathbb{P}$-a.s.. It follows that $g\mathbf{1}_{\{\tau_n<T\}}\geq f\mathbf{1}_{\{\tau_n<T\}}$, $\bar{\mathbb{Q}}$-a.s. since $\bar{\mathbb{Q}}\sim\mathbb{P}$. We deduce that $$\mathbb{E}^{\bar{\mathbb{Q}}}[f \mathbf{1}_{\{\tau_n<T\}} ]\leq \mathbb{E}^{\bar{\mathbb{Q}}}[g \mathbf{1}_{\{\tau_n<T\}} ],\nonumber$$ which implies that $(\ref{addgn})$ holds. By $(\ref{ineassum})$, $(\ref{addgn})$ and the fact that $g^n$ is $\mathcal{F}_{\tau_n}$-measurable, we can conclude that $$\label{ineg}
\mathbb{E}^{\mathbb{Q}}[g^n]=\mathbb{E}^{\bar{\mathbb{Q}}}[g^n]\leq \mathbb{E}^{\bar{\mathbb{Q}}}[g]\leq x.$$
For each fixed $n\in\mathbb{N}$, consider a pair $(\phi_T^{0,n},\phi_T^{1,n})\notin \mathcal{U}_x(\lambda_n, S^{\tau_n})$ where we consider the stopped process $S^{\tau_n}$ as the underlying price process with transaction costs $\lambda_n$ such that $\phi_t^{i,n}=\phi_{\tau_n}^{i,n}$ for $\tau_{n}\leq t\leq T$. By the characterization $(\ref{chaUacp})$ of $\mathcal{U}_x(\lambda_n, S^{\tau_n})$, for any constant $a_n>0$, there exits a $\tilde{S}^n\in\mathcal{S}(\lambda_n, S^{\tau_n})$ such that for any $\widetilde{X}^{\max,\tilde{S}^n}\in\mathcal{X}(\tilde{S}^{n}, a_n)$ with $(\phi_T^{0,n}, \phi_T^{1,n})\succeq (-\widetilde{X}_T^{\max,\tilde{S}^n}, 0)$, we have $$\mathbb{E}[(\phi_T^{0,n}+\widetilde{X}_T^{\max,\tilde{S}^n})Z_T^{0,n}+\phi_T^{1,n}Z_T^{1,n}]> x+\mathbb{E}[\widetilde{X}_T^{\max,\tilde{S}^n}Z_T^{0,n}],\nonumber$$ for some $(Z^{0,n}_T, Z^{1,n}_T)\in\bar{\mathcal{Z}}(\lambda_n, S^{\tau_n})$.
In particular, we can choose $\widetilde{X}_t^{\max,\tilde{S}^n}\equiv a_n$ for $0\leq t\leq T$. It follows that $\phi_T^{0,n}+\phi_T^{1,n}\tilde{S}^n_T\geq -a_n$ $\mathbb{Q}^n$-a.s. where $\frac{d\mathbb{Q}^n}{d\mathbb{P}}=Z_T^{0,n}$, and hence it is integrable. It follows that $$\begin{split}
\mathbb{E}[(\phi_T^{0,n}+\widetilde{X}_T^{\max,\tilde{S}^n})Z_T^{0,n}+\phi_T^{1,n}Z_T^{1,n}]=&\mathbb{E}^{\mathbb{Q}^n}[(\phi_T^{0,n}+\phi_T^{1,n}\tilde{S}^n_T)+\widetilde{X}_T^{\max,\tilde{S}^n}]\\
=&\mathbb{E}^{\mathbb{Q}^n}[\phi_T^{0,n}+\phi_T^{1,n}\tilde{S}^n_T]+\mathbb{E}^{\mathbb{Q}^n}[\widetilde{X}_T^{\max,\tilde{S}^n}]\\
>&x+\mathbb{E}^{\mathbb{Q}^n}[\widetilde{X}_T^{\max,\tilde{S}^n}],\nonumber
\end{split}$$ which implies that $$\mathbb{E}^{\mathbb{Q}^n}[\phi_T^{0,n}+\phi_T^{1,n}\tilde{S}^n_T]>x.\nonumber$$ In the case that $\mathbb{Q}^n$ is only absolutely continuous with respect to $\mathbb{P}$, but not equivalent to $\mathbb{P}$, The above argument asserts that any $\lambda^n$-CPS for $S$ is $\lambda^n$-CPS for $S^{\tau_n}$. Therefore, there exists some $(\bar{Z}^0_T, \bar{Z}^1_T)\in\bar{\mathcal{Z}}(\lambda^n, S^{\tau_n})$ such that $\bar{Z}_T^0>0$ a.s.. For $0<\beta<1$ sufficiently small, define $\hat{Z}^n=\beta\bar{Z}_T+(1-\beta)Z^{n}_T$. We obtain that $\hat{Z}_T^{0,n}>0$ a.s.. Define $$\hat{S}^n_t=\frac{\hat{Z}^{1,n}_t}{\hat{Z}^{0,n}_t},\ \ \ \text{and}\ \ \ \frac{d\hat{\mathbb{Q}}^n}{d\mathbb{P}}=\hat{Z}_T^{0,n}.\nonumber$$ We have $(\hat{\mathbb{Q}}^n, \hat{S}^n)$ is $\lambda^n$-CPS and also $$\mathbb{E}^{\hat{\mathbb{Q}}^n}[\phi_T^{0,n}+\phi_T^{1,n}\tilde{S}^n_T]>x,\nonumber$$ which can not satisfy $(\ref{ineg})$. Therefore, we obtain that if $(\ref{ineg})$ holds, there exits a pair $(\phi_T^{0,n}, \phi_T^{1,n})\in\mathcal{U}_x(\lambda^n, S^{\tau_n})$ such that $(\phi_0^{0,n}, \phi_0^{1,n})=(x,0)$ and $(\phi_T^{0,n}, \phi_T^{1,n})=(\phi_{\tau_n}^{0,n}, \phi_{\tau_n}^{1,n})=(g^n, 0)$.
At last, by taking the limit $n\rightarrow\infty$ and the convex combinations of $(\phi^{0,n}, \phi^{1,n})$, similar to the proof of Lemma $\ref{fatoulem}$, we can conclude that the limit $(\phi^0, \phi^1)\in\mathcal{U}_x(\lambda,S)$, which completes the proof.
For the proof of Proposition $\ref{dualprop}$, we still need some auxiliary results from Lemma $\ref{clK}$ to Lemma $\ref{lemma11}$ below.
\[clK\] Under Assumptions $\ref{Assum}$, $\ref{assE}$ and $\ref{assNo}$, we have that $$\bar{\mathcal{K}}=\{(x,q)\in\mathbb{R}^{1+N}:\mathcal{H}(x,q)\neq \emptyset\},\nonumber$$ where $\bar{\mathcal{K}}$ is the closure of the set $\mathcal{K}$ in $\mathbb{R}^{1+N}$.
Fix any $(x,q)\in\bar{\mathcal{K}}$, and let $(x^n,q^n)_{n\geq 1}$ be a sequence in $\mathcal{K}$ that converges to $(x,q)$. We need to verify that $\mathcal{H}(x,q)\neq\emptyset$. Choose a sequence $V_T^n\in\mathcal{H}(x^n,q^n)$ with $V_T^n=V(\phi^{0,n},\phi^{1,n})_T$ and $(\phi^{0,n}, \phi^{1,n})\in\mathcal{A}_{x^n}$, $n\geq 1$. Lemma $\ref{ineqC}$ gives that $$\label{lemineqC}
\mathbb{E}^{\mathbb{Q}}[V(\phi^{0,n}, \phi^{1,n})_T+q^n\cdot\mathcal{E}_T]\leq x^n+\mathbb{E}^{\mathbb{Q}}[q^n\cdot\mathcal{E}_T],\ \ \forall \mathbb{Q}\in\mathcal{M}.$$ In addition, the fact that $x^n\rightarrow x$ and $q^n\rightarrow q$ imply that that there exists finite constants $k_1$ and $k_2$ such that $x^n<k_1$ and $(q^n)^i<k_2$, $1\leq i\leq N$, for $n$ large enough. We deduce that $q^n\cdot\mathcal{E}_T \leq k_2\sum_{i=1}^{N}\mathcal{E}_T^i$. By Lemma $\ref{boundE}$, it follows that there exists a constant $\hat{a}>0$ and for each $\tilde{S}\in\mathcal{S}(\lambda)$, there exists a $\hat{X}^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S},\hat{a})$ such that $V(\phi^{0,n}, \phi^{1,n})_T+\hat{X}_T^{\max,\tilde{S}}\geq 0$ for $n$ large enough. Lemma $\ref{lemm}$ and Lemma $A1.1$ of [@Sch94] imply that we can find the convex combinations of $\phi_T^{0,n}$ and $\phi_T^{1,n}$ converging almost surely to random variables $\phi_T^{0}$ and $\phi_T^{1}$ respectively. Moreover, it is clear that $V(\phi^0,\phi^1)_T+q\cdot\mathcal{E}_T\geq 0$ a.s. where $V(\phi^0,\phi^1)_T=\phi_T^0+(\phi_T^1)^+(1-\lambda)S_T-(\phi_T)^-S_T$. Fatou’s Lemma and $(\ref{lemineqC})$ therefore imply that $$\begin{split}
&\mathbb{E}^{\mathbb{Q}}[V(\phi^0,\phi^1)_T+q\cdot\mathcal{E}_T]\leq \lim_{n\rightarrow\infty}\mathbb{E}^{\mathbb{Q}}[V(\phi^{0,n},\phi^{1,n})_T+q^n\cdot\mathcal{E}_T]\\
\leq &\lim_{n\rightarrow\infty}\Big(x^n+\mathbb{E}^{\mathbb{Q}}[q^n\cdot\mathcal{E}_T]\Big)=x+\mathbb{E}^{\mathbb{Q}}[q\cdot\mathcal{E}_T],\ \ \forall \mathbb{Q}\in\mathcal{M}.\nonumber
\end{split}$$ It follows that $$\mathbb{E}^{\mathbb{Q}}[V(\phi^0,\phi^1)_T]\leq x,\ \ \forall \mathbb{Q}\in\mathcal{M}.\nonumber$$ Lemma $\ref{charac}$ guarantees the existence of acceptable portfolio $(\hat{\phi}^0,\hat{\phi}^1)\in\mathcal{A}_x$ such that $V(\hat{\phi}^0,\hat{\phi}^1)_T\geq V(\phi^0,\phi^1)_T\geq -q\cdot\mathcal{E}_T$. Therefore, we obtain that $V(\hat{\phi}^0,\hat{\phi}^1)_T\in\mathcal{H}(x,q)$, which completes the proof.
For a vector $p\in\mathbb{R}^N$, we define the set $$\label{auxiM}
\mathcal{M}(p)=\{\mathbb{Q}\in\mathcal{M}: \mathbb{E}^{\mathbb{Q}}[\mathcal{E}_T]=p\}$$ From its definition, $\mathcal{P}$ is the intersection of $\mathcal{L}$ with the hyperplane $y\equiv 1$ which defines the set of arbitrage-free prices of the contingent claim $\mathcal{E}_T$.
\[lemma9\] Assume that all conditions of Proposition $\ref{dualprop}$ hold and let $p\in\mathbb{R}^N$. The set $\mathcal{M}(p)$ is not empty if and only if $p\in\mathcal{P}$. In particular, $$\bigcup_{p\in\mathcal{P}}\mathcal{M}(p)=\mathcal{M}.\nonumber$$
Under Assumptions $\ref{assE}$ and $\ref{assNo}$, Lemma $\ref{lemma9}$ follows directly from the proof Lemma $8$ of [@kram04], if we replace the set $\mathcal{M}'(p)$, Lemma $4$, Lemma $5$ and Lemma $6$ in [@kram04] by the set $\mathcal{M}(p)$, Lemma $\ref{ineqC}$, Lemma $\ref{charac}$ and Lemma $\ref{clK}$ in this paper.
\[lemma10\] Under the assumptions of Proposition $\ref{dualprop}$ and $p\in\mathcal{P}$, the density process of any $\mathbb{Q}\in\mathcal{M}(p)$ belongs to $\mathcal{Y}(1,p)$.
According to the definition of CPS and Proposition $2.3$ of [@Sch2], it is clear that the density process of $\mathbb{Q}\in\mathcal{M}$ belongs to $\mathcal{Y}(1)$ defined by $(\ref{dualY})$ . The conclusion follows by Lemma $\ref{ineqC}$ and the definition of $\mathcal{M}(p)$.
\[lemma11\] Under the assumptions of Proposition $\ref{dualprop}$, a nonnegative random variable $g$ belongs to $\mathcal{C}(x,q)$ where $(x,q)\in\mathcal{K}$ if and only if $$\label{needto}
\mathbb{E}^{\mathbb{Q}}[g]\leq x+p\cdot q,\ \ \forall p\in\mathcal{P}\ \text{and}\ \mathbb{Q}\in\mathcal{M}(p).$$
Suppose $g\in\mathcal{C}(x,q)$, Lemma $\ref{ineqC}$ implies the inequality $(\ref{needto})$. On the other hand, consider the random variable $\beta\triangleq g-q\cdot\mathcal{E}_T$. It follows from $(\ref{needto})$ that $$\begin{split}
\sup_{\mathbb{Q}\in\mathcal{M}}\mathbb{E}^{\mathbb{Q}}[\beta]&=\sup_{p\in\mathcal{P}}\sup_{\mathbb{Q}\in\mathcal{M}(p)}\mathbb{E}^{\mathbb{Q}}[\beta]\\
&=\sup_{p\in\mathcal{P}}\sup_{\mathbb{Q}\in\mathcal{M}(p)}(\mathbb{E}^{\mathbb{Q}}[g]-q\cdot p)\leq x.\nonumber
\end{split}$$ Assumption $\ref{assE}$ and Lemma $\ref{boundE}$ imply the existence of a constant $\hat{a}>0$ such that for each $\tilde{S}\in\mathcal{S}$, there exists a $\hat{X}^{\max,\tilde{S}}\in\mathcal{X}(\tilde{S},\hat{a})$ and $\beta\geq -\hat{X}_T^{\max,\tilde{S}}$. Lemma $\ref{charac}$ guarantees the existence of acceptable portfolio with $\phi^0_0=x$, $\phi_0^1=0$ and $V(\phi^0,\phi^1)_T=\phi_T^0\geq \beta$. We therefore obtain that $$0\leq g\leq V(\phi^0,\phi^1)_T+q\cdot\mathcal{E}_T,\nonumber$$ which implies that $V(\phi^0,\phi^1)_T\in\mathcal{H}(x,q)$ and $g$ belongs to $\mathcal{C}(x,q)$.
We are now ready to proceed to complete the proof of Proposition $\ref{dualprop}$.
We first prove the assertion $(i)$. Assume that $(x,q)\in\mathcal{K}$. We can find a constant $\delta>0$ such that $(x-\delta, q)\in\mathcal{K}$ since $\mathcal{K}$ is open. Consider $V_T=V(\phi^0,\phi^1)_T\in\mathcal{H}(x-\delta,q)$, it is clear that $\widetilde{V}\triangleq V(\phi^0+\delta,\phi^1)$ is in $\mathcal{H}(x,q)$ and $\delta\leq \widetilde{V}_T+q\cdot\mathcal{E}_T$ which implies that $\delta\in\mathcal{C}(x,q)$.
Let $(x,q)\in\mathcal{K}$. If $g\in\mathcal{C}(x,q)$, $(\ref{bipo1})$ holds true by the definition of $\mathcal{D}(y,r)$, $(y,r)\in\mathcal{L}$. On the other hand, consider a nonnegative random variable such that $(\ref{bipo1})$ holds. It follows that $g$ satisfies $(\ref{needto})$ by Lemma $\ref{lemma10}$. Lemma $\ref{lemma11}$ then implies that $g$ belongs to $\mathcal{C}(x,q)$.
It is clear that $k\mathcal{D}(y,r)=\mathcal{D}(ky,kr)$ for any $k>0$ and $(y,r)\in\mathcal{L}$. Hence, to verify the assertion $(ii)$, it is enough to consider the case that $(y,r)=(1,p)$ for some $p\in\mathcal{P}$. Due to Lemma $\ref{lemma10}$, there exists a process $Y_t=\mathbb{E}\Big[\frac{d\mathbb{Q}}{d\mathbb{P}}\Big|\mathcal{F}_t\Big]$ with $\mathbb{Q}\in\mathcal{M}(p)$, which satisfies $Y_T\in\mathcal{D}(1,p)$ and $Y_T>0$ a.s..
For any $h\in\mathcal{D}(1,p)$, $(\ref{bipo2})$ holds by the definition of $\mathcal{D}(1,p)$. Conversely, consider any nonnegative random variable $h$ satisfying $(\ref{bipo2})$. In particular, we have that $$\mathbb{E}^{\mathbb{Q}}[gh]\leq 1,\ \ \forall g\in\mathcal{C}(1,0).\nonumber$$ Since $\mathcal{C}(1,0)=\mathcal{V}_1^{\text{adm}}$ which is defined in $(\ref{0admV})$, Lemma $A.1$ in [@Chris22] asserts the existence of an optional strong supermartingale $(Y^0, Y^1)\in\mathcal{Z}(1)$ such that $h\leq Y_T^0$. Let us define the process $\tilde{Y}$ by $$\tilde{Y}_t=\left\{
\begin{array}{rl}
Y_t^0,&\ \ \ \ t<T,\\
h,&\ \ \ \ t=T.
\end{array}\right.\nonumber$$ It follows that $\tilde{Y}\in\mathcal{Y}(1,p)$. Therefore we obtain that $h\in\mathcal{D}(1,p)$.
Once we build the bipolar results in Proposition $\ref{dualprop}$, Theorem $\ref{mainthm}$ follows the proof of Theorem $1$ and Theorem $2$ of [@kram04] if we replace the one-dimensional duality theory in [@kram99] by Theorem $3.2$ in [@Chris22] under proportional transaction costs.
Proof of Theorem $\ref{theorem1}$
---------------------------------
By Theorem $4$ in Appendix I in [@Meyer], every optional strong supermartingale is indistinguishable from a [làdlàg ]{}process. Without loss of generality, we can assume all optional strong supermartingales are [làdlàg ]{}. In particular, we can assume that $\hat{S}=\frac{Y^{1,\ast}}{Y^{0,\ast}}$ is [làdlàg ]{}. Fix $(x,q)\in\mathcal{K}$, for any $(y,r)\in\partial u(x,q)$, by the self-financing condition and integration by parts formula, we deduce that $$\begin{split}
Y_t^{0,\ast}(y, r)\phi_t^{0,\ast}(x,q)+Y^{1,\ast}_t(y,r)\phi_t^{1,\ast}(x,q)=&Y^{0,\ast}_t(y,r)(\phi_t^{0,\ast}(x,q)+\phi_t^{1,\ast}(x,q)\hat{S}_t)\\
=&Y^{0,\ast}_t(y,r)(x+(\phi^{1,\ast}\cdot \hat{\mathbf{S}})_t+K_t)\nonumber
\end{split}$$ where $(K_t)_{0\leq t\leq T}$ is a non-increasing predictable process defined by $$\begin{split}
K_t\triangleq &\int_0^t(\hat{S}_u-S_u)d\phi^{1,\ast, \uparrow, c}_u(x,q)+ \int_0^1((1-\lambda)S_u-\hat{S}_u)d\phi_u^{1,\ast,\downarrow, c}(x,q)\\
&+\sum_{0<u\leq t}(\hat{S}_{u}^p-S_{u-})\triangle \phi_u^{1,\ast, \uparrow}(x,q)+\sum_{0<u\leq t}((1-\lambda)S_{u-}-\hat{S}_u^p)\triangle \phi_u^{1,\ast,\downarrow}(x,q)\\
&+\sum_{0\leq u<t}(\hat{S}_u- S_u)\triangle_+\phi_u^{1,\ast,\uparrow}(x,q)+\sum_{0\leq u<t}(\hat{S}_u-(1-\lambda)S_u)\triangle_+\phi^{1,\ast,\downarrow}_u(x,q)\nonumber
\end{split}$$ for $t\in[0,T]$. Therefore, to show that the equality $(\ref{equiphi})$ holds is equivalent to show that $(\ref{behavphi})$ holds.
Under Assumption $\ref{importantnew}$, for some $(y,r)\in\partial u(x,q)$, there exists a minimizing sequence $Z^n(y, r)=(Z^{0,n}(y, r), Z^{1,n}(y, r))$ in $\mathcal{B}(1)$ such that $$\label{usenew}
\underset{n\rightarrow\infty}{\lim\inf}\mathbb{E}[Z_T^{0,n}(y, r)\mathcal{E}_T]=\frac{r}{y}.$$ By Lemma $\ref{seplemma}$, there exists a sandwiched strong supermartingale deflator $\mathbf{Y}^{\ast}(y,r)=(Y^{\ast, p}(y,r), Y^{\ast}(y,r))$ such that for the same minimizing sequence, we have $$(yZ^{0,n}_{\tau-}(y,r), yZ^{1,n}_{\tau-}(y,r))\xrightarrow{\mathbb{P}}(Y^{0,\ast,p}_{\tau}(y,r), Y^{1,\ast,p}_{\tau}(y,r)), \nonumber$$ and $$(yZ^{0,n}_{\tau}(y,r), yZ^{1,n}_{\tau}(y,r))\xrightarrow{\mathbb{P}}(Y^{0,\ast}_{\tau}(y,r), Y^{1,\ast}_{\tau}(y,r)), \nonumber$$ as $n\rightarrow\infty$ for all $[0,T]$-valued stopping time $\tau$, where $Y^{0,\ast}(y,r)$ is the dual optimizer to the problem $(\ref{dualv})$.
Denote $\tilde{S}^n\triangleq\frac{Z^{1,n}(y, r)}{Z^{0,n}(y, r)}$, $\tilde{S}^n$ stays in the bid-ask spread $[(1-\lambda)S, S]$ and it is hence clear that $\tilde{S}^n\in\mathcal{S}$ under the transaction costs $\lambda$. Using the integration by parts formula again, we get $$\begin{split}
\phi_t^{0,\ast}(x,q)+\phi_t^{1,\ast}\tilde{S}_t^n=&\phi_t^{0,\ast}(x,q)+\int_0^t\phi_u^{1,\ast}(x,q)d\tilde{S}^n_u+\int_0^t\tilde{S}_u^{n}d\phi_u^{1,\ast,c}(x,q)\\
&+\sum_{0<u\leq t}\tilde{S}_{u-}^n\triangle \phi_u^{1,\ast}(x,q)+\sum_{0\leq u<t}\tilde{S}_u^{n}\triangle_+\phi_u^{1,\ast}(x,q)\nonumber
\end{split}$$ so that we can write $$\phi_t^{0,\ast}(x,q)+\phi_t^{1,\ast}\tilde{S}_t^n=x+\int_0^t\phi_u^{1,\ast}(x,q)d\tilde{S}_u^n+K_t^n,\nonumber$$ where $$\begin{split}
K_t^n\triangleq &\int_0^t(\tilde{S}_u^n-S_u)d\phi_u^{1,\ast,\uparrow, c}(x,q)+\int_0^t((1-\lambda)S_u-\tilde{S}_u^n)d\phi_u^{1,\ast,\downarrow, c}(x,q)\\
&+\sum_{0<u\leq t}(\tilde{S}_{u-}^{n}-S_{u-})\triangle \phi_u^{1,\ast,\uparrow}(x,q)+\sum_{0<u\leq t}((1-\lambda)S_{u-}-\tilde{S}_{u-}^{n})\triangle \phi_u^{1,\ast,\downarrow}(x,q)\\
&+\sum_{0\leq u< t}(\tilde{S}_{u}^{n}-S_{u})\triangle_+ \phi_u^{1,\ast,\uparrow}(x,q)+\sum_{0\leq u< t}((1-\lambda)S_{u}-\tilde{S}_{u}^{n})\triangle_+ \phi_u^{1,\ast,\downarrow}(x,q).\nonumber
\end{split}$$ is a non-increasing predictable process.
Since $\phi^{1,\ast}(x,q)$ is predictable and of finite variation, it is clear from integration by parts that $Z^{0,n}(\bar{y}, \bar{r})(x+\phi^{1,\ast}(x,q)\cdot \tilde{S}^n)$ is a local martingale. For the choice of $\tilde{S}^n\in\mathcal{S}$, by the definition of acceptable portfolio, there exists a maximal element $X^{\max,\tilde{S}^n}$ such that $$x+\int_0^t\phi_u^{1,\ast}(x,q)d\tilde{S}^n_u+X_t^{\max,\tilde{S}^n}\geq V(\phi^{0,\ast}(x,q), \phi^{1,\ast}(x,q))_t+X_t^{\max, \tilde{S}^n}\geq 0.\nonumber$$ Also, denote the measure $\frac{d\mathbb{Q}^n}{d\mathbb{P}}=Z_T^{0,n}(y, r)$, we have $\mathbb{Q}^n\in\mathcal{M}(\tilde{S}^n)$. Consider the subset $$\mathcal{M}'(\tilde{S}^n)\triangleq \{\mathbb{Q}\in\mathcal{M}(\tilde{S}^n):\ \text{$X^{\max,\tilde{S}^n}$ is a UI martingale under $\mathbb{Q}$}\}.\nonumber$$ There exists a sequence $(\mathbb{Q}^{n,m})_{m=1}^{\infty}$ in $\mathcal{M}'(\tilde{S}^n)$ converging to $\mathbb{Q}^n$ in the norm topology of $\mathbb{L}^1(\Omega, \mathcal{F}, \mathbb{P})$. And for each $\mathbb{Q}^{n,m}\in\mathcal{M}'(\tilde{S}^n)$, $(x+\phi^{1,\ast}(x,q)\cdot \tilde{S}^n+X^{\max,\tilde{S}^n})$ is a true supermartingale under $\mathbb{Q}^{n,m}$. Hence, we can derive that $$\begin{split}
&\mathbb{E}^{\mathbb{Q}^{n,m}}\Big[x+\int_0^T\phi_u^{1,\ast}(x,q)d\tilde{S}^n_u+q\mathcal{E}_T\Big]\\
=&\mathbb{E}^{\mathbb{Q}^{n,m}}\Big[x+\int_0^T\phi_u^{1,\ast}(x,q)d\tilde{S}^n_u+X_T^{\max,\tilde{S}^n}\Big] -\mathbb{E}^{\mathbb{Q}^{n,m}}[X_T^{\max,\tilde{S}^n}] +\mathbb{E}^{\mathbb{Q}^{n,m}}[q\mathcal{E}_T]\\
\leq& x+\mathbb{E}^{\mathbb{Q}^{n,m}}[q\mathcal{E}_T].\nonumber
\end{split}$$
Following the proof of Lemma $\ref{ineqC}$ and by passing the limit as $m\rightarrow \infty$, we obtain that $$\mathbb{E}\Big[Z^{0,n}_T(y, r) (x+\int_0^T \phi_u^{1,\ast}(x,q)d\tilde{S}_u^n+q\mathcal{E}_T) \Big]\leq x+\mathbb{E}[Z_T^{0,n}(y, r)q\mathcal{E}_T].\nonumber$$ By Fatou’s lemma, we obtain that $$\begin{split}
xy+qr=&\mathbb{E}[Y^{0,\ast}_T(y, r)(\phi_T^{0,\ast}(x,q)+q\mathcal{E}_T)]\leq \underset{n\rightarrow\infty}{\lim\inf}\mathbb{E}[yZ_T^{0,n}(y, r)(\phi_T^{0,\ast}(x,q)+q\mathcal{E}_T) ]\\
\leq& \underset{n\rightarrow\infty}{\lim\inf}\mathbb{E}[yZ_T^{0,n}(y, r)K_T^n]+xy+\lim_{n\rightarrow\infty}\mathbb{E}[yZ_T^{0,n}(y, r)q\mathcal{E}_T]\\
=& \underset{n\rightarrow\infty}{\lim\inf}\mathbb{E}[yZ_T^{0,n}(y, r)K_T^n]+xy+qr.\nonumber
\end{split}$$
Therefore, it follows that $Z_T^{0,n}(y, r)K_T^n$ converges to $0$ in $\mathbb{L}^1(\mathbb{P})$ since $K_T^n\leq 0$. We can mimic the proof of Theorem $3.5$ of [@Chris22] and show that $K_T^n$ converges to $K_T$ almost surely, and hence $K_T=0$. As $K_0=0$ and $K_t$ is a non-increasing process, $(\ref{behavphi})$ is verified and hence $(\ref{equiphi})$ also holds true.
Proof of Theorem $\ref{theosandsha}$
------------------------------------
Using the modified acceptable portfolios in Definition $\ref{sandSaccpt}$, we can now proceed to verify the existence of the sandwiched shadow price.
Under all assumptions of Theorem $\ref{theorem1}$, for some $(y,r)\in\partial u(x,q)$ in Assumption $\ref{importantnew}$, let $Z^n(y, r)$ be the minimizing sequence which satisfies $(\ref{marginalprice})$, $(\ref{con111})$, $(\ref{con222})$. For any $X(\phi^0,\phi^1)_T\in\mathcal{H}(x,q;\hat{\mathbf{S}})$, using Definition $\ref{sandSaccpt}$ of acceptable portfolios under the sandwiched shadow price $\hat{\mathbf{S}}$, we deduce that $$\phi_T^0+\phi_T^1\tilde{S}_T^n+q\mathcal{E}_T=\phi_T^0+q\mathcal{E}_T\geq V(\phi^0, \phi^1)_T+q\mathcal{E}_T\geq 0.\nonumber$$ where $\tilde{S}^n\triangleq \frac{Z^{1,n} (y,r)}{Z^{0,n}(y,r)}\in\mathcal{S}$ under transaction costs $\lambda$. Fatou’s lemma implies that $$\label{newineq1}
\begin{split}
\mathbb{E}\Big[Y_T^{0,\ast}(y, r)(x+\int_0^T\phi_u^{1}d\hat{\mathbf{S}}_u+q\mathcal{E}_T)\Big]&=\mathbb{E}[Y_T^{0,\ast}(y, r)(\phi_T^{0}+\phi_T^{1}\hat{S}_T +q\mathcal{E}_T)]\\
&\leq \underset{n\rightarrow\infty}{\lim\inf}\mathbb{E}[yZ_T^{0,n} (y,r)(\phi_T^0+q\mathcal{E}_T)].
\end{split}$$
Again, Definition $(\ref{sandSaccpt})$ gives the existence of some $X^{\max,\tilde{S}^n}\in\mathcal{X}(\tilde{S}^n, a)$ for some constant $a>0$ such that $\phi_T^0+X_T^{\max,\tilde{S}^n}\geq 0$. By the similar proof of Theorem $\ref{theorem1}$ above, we deduce that $$\label{newineq2}
\lim_{n\rightarrow\infty}\mathbb{E}[yZ_T^{0,n} (y,r)(\phi_T^0+q\mathcal{E}_T)]\leq xy+qr=\mathbb{E}[Y^{0,\ast}_T(y, r)(\phi_T^{0,\ast}(x,q)+q\mathcal{E}_T)].$$
Fenchel’s inequality implies that $$\begin{split}
\mathbb{E}\Big[U(x+\int_0^T\phi_u^1d\hat{\mathbf{S}}_u+q\mathcal{E}_T)\Big]&\leq \mathbb{E}\Big[\tilde{U}(Y_T^{0,\ast}(y, r))+Y_T^{0,\ast}(y, r)(x+\int_0^T\phi_u^1d\hat{\mathbf{S}}_u+q\mathcal{E}_T)\Big]\\
&=\mathbb{E}[\tilde{U}(Y_T^{0,\ast}(y, r))+Y_T^{0,\ast}(y, r)(\phi_T^0+\phi_T^1\hat{S}_T+q\mathcal{E}_T)]\\
&\leq \mathbb{E}[\tilde{U}(Y_T^{0,\ast}(y,r))+Y_T^{0,\ast}(y,r)(\phi_T^{0,\ast}(x,q)+\phi_T^{1,\ast}(x,q)\hat{S}_T+q\mathcal{E}_T)]\nonumber
\end{split}$$ using $(\ref{newineq1})$ and $(\ref{newineq2})$. Therefore, by $(\ref{ok1})$, it is easy to see that $$\begin{split}
\mathbb{E}[U(X(\phi^0,\phi^1)_T+q\mathcal{E}_T)]=&\mathbb{E}\Big[U(x+\int_0^T\phi_u^1d\hat{\mathbf{S}}_u+q\mathcal{E}_T)\Big]\\
\leq &\mathbb{E}[\tilde{U}(Y_T^{0,\ast}(y, r))+Y_T^{0,\ast}(y, r)(\phi_T^{0,\ast}(x,q)+\phi_T^{1,\ast}(x,q)\hat{S}_T+q\mathcal{E}_T)]\\
=&\mathbb{E}[U(\phi_T^{0,\ast}(x,q)+\phi_T^{1,\ast}(x,q)\hat{S}_T+q\mathcal{E}_T)]\\
=&\mathbb{E}[U(V(\phi^{0,\ast}(x,q), \phi^{1,\ast}(x,q))_T+q\mathcal{E}_T)],\nonumber
\end{split}$$ and the proof is complete.
Proof of Theorem $\ref{classical1}$
-----------------------------------
Fix $(x,q)\in\mathcal{K}$ and consider some $(y,r)\in\partial u(x,q)$. Suppose now that $(Y^{0,\ast}(y,r), Y^{1,\ast}(y,r))\in\mathcal{B}(y)$ and $Y^{0,\ast}(y,r)\in y\mathcal{M}(\frac{r}{y})$. The process $(Y^{0,\ast,p}(y,r), Y^{1,\ast,p}(y,r))$ coincides with $(Y^{0,\ast}(y,r), Y^{1,\ast}(y,r))$ and $\hat{S}\triangleq \frac{Y^{1,\ast}(y,r)}{Y^{0,\ast}(y,r)}\in\mathcal{S}$ under transaction costs, moreover, we have $$x+\int_0^t\phi_u^{1,\ast}(x,q)d\hat{\mathbf{S}}_u=x+\int_0^t\phi_u^{1,\ast}(x,q)d\hat{S}_u.\nonumber$$
We claim that $Y^{0,\ast}(y,r)\in\mathcal{Y}(y,r;\hat{S})$. The proof of Proposition $3.7$ of [@Chris22] already asserts that $Y^{0,\ast}(y,r)\in\mathcal{Y}(y;\hat{S})$ and it is enough to verify that $$\mathbb{E}[Y_T^{0,\ast}(y,r)(X_T+q \mathcal{E}_T)]\leq xy+q r\nonumber$$ for any $X_T\in\mathcal{H}(x,q;\hat{S})$ and $(x,q)\in\mathcal{K}(\hat{S})$. As $X(\phi^0, \phi^1)_T\in\mathcal{V}_x(\hat{S})$ for some $(\phi^0, \phi^1)\in\mathcal{A}_x(\hat{S})$, we obtain that $$X_T=x+\int_0^T\phi_u^1d\hat{S}_u=X_T'-X_T^{\max},\ \ \text{where}\ X',X^{\max}\in\mathcal{X}(\tilde{S}).\nonumber$$
Consider the set $$\mathcal{M}'(\hat{S})\triangleq \{\mathbb{Q}\in\mathcal{M}(\hat{S}):\ \text{$X^{\max}$ is a UI martingale under $\mathbb{Q}$}\}.\nonumber$$ We have that $(x+\int_0^t\phi_u^1d\hat{S}_u+X_t^{\max})_{0\leq t\leq T}$ is a nonnegative supermartingale under each $\mathbb{Q}\in\mathcal{M}'(\hat{S})$. Similar to the proof of Lemma $\ref{ineqC}$, we can choose a sequence $\mathbb{Q}^n$ converging to $\mathbb{Q}$ in the norm topology where $\frac{d\mathbb{Q}}{d\mathbb{P}}\triangleq \frac{1}{y}Y_T^{0,\ast}(y,r)$. By passing to the limit as $n\rightarrow\infty$ and under Assumption $\ref{assE}$, it follows that $$\mathbb{E}\Big[Y_T^{0,\ast}(y,r)(x+\int_0^T\phi_u^1d\hat{S}_u+q\mathcal{E}_T)\Big]\leq xy+\mathbb{E}[Y_T^{0,\ast}(y,r)q\mathcal{E}_T]=xy+qr,\nonumber$$ since $Y^{0,\ast}(y,r)\in y\mathcal{M}(\frac{r}{y})$. Therefore, the claim $Y^{0,\ast}(y,r)\in\mathcal{Y}(y,r;\hat{S})$ holds.
Fix $(x,q)\in\mathcal{K}$ and consider $(y,r)\in\partial u(x,q)$. It is easy to see that $$\begin{split}
\mathbb{E}[\tilde{U}(Y_T^{0,\ast}(y,r))]+xy+q r=&v(y,r)+xy+q r=u(x,q)\leq u(x,q;\hat{S})\\
\leq &v(y,r;\hat{S})+xy+q r\leq \mathbb{E}[\tilde{U}(Y_T^{0,\ast}(y,r))]+xy+q r\nonumber
\end{split}$$ since $Y_T^{0,\ast}(y,r)\in\mathcal{Y}(y,r)$. Therefore, we obtain that $u(x,q)=u(x,q;\hat{S})$ together with $(y,r)\in\partial u(x,q;\hat{S})$ and $Y_T^{0,\ast}(y,r)$ is the optimal solution to $v(y,r;\hat{S})$ defined by $(\ref{shadowdualv})$. As a consequence, we obtain that $(\phi^{0,\ast}(x,q), \phi^{1,\ast}(x,q))$ is the optimal solution to the utility maximization problem $(\ref{frictionless})$ in the market $\hat{S}$ and $\hat{S}$ is a classic shadow price process.
[^1]: E. Bayraktar is supported in part by the National Science Foundation under grant DMS-0955463 and the Susan M. Smith Professorship.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Kane and Wells recently argued that collider data point to a Higgsino-like lightest supersymmetric partner which would explain the dark matter in our Galactic halo. They discuss direct detection of such dark-matter particles in laboratory detectors. Here, we argue that such a particle, if it is indeed the dark matter, might alternatively be accessible in experiments which search for energetic neutrinos from dark-matter annihilation in the Sun. We provide accurate analytic estimates for the rates which take into account all relevant physical effects. Currently, the predicted signal falls roughly one to three orders of magnitude below experimental bounds, depending on the mass and coupling of the particle; however, detectors such as MACRO, super-Kamiokande, and AMANDA will continue to take data and should be able to rule out or confirm an interesting portion of the possible mass range for such a dark-matter particle within the next five years.'
address:
- 'Randall Physics Laboratory, University of Michigan, Ann Arbor, Michigan 48109-1120'
- 'Department of Physics, Columbia University, 538 West 120th St., New York, New York 10027'
author:
- 'Katherine Freese[^1]'
- 'Marc Kamionkowski[^2]'
date: September 1996
title: Indirect Detection of a Light Higgsino Motivated by Collider Data
---
\#1\#2[3.6pt]{} \#1[0= 0=0 1= 1=1 0>1 \#1 / ]{} \#1[\#1]{}
INTRODUCTION
============
Recently, Kane and Wells (KW) proposed a Higgsino-like supersymmetric (SUSY) particle as a cold-dark-matter candidate [@kanewells]. This proposal was motivated [@ambrosanio] by a supersymmetric interpretation of the CDF $ee\gamma \gamma + \slashchar{E}_T$ event at Fermilab [@CDF] and/or the reported $Z \rightarrow b \bar b$ excess at LEP. In general, the lightest SUSY particle (LSP) is stable (if $R$-parity is conserved), and the other (heavier) SUSY particles eventually decay to the LSP. KW suggest the following explanation of the CDF event: $\tilde e^+(\rightarrow e^+ \chi_2) \tilde e^-(\rightarrow e^- \chi_2)$, followed by the photino-like second-lightest neutralino $\chi_2$ decaying radiatively into the lightest neutralino (and LSP) $\chi$ and a photon. \[Note that for this interpretation to work, the gaugino mass parameters must satisfy $M_1 \sim M_2$ rather than the gauge unification condition $M_1 =(5/3)\tan^2\theta_W M_2$.\] At present there has been only one such event, so that it is premature to claim discovery of supersymmetry. However, the evidence is certainly intriguing. It is therefore interesting to investigate the feasibility of discovering this particular dark-matter candidate with existing and forthcoming detectors. The hope is to either rule out this candidate or to detect it very soon: the payoff, namely the discovery of supersymmetry and of the dark matter, would be enormous.
The proposed dark-matter candidate is Higgsino-like: $$\chi \sim {\rm sin}\beta \tilde H_d^0 + {\rm cos}\beta
\tilde H_u^0 + \delta \tilde Z,$$ with $\delta<0.1$. In the models they consider, the LSP interactions with light fermions $f$ are due primarily to $Z^0$ exchange and are therefore approximated well by a $\chi\chi \bar
f f$ low-energy effective Lagrangian with a coupling proportional to $\cos
2\beta$, where $\tan\beta$ is the ratio of Higgs vacuum expectation values. The dark-matter phenomenology can therefore be parameterized simply by the LSP mass, $m_\chi$, and $\cos2\beta$. In order for the SUSY explanation of the CDF $ee\gamma\gamma+
\slashchar{E}_T$ event to work, the LSP mass should be in the range $30\, {\rm GeV} \la m_\chi \la 55\, {\rm GeV}$, and the radiative decay of the second-to-lightest neutralino requires $\tan\beta<2$ [@ambrosanio]. The model is consistent with constraints to the $Z^0$ invisible width [@width] if $\tan\beta$ is further restricted to be even closer to unity, or equivalently, $\cos2\beta$ closer to zero. The cosmological abundance of the LSP is inversely proportional to its annihilation cross section. In this model, annihilation occurs predominantly through an $s$-channel $Z^0$ exchange, so the annihilation cross section is proportional to $\cos^2 2\beta$ and depends also on the mass. In the regime of supersymmetric parameter space studied prior to the work of Kane and Wells, Higgsinos were not found to be viable dark-matter candidates because they annihilated too efficiently in the early Universe to retain a relic density of any significance. Here, on the other hand, the smaller values of tan$\beta$ for this candidate lead to a smaller annihilation cross section and hence to a cosmologically interesting relic abundance (note that there is also no coannihilation with charginos in this case since the Higgsinos are much lighter than the charginos). The requirement that $\Omega_\chi
h^2\la1$ (which derives from a conservative lower limit of 10 Gyr to the age of the Universe) fixes $\cos^2 2\beta\ga 0.002$. Finally, although the CDF event does not require it, the $Z\rightarrow b\bar b$ anomaly favors a value of $m_\chi\la 40$ GeV. The range of models considered by KW which satisfy these constraints are shown in their Fig. 1. The parameter space is restricted primarily by their favored range for the relic abundance $0.1
\la\Omega_\chi h^2 \la 0.5$ and by the $Z^0$ invisible width. This leaves an irregularly shaped region of favored parameter space which spans roughly the mass range $30 \, {\rm GeV}\, \la
m_\chi \la 40\, {\rm GeV}$ and $1.05 \la \tan\beta \la
1.4$, or equivalently, $0.002\la\cos^2 2\beta \la 0.11$.
KW studied the prospects for direct detection [@witten] of these particles in laboratory dark-matter detectors and found them promising for next-generation detectors [@labdetectors]. In this paper, we point out that for the Higgsino-like particle they consider, searches for energetic neutrinos from LSP annihilation in the Sun [@SOS] may be an equally or more promising avenue toward detection. The point is that the Higgsinos they consider have primarily axial-vector (rather than scalar) couplings to nuclei, so they couple to the spin (rather than the mass) of a nucleus. The isotopic fraction of terrestrial nuclei with spin is generally small. On the other hand, about 75% of the mass of the Sun is composed of nuclei with spin (i.e., protons). Although detection of weakly-interacting massive particles (WIMPs) with scalar interactions is probably more promising with laboratory detectors, detection of WIMPs with axial-vector interactions is generally more promising with astrophysical-neutrino detectors [@bernard; @jkg; @taorich]. For example, Dirac neutrinos, which have scalar-like interactions, are currently ruled out as the Galactic dark matter over most of the plausible mass range by direct-detection experiments [@heidelberg]. On the other hand, Majorana neutrinos, which have axial-vector interactions, are ruled out over a large mass range by null searches at Kamiokande [@kamiokande], but are quite inaccessible to direct searches [@taorich]. In fact, Rich and Tao found that the current bounds to the axial-vector WIMP-nucleon interaction strength from energetic-neutrino searches were roughly three orders of magnitude stronger than current limits from direct searches for WIMPs with masses near 30–40 GeV [@taorich]. In this paper we consider both indirect and direct detection of the newly proposed [@kanewells] Higgsino dark-matter candidate.
In Section II, we calculate the energetic-neutrino rates and discuss the prospects for indirect detection. In Section III, we review the prospects for direct detection. In Section IV, we discuss the results and make some concluding remarks.
ENERGETIC-NEUTRINO RATES
========================
If these particles are indeed present with a halo density $\rho_\chi \sim 0.3\, {\rm GeV}\,{\rm cm}^{-3}$, then some passing through the Sun would lose enough energy to be captured. They then sink to the core of the Sun, and there build up enough density to start annihilating with each other. Among the annihilation products are ordinary neutrinos which would be observable in various existing detectors here on Earth. The energies of the neutrinos have a broad distribution centered roughly at a third of the LSP mass. The detectors with data already taken include Kamiokande [@kamiokande], IMB [@imb], MACRO [@macro], Frejus [@frejus], Baksan [@baksan], and those being deployed now include AMANDA [@amanda], NESTOR [@nestor], and super-Kamiokande. The best technique for inferring the existence of these neutrinos is as follows: muon neutrinos interact in the rock outside of the detector, and give rise to upward-going muons which can be registered in the detector. \[Note that both muon neutrinos as well as muon antineutrinos are produced by the annihilation; these give rise to upward-going muons and antimuons. Both have been included in all our calculations and estimates, and we use the words ‘neutrinos’ and ‘muons’ to refer to the sum of particles and antiparticles.\] The muon-energy thresholds for IMB and MACRO are roughly 2 GeV, 1.7 GeV for Kamiokande, and roughly 1 GeV for Baksan. At present, IMB and Kamiokande constrain the flux of energetic neutrinos from the Sun with energies $\ga 2$ GeV to be $$\Gamma_{\rm det} \la 2.1 \times 10^{-2} {\rm m}^{-2} {\rm yr}^{-1}.
\label{constrainedrate}$$ In addition to the muon neutrinos, there is a comparable flux of electron neutrinos produced by the Higgsino annihilation. For the relatively low energies considered here, the efficiency for detection of electron neutrinos may be comparable to that described above for muon neutrinos. Thus the sensitivity to energetic neutrinos from Higgsino annihilation could be improved if one takes both channels into account. In this paper we focus on limits on the Higgsino particle that can be obtained from considering the production of muons only. Note that the Earth is composed primarily of spinless nuclei, so axially-coupled WIMPs will not be captured, and we expect no energetic-neutrino signal from the Earth for this dark-matter candidate.
Calculation of the predicted flux of neutrino-induced muons is straightforward but lengthy. It must take into account the complete capture-rate calculation, which includes the elastic-scattering cross section and the proper kinematic factors, and the time scale for equilibration between capture and annihilation. The neutrinos will be produced by decays of $b$ and $c$ quarks and $\tau$ leptons to which the LSPs annihilate. An accurate calculation of the neutrino spectrum must take into account the branching ratios for annihilation into various final states, hadronization and slowing of heavy hadrons, the three-body fermion-decay kinematics, and slowing and absorption of neutrinos in the Sun [@ritz; @jerrymarc]. If the LSP has only an axial-vector coupling to nuclei, the result for the flux of neutrino-induced muons (for a local halo density of 0.3 GeV cm$^{-3}$ and velocity dispersion of 270 km s$^{-1}$) can be written (Eq. (9.55) in Ref. [@jkg])[^3] $$\begin{aligned}
\Gamma_{\rm det} & = & (1.65\times 10^{-4}\,{\rm m}^{-2}\,{\rm
yr}^{-1})\, \sigma_{40} \nonumber \\
& \times &\tanh^2(t_\odot/\tau_\odot)\,
(m_\chi/{\rm GeV})\, S(m_\chi/m_p)\, \xi(m_\chi),\end{aligned}$$ where $\sigma_{40}$ is the cross section for LSP-proton elastic scattering due to axial-vector interactions in units of $10^{-40}$ cm$^2$, $S(m_\chi/m_p)$ (where $m_p$ is the proton mass) is a kinematic suppression factor, $\tau_\odot$ is the capture-annihilation equilibration timescale and $t_\odot$ is the age of the Sun, and $\xi(m_\chi)$ is a measure of the second moment of the neutrino energy distribution. We now discuss each of these factors.
In the models we are considering, the Higgsino-quark interaction is due primarily to $Z^0$ exchange. The cross section for scattering from a nucleus of mass $m_N$ with an unpaired proton is approximated by [@kanewells] $$\begin{aligned}
\sigma_N & = &{2 m_\chi^2 m_N^2 \over \pi (m_\chi+ m_N)^2 } \, G_F^2
\cos^2 2\beta \nonumber \\
&\times & \lambda^2 J(J+1)\,(\Delta
d+\Delta s -\Delta u)^2;
\label{elastic}\end{aligned}$$ for unpaired-neutron nuclei switch $\Delta d$ and $\Delta u$ (our $\Delta d$ and $\Delta u$ are switched relative to those in KW). For the LSP-proton elastic scattering cross section, we take $m_N = m_p$ (proton mass) and the Lande factor $\lambda^2 J(J+1) = 3/4$ in Eq. (\[elastic\]). Here the $\Delta q$’s are the fraction of spin in the proton carried by each quark. Using values from a recent compilation [@smc] which give $\Delta d+ \Delta s -\Delta u = 1.24$, the LSP-proton elastic scattering cross section evaluates to $\sigma_{40}\simeq 340 \cos^2
2\beta$. Using Eqs. (9.21–22) in Ref. [@jkg], the kinematic factor $S(m_\chi/m_p)$ falls in the range 0.5–0.6 for LSP masses between 30 and 40 GeV. Furthermore, the equilibration time scale is given by (Eq. (9.8) in Ref. [@jkg]) $${t_\odot \over \tau_\odot} = 330 \, \left({ C \over
\sec^{-1} }\right)^{1/2} \left( {\langle\sigma_A v\rangle
\over {\rm cm}^3\,
{\rm s}^{-1} }\right)^{1/2} \left( {m_\chi \over 10 {\rm GeV}} \right)
^{3/4},
\label{equilibrationtimescale}$$ where the capture rate is (Eq. (9.19) in Ref. [@jkg]), $$C=(1.3\times 10^{25}\,{\rm s}^{-1})\,
\sigma_{40}\, S(m_\chi/m_p)\,(m_\chi/{\rm GeV})^{-1}.$$ Here, $\langle\sigma_A
v\rangle$ is the thermally averaged annihilation cross section times relative velocity in the limit $v\rightarrow 0$ (i.e., the $s$-wave contribution). If the neutralino-quark interaction occurs predominantly via $Z^0$ exchange, the $v\rightarrow0$ annihilation cross section is $$\langle \sigma_A v \rangle = {G_F^2 \, \cos^2 2\beta \over 8 \pi}
\, \sum_f\, c_f \, m_f^2,
\label{annihilationxsection}$$ where the sum is over the $\tau$ lepton and $b$ and $c$ quarks to which the Higgsinos annihilate predominantly, $m_f$ is the fermion mass, and $c_f$ is a color factor (3 for quarks and 1 for leptons). Eq. (\[annihilationxsection\]) evaluates to $\langle \sigma_A v
\rangle \simeq 4.42\times10^{-27}\, {\rm cm}^3\, {\rm s}^{-1}\, \cos^2
2\beta$. Inserting our expressions for the annihilation cross section and capture rate into Eq. (\[equilibrationtimescale\]), we get $t_\odot / \tau_\odot \simeq 2600 \, (m_\chi/35\,{\rm
GeV})^{1/4}\cos^2 2\beta$. For $\cos^2 2\beta\ga0.002$ and the Higgsino masses of interest, $t_\odot \ga \tau_\odot \ga 5$, so we may safely set $\tanh^2(t_\odot/\tau_\odot)$ equal to unity.
Finally, there is $\xi(m_\chi)$, a measure of the second moment of the neutrino energy distribution. This depends on the branching ratios for LSP-LSP annihilation into various annihilation channels. It is explicitly given by $$\begin{aligned}
\xi(m_\chi)=\sum_F B_F [& 3.47 & \Nzsq_{F,\nu}(m_\chi) \nonumber
\\ & + & 2.08
\Nzsq_{F,\bar\nu} (m_\chi)],
\label{xieqn}\end{aligned}$$ where the sum is over all final states to which the LSPs can annihilate, and $B_F$ is the branching ratio for annihilation to each channel, which can be calculated with Eq. (\[annihilationxsection\]). The $\Nzsq_{F,i}$ is the second moment of the energy distribution (scaled by $m_\chi^2$) of neutrino type $i$ from final state $F$. Analytic expressions for $\Nzsq$, which take into account hadronization, stopping of heavy hadrons, the three-body decay kinematics, and slowing and absorption of neutrinos as they pass through the Sun, are given in Ref. [@jerrymarc], and these are accurate for the low-mass Higgsinos considered here [@joakim]. As indicated in Eq. (\[xieqn\]), $\xi(m_\chi)$ depends on the annihilation branching ratios, and the range of possible values is indicated in Fig. 33 in Ref. [@jkg]. For LSPs with masses in the range 30–40 GeV, $\xi$ takes on its largest value ($\sim0.11$) for annihilation into $\tau$ leptons ($B_\tau=1$) and its smallest value ($\sim0.034$) for annihilation into $b$ quarks ($B_b=1$). If, as assumed by KW, annihilation occurs via the $Z^0$ and $\tan\beta$ is near unity, then $B_F\propto c_f m_f^2$ \[c.f., Eq. (\[annihilationxsection\])\]. Therefore, Higgsinos should annihilate primarily to $b$ quarks, so we will take $\xi(m_\chi)=0.034$. It should be kept in mind, however, that there will always be some nonzero annihilation branch into $\tau$ leptons, and in some models, if annihilation via $t$- and $u$-channel exchange of a stau is larger, it may be significant. More generally, one can approximate $\xi$ for arbitrary annihilation branching ratios by noting that annihilation always occurs almost entirely to $\tau$ leptons and $b$ and $c$ quarks (since the other quarks are so much lighter). Furthermore, the value one would obtain for $\xi$ for the case of annihilation predominantly into $c$ quarks (i.e., if for some reason $B_c=1$) is quite close to the value one obtains for the most likely case of annihilation predominantly to $b$ quarks (i.e., when $B_b=1$), especially at low energies. Therefore, for LSPs with masses near 35 GeV, and for arbitrary branching ratios to $\tau$ leptons and $b$ and $c$ quarks, we can write $\xi
\simeq 0.11 \,B_\tau + 0.034\,(1-B_\tau)$. Therefore, by taking $B_\tau=0$, we are using a conservative lower limit for $\xi$, and it could conceivably be a factor of 3 larger.
Putting together all the factors, the detection rate is given by $$\begin{aligned}
\Gamma&&_{\rm det} \simeq (2.7\times 10^{-2}\,{\rm
m}^{-2}\,{\rm yr}^{-1})\, (m_\chi/35\, {\rm GeV})\, \cos^2
2\beta \nonumber \\
&&\times \left( { \Delta d +\Delta s -\Delta u \over 1.24}
\right)^2 [3.2\, B_\tau + (1-B_\tau)]
\left[ {S(m_\chi/m_p)
\over 0.55} \right] \nonumber \\
&& \times \tanh^2[2600 \, (m_\chi/35\,{\rm
GeV})^{1/4}\cos^2 2\beta],
\label{indirectrate}\end{aligned}$$ where we have included the dependence on the model parameters $m_\chi$ and $\cos^2 2\beta$, spin content of the proton, on the annihilation branch to $\tau$ leptons, the kinematic factor, and the equilibration timescale, although as indicated above, the dependence on these last two factors will be very weak. Eq. (\[indirectrate\]) is obtained assuming no muon energy thresholds (i.e. all muons can be detected); this assumption is a good approximation for detectors with thresholds near a few GeV (e.g., MACRO, Kamiokande, and Baksan), since these energies are negligible compared with the Higgsino mass. However, for detectors with higher thresholds (e.g., AMANDA and NESTOR), a good fraction of the signal may be below threshold.
The relic density is inversely proportional to $\cos^2 2\beta$. Since the count rate scales as cos$^2 2 \beta$, for a given mass the rate drops with increasing $\Omega_\chi h^2$. Furthermore, for a given relic abundance, the rate drops with increasing mass (see Fig. 1 in KW for the dependence of tan$\beta$ on mass and abundance.) As an example, for $\Omega_\chi h^2 = 0.3$, the predicted count rates are $\Gamma_{\rm det} \sim 2.6 \times 10^{-3}$ for $m_\chi =
30$ GeV and $\Gamma_{\rm det} \sim 1.1 \times 10^{-3}$ for $m_\chi
= 35$ GeV.
In general, if, as is likely, annihilation occurs predominantly to $b$ quarks, then the predicted rates fall roughly one to three orders of magnitude below currently published limits for the $\cos^2 2\beta$ range of interest, $0.002 \lesssim {\rm cos}^2 2\beta \lesssim 0.11$. However, the accumulated exposure of Baksan is greater than that of Kamiokande, so even better sensitivities (perhaps by a factor of two) have probably already been achieved. Also, by performing an analysis of the data which takes into account the predicted angular and energy distribution of the neutrino-induced muons, one should be able to improve the sensitivity with existing data [@edsjo]. More significantly, the MACRO collaboration expects to reach a sensitivity a factor of ten or so better than current bounds within the next five years as it continues to take data.
Future detectors such as super-Kamiokande, AMANDA, and NESTOR may have exposures orders of magnitude larger than current detectors. Whether they can improve on current sensitivities to light Higgsinos will depend on the thresholds of these detectors. These future detectors may have much larger thresholds (e.g. 10–30 GeV) so that the calculations described above would have to be redone. Although the neutrino energy distribution is centered roughly at a third the LSP mass, the probability of detecting a neutrino is proportional to the square of the neutrino energy. Therefore, the upward-muon signal is due to a large extent to the high-energy tail of the neutrino distribution. If so, there may still be a significant signal even for thresholds as high as 10–30 GeV, depending on the model parameters. In addition, for larger thresholds, the annihilation channel to tau particles via intermediate staus becomes important because the neutrinos produced via this channel are stiffer and hence capable of being above these larger thresholds.
PROSPECTS FOR DIRECT DETECTION
==============================
Let us now briefly review the prospects for direct detection in a laboratory detector—$^{73}$Ge detector, for example [@kanewells]. The rates are controlled by the cross section for Higgsino elastic scattering from a $^{73}$Ge nucleus. (Note that the scattering rate from the naturally abundant isotope, $^{76}$Ge, is very small since this isotope has no spin.) The cross-section for scattering from a nucleus with an unpaired neutron, such as $^{73}$Ge, was given in Eq. (\[elastic\]) (with $\Delta d$ and $\Delta u$ reversed). In the single-particle shell model, the Lande factor for $^{73}$Ge evaluates to $\lambda^2 J(J+1)\simeq0.3$. However, it should be kept in mind that the odd-group model predicts a number 80% smaller and more detailed calculations suggest it may be 2% [@ted] to 40% [@dimitrov] smaller than in the single-particle shell model. Using the simplest (and most optimistic) value, Eq. (\[elastic\]) evaluates to $\sigma_{73}^{40}\simeq 6.5\times10^4
\cos^2 2\beta [\lambda^2 J(J+1)/0.3]$, in units of $10^{-40}$ cm$^2$, for Higgsinos with masses 30–40 GeV.
In order of magnitude, the rate for scattering from $^{73}$Ge is $R\sim f_{73}\sigma_{73} \rho
v / (m_\chi
m_N)$ where $\rho$ is the local halo density, $v$ is the halo velocity dispersion, and $f_{73}$ is the isotopic fraction of $^{73}$Ge in the sample. A careful calculation must include the velocity distribution (and its yearly modulation) of halo dark-matter particles incident on the detector and the proper form-factor suppression for spin-dependent scattering from $^{73}$Ge [@ted; @dimitrov]. For a Higgsino of mass 40 GeV, the event rate (including all relevant physical effects) for scattering in $^{73}$Ge may be obtained from the differential event rate for detection of an axially-coupled WIMP plotted in Fig. 22 in Ref. [@jkg]. The result (averaged over the yearly modulation) for scattering in natural germanium ($f_{73}=0.078$) is $$\begin{aligned}
R &\simeq & (1.2\times 10^{-5}\,{\rm kg}^{-1}\, {\rm day}^{-1}) \,
f_{73}\,\sigma_{73}^{40} \nonumber
\\ & \simeq& 0.0624\, \cos^2 2\beta\, [\lambda^2 J(J+1)/0.3]
\label{directrate}\end{aligned}$$ This result agrees (well within the nuclear-physics uncertainties) with the results shown in Fig. 2 in KW, although our numbers are slightly smaller. Eq. (\[directrate\]) is obtained assuming no thresholds. However, finite energy thresholds in realistic experiments will cut out a significant fraction of events and lower the predicted detection rate accordingly.
Now consider, for illustration, the Cryogenic Dark Matter Search (CDMS) experiment [@cdms], which will first run with 1 kg of natural germanium. After background rejection from demanding ionization-calorimetry coincidence, there will still be a background event rate of roughly $R_b \simeq 1\,
{\rm kg}^{-1}\, {\rm day}^{-1}$. After a one-year exposure ($E=$ 365 kg-days), the $3\sigma$ sensitivity of the experiment will be roughly $S\simeq 3\sqrt{R_b/E} \simeq 0.16\, {\rm kg}^{-1}\, {\rm day}^{-1}$. They also plan to run a similar experiment with roughly 0.5 kg of enriched $^{73}$Ge. Assuming the same background-event rate, this would improve the sensitivity to scattering from $^{73}$Ge to 0.017 kg$^{-1}$ day$^{-1}$ when compared with the prediction above \[c.f., Eq. (\[directrate\])\] for scattering in natural germanium.
For $0.002\la \cos^2 2\beta \la 0.11$, the predicted rate for Higgsino scattering in natural germanium is 0.0001–0.007 kg$^{-1}$ day$^{-1}$ (again, this is only slightly smaller than the results shown in Fig. 2 of KW). Therefore, even the most optimistic models (with the most optimistic nuclear-physics and energy-threshold assumptions) seem to fall roughly a factor of two below this forecasted CDMS sensitivity.
DISCUSSION
==========
The predicted rates for both direct \[Eq. (\[directrate\])\] and indirect \[Eq. (\[indirectrate\])\] detection of axially-coupled WIMPs are proportional to the WIMP-nucleon coupling—in this case, $\cos^2 2\beta$. For a given WIMP mass, we can therefore compare the forecasted enriched-$^{73}$Ge CDMS sensitivity with the current upward-muon limit. Doing so, we find that the enriched-$^{73}$Ge sensitivity (0.017 kg$^{-1}$ day$^{-1}$) will improve on the [*current*]{} limit to the upward-muon flux ($2.1\times10^{-2}$ m$^{-2}$ yr$^{-1}$) roughly by a factor of 4. When we compare this with the forecasted factor-of-ten improvement expected in MACRO, it appears that the sensitivity of indirect-detection experiments looks favorable. Before drawing any conclusions, however, it should be noted that the sensitivity in detectors with other nuclei with spin may be significantly better. We therefore conclude that the two schemes will be competitive for detection of axially-coupled WIMPs. Realistically, we must also emphasize that the forecasted sensitivities of these current experiments will probe only the most optimistic region of KW’s favored parameter space. It will require much larger low-background laboratory detectors or astrophysical-neutrino observatories to probe a good fraction of the interesting light-Higgsino models.
We should re-emphasize that our estimates for both direct- and indirect-detection rates may actually be conservative. If our halo is flattened (as halos of many spirals seem to be), then the local halo density could be twice as large. If the LSP has a considerable annihilation branch to $\tau$ leptons, the neutrino rates could be up to a factor of 3 larger. It should also be kept in mind that—although KW argued that they should be small—the LSP-quark interaction will have some squark-exchange contributions in addition to the $Z^0$ contributions. This will probably expand the viable parameter space. It also implies that there may realistically be additional contributions to the axial-vector coupling to nuclei, and there may be some scalar interaction. This could enhance rates for both direct and indirect detection, although much more dramatically for direct detection. It would also result in a neutrino signal from the center of the Earth.
There are also uncertainties which might reduce the rates. Variations within current experimental constraints in the spin content of the nucleon could either increase or decrease both direct- and indirect-detection rates, perhaps dramatically. The Lande factor is likely to be smaller than the value from the single-particle shell model which we used, and if so, the direct-detection rates will be lowered accordingly. There may be sizeable errors in the form factor for spin-independent scattering which would affect the direct rates. Finally, we have assumed 100% direct-detection efficiency. However, for the Higgsino mass range of interest, a significant fraction of the recoils could be below the detection threshold, and the sensitivity to an LSP passing through the detector would be degraded accordingly.
Within the next five years, detectors at LEP and Fermilab will be able to confirm or rule out the $ee\gamma\gamma + \slashchar{E}_T$ event that motivated this Higgsino dark-matter candidate. Much of the available parameter space for the SUSY interpretation will in fact be tested within one year. If the SUSY interpretation of this event is indeed correct, then this Higgsino particle exists, but one still does not know how long it lives; the fact that it escapes the detector only proves that its lifetime is longer than $\sim 10^{-8}$ seconds. Proving that this particle is in fact the dark matter in the halo of our Galaxy would require that it be detected either directly or indirectly due to its annihilation in the Sun. Of course these techniques may be used to rule out the existence of the particle instead. In this paper, we have suggested that the timescale for indirect detection may be comparable to that for experimental verification or disproof of the motivating event.
To conclude, we have provided estimates of the rates for indirect detection of Higgsino dark-matter candidates motivated by collider data. Our calculations take into account all relevant physical effects; the accuracy of our estimates should be well within the irreducible astrophysics and particle-physics uncertainties inherent in any such calculation. Our results—obtained assuming only a $Z^0$-exchange contribution to the LSP-quark interaction—suggest that indirect detection may provide a realistic alternative avenue toward verification or falsification of this Higgsino dark-matter candidate.
We would like to thank C. Akerib, E. Diehl, R. Gaitskell, G. Kane, C. Kolda, and G. Tarle for helpful discussions. MK thanks the Theory Division at CERN for hospitality, and MK and KF thank the Fermilab Theoretical Astrophysics Center (where part of this work was completed) for hospitality. This work was supported at Columbia by the D.O.E. under contract DEFG02-92-ER 40699, NASA under NAG5-3091, and the Alfred P. Sloan Foundation, and at Michigan by NSF-PHY9407194. Portions of this work (KF) were completed at the Aspen Center for Physics.
G. L. Kane and J. D. Wells, Phys. Rev. Lett. [**76**]{}, 4458 (1996).
S. Ambrosanio et al., Phys. Rev. Lett [ **76**]{}, 3498 (1996).
S. Park, “Search for New Phenomena at CDF,” 10th Topical Workshop on Proton-Antiproton Collider Physics, edited by R. Raja and J. Yoh (AIP Press, 1996).
LEP Collaborations report, CERN-PPE/95-172.
M. W. Goodman and E. Witten, Phys. Rev. **D31, 3059 (1986); I. Wasserman, [Phys. Rev.]{} [ **D33**]{}, 2071 (1986); K. Freese, J. Frieman, and A. Gould, Phys. Rev. **D37, 3388 (1988); K. Griest, Phys. Rev. **D38, 2357 (1988); FERMILAB-Pub-89/139-A (E); A. Drukier, K. Freese, and D. Spergel, Phys. Rev. **D33, 3495 (1986).********
See, e.g., J. Low Temp. Phys. **93 (1993).**
J. Silk, K. Olive, and M. Srednicki, Phys. Rev. Lett. [**55**]{}, 257 (1985); K. Freese, [ Phys. Lett.]{} [**B167**]{}, 295 (1986); L. M. Krauss, K. Freese, D. N. Spergel, and W. H. Press, [ Astrophys. J.]{} [**299**]{}, 1001 (1985); L. M. Krauss, M. Srednicki, and F. Wilczek, [Phys. Rev.]{} [**D33**]{}, 2079 (1986); T. Gaisser, G. Steigman, and S. Tilav, [Phys. Rev.]{} [**D34**]{}, 2206 (1986); M. Kamionkowski, [Phys. Rev.]{} [**D44**]{}, 3021 (1991).
M. Kamionkowski, K. Griest, G. Jungman, and B. Sadoulet, Phys. Rev. Lett. [**74**]{}, 5174 (1995).
G. Jungman, M. Kamionkowski, and K. Griest, Phys. Rep. [**267**]{}, 195 (1996).
J. Rich and C. Tao, DAPNIA/SPP 95-01.
M. Beck, Nucl. Phys. (Proc. Suppl.) [ **B35**]{}, 150 (1994); M. Beck et al., Phys. Lett. [**B336**]{}, 141 (1994); S. P. Ahlen et al., [Phys. Lett.]{} [ **B195**]{}, 603 (1987); D. O. Caldwell et al., [ Phys. Rev. Lett.]{} [**61**]{} (1988) 510.
M. Mori et al. (Kamiokande Collaboration), [Phys. Lett.]{} [**B289**]{}, 463 (1992); M. Mori et al. (Kamiokande Collaboration), [Phys. Rev.]{} [**D48**]{}, 5505 (1993).
J. M. LoSecco et al. (IMB Collaboration), Phys. Lett. **B188, 388 (1987).**
E. Diehl, Ph.D. Thesis, U. of Michigan (1994); E. Diehl, G. Kane, C. Kolda, and J. Wells, Phys. Rev. [**D50**]{}, 4223 (1994).
Frejus Collaboration, presented by H. J. Daum, Topical Seminar on Astrophysics and Particle Physics, San Miniato, Italy, 1989.
M. M. Boliev et al., Bull. Acad. Sci. USSR, Phys. Ser. [**55**]{}, 126 (1991) \[Izv. Akad. Nauk. SSSR, Fiz. [**55**]{}, 748 (1991)\].
D. M. Lowder et al., [Nature]{} [**353**]{}, 331 (1991).
L. Resvanis, Europhys. News [**23**]{}, 172 (1992).
S. Ritz and D. Seckel, Nucl. Phys. [**B304**]{}, 877 (1988).
G. Jungman and M. Kamionkowski, Phys. Rev. [**D51**]{}, 328 (1995).
D. Adams et al. (The Spin Muon Collaboration), CERN-PPE/94-57 (1994).
J. Edsjö, private communication.
L. Bergstrom, J. Edsjö, and M. Kamionkowski, in preparation.
M. T. Ressell et al., Phys. Rev. [**D48**]{}, 5519 (1992).
V. Dimitrov, J. Engel, and S. Pittel, Phys. Rev. [**D51**]{}, 291 (1995).
P. D. Barnes et al., J. Low Temp. Phys. [**93**]{}, 79 (1993); T. Shutt et al., Phys. Rev. Lett. [**61**]{}, 3425 (1992); ibid. 3531 (1992).
[^1]: [email protected]
[^2]: [email protected]
[^3]: Note that there is a factor of $\xi(m_\chi)$ missing and the $\tanh(t_\odot/\tau_\odot)$ should be $\tanh^2(t_\odot/\tau_\odot)$ in Eq. (9.55) in Ref. [@jkg].
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Tijana Radivojević
- Jonatha Anselmi
- Enrico Scalas
title: |
Ergodic transition in a simple model of\
the continuous double auction
---
Model {#model .unnumbered}
=====
In our model, following [@cont10], prices assume $N$ integer values from $1$ to $N$. A price can be regarded as a [*class*]{} where orders are placed. This way of representing prices is a faithful representation of what happens in real markets due to price discretization. The only unrealistic feature is the presence of an upper limit to prices which we keep to ensure partial analytical tractability. Note that orders can be considered as [*objects*]{} to be classified by prices at which they are placed (see [@garibaldi2010] for a general discussion on the problem of allocating objects to classes). We only consider two kinds of orders, namely, limit orders and market orders.
{width="0.7\columnwidth"} \[book\]
As discussed above, limit orders can be of two types: limit bid orders, i.e. orders to buy [*one*]{} share and limit ask orders, i.e. orders to sell [*one*]{} share. Market orders are also of two types as either the best limit bid order or the best limit ask order are accepted when a transaction occurs. When a market order arrives, only one limit order will be removed either from the best bid category or the best ask category, provided that these categories contain at least one order. In other words, we assume that all the orders are characterized by quantities equal to one share. We shall further assume that limit ask orders arrive following a Poisson process with rate $\lambda_a$ and that limit bid orders arrive at a rate $\lambda_b$. These orders are written in the book and they wait until a market order arrives. Remember that the best limit bid orders, namely the limit bid orders offered at the highest price are always strictly smaller than the best limit ask orders, that is the limit ask orders offered at the lowest price. Market orders to buy, accepting one of the best ask orders, arrive in the market separated by exponentially distributed waiting times with parameter $\mu_b$, whereas market orders to sell arrive with an exponential distribution of waiting times with parameter $\mu_a$. If the book is empty, or if the appropriate side of the book is empty, market orders are not executed. It is necessary to remark that order inter-arrival times are not exponentially distributed in real markets [@scalas06b] due to the non-stationary behavior of humans [@barabasi10]. However, to ensure analytical tractability, here we assume that inter-arrival times are exponentially distributed. We shall assume that limit ask orders are uniformly placed in the price classes from ${p}_b + 1$ to ${p}_b + n$, where ${p}_b$ is the class of the current best bids. Conversely, limit bid orders are uniformly placed in the price classes from ${p}_a - n$ to ${p}_a -1$, where ${p}_a$ is the class of the current best asks. As mentioned above, the accessible system states are limited by the condition ${p}_b < {p}_a$. When ${p}_a$ is between $1$ and $n$ (${p}_b$ between $N-n+1$ and $N$), the bid (respectively, ask) interval is restricted correspondingly. For instance, if ${p}_a = 1$, no bids are possible. Finally, if no orders are present in the book, the next bid, $b$, will be uniformly chosen in $p-n \leq b \leq p$ and the next ask, $a$ in $p \leq a \leq p+n$, where $p$ is the price of the last trade. The specification of an initial price (which can be interpreted as the opening auction price) is then sufficient to start the auction. Note that the model outlined above is essentially the same as in [@smith02]. In terms of agent-based models, it is a [*zero intelligence agent-based model*]{} [@gode]. However, our version does not suffer from using odd mathematical objects such as uniform distributions over semi-infinite intervals. A preliminary discussion of our model was presented in [@RAS].
The trade price process $P(t)$ is a continuous-time random walk that we wish to characterize. Let $T_i$ denote the epoch of the $i$-th trade; in particular, we are interested in the behaviour of the following random variable $$\label{logreturn}
R_i = \log(P(T_{i+1})/P(T_i)),$$ called the [*tick-by-tick*]{} logarithmic return. This is the usual variable used in statistical finance and financial econometrics for the analysis of tick-by-tick data [@scalas06a]. It turns out that the behaviour of $R_i$ crucially depends on the presence or absence of statistical equilibrium in the supply mechanism.
Main results {#main-results .unnumbered}
============
Indeed, there are two main regimes in this model and they are triggered by an ergodic transition. Let us denote by $A(t)$ the total number of limit ask orders and by $B(t)$ the total number of limit bid orders present in the book. By definition, these two random processes are independent and they are $M/M/1$ queues with rates $\lambda_a$ and $\mu_b$ and $\lambda_b$ and $\mu_a$, respectively [@queue]. $M/M/1$ queues are the continuous-time equivalent of birth and death Markov chains. The conditions for the existence of statistical equilibrium (ergodicity) are given by the following inequalities $$\label{askqueue}
\lambda_a < \mu_b,$$ and $$\label{bidqueue}
\lambda_b < \mu_a.$$ The intuitive meaning of these conditions is as follows. If the rate of arrival for limit orders is larger than the rate of market orders, then the number of orders in the book eventually explodes. However, in this case, prices will be able to fluctuate only among a few values. When the rate of market orders is larger than the rate of limit orders, the number of orders in the book remains finite and prices are free to fluctuate over the whole available range. In the ergodic regime, the invariant (and equilibrium) distributions of $A(t)$ and $B(t)$ are given by two geometric distributions $$\label{askdistribution}
\mathbb{P}(A = a) = \left( \frac{\lambda_a}{\mu_b} \right)^a \left(1 - \frac{\lambda_a}{\mu_b} \right),$$ and $$\label{biddistribution}
\mathbb{P}(B = b) = \left( \frac{\lambda_b}{\mu_a} \right)^b \left(1 - \frac{\lambda_b}{\mu_a} \right).$$ Given the independence between $A(t)$ and $B(t)$, the joint probability density is $$\mathbb{P} (A = a, B= b) = \left( \frac{\lambda_a}{\mu_b} \right)^a \left(1 - \frac{\lambda_a}{\mu_b} \right) \left( \frac{\lambda_b}{\mu_a}
\right)^b \left(1 - \frac{\lambda_b}{\mu_a} \right),$$ from which one can find the probability of finding an empty book $$\label{emptybook}
\mathbb{P} (A = 0, B= 0) = \left(1 - \frac{\lambda_a}{\mu_b} \right) \left(1 - \frac{\lambda_b}{\mu_a} \right).$$ For our further analysis, we shall focus on the case of a symmetric auction, assuming $\lambda_a = \lambda_b = \lambda$ and $\mu_a = \mu_b = \mu$ and we shall consider the ratio $\rho = \lambda / \mu$ as the basic order parameter of the model. In fact, there is no reason for a random auction to be unbalanced towards selling or buying. As discussed above, if $\rho < 1$, we are in the ergodic regime, whereas for $\rho \geq 1$, we are in a regime where the orders accumulate and $A(t), B(t) \to \infty$, for $t \to \infty$. The two regimes give rise to two radically different behaviours for the tick-by-tick log-returns . This is qualitatively shown in Fig. \[timeseries\], where we report the behavior of prices and log-returns in a Monte Carlo simulation for $\rho = 0.8$ (Fig. \[timeseries\] a), $\rho = 1.2$ (Fig. \[timeseries\] b) and for $\rho = 6$ (Fig. \[timeseries\] c). One can see by eye that, in the ergodic regime, high and low log-returns are clustered, whereas, in the non-ergodic one, such a volatility clustering does not occur. Fig. \[timeseries\] a clarifies the origin of clustering. When the price is lower, log-returns are higher and the price process has the persistence behavior typical of random walks which immediately leads to clusters of low and high volatility as the price slowly moves up and down, respectively. The comparison between Fig. \[timeseries\] b and Fig. \[timeseries\] c shows that there are two sub-regimes in the non-ergodic case. If $1 \leq \rho < n$, even if $A(t)$ and $B(t)$ diverge, the limit orders belonging to the best bid and the best ask can be removed by market orders and prices can fluctuate among a set, whereas if $\rho \geq n$, then after a transient, the number of limit orders belonging to the best bid and the best ask diverges and prices can only fluctuate between two values. In this condition, the price process becomes a random telegraph process. This behavior is justified by the fact that the process of the number of orders at the best bid price (respectively, best ask) can be coupled with the state of an $M/M/1$ queue with arrival rate $\lambda_b/n$ ($\lambda_a/n$) and service rate $\mu_a$ ($\mu_b$); this is so because limit orders, upon arrival, distribute uniformly over the $n$ best prices. If $\lambda_b/(\mu_a n) = \rho/n \ge 1$, then the number of orders at the best bid price converges to infinity, as $t \to \infty$, meaning that all trades will occur at the price where the bids accumulate. If $\lambda_b/(\mu_a n)= \rho/n <1$, then the queue of the best bids eventually empties with probability one, meaning that there is a positive probability that the trading price changes. However, if we are in the region $\frac{1}{n}\le\tfrac{\lambda_b}{\mu_a n}=\frac{\rho}{n}<1$, then we know that $A(t),B(t)\to\infty$, as $t\to\infty$. This means that the queue of the bids at some price will eventually never empty, which means that trades at lower prices will never occur. By symmetry, the same argument holds for asks.
The transition between regimes can be detected studying the moments of log-returns. In table \[stat\], we give the descriptive statistics for log-returns, including mean, standard deviation, skewness and kurtosis as well as the autocorrelation coefficient at the first lag of absolute log-returns, $c_1$, for different values of the parameter $\rho$. These statistics are computed on 1000 simulation runs with $10^6$ events.
$\rho=0.1$ $\rho=0.8$ $\rho=1$ $\rho=1.2$ $\rho=6$
---------- ------------------------ ------------------------ ------------------------ ------------------------ ------------------------
mean $-3.298\cdot 10^ {-6}$ $-7.472\cdot 10^ {-7}$ $-2.068\cdot 10^ {-7}$ $-1.918\cdot 10^ {-9}$ $-3.294\cdot 10^ {-9}$
st. dev. 0.0921 0.0782 0.0179 0.0075 0.0028
skewness 0.896 0.287 -0.034 $3.443\cdot 10^{-4}$ $-3.8\cdot 10^{-7}$
kurtosis 227.089 277.33 116.679 3.373 2.004
$c_1$ 0.515 0.538 0.304 0.164 $1.06\cdot 10^{-5}$
: Descriptive statistics of log-returns and the autocorrelation coefficient at the first lag of absolute log-returns for different $\rho$, made on 1000 runs of simulation of $10^6$ events.[]{data-label="stat"}
Figure \[errorbars\] shows standard deviation, kurtosis and $c_1$ in more detail, namely mean values and error bars are given for these three quantities estimated from 1000 runs. One can see that these quantities increase in the non-ergodic case.
Our findings can be compared with those presented in a recent study of financial stylized facts [@brandouy], where the authors find that higher rate of limit orders stabilizes the market by decreasing the standard deviation of returns. In Fig. \[acf\], we plot the sample autocorrelation for the tick-by-tick absolute log-return series for $\rho=0.8$, $\rho=1.2$ and for $\rho=6$. A slow decay of the ACF in ergodic case, showing long range-memory, is in agreement with the stylized facts found in financial data. One can see that this decay is much faster if $\rho$ increases.
In Fig. \[kurtosis\], we plot the complementary cumulative distribution of kurtosis values (Fig. \[kurtosis\] (a)) and of $c_1$ (Fig. \[kurtosis\] (b)) for 1000 simulations and for different values of $\rho$ in order to corroborate the observation made above: there is a jump in the kurtosis of logarithmic returns as well as a jump in the first-lag autocorrelation of absolute log-returns as $\rho$ moves from values larger than $1$ to values smaller than $1$.
Conclusions {#conclusions .unnumbered}
===========
Acknowledgments {#acknowledgments .unnumbered}
===============
We wish to thank Mauro Politi for inspiring discussions. This work was supported by the Italian grant PRIN 2009, 2009H8WPX5\_002, [*Finitary and non-finitary probabilistic methods in economics*]{}.
[99]{}
R. Cont, S. Stojkov, and R. Tareja, A Stochastic Model for Order Book Dynamics, Operations Research [**58**]{}, 549–563 (2010).
U. Garibaldi and E. Scalas, [*Finitary Probabilistic Methods in Econophysics*]{}, Cambridge University Press, Cambridge UK, (2010).
E. Scalas, T. Kaizoji, M. Kirchler, J. Huber, and A. Tedeschi, Waiting times between orders and trades in double-auction markets, Physica A [**366**]{}, 463–471 (2006).
A.-L. Barabási, [*Bursts: The Hidden Pattern Behind Everything We Do*]{}, Dutton Adult, Boston (2010).
E. Smith, J. Doyne Farmer, L. Gillemot, and S. Krishnamurthy, Statistical theory of the continuous double auction, Quant. Financ. [**3**]{}, 481–514, (2002).
D.K. Gode and S. Sunder, Allocative efficiency of markets with zero-intelligence traders: Market as a partial substitute for individual rationality. Journal of Political Economy [**101**]{}, 119 – 137 (1993).
T. Radivojević, J. Anselmi, and E. Scalas, A stylized model for the continuous double auction. In: [*Managing Market Complexity*]{}, Lecture Notes in Economics and Mathematical Systems [**662**]{} 115–125, Springer, Berlin (2012).
E. Scalas, The application of continuous-time random walks in finance and economics, Physica A [**362**]{}, 225-–239 (2006).
S. Asmussen, [*Applied Probability and Queues*]{}, Springer, Berlin, (2003).
O. Brandouy, A. Corelli, I. Veryzhenko, and R. Waldeck, A re-examination of the “zero is enough” hypothesis in the emergence of financial stylized facts, J. Econ. Interact. Coord [**7**]{}, 223–248 (2012).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we consider Dirac operators in $\R^n$, $n\ge2$, with a potential $V$. Under mild decay and continuity assumptions on $V$ and some spectral assumptions on the operator, we prove a limiting absorption principle for the resolvent, which implies a family of Strichartz estimates for the linear Dirac equation. For large potentials the dynamical estimates are not an immediate corollary of the free case since the resolvent of the free Dirac operator does not decay in operator norm on weighted $L^2$ spaces as the frequency goes to infinity.'
address:
- |
Department of Mathematics\
University of Illinois\
Urbana, IL 61801, U.S.A.
- |
Department of Mathematics\
University of Cincinnati\
Cincinnati, OH 45221 U.S.A.
- |
Department of Mathematics\
Rose-Hulman Institute of Technology\
Terre Haute, IN 47803, U.S.A.
author:
- 'M. Burak Erdoan, Michael Goldberg, William R. Green'
title: Limiting absorption principle and Strichartz estimates for Dirac operators in two and higher dimensions
---
[^1]
Introduction
============
In this paper we obtain limiting absorption principle bounds and Strichartz estimates for the linear Dirac equation in dimensions two and higher with potential: $$\begin{aligned}
\label{dirac}
i\partial_t \psi(x,t) = (D_m +V(x)) \psi(x,t), \,\,\,\, \psi(x,0)= \psi_0(x).\end{aligned}$$ Here $x\in \R^n$ and $\psi(x,t) \in \mathbb{C}^{2^{N}}$ where $N=\lfloor \frac{n+1}{2}\rfloor$. The $n$-dimensional free Dirac operator $D_m$ is defined by $$\begin{aligned}
\label{eq:Dm}
D_m= -i \alpha \cdot \nabla+ m\beta = -i \sum_{k=1}^{n} \alpha_k \partial_k + m\beta,\end{aligned}$$ where $m\geq 0$ is a constant, and the $2^{N}\times 2^{N}$ Hermitian matrices $\alpha_j$ satisfy the anti-commutation relationships $$\begin{aligned}
\label{matrixeq}
\left\{\begin{array}{ll}
\alpha_j \alpha_k+\alpha_k\alpha_j =2\delta_{jk}
\mathbbm 1_{\mathbb C^{2^{N}}}
& j,k \in\{1,2,\dots, n\}\\
\alpha_j \beta+\beta \alpha_j=
\mathbb O_{\mathbb C^{2^{N}}}\\
\beta^2 = \mathbbm 1_{\mathbb C^{2^{N}}}
\end{array}
\right.\end{aligned}$$ Physically, $m$ represents the mass of the quantum particle. If $m=0$ the particle is massless and if $m>0$ the particle is massive. We note that dimensions $n=2,3$ are of particular physical interest. Following standard conventions, we define the free Dirac operator in dimension two with the Pauli spin matrices $$\alpha_1=\left[\begin{array}{cc} 0 & -i\\ i & 0
\end{array}\right],\
\alpha_2=\left[\begin{array}{cc} 0 & 1\\ 1 & 0
\end{array}\right],\
\beta=\left[\begin{array}{cc} 1 & 0\\ 0 & -1
\end{array}\right].$$ In dimension three we use $$\begin{aligned}
\beta=\left[\begin{array}{cc} I_{\mathbb{C}^2} & 0\\ 0 & -I_{\mathbb{C}^2}
\end{array}\right], \ \alpha_i=\left[\begin{array}{cc} 0 & \sigma_i \\ \sigma_i & 0
\end{array}\right],
\end{aligned}$$ $$\begin{aligned}
\sigma_1=\left[\begin{array}{cc} 0 & -i\\ i & 0
\end{array}\right],\
\sigma_2=\left[\begin{array}{cc} 0 & 1\\ 1 & 0
\end{array}\right],\
\sigma_3=\left[\begin{array}{cc} 1 & 0\\ 0 & -1
\end{array}\right].\end{aligned}$$ In higher dimensions $n>3$, one can create a full set of anti-commuting matrices $\alpha_j$ iteratively, see [@KS] for an explicit construction.
The Dirac equation arose as an attempt to reconcile the theories of relativity and quantum mechanics and describe the behavior of subatomic particles at near luminal speeds. The relativistic relationship between energy, momentum and mass, $E^2=c^2p^2+m^2c^4$ can be combined with the quantum-mechanical notions of energy $E=i\hbar \partial_t$ and momentum $p=-i\hbar \nabla$ to obtain a Klein-Gordon equation $$-\hbar^2 \partial_{t}^2 \psi(x,t)=-c^2\hbar^2 \Delta \psi(x,t)
+m^2c^4 \psi(x,t).$$ Here $\hbar$ is Planck’s constant and $c$ is the speed of light. However, the Klein-Gordon does not preserve $L^2$ norm of the initial data and is incompatible with quantum mechanical interpretations of the wave function. By considering $E$ directly, one arrives at the non-local equation \[eq:nonloc\] i\_t (x,t)= (x,t). In our mathematical analysis, we rescale so that we may take the constants $\hbar$ and $c$ to be one. Dirac’s insight was to rewrite the right hand side in terms of the first order operator $D_m=-i\alpha \cdot \nabla+m\beta$. This leads to the free Dirac equation, with $V=0$, a system of coupled hyperbolic equations with $\alpha, \beta$ required to be matrices. Dirac’s modification allows one to account for the spin of quantum particles, as well as providing a way to incorporate external electro-magnetic fields in a manner compatible with the relativistic theory where the Klein-Gordon and cannot. In addition, we note that has infinite speed of propagation, which is in contrast with the causality principle in relativity. In dimension $n=3$, the Dirac equation models the evolution of spin $1/2$ particles, while in dimension $n=2$ the massless Dirac equation is of considerable interest due to its connection to graphene, see, e.g., [@FW].
Formally, the Dirac equation is a square root of a system of Klein-Gordon or wave equations when $m>0$ and $m=0$ respectively. One consequence is that the spectrum of the free Dirac operators is unbounded in both the positive and negative directions. In particular, the continuous spectrum of $D_m$ is $(-\infty,-m]\cup[m,\infty)$. By Weyl’s criterion, the continuous spectrum of the perturbed Dirac operator is also $(-\infty,-m]\cup[m,\infty)$ for a large class of potentials. The absence of embedded eigenvalues in the continuous spectrum in general dimensions was established in [@BC1] for the class of potentials we are interested in by adapting the argument of [@BG1] for three dimensions. This result was used to study linearizations about a solitary wave for a non-linear equation. For other results in this direction for small dimensions and specific classes of potentials see [@Roze; @BG1; @Vog; @GM]. Finally, there is no singular continuous spectrum, see [@GM]. For a further background on the Dirac equation see [@Thaller].
We denote the perturbed Dirac operator by $H:=D_m+V$, then $e^{-itH}$ is formally the solution operator to . For the class of potentials considered in Theorem \[thm:main\], we note that $H$ is self-adjoint by the Kato-Rellich theorem. We denote $a-:=a-\eps$ for a small, but fixed $\eps>0$. Further, we write $A\les B$ to indicate there is a fixed absolute constant $C>0$ so that $A\leq C B$.
\[thm:main\] Let $V$ be a $2^{N}\times 2^{N}$ real Hermitian matrix for all $x\in\R^n$, $n\ge2$, with continuous entries satisfying $|V_{ij}(x)| \les \la x\ra^{-1- } $ when $m=0$, and $|V_{ij}(x)| \les \la x\ra^{-2- } $ when $m>0$. Furthermore, assume that threshold energies are regular. Then, with $P_c$ being the projection onto the continuous spectrum, $$\label{eq:strichmassive} \|\la \nabla\ra^{-\theta} e^{-itH} P_c f\|_{L_t^p(L_x^q)}
\les \|f\|_{L^2(\R^n)}$$ in the case $m > 0$, provided that $$\theta \geq \frac12 + \frac1p - \frac1q \quad \text{and} \quad
\frac{2}{p}+\frac{n}{q}=\frac{n}{2}, \,\,\,\,\, 2\le q<\frac{2n}{n-2}.$$ In the case $m=0$, the bound is $$\label{eq:strichmassless} \| |\nabla|^{-\theta} e^{-itH} P_c f\|_{L_t^p(L_x^q)}
\les \|f\|_{L^2(\R^n)}$$ provided that $$\theta = \frac{n}2-\frac1p-\frac{n}q \quad \text{and} \quad
\frac{2}{p}+\frac{n-1}{q}\leq\frac{n-1}{2}, \,\,\,\,\, p > 2,\ 2 \leq q < \infty.$$
The combinations of $(p,q , \theta)$ stated above are the same ones found in Strichartz estimates for the free massive ($m>0$) and massless ($m=0$) Dirac equation, respectively. Note that the range of admissible Strichartz exponents $(p,q)$ match those for the Schrödinger equation in the massive case, and the derivative is not homogeneous. This reflects the fact that the low energy behavior of the Dirac system is comparable to the Schrödinger equation, while the high energy behavior is closer to the wave equation (which requires differentiability of initial data). See the Appendix of [@DF07] for a derivation of Strichartz estimates for the free evolution $e^{-itD_m}$. The free massless Dirac system has the same scaling properties and admissible combinations as the free wave equation, which are proved in [@KT] for the wave equation.
These families of perturbed Strichartz estimates are a consequence of the uniform resolvent estimates that we prove. Much of the paper is devoted to proving the following resolvent bounds, which hold for any subset of the continuous spectrum of $H$.
\[thm:limap\] Let $V$ be a $2^{N}\times 2^{N}$ real Hermitian matrix for all $x\in\R^n$, $n\ge2$, with continuous entries satisfying $|V_{ij}(x)| \les \la x\ra^{-1- }$. Then for $\sigma > \frac12$ there exists $\lambda_1<\infty$ so that $$\label{eq:limap1}
\sup_{|\lambda| > \lambda_1 } \
\| \la x\ra^{-\sigma} (H - (\lambda + i0))^{-1}
\la x\ra^{-\sigma}\|_{2\to2} \les 1.$$ Under the assumption that the threshold energies are regular, and if $m>0$ the stronger decay condition $|V_{ij}(x)| \les \la x\ra^{-2-}$, this bound can be extended as follows $$\sup_{|\lambda| >m} \
\| \la x\ra^{-\sigma} (H - (\lambda + i0))^{-1}
\la x\ra^{-\sigma}\|_{2\to2} \les 1,$$ provided that $\sigma>\frac12$ when $m=0$, and $\sigma>1$ when $m>0$.
We note that the proof of the high energy limiting absorption principle does not require $V$ to be real or Hermitian. Since $V$ is assumed to be bounded and the free Dirac operator $D_m$ is self-adjoint, $H$ has the same domain as $D_m$ and for unit functions $\eta$ in the domain the quadratic form $\la H\eta, \eta\ra$ is confined to a strip of finite width around the real axis.
The immediate consequence of is that there cannot be any embedded eigenvalues or resonances on $(-\infty,-\lambda_1)\cup(\lambda_1,\infty)$ for $\lambda_1=\lambda_1(V) $ sufficiently large. A perturbation argument shows that the eigenvalue-free zone extends to a sector of the complex plane.
\[cor:e-free\] Under the hypotheses of Theorem \[thm:limap\], there exist $\lambda_1 < \infty$ and $\delta > 0$ depending on $V$, $m$, and $\sigma > \frac12$ so that \[eqn:limap2\] \_ x\^[-]{} (H - (+ i))\^[-1]{} x\^[-]{}\_[22]{} 1. As a result, there is a compact subset of the complex plane outside of which the spectrum of $H$ is confined to the real axis.
Our results apply to a broad class of electric potentials $V(x)$ and require no implicit smallness condition, only that $V$ is bounded, continuous and satisfies a mild polynomial decay at infinity. The potentials need not be small, radial, or smooth. Our results apply for the potentials that naturally arise when linearizing about soliton solutions for the non-linear Dirac equation.
There is a rich history of results on limiting absorption principles and mapping estimates of dispersive equations. Much of this history is focused on the analysis of the Schrödinger, wave or Klein-Gordon equation. We refer the reader to [@GV; @RS; @GS; @GST; @DF07; @Stef; @EGS; @BT; @MMT; @EGS2; @FV; @DFVV; @RT2], for example. There are far fewer results in the case of the Dirac system, due to its more complicated mathematical structure.
It is known that the Dirac resolvent does not decay in the spectral parameter, [@Yam]. That is, the bound does not decay as $\lambda\to \infty$. This is a stark contrast to the Schrödinger resolvent in which one obtains a decay in the spectral parameter $\lambda$ as $\lambda \to \infty$. The bootstrapping argument of Agmon, [@Agmon], produces uniform bounds on the resolvent operators only on compact subsets of the purely absolutely continuous spectrum. Limiting absorption principles have been studied to establish the limiting behavior of resolvents as one approaches the spectrum, see [@Vog; @BaHe; @BG1]. The work of Georgescu and Mantoiu provides resolvent bounds on compact subsets of the spectrum, [@GM]. Other limiting absorption principles have been established, often in service of providing dispersive, smoothing or Strichartz estimates, [@Bouss1; @DF; @BG]. Very recently, [@CG], established a limiting absorption principle for the free massless Dirac operator in dimensions $n\geq 2$.
One consequence of the resolvent bounds in Theorem \[thm:limap\] is the family of Strichartz estimates given in Theorem \[thm:main\]. Strichartz estimates have been used to study non-linear Dirac equations, [@MMNO; @CTS; @BH3; @BH; @BC1; @BC2]. These are often adapted to the problem by localizing in frequency or considering specialized potentials. Strichartz estimates may be obtained by establishing a virial identity see, for example, [@BDF; @C], which consider magnetic potentials with a certain smallness condition. The first and third author proved a class of Strichartz estimates for the two-dimensional Dirac equation, [@EGDirac2d], by first establishing dispersive estimates of the two-dimensional Dirac propagator.
The paper is organized as follows: In Section \[sec:setup\] we show how the Strichartz estimates in Theorem \[thm:main\] follows from the resolvent bounds in Theorem \[thm:limap\]. The bulk of the paper is then devoted to proving Theorem \[thm:limap\].
In Section \[sec:free\] we present the basic properties of the free resolvents of Dirac and Schrödinger operators. The small energy case of Theorem \[thm:limap\] is then treated in Section \[sec:small\]. In Section \[sec:high\_limap\], we treat the case of large energies by adapting an intricate argument originally devised in [@EGS; @EGS2] for Schrödinger operators in dimensions $n \geq 3$ with a non-smooth magnetic potential. A brief argument in Section \[sec:complex\] derives Corollary \[cor:e-free\] from the main high-energy bounds.
The basic setup {#sec:setup}
===============
The Strichartz estimates stated in Theorem \[thm:main\] will be proved using Proposition \[prop:Katoth\] below, which is essentially Theorem 4.1 in [@RS]. It is based on Kato’s notion of smoothing operators, see [@Kato]. We recall that for a self-adjoint operator $H$, an operator $\Gamma$ is called $H$-smooth in the sense of Kato if for any $f\in {\mathcal D}(H)$ $$\label{eq:Csmooth} \|\Gamma e^{-it H} f\|_{L^2_t L^2_x}\le
C_{\Gamma}(H) \|f\|_{L^2_x}.$$ Let $\Omega\subset \R$ and let $P_\Omega$ be a spectral projection of $H$ associated with a set $\Omega$. We say that $\Gamma$ is $H$-smooth on $\Omega$ if $\Gamma P_{\Omega}$ is $H$-smooth. It is not difficult to show (see e.g. [@RS4 Theorems XIII.25 and XIII.30]) that, $\Gamma$ is $H$-smooth on $\Omega$ if $$\label{eq:smon} \sup_{\lambda \in \Omega} \| \Gamma
[R^+_{H}(\lambda ) -R_H^-(\lambda) ] \Gamma^* \|_{L^2 \to L^2 } \le C_{\Gamma}(H,
\Omega).$$
Given the known Strichartz bounds for the free Dirac equation, the following proposition and Theorem \[thm:limap\] imply Theorem \[thm:main\]. For brevity we state only the $m>0$ case.
\[prop:Katoth\] Let $H_0=D_m$, $m>0$, and $H= H_0 + V$, where $|V(x)|\lesssim \la x\ra^{-2\sigma}$. Assume that $w(x):=\la x \ra^{-\sigma}$ is $H_0$-smooth and $H$-smooth on $\Omega$ for some $\Omega\subset \R$. Assume also that the unitary semigroup $e^{-it H_0}$ satisfies the estimate $$\label{eq:strH0} \big\| \la\nabla\ra^{-\theta} e^{-it H_0} \big\|_{L^2 \to L^q_t L^r_x} < \infty$$ for some $q\in (2,\infty]$, $r\in [1,\infty]$, and $\theta \in \R$. Then the semigroup $e^{-itH}$ associated with $H = H_0 + V$, restricted to the spectral set $\Omega$, also verifies the estimate , i.e., $$\label{eq:strHk} \big\|\la\nabla\ra^{-\theta} e^{-it H}P_\Omega \big\|_{L^2 \to L^q_t L^r_x} < \infty.$$
For completeness we supply the proof following [@RS]. We have $$e^{-it H}P_\Omega f = e^{-it H_0}P_\Omega f - i \int_0^t e^{-i (t-s) H_0} V e^{-is H }P_\Omega f ds.$$ By Christ-Kiselev Lemma [@CK], it suffices to prove that $$\Big\|\la\nabla\ra^{-\theta} \int_0^\infty e^{-i (t-s) H_0} V e^{-is H }P_\Omega f ds \Big\|_{ L^q_t L^r_x}\lesssim \|f\|_{L^2}.$$ Using , we bound the left hand side by \[eq:temp11\] \_0\^e\^[i s H\_0]{} V w\^[-1]{} w e\^[-is H ]{}P\_f ds \_[L\^2 ]{}. Since $w$ is $H_0$ smooth and $H$-smooth on $\Omega$, and $|Vw^{-1}|\les w$, we have $$\|we^{-itH}P_\Omega f \|_{L^2_t L^2_x}\lesssim \|f\|_{L^2_x}, \,\,\,\,\, \|Vw^{-1} e^{-it H_0} f\|_{L^2_t L^2_x}\lesssim \|f\|_{L^2_x},$$ and its dual $$\Big\| \int_0^\infty e^{i s H_0} V w^{-1} g(s,x) ds \Big\|_{L^2} \les \|g\|_{L^2_sL^2_x}.$$ Composing these two inequalities suffices to bound by $\|f\|_{L^2}$.
Properties of the Free Resolvent {#sec:free}
================================
The following identity,[^2] which follows from , \[dirac\_schro\_free\] (D\_m-1)(D\_m+1) =(-i+m-1) (-i+m+1) =(-+m\^2-\^2) allows us to formally define the free Dirac resolvent operator $\mathcal R_0(\lambda)=(D_m-\lambda)^{-1}$ in terms of the free resolvent $R_0(\lambda)=(-\Delta-\lambda)^{-1}$ of the Schrödinger operator for $\lambda$ in the resolvent set: $$\begin{aligned}
\label{eqn:resolvdef}
\mathcal R_0(\lambda)=(D_m+\lambda) R_0(\lambda^2-m^2).\end{aligned}$$ We first discuss the properties of Schrödinger resolvent $R_0$. There are two possible continuations to the positive halfline, namely $$R_0^\pm(\lambda^2)=\lim_{\eps\to 0^+} R_0(\lambda^2\pm i\eps),\,\,\,\,\,\,\lambda>0,$$ where the limit is in the operator norm from $L^2_\sigma $ to $L^2_{-\sigma}$, $\sigma>\frac12$. Here $L^2_\sigma$ denotes the weighted $L^2$ space with norm $$\|f\|_{L^2_\sigma} := \| \la \cdot\ra^{\sigma} f \|_{L^2}.$$ Existence of the limits $R_0^\pm(\lambda^2)$ is known as the limiting absorption principle. In fact $R_0^\pm(\lambda^2)$ varies continuously in $\lambda$ over the interval $(0,\infty)$. In dimensions $n\geq 3$ the continuity extends to $\lambda \in [0,\infty)$ with a uniform bound \[eq:lapScunif\] R\_0\^(\^2)\_[L\^2\_L\^2\_[-]{}]{}C\_[,n]{}, 0 provided $\sigma > 1$. In two dimensions the free Schrödinger operator has a threshold resonance and consequently $R_0^\pm(\lambda^2)$ is unbounded as $\lambda$ approaches zero. However there is still a useful uniform estimate, \[eq:lapSc\] R\_0\^(\^2)\_[L\^2\_L\^2\_[-]{}]{}+ R\_0\^(\^2)\_[L\^2\_L\^2\_[-]{}]{}C\_[,n]{}, >0,>12, which is true in all dimensions $n\geq 2$. This bound for large $\lambda$ is largely due to scaling considerations. The bound for small $\lambda$ will be proved in the next section.
Using the limiting absorption bounds for Schrödinger and , we obtain for $n\geq 2$ \[eq:freeDiracLAP\] R\_0()\_[L\^2\_L\^2\_[-]{}]{} C\_[,\_0,n]{}, || >\_0>m,>12. An analogous uniform bound holds on the entire interval $|\lambda|>m$ if $n\geq 3$ and $\sigma > 1$. In the case $m=0$ we have the following stronger uniform bound for $n\geq 2$ \[eq:masslessLAP\] R\_0()\_[L\^2\_L\^2\_[-]{}]{} C\_[, n]{}, || >0,>12. In particular, two dimensional massless free Dirac operator does not have a threshold resonance.
The kernel of the free resolvent $R_0^+(\lambda^2)$ in $\R^n$ is given by[^3] $$R_0^+(\lambda^2)(x,y) = C_n\,
\frac{\lambda^\frac{n-2}{2}}{|x-y|^{\frac{n-2}{2}}}
H_{\frac{n-2}{2}}^+(\lambda|x-y|)$$ where $H_\nu^+$ is a Hankel function. There is the scaling relation $$\label{eq:Rscale}
R_0^+(\lambda^2)(x,y)= \lambda^{n-2} R_0^+(1)(\lambda x,\lambda y)
\quad \forall\;\lambda>0$$ and the representation, see the asymptotics of $H_\nu^+$ in [@AbSteg], $$\label{eq:Hsplit}
R_0^+(1) (x,y) = \frac{e^{i|x-y|}}{|x-y|^{\frac{n-1}{2}}}
a(|x-y|) + \frac{b(|x-y|)}{|x-y|^{n-2}}$$ provided $n\ge2$. Here $$\label{eq:a}
|a^{(k)}(r)| \les r^{-k} \quad\forall\;k\ge0, \quad a(r)=0 \quad\forall\; 0< r<\f12$$ and $b(r)=0$ for all $r>\f34$, with $$\begin{aligned}
& |b^{(k)}(r)| \les 1 \quad\forall\;k\ge0, \quad n\text{\ \
odd} \label{eq:bodd}\\
&\left. \begin{array}{lll} |b^{(k)}(r)| &\les 1 &
\forall\;0\le k< n-2 \\
|b^{(k)}(r)| &\les r^{n-k-2}|\log r| &\forall\;k\ge n-2
\end{array}\right\}
\; n\ge 2\text{\ \ even} \label{eq:beven}\end{aligned}$$ for all $r>0$. In dimension $n=2$ we will need the following more detailed expansion of $b$ \[eq:2db\] b(r) = (-((r/2) + ) + ) + E(r), where $$\big| \mathcal E^{(k)}(r) \big| \les \bigg| \frac{d^k}{dr^k} (r^2\log r) \bigg| \qquad k=0,1,2.$$
In order to gain sharp control over the scaling behavior as $\lambda \to \infty$ we discuss the $\sigma=\frac12$ endpoint of the limiting absorption principle. As in Chapter XIV of [@Hor] define $$\|f\|_{B}:=\sum_{j=0}^\infty 2^{\frac{j}{2}} \|f\|_{L^2(D_j)}, \quad
\|f\|_{B^*} := \sup_{j\ge0} 2^{-\frac{j}{2}} \|f\|_{L^2(D_j)},$$ where $D_j=\{ x\::\: |x|\sim 2^j\}$ for $j\ge1$ and $D_0=\{|x|\le 1\}$. For each $\sigma > \frac12$, there are containment relations $L^2_\sigma \subset B$ and $B^* \subset L^2_{-\sigma}$. It is known that $\|R_0(1)\|_{B\to B^*} <\infty$.
Note that \[eq:BstarB\] Vf\_Bf\_[B\^\*]{} \_[j=0]{}\^2\^j V\_[L\^(D\_j)]{} x\^[1+]{} V\_[L\^]{} f\_[B\^\*]{}. Also recall that, by Lemma 3.1 in [@EGS2], we have the following scaling relations for any $\lambda\ge 1$ $$\|f( \lambda^{-1}\cdot) \|_B \les \lambda^{\frac{n+1}{2}}
\|f\|_B,\quad \|g(\lambda \cdot) \|_{B^*} \les
\lambda^{-\frac{n-1}{2}} \|g\|_{B^*},$$ provided the right-hand sides are finite. This and immediately imply the following statement. In what follows, $R_0$ stands for either of $R_0^{\pm}$.
\[prop:R0scale\] For all $\lambda\ge1$, we have $$\|R_0(\lambda^2)\|_{B\to B^*} \les \lambda^{-1} \|R_0(1)\|_{B\to
B^*}.$$
First, from $$(R_0^+(\lambda^2) f)(x) = \lambda^{-2} [R_0^+(1)
f(\cdot\lambda^{-1})] (\lambda x)$$ Hence, by the previous lemma, $$\|R_0^+(\lambda^2) f\|_{B^*} \les \lambda^{-2}
\lambda^{-\frac{n-1}{2}} \| R_0^+(1)
f(\cdot\lambda^{-1}) \|_{B^*}
\les \lambda^{-1} \|R_0^+(1)f\|_{B^*}$$ as claimed.
Energies close to $\{-m,m\}$ {#sec:small}
============================
In this section, assuming the regularity of threshold energies, we prove Theorem \[thm:limap\] when the spectral parameter $\lambda$ is sufficiently close to the threshold energy $\lambda=m$, respectively $\lambda=-m$. We consider the positive portion of the spectrum $[m,\infty)$, the negative part can be controlled similarly. That is, for sufficiently small $\lambda_0$ $$\begin{aligned}
\label{eq:sm1}
&\sup_{m<\lambda<\lambda_0} \|w (\mR^+_V(\lambda)-\mR^-_V(\lambda)) w \|_{2\to2} < \infty,\end{aligned}$$ where $w=\la x\ra^{-\sigma} $ for $\sigma >1$. In fact, we prove that $$\begin{aligned}
\label{eq:sm2}
&\sup_{m<\lambda<\lambda_0} \|w \mR^\pm_V(\lambda) w \|_{2\to2} < \infty,\end{aligned}$$ provided that $\sigma > 1$ when $m > 0$, and provided $\sigma > \frac12$ when $m=0$. A similar statement holds for negative energies.
We refer to reader to [@EGDirac2d] for the case $n=2$ and $m>0$, as this argument is substantially different from the other cases. In all the remaining cases we have $$\begin{gathered}
\|w \mR^\pm_V(\lambda) w \|_{2\to2} = \|w (1+\mR_0^\pm(\lambda )V)^{-1}w^{-1} w \mR^\pm_0(\lambda )
w
\|_{2\to2} \\
\le \|w (1+\mR_0^\pm(\lambda )V)^{-1}w^{-1} \|_{2\to2} \| w \mR^\pm_0(\lambda )
w
\|_{2\to2},\end{gathered}$$ so it suffices to show that $$\begin{aligned}
\label{eq:Z0small}
&\sup_{m< \lambda<\lambda_0} \|w (1+\mR_0^\pm(\lambda )V)^{-1}w^{-1} \|_{2\to2} <\infty, \\
&\sup_{m< \lambda<\lambda_0} \| w \mR^\pm_0(\lambda )w\|_{2\to2} < \infty. \label{eq:R0small}\end{aligned}$$ In dimensions $n\geq 3$, is an immediate consequence of the fact that the free Dirac operator is regular at the threshold, provided $\sigma > 1$. We will show below that is also true when $m=0$ and $n\geq 2$ with $\sigma > \frac12$. The $n=2$ case is somewhat surprising because the threshold is not regular for the free Schrödinger operator, nor for the free Dirac operator with $m > 0$.
In dimensions $n\geq 3$, let $G=\mR_0^\pm(m)= (D_m+m)R_0(0)$. In the case $n=2$, $m=0$, define $G= D_0 G_0$, where $$G_0f(x)=-\frac{1}{2\pi}\int_{\R^2} \log|x-y|f(y)\,dy \label{G0 def}.$$ Let $B_\lambda^\pm=\mR_0^\pm(\lambda)-G$. We assume that the threshold $m$ is a regular point of the spectrum, namely the boundedness of the operators \[eq:z0GL\] w(I + GV)\^[-1]{}w\^[-1]{} = (I+wGVw\^[-1]{})\^[-1]{}: L\^2L\^2. By a standard Fredholm alternative argument, is equivalent to the absence of resonances and eigenfunctions at $m$. We now prove that under suitable conditions \[eq:z0BL\] wB\^\_Vw\^[-1]{}\_[22]{} wB\^\_w \_[22]{} 0 m\^+. This and imply by summing the Neumann series directly, and it implies since $w G w$ is $L^2$-bounded for $\sigma > 1$ if $m > 0$ and $\sigma > \frac12$ if $m=0$.
To prove the bound recall the properties of the kernel of $B^\pm_\lambda$: with $\lambda =\sqrt{m^2+z^2}$, we have $$\begin{gathered}
\label{eq:Bexp}
B_\lambda^\pm=\mR_0^\pm(\lambda)-G=\big(D_m+\sqrt{m^2+z^2}\big)R_0(z^2) - (D_m+m )R_0(0) \\
=\big(\sqrt{m^2+z^2} -m\big)R_0(z^2) + m (\beta+I) \big[R_0(z^2)- R_0(0)\big] -i\alpha\cdot\nabla \big[R_0(z^2)- R_0(0)\big]\end{gathered}$$ in dimensions $n \geq 3$. When $m=0$ we have \[eq:Bexpm0\] B\_\^ = zR\_0(z\^2) - i and this holds for $n=2$ by replacing $R_0(0)$ with $G_0$.
By the limiting absorption principle for the free Schrödinger operator, the second summand of goes to zero as $z\to 0$ as operators from $L^2_\sigma $ to $L^2_{-\sigma}$, provided that $\sigma>1$, see . This can also be proved using the limiting absorption bound at frequency 1 and scaling, similar to the remaining cases that we discuss below. The remaining terms are identical to those in or better, since $0\leq \sqrt{m^2 + z^2} - m \leq z$. We will prove that both terms of go to zero for $\sigma>\frac12$ for dimensions $n\geq 2$.
For the first term, using the scaling relation and the representation we have $$z R_0(z^2)
= zR_0(z^2)(x,y)\widetilde\chi(z|x-y|) + z\frac{b(z|x-y|) }{|x-y|^{n-2}},$$ where $\widetilde\chi$ is a smooth cutoff for the complement of the unit ball. Using and , the low energy term can be bounded as follows $$\Big| z\frac{b(z|x-y|) }{|x-y|^{n-2}} \Big| \les z\frac{(z|x-y|)^{0-}}{|x-y|^{n-2}} \chi(z|x-y|)\les z\frac{(z|x-y|)^{-1+}}{|x-y|^{n-2 }}= \frac{z^{0+}}{|x-y|^{n-1- }}.$$ By the weighted version of the Schur’s test, this operator is $O(z^{0+}) $ as $z\to 0$ as an operator from $L^2_\sigma $ to $L^2_{-\sigma}$, provided that $\sigma>1/2$. We can rewrite the high energy term using the scaling relation : $$zR_0(z^2)(x,y)\widetilde\chi(z|x-y|) = z^{n-1} \big[R_0(1)\widetilde \chi\big](zx,zy).$$ Therefore, with $\chi$ be a smooth cutoff for $1/10$ neighborhood of the origin, $$\begin{aligned}
\big\|\la x\ra^{-\sigma} z\Big[R_0(z^2)&(x,y)\widetilde\chi(z|x-y|) \Big] \la y\ra^{-\sigma}\big\|_{L^2\to L^2} \\
&\les z^{-1}\big\| \la x/z\ra^{-\sigma} \big[R_0(1)\widetilde \chi\big](x,y) \la y/z\ra^{-\sigma}\big\|_{L^2\to L^2}\\
&\les z^{2\sigma-1} \big\| (z+|x|)^{-\sigma} \big[R_0(1)\widetilde \chi\big](x,y) (|y|+z)^{-\sigma}\big\|_{L^2\to L^2} \\
&\les z^{2\sigma-1} \big\| \la x\ra^{-\sigma} \big[R_0(1)\widetilde \chi\big](x,y) \la y\ra^{-\sigma}\big\|_{L^2\to L^2} \\
& \qquad + z^{2\sigma-1} \big\| (z+|x|)^{-\sigma} \chi(|x|) \big[R_0(1)\widetilde \chi\big](x,y) \la y\ra^{-\sigma}\big\|_{L^2\to L^2}\\
& \qquad \qquad + z^{2\sigma-1} \big\| \la x\ra^{-\sigma} \big[R_0(1)\widetilde \chi\big](x,y) (z+|y|)^{-\sigma} \chi(|y|) \big\|_{L^2\to L^2}.\end{aligned}$$ The first summand converges to zero provided that $\sigma>\frac12$ by the limiting absorption bound for the free Schrödinger operator for $\lambda=1$. Using the representation and the bound , and considering the Hilbert Schmidt norms, the second and third summands can be bounded by the square root of $$z^{4\sigma-2}\int \la x\ra^{-2\sigma}\frac{1}{\la x-y\ra^{ n-1} } (z+|y|)^{-2\sigma} \chi(|y|) dx dy \les z^{0+} \to 0,$$ provided that $\sigma>\frac12$.
We now consider the second summand in . In dimensions $n \geq 3$ we may use the scaling relation and the representation to write $$\nabla \big[R_0(z^2)- R_0(0)\big]
= \nabla \Big[R_0(z^2)(x,y)\widetilde\chi(z|x-y|) + \frac{b(z|x-y|)-b(0)}{|x-y|^{n-2}} \Big],$$ where $\widetilde\chi$ is a smooth cutoff for the complement of the unit ball. Using and , the low energy term can be bounded as follows $$\Big| \nabla \frac{b(z|x-y|)-b(0)}{|x-y|^{n-2}} \Big| \les \frac{z^{0+}}{|x-y|^{n-1-}},$$ which goes to zero as $z\to 0$ as an operator from $L^2_\sigma $ to $L^2_{-\sigma}$, provided that $\sigma>1/2$. If $n = 2$ we use to claim an analogous bound $$\big| \nabla\big[ b(z|x-y|) - G_0\big] \big| \les \frac{z^{0+}}{|x-y|^{1-}}.$$ Turning our attention to the high energy term, we use the scaling relation to write $$\nabla \Big[R_0(z^2)(x,y)\widetilde\chi(z|x-y|) \Big]= z^{n-1} \nabla\big[R_0(1)\widetilde \chi\big](zx,zy).$$ Therefore $$\begin{aligned}
\big\|\la x\ra^{-\sigma}\nabla \Big[R_0(z^2)&(x,y)\widetilde\chi(z|x-y|) \Big] \la y\ra^{-\sigma}\big\|_{L^2\to L^2} \\
&\les z^{-1}\big\| \la x/z\ra^{-\sigma} \nabla\big[R_0(1)\widetilde \chi\big](x,y) \la y/z\ra^{-\sigma}\big\|_{L^2\to L^2}\\
&\les z^{2\sigma-1} \big\| (z+|x|)^{-\sigma} \nabla\big[R_0(1)\widetilde \chi\big](x,y) (|y|+z)^{-\sigma}\big\|_{L^2\to L^2} \\
&\les z^{2\sigma-1} \big\| \la x\ra^{-\sigma} \nabla\big[R_0(1)\widetilde \chi\big](x,y) \la y\ra^{-\sigma}\big\|_{L^2\to L^2} \\
&\qquad+ z^{2\sigma-1} \big\| (z+|x|)^{-\sigma} \chi(|x|) \nabla\big[R_0(1)\widetilde \chi\big](x,y) \la y\ra^{-\sigma}\big\|_{L^2\to L^2}\\
&\qquad\qquad+ z^{2\sigma-1} \big\| \la x\ra^{-\sigma} \nabla\big[R_0(1)\widetilde \chi\big](x,y) (z+|y|)^{-\sigma} \chi(|y|) \big\|_{L^2\to L^2},\end{aligned}$$ where $\chi$ is a smooth cutoff for $1/10$ neighborhood of the origin. The first summand converges to zero provided that $\sigma>\frac12$ by the limiting absorption bound for the free Schrödinger operator for $\lambda=1$. Using the representation and the bound , and considering the Hilbert Schmidt norms, the second and third summands can be bounded by the square root of $$z^{4\sigma-2}\int \la x\ra^{-2\sigma}\frac{1}{\la x-y\ra^{ n-1} } (z+|y|)^{-2\sigma} \chi(|y|) dx dy \les z^{0+} \to 0,$$ provided that $\sigma>\frac12$.
The high energies limiting absorption principle {#sec:high_limap}
===============================================
Let us briefly consider intermediate energies, i.e., $\lambda\in I:= [\lambda_0,\lambda_1]\subset (-\infty,-m) \cup (m,\infty)$. It was shown in [@GM], see Theorem 1.6, that the resolvent of $H$ satisfies the limiting absorption principle uniformly in $\lambda$: $$\sup_{\lambda\in I}\|\la x\ra^{-\sigma}
\mR_V^\pm(\lambda) \la x\ra^{-\sigma} \|_{L^2\to L^2} \le C_I,$$ provided that there are no embedded eigenvalues, $\sigma>\frac12$, and $|V(x)| \les \la x\ra^{-1-}$.
In this section we complete the proof of Theorem \[thm:limap\] by considering energies sufficiently far from threshold, in the non-compact interval $|\lambda| \in (\lambda_1, \infty)$. In other words, we establish a limiting absorption principle for the perturbed Dirac resolvent $\mR_V^\pm (\lambda)$ at high energies: \[eqn:lap\] \_[||>\_0]{} \_V\^()\_[L\^[2,]{}L\^[2,-]{}]{} 1, > 12. In fact we control the slightly stronger operator norm from $B$ to $B^*$, and show that embedded eigenvalues are absent in this part of the spectrum.
Recall that (with $z=\sqrt{\lambda^2-m^2}$) we have \[eq:res\_iden\] \_V\^()= \_0\^()\^[-1]{}\
= \_0\^() \^[-1]{}, where $$L_z= V (D_m+\lambda).$$ Scaling arguments nearly identical to Proposition \[prop:R0scale\] show that $\|\mR_0^\pm(\lambda)\|_{B \to B^*} \les 1$ for all $\lambda \geq \lambda_0 > m$.
Since multiplication by $V$ maps $B^*$ to $B$ (see ), the limiting absorption principle for $R_0$ implies a bound $\|L_zR_0^\pm(z^2)\|_{B \to B} < C\|\la x\ra^{1+}V\|_{L^\infty}$ uniformly in $z \ge z_0 > 0$. If $V$ is sufficiently small then the operator norm of $L_zR_0^\pm(z^2)$ is less than $1$ for all $z \geq z_0$. Then one can conclude $$\sup_{z \geq z_0} \big\| \big[I + L_z R_0^\pm(z^2)\big]^{-1}\big\|_{B\to B} \lesssim 1.$$
The main goal of this section is to show that the same bound on $[I + L_zR_0^\pm(z^2)]^{-1}$ holds even when $V$ is not small. This cannot be proved directly from the size of $L_zR_0^\pm(z^2)$, which need not become small as $z \to \infty$. Instead, the following crucial lemma shows that the Neumann series $$\label{eq:geomser}
\big[I + L_z R_0^\pm(z^2)\big]^{-1} = \sum_{\ell=0}^\infty (-1)^\ell
\big(L_z R_0^\pm(z^2)\big)^\ell$$ is absolutely convergent for large $z$ due to the behavior of later terms in the series.
\[lem:largem\] Assume the entries of $V$ are continuous and satisfy $|V_{ij}(x)|\les \la x\ra^{-1-}$. There exist sufficiently large $M = M( V)$ and $z_1 =
z_1( V)$ such that $$\label{eq:smallprod}
\sup_{z>z_1}\| (L_z R_0^\pm(z^2))^M \|_{B \to B} \le \f12.$$
Assuming for the moment that holds, the operator inverse in is bounded uniformly in $z > z_1$, thus we conclude for $\lambda_0=\lambda_0(V)$ sufficiently large. The remainder of this section is devoted to the proof of Lemma \[lem:largem\]. The method will be similar to the one in [@EGS2].
The directed free resolvent
---------------------------
The first step is to decompose the free Schrödinger resolvent into a large number of pieces according to the size of $|x-y|$ and where $\frac{x-y}{|x-y|}$ lies on the unit sphere. This section presents a limiting absorption estimate for these truncated free resolvent kernels and for their first-order derivatives. The constants will not depend on the parameters of truncation, which gives us the freedom to choose those values later on. Similar estimates were obtained in [@EGS2] in dimensions $n\geq 3$, with derivatives of up to second order. We emphasize here the steps where $n \geq 2$ and the number of derivatives are most prominent, and refer the reader to [@EGS2] for technical details that are shared by both arguments.
For any $\delta\in(0,1)$, let $\Phi_\delta$ be a smooth cut-off function to a $\delta$-neighborhood of the north pole in $S^{n-1}$. Also, for any $d\in(0,\infty)$, $\eta_d(x)=\eta(|x|/d)$ denotes a smooth cut-off to the set $|x|>d$. In what follows, we shall use the notation $$R_{d,\delta}(\lambda^2)(x,y)=[R_0(\lambda^2) \eta_d \Phi_\delta]
(x,y) = R_0(\lambda^2)(x,y) \eta_d(|x-y|)
\Phi_\delta\Big(\frac{x-y}{|x-y|}\Big).$$ Note that this operator obeys the same scaling as $R_0$, see . More precisely, $$R_{d,\delta}(\lambda^2)(x,y) = \lambda^{n-2}
R_{d\lambda, \delta}(1)(\lambda x,\lambda y).$$ Thus, Proposition \[prop:R0scale\] applies to $R_{d,\delta}(\lambda^2)$ in the form $$\label{eq:calRscale}
\|R_{d,\delta}(\lambda^2)\|_{B\to B^*} \les \lambda^{-1}
\|R_{d\lambda, \delta}(1)\|_{B\to B^*}$$ for all $\lambda\ge1$ or, more generally, $$\label{eq:dercalRscale}
\| D^{\alpha}R_{d,\delta}(\lambda^2)\|_{B\to B^*} \les
\lambda^{-1+|\alpha|}
\|D^{\alpha}R_{d\lambda, \delta}(1)\|_{B\to B^*}$$ for all multi-indices $\alpha$ and $\lambda\ge1$.
We sketch a proof of a limiting absorption bound for $R_{d,\delta}$ and its derivatives of order at most one uniformly in the parameters $d,\delta\in(0,1)$, see Proposition \[prop:dir\_res\] below. This will be based on the oscillatory integral estimate in Lemma \[lem:VCP\], which was proved in [@EGS2] for $n\geq 3$ and $\frac{n-1}{2}\le p\le \frac{n+3}{2}$. The extension here to dimension $n=2$ with a smaller range of $p$ can be obtained from the proof of Lemma 3.4 in [@EGS2] by minor changes in the case analysis. More specifically, there is a step where one can replace the inequality $3(\frac{n-1}{2}) \geq \frac{n+3}{2}$ (which is true only if $n \geq 3$) with the slightly weaker bound $3(\frac{n-1}{2}) \geq \frac{n+1}{2}$.
\[lem:VCP\] Let $\chi$ denote a smooth cut-off function to the region $1<|x|<2$. With $a(r)$ as in , define $$\label{eq:pint}
(T_{\delta,p, R_1, R_2} f)(x) = \int \chi\big(\frac{x}{R_1}\big)
\frac{e^{i|x-y|}}{|x-y|^p}
a(|x-y|)
\Phi_\delta\Big(\frac{x-y}{|x-y|}\Big) \chi\big(\frac{y}{R_2}\big) f(y)\, dy.$$ Then, for any $n\ge2$, and $\frac{n-1}{2}\le p\le \frac{n+1}{2}$, $$\label{eq:Tbound}
\| T_{\delta,p, R_1, R_2} f \|_2 \le C_n\, \delta^{p-\frac{n-1}{2}}
\sqrt{R_1 R_2}\, \|f\|_2$$ for all $R_1, R_2\ge1$, $\delta\in(0,1)$. The constant $C_n$ only depends on $n\ge2$.
\[prop:dir\_res\] Let $n\ge2$. Then for any $d\in (0,\infty)$, $\delta\in (0,1)$, and $\lambda\ge 1$ there is the bound $$\label{eq:limap}
\| D^{\alpha}R_{d,\delta}(\lambda^2)f \|_{B^*} \le
C_n\lambda^{-1+|\alpha|} \|f\|_B$$ for any $0\le|\alpha|\le 1$. The constant $C_n$ depends only on the dimension $n\ge2$.
In view of and it suffices to prove this estimate for $\lambda=1$. We need to prove that for any $0\le|\alpha|\le 1$ $$\label{eq:star}
\| \chi(\cdot/R_1) D^\alpha R_{d,\delta}(1)\, \chi(\cdot/R_2) f
\|_2 \le C_n\, \sqrt{R_1 R_2}\, \|f\|_2$$ where $R_1, R_2\ge 1$ are arbitrary. We write $$\label{eq:T0T1}
R_{d,\delta}(1)=R_0^+(1) \eta_d\Phi_\delta = T_0 + T_1$$ where the kernels of $T_0, T_1$ are $$\label{eq:Tdef}\begin{split}
T_0(x,y) &=
\frac{b(|x-y|)}{|x-y|^{n-2}}
\eta_d(|x-y|) \Phi_\delta(x,y),\\
T_1(x,y) &= \frac{e^{i|x-y|}}{|x-y|^{\frac{n-1}{2}}}
\eta_d(|x-y|)a(|x-y|)
\Phi_\delta(x,y),
\end{split}$$ respectively, see . The modified function $\eta_d(r)a(r)$ satisfies all decay estimates in with constants independent of the choice of $d$.
We begin by showing that $\widehat{T_0 f} = m_0 \hat{f}$ where $|m_0(\xi)|\les \la
\xi\ra^{-1}$. This will imply for $T_0$. By definition $$m_0(\xi) = \int_0^\infty\int_{S^{n-1}} rb(r) \eta_d(r) e^{-ir
\omega\cdot\xi} \Phi_\delta(\omega) \,\sigma(d\omega) \,dr.$$ Since $b(r)=0$ if $r>\f34$, $|m_0(\xi)|\les 1$. Hence we may assume that $|\xi|\ge1$. If $|\xi_n| \ge |\xi|/10$, then $|\omega \cdot \xi| \ges
|\xi|$ and $$|m_0(\xi)|\les \int_{S^{n-1}} \Phi_\delta(\omega) \la
\omega\cdot\xi\ra^{-2}\,\sigma(d\omega)\les \delta^{n-1} |\xi|^{-2},$$ where we have used that $$\Big| \int_0^\infty e^{-ir\rho} rb(r)\eta_d(r)\chi(r)\,dr\Big |\les
\la \rho\ra^{-1}.$$ This follows from and after an integration by parts. Now suppose that $|\xi_n| \le
|\xi|/10$. Set $\xi=|\xi| \hat{\xi}$ and change integration variables as follows: $$\begin{aligned}
&\int_{S^{n-1}}\int_0^\infty rb(r) \eta_d(r) \chi(r) e^{-ir|\xi|
\omega\cdot\hat{\xi}}\,dr\, \Phi_\delta(\omega) \,\sigma(d\omega)\\
& = \int_{\R^{n-1}} \int_0^\infty rb(r) \eta_d(r) \chi(r)
e^{-ir|\xi| u_1}\,dr \,{\widetilde}\Phi_\delta(u_1,\ldots,u_{n-1})\,
du_1du_2\ldots du_{n-1}\\
& = \delta^{n-2} \int_0^\infty \int_\R rb(r) \eta_d(r) \chi(r)
e^{-ir|\xi| u_1}\Psi_\delta(u_1)\, du_1 dr,\end{aligned}$$ where $(u_1, \ldots, u_{n-1})$ is a parametrization of the support of $\Phi_\delta$, aligning $u_1$ with $\hat{\xi}$. The function $\Psi_\delta$ is a smooth cut-off supported on an interval of length $\sim \delta$ resulting from the integration of ${\widetilde}{\Phi}_\delta$. Thus, $$|m_0(\xi)|\les \delta^{n-2}\int_0^1
|\widehat{\Psi_\delta}(r|\xi|)|\,dr\les
\delta^{n-2}|\xi|^{-1}\| \widehat{\Psi_\delta}(u)\|_{L^1_u}\les
\delta^{n-2}|\xi|^{-1}.$$ In conclusion, $|m_0(\xi)|\les \la \xi\ra^{-1}$ as claimed.
Next, consider $T_1$. By the Leibniz rule, $$\begin{aligned}
D^\alpha_x T_1(x,y) &= \sum_{\beta\le \alpha} c_{\alpha,\beta}\,
D^{\alpha-\beta}_x \Big[\frac{e^{i|x-y|}}{|x-y|^{\frac{n-1}{2}}}
\eta_d(|x-y|) a(|x-y|)\Big] D^{\beta}_x
\Phi_\delta(x,y) \nonumber\\
& = \sum_{\beta\le \alpha} \delta^{-|\beta|}\,c_{\alpha,\beta}\,
\frac{e^{i|x-y|}}{|x-y|^{\frac{n-1}{2}+|\beta|}}
a_{\alpha,\beta, d}(|x-y|)
\Phi_{\delta,\beta}(x,y),\label{eq:newT1}\end{aligned}$$ where $\Phi_{\delta,\beta}= \delta^{|\beta|} D^\beta \Phi_\delta$ is a modified angular cut-off and $a_{\alpha,\beta, d}$ satisfies the same bounds as $a$, see , with constants that do not depend on $d$. The estimate for $T_1$ follows from Lemma \[lem:VCP\] with $p=\frac{n-1}{2}+|\beta|$.
A partition of unity $\{\Phi_i\}$ over $S^{n-1}$ induces a directional decomposition of the free resolvent, namely $$\label{eq:decomp}
R_0(\lambda^2) = \sum_{i} R_i(\lambda^2) + R_d(\lambda^2)$$ where $R_i(\lambda^2) := R_{d,\delta}(\lambda^2)$ with $\Phi_i$ playing the role of $\Phi_\delta$ from above. Moreover, $R_d(\lambda^2)(x)=(1 - \eta_d(|x|))R_0(\lambda^2)(|x|)$ is the “short range piece”. We have the following $L^2$ bound for $R_d(\lambda^2)$:
\[lem:R\_d\] With $R_d^+(\lambda^2)$ defined as above, the mapping estimate $$\label{eq:R_d}
\|D^\alpha R_d(\lambda^2)f \|_2 \le C_n\, \lambda^{-2+|\alpha|}
\la d\lambda\ra \|f\|_2$$ holds uniformly for every choice of $d \in (0,\infty)$, $0\le|\alpha|\le1$, and $\lambda \ge 1$.
By the scaling relation , for any $\alpha$, $$\| D^\alpha R_0(\lambda^2) \chi_{[|x|<d]} \|_{2\to 2}= \lambda^{-2+|\alpha|} \|
D^\alpha R_0(1) \chi_{[|x|<\lambda d]} \|_{2\to2}$$ where $\chi_{[|x|<\rho]}=\chi(|x|/\rho)$ is a smooth cut-off to the set $|x|<\rho$ with $\rho>0$ arbitrary. The notation is somewhat ambiguous here; we are seeking an estimate for the convolution operator with kernel $D^\alpha R_0(1)\chi_{[|x| < \lambda d]}$. The lemma is proved by showing that the Fourier transform of $R_0(1)\chi_{[|x| < \rho]}$ is bounded point-wise by $\japanese[\rho]\japanese[\xi]^{-1}$.
Consider first the case $\rho \le 1$. The decomposition implies that $$\int_{\R^n} |R_0(1)(x)\chi(|x|/\rho)|\,dx \les \rho^2\log(\rho).$$ Furthermore, since $(\Delta+1)R_0(1)$ is a point mass at the origin, the distribution $\Delta[R_0(1)\chi_{[|x| < \rho]}]$ consists of a point mass plus a function of $L^1$ norm $\lesssim |\log(\rho)|$. This implies that the Fourier transform of $R_0(1)\chi_{[|x| < \rho]}$ is bounded by $|\log(\rho)| |\xi|^{-2}$. The desired Fourier transform estimate follows by interpolating these two bounds.
When $\rho > 1$, it is more convenient to estimate $$\rho^n \Big| \int \big[\PV \frac{1}{|\eta|^2-1} +
i \sigma_{S^{n-1}}(d\eta)\big]\hat{\chi}((\xi-\eta)\rho)\, d\eta \Big|.$$ A standard calculation shows this to be less than $\rho\japanese[\rho(|\xi|^2 - 1)]^{-1} < \rho\japanese[\xi]^{-2}$.
Proof of Lemma \[lem:largem\]
-----------------------------
Decomposing each free resolvent in the $M$-fold product $(L_zR_0(z^2) )^M$ as in yields the identity $$\label{eq:decomp2}
(L_zR_0(z^2) )^M = \sum_{i_1 \ldots i_M} \prod_{k=1}^M
\big(L_z R_{i_k}(z^2) \big).$$ The indices $i_k$ may take numerical values corresponding to the partition of unity $\{\Phi_i\}$, or else the letter $d$ to indicate a short-range resolvent. There are two main types of products represented here, namely:
- [*Directed Products*]{}, where the support of functions $\Phi_{i_k}$ and $\Phi_{i_{k+1}}$ are separated by less than $10\delta$ for each $k$. A product is also considered to be directed if it has this property once all instances of $i_k = d$ are removed. The term $(L_zR_d(z^2) )^M$ is a vacuous example of a directed product.
- All other terms not meeting the above criteria are [*Undirected Products*]{}. An undirected product must contain two adjacent numerical indices (i.e., after discarding all instances where $i_k = d$) for which the corresponding functions $\Phi_i$ have disjoint support with distance at least $10\delta$ between them.
\[lem:combinatorics\] For any $\delta>0$, there exists a partition of unity $\{\Phi_i\}$ with approximately $\delta^{1-n}$ elements, having ${\rm diam}\,{\rm
supp}\,(\Phi_i) < \delta$ for each $i$ and admitting no more than $\delta^{1-n}(C_n)^M$ directed products of length $M$ in .
The first claim is a standard fact from differential geometry. For the second claim note that there are $\les \delta^{1-n}$ choices for the first element in a directed product, but only $C_n$ choices at each subsequent step.
The iterated resolvent $(L_zR_0(z^2))^M$ is an oscillatory integral operator with phase $e^{iz\sum_{k=1}^M |x_{k} - x_{k-1}|}$, where $x_0 = y$ and $x_M = x$. Loosely speaking, there is a region of stationary phase where $\sum_{k=1}^M |x_k - x_{k-1}|
\approx |x-y|$. The integral kernel of a directed product is supported here, hence one cannot gain any benefit from oscillation as $z \to \infty$ beyond the bounds for individual resolvents in Proposition \[prop:dir\_res\] and Lemma \[lem:R\_d\]. Those bounds do not decrease to zero in the limit of large $z$. It appears that the operator norm of a directed product does not decrease to zero either.
Never the less, one can show that long directed products have an operator norm that is small enough for our purposes. The geometric idea is relatively simple: If $\delta < \frac{1}{20M}$, then all the angular cutoffs $\Phi_{i_k}$ have support within a single hemisphere. The convolution operators $R_{i_k}(z^2)$ are therefore biased consistently to one side. This introduces a gain from the product $\prod_{k=1}^M V(x_k)$, as only a handful of $x_k$ can be located near the origin, and $V(x_k)$ is small everywhere else. The following lemma is adapted from Lemma 4.8 of [@EGS2]. There is a small but significant difference in the structure of the perturbation. Here $L_z = V(D_m + \lambda)$ has the property that $L_zR_0(z^2)$ is a bounded operator on the space $B$. The symmetrized version $(L_z + L_z^*)R_0(z^2)$ does not map $B$ to itself unless $V(x)$ is assumed to be differentiable. This explains the use of more elaborate function spaces in [@EGS2] and the need for bounds on the second derivative of the truncated free resolvent (which ultimately restricts the dimension to $n \geq 3)$.
Now we may state the bound for directed products involving $L_z$ in dimensions $n \geq 2$.
\[lem:directedprod\] Given any $r > 0$, there exists a distance $d = d(r) > 0$ such that each directed product in satisfies the estimate $$\bignorm[\prod_{k=1}^M \big(L_z R_{i_k}(z^2) \big)f][B]
\le
C_{n,V,r}\, r^{M} \norm[f][B]$$ uniformly over all $z> d^{-1}$ and all choices of $M$ and $\delta$ satisfying $\delta \le \frac{1}{20M}$.
Consequently, given any $c>0$, there exists a number $M= m(c,V)$ and a partition of unity governed by $\delta = \frac{1}{20M}$ so that the sum over all directed products achieves the bound $$\label{eq:directedprod}
\sum_{\substack{i_1 \ldots i_M \\ directed}} \bignorm[\prod_{k=1}^M
\big(L_zR_{i_k}(z^2))][B \to B] \le
{\textstyle \frac{c}{2}}$$ uniformly in $z > d^{-1}$.
In this proof, we will keep track of the superscripts $\pm$ on the resolvents. Also, we will write $\|V\|_{B^*\to B}=C_V$. There is no loss of generality if we assume that $r < C_nC_V$.
After a rotation, we may assume that every function $\Phi_{i_k}$ which appears in the product has support within a half-radian neighborhood of the north pole, where $x_n > \frac23$. If $f \in B$ is supported on the half space $\{x_n
> A\}$, then the support of $R_{i_k}^+(\lambda^2)f$ must be translated upward to $\{x_n > A + \frac23d\}$. The short-range resolvent $R_d^+(\lambda^2)$ does not have a preferred direction; however if $f \in B$ is supported on $\{x_n > A\}$ then ${\rm supp}\,R_d^+(\lambda^2)f
\subset \{x_n > A - 2d\}$.
The purpose of keeping track of supports is that if $f \in
B^* $ is supported away from the origin, in the set $\{|x| >
A\}$, then the estimate in can be improved to $$\norm[Vf][B] \les \norm[\la x \ra^{1+}V\chi_{[|x| > A]}][L^\infty]
\norm[f][B^*]\les \la A\ra^{0-} \norm[f][B^*].$$ Note that we can choose $A=A(n,V,r)$ so that $$\label{eq:distantL}
\norm[Vf][B]\leq \frac{r^2}{C_n^2C_V}\|f\|_{B^*},$$ provided that $f$ is supported in the set $\{|x| >
A\}$.
Let $\chi$ be a smooth function supported on the interval $[-1,
\infty)$ such that $\chi(x_n) + \chi(-x_n) = 1$. We will initially estimate the operator norm of $\big(\prod_k (L_zR_{i_k}^+(z^2))
\big) \chi(x_n)$. Multiplication by $\chi(x_n)$ is bounded operator of approximately unit norm on $B^*$.
The support of $\chi(x_n) f$ lies in the half-space $\{x_n > -1\}$. Suppose every one of the indices $i_k$ is numerical. Then each application of an operator $L_zR_{i_k}^+(z^2)$ translates the support upward by $\frac23d$. For the first $\frac{3A}{2d}$ steps the operator norm of $L_zR_{i_k}^+(z^2)$ is bounded by $C_nC_V$. Thereafter it is possible to use the stronger bound of because the support will have moved into the half-space $\{x_n > A\}$. The combined estimate is $$\label{eq:alldirected}
\begin{aligned}
\bignorm[\prod_{k=1}^m (L_zR_{i_k}^+(z^2))\,\chi(x_n)f][B]
&\le (C_nC_V)^{m} \Big(\frac{r^2}{(C_nC_V)^2}\Big)^{m-\frac{3a}{2d}}
\norm[f][B] \\
& = (C_n C_V)^{-m}(r^{-1}C_n C_V)^{\frac{3a}{d}} r^{2m} \norm[f][B].
\end{aligned}$$ This is also valid for small $m$ by our assumption that $r < C_nC_V$.
If each directed resolvent $R_{\Phi_i}^+(\lambda^2)$ is seen as taking one step forward, then the short-range resolvent $R_d^+(\lambda^2)$ may take as many as three steps back. Suppose a directed product includes exactly one index $i_k = d$. This will have the most pronounced effect if it occurs near the beginning of the product, delaying the upward progression of supports by a total of $4$ steps. In this case one combines , and Lemma \[lem:R\_d\] to obtain $$\bignorm[\prod_{k=1}^m (L_zR_{i_k}^+(z^2))\,\chi(x_n)f][B]
\le (C_nC_V)^{m} d \Big(\frac{r^2}{(C_nC_V)^2}\Big)^{m-(\frac{3A}{2d}+4)}
\norm[f][B].$$ Notice that this estimate agrees with the one in up to a factor of $d(r^{-1}C_nC_V)^8$. By setting $d = d(r) = \big(\frac{r}{C_nC_V}\big)^8$, the bound in is strictly larger. Similar arguments yield the same result for any directed product with one or more instances of the short-range resolvent $R_d^+(\lambda^2)$.
To remove the spatial cutoff, write $$\prod_{k=1}^m L_zR_{i_k}^+(z^2) \ = \ \Big(\prod_{k=1}^{m/2}
L_zR_{i_k}^+(z^2)\Big)(\chi(x_n) + \chi(-x_n))
\Big(\prod_{k=\frac{m}2+1}^m L_zR_{i_k}^+(z^2)\Big).$$ Consider the $\chi(x_n)$ term. By , the first half of the product carries an operator norm bound of $(C_n\, C_V)^{-\frac{m}{2}}
(r^{-1}C_nC_V)^{\frac{3A}{d(r)}}r^m$. The second half contributes at most $(C_n C_V)^{m/2}$. Put together, this product has an operator norm less than $C_{n,V,r}\,r^m$, where $C_{n,V,r} = (r^{-1}C_nC_V)^{\frac{3A}{d(r)}}$.
The $\chi(-x_n)$ term has nearly identical estimates, by duality. The adjoint of any directed resolvent $R_{\Phi}^+(z^2)$ is precisely $R_{{\widetilde}{\Phi}}^-(z^2)$, with ${\widetilde}{\Phi}$ being the antipodal image of $\Phi$. Because the order of multiplication is reversed, one applies the geometric argument above to the adjoint operators $ R_{i_k}^-(z^2) L_z^*$ (modulo the antipodal map).
According to Lemma \[lem:combinatorics\] there are at most $\delta^{1-n}
(C_n)^m$ directed products of length $m$. To prove , it therefore suffices to let $r = \frac{1}{2C_n}$, and $\delta = \frac{1}{20m}$ so that the sum of the operator norms of all directed products is bounded by $20^{n-1} C_{n,V} m^{n-1}2^{-m}$. This can be made smaller than $\frac{c}{2}$ by choosing $m$ sufficiently large.
As for the undirected products, recall that their defining feature is the presence of adjacent resolvents $R_{i}^+(\lambda^2)$ oriented in distinct directions. The resulting oscillatory integral has no region of stationary phase, and therefore exhibits improved bounds at high energy provided the potential $V(x)$ is smooth. The following lemma follows from Lemma 4.9 in [@EGS2] by minor changes in the proof.
\[lem:nonstatphase\] Let $\Phi_1$ and $\Phi_2$ be chosen from a partition of unity of $S^{n-1}$ so that their supports are separated by a distance greater than $10\delta$. Suppose $V \in C^\infty(\R^n)$ with compact support. Then for each $j \ge 0$, and any $N\ge1$, $$\label{eq:Phi1Phi2}
\big\| L_zR_{d,\Phi_2}^+(z^2) (L_z R_d^+(z^2))^j L_z
R_{d,\Phi_1}^+(z^2) \big\|_{B\to B} =
\O(z^{-N})$$ as $z\to\infty$ and similarly for $R^-(z^2)$.
Note that under the conditions of Lemma \[lem:nonstatphase\] each undirected product in satisfies the bound $$\label{eq:nonstatphase}
\bignorm[\prod_{k=1}^M \big(L_z R_{i_k}(z^2)\big)][B \to B]
= \O(z^{-N})$$ for any $N\ge1$. We now show by approximation that vanishing still holds for merely continuous $V$, but without any control over the rate.
\[lem:RiemannLebesgue\] Let $\Phi_1$ and $\Phi_2$ be chosen as in Lemma \[lem:nonstatphase\]. Suppose $V$ is a continuous function with $V\in Y$. Then each undirected product in satisfies the limiting bound $$\label{eq:RiemannLebesgue}
\lim_{z \to \infty} \bignorm[ \prod_{k=1}^M
\big(L_z R_{i_k}(z^2)\big)][B\to B] = \ 0.$$
For any small $\gamma > 0$, there exists a smooth approximation $V_\gamma \in C^\infty(\R^n)$ of compact support so that $\norm[V -
V_\gamma][B^*\to B] < \gamma$ and $\|V_\gamma\|_{B^*\to B}<2\|V\|_{B^*\to B}$. Define the operator $L_{z,\gamma}$ accordingly. We have $$\bignorm[\prod_{k=1}^M \big(L_z R_{i_k}(z^2)\big) -
\prod_{k=1}^M \big(L_{z,\gamma}R_{i_k}(z^2)\big)][B\to B]
\les \gamma (2\|V\|_{B^*\to B})^{M-1}$$ uniformly in $z\ge1$. Thus, by , $$\limsup_{z \to \infty} \bignorm[\prod_{k=1}^M
\big(L_z R_{i_k}(z^2)\big)][B\to B] \les \gamma
(2\|V\|_{B^*\to B})^{M-1}.$$ Sending $\gamma\to0$ finishes the proof.
Lemma \[lem:directedprod\] provides a recipe for selecting a value of $M$, together with a partition of unity $\{\Phi_i\}$ and a short-range threshold $d$, so that the sum over all directed products in will be an operator of norm less than $\frac{1}{4}$. This fixes the number of undirected products as approximately $\delta^{M(1-n)} = (20 M)^{M(n-1)}$. For each of these, Lemma \[lem:RiemannLebesgue\] asserts that its operator norm tends to zero as $\lambda \to \infty$. The same is true for the finite sum over all undirected products of length $M$. In particular it is less than the directed product estimate provided $z >
z_1(M)$ is sufficiently large.
Extension to the complex plane {#sec:complex}
==============================
This section provides a short proof of Corollary \[cor:e-free\] via a perturbation argument. We first record a strong statement of continuity in the limiting absorption principle.
\[prop:continuity\] Let $\lambda_n \to 1$ and $\eps_n \to 0^+$. Then for each $\sigma > \frac12$, $$\lim_{n\to \infty}
\|R_0(\lambda_n + i\eps_n) - R_0^+(1)\|_{L^2_\sigma \to L^2_{-\sigma}} = 0.$$
By the triangle inequality and scaling relations, $$\begin{aligned}
\|R_0(\lambda_n + i\eps_n) &- R_0^+(1)\|_{L^2_\sigma \to L^2_{-\sigma}} \\
&\leq \|R_0(\lambda_n + i\eps_n) - R_0^+(\lambda_n)\|_{L^2_\sigma \to L^2_{-\sigma}}
+ \|R_0^+(\lambda_n) - R_0^+(1)\|_{L^2_\sigma \to L^2_{-\sigma}} \\
&\les \|R_0(1+i{\scriptstyle \frac{\eps_n}{\lambda_n}}) - R_0^+(1)\|_{L^2_\sigma \to L^2_{-\sigma}}
+ \|R_0^+(\lambda_n) - R_0^+(1)\|_{L^2_\sigma \to L^2_{-\sigma}}.\end{aligned}$$ Both of the differences in the last line converge to zero by the limiting absorption principle and continuity of the resolvent respectively.
By the standard resolvent identities and $$\begin{aligned}
\mR_V(\lambda + i\gamma) &= \mR_0(\lambda+i\gamma) \big[I + V\mR_0(\lambda+i\gamma)\big]^{-1}\\
&= (D_m+\lambda+i\gamma)R_0((\lambda+i\gamma)^2 - m^2)
\big[I+V\mR_0(\lambda+i\gamma)\big]^{-1}\end{aligned}$$ The free Dirac resolvent is controlled by rescaling by $\lambda$. For the gradient term, $$\| \nabla R_0((\lambda+i\gamma)^2-m^2)\|_{L^2_\sigma \to L^2_{-\sigma}}
\leq \| \nabla R_0((1+i{\scriptstyle \frac{\gamma}{\lambda}})^2 - ({\scriptstyle \frac{m}{\lambda}})^2)\|_{L^2_\sigma \to L^2_{-\sigma}} \les 1,$$ with the last inequality following from continuity of the Schrödinger resolvent at 1. Similarly, $$\begin{gathered}
|m+\lambda+i\gamma| \,\| R_0((\lambda+i\gamma)^2-m^2)\|_{L^2_\sigma \to L^2_{-\sigma}}\\
\leq \Big(1+\frac{m}{\lambda} + \frac{\gamma}{\lambda}\Big)
\| R_0((1+i{\scriptstyle \frac{\gamma}{\lambda}})^2 - ({\scriptstyle \frac{m}{\lambda}})^2)\|_{L^2_\sigma \to L^2_{-\sigma}}
\les 1\end{gathered}$$ provided $\frac{\gamma}{\lambda}$ and $\frac{m}{\lambda}$ are sufficiently small. Boundedness of $\mR_V(\lambda+i\gamma)$ therefore rests on the behavior of $[I+V\mR_0(\lambda+i\gamma)]^{-1}$.
The crucial bound shows that $$\| [I + V\mR_0^+(\lambda)]^{-1} \|_{L^2_\sigma \to L^2_{\sigma}} \les 1 + \| V\mR_0^+(\lambda) \|^{M-1}_{L^2_\sigma \to L^2_\sigma}
\leq C(V)$$ for all $\lambda > \lambda_0$. We would like to expand $$\label{eqn:gamma_perturbed}
\big[I + V\mR_0(\lambda+i\gamma)\big]^{-1} = \big[ (I + V\mR_0^+(\lambda)) +
V\big(\mR_0(\lambda+i\gamma) - \mR_0^+(\lambda)\big) \big]^{-1}$$ via a Neumann series, and this can be done provided the operator norm of $V(\mR_0(\lambda+i\gamma) - \mR_0^+(\lambda))$ is less than $\frac{1}{C(V)}$.
The difference of Dirac resolvents can be expanded out as $$\begin{aligned}
\mR_0(\lambda+i\gamma) - \mR_0^+(\lambda)
&= (D_m + \lambda + i\gamma)R_0((\lambda+i\gamma)^2-m^2) - (D_m + \lambda)R_0^+(\lambda^2-m^2)
\\
&= i\gamma R_0^+(\lambda^2 - m^2) + (D_m + \lambda + i\gamma)
\big(R_0((\lambda+i\gamma)^2-m^2)-R_0^+(\lambda^2-m^2)\big).\end{aligned}$$ Each of these terms can be bounded using the same scaling arguments as above, with the end result $$\begin{gathered}
\|\mR_0(\lambda+i\gamma) - \mR_0^+(\lambda)\|_{L^2_\sigma \to L^2_{-\sigma}} \\
\les \frac{\gamma}{\lambda}\|R_0^+(1 - ({\scriptstyle \frac{m}{\lambda}})^2)\|_{L^2_\sigma \to L^2_\sigma}
+ \Big(1 + \frac{m}{\lambda} + \frac{\gamma}{\lambda}\Big)
\|R_0((1 + i{\scriptstyle \frac{\gamma}{\lambda}})^2 - ({\scriptstyle \frac{m}{\lambda}})^2)
- R_0^+(1-({\scriptstyle \frac{m}{\lambda}})^2) \|_{L^2_\sigma \to L^2_\sigma}.\end{gathered}$$
Proposition \[prop:continuity\] asserts the existence of constants $\delta > 0$ and $K < \infty$ such that if $0 \leq \frac{\gamma}{\lambda}, \frac{m}{\lambda} < \delta$, then $\|R_0^+( 1 - (\frac{m}{\lambda})^2)\| \leq K$ and $\|R_0((1 + i\frac{\gamma}{\lambda})^2 - (\frac{m}{\lambda})^2)
- R_0^+(1-(\frac{m}{\lambda})^2) \| \leq \frac{1}{10\|V\| C(V)}$. The latter inequality holds because both arguments in the free resolvent reside in a small neighborhood of 1. With the additional restrictions that $\delta < \min(1, \frac{1}{5K \|V\|C(V)})$, it follows that $$\|\mR_0(\lambda+i\gamma) - \mR_0^+(\lambda)\|_{L^2_\sigma \to L^2_{-\sigma}} < \frac{1}{2\|V\|C(V)}.$$ Composing with pointwise multiplication by $V$ completes the estimate for .
When $|\gamma| > \|V\|$, the perturbed resolvent can be estimated by much more elementary means. Here we can use self-adjointness of $D_m$ to bound $\|\mR_0(\lambda + i\gamma)\|_{L^2 \to L^2} \leq \frac{1}{|\gamma|}$, after which it follows that the Neumann series for $[I + V\mR_0(\lambda + i\gamma)]^{-1}$ converges even in the space of bounded operators on $L^2$. One concludes that $\|\mR(\lambda + i\gamma)\|_{L^2 \to L^2} \leq \frac{1}{|\gamma|(1-|\gamma|\, \|V\|)}$.
Within the cones $0 < |\gamma| < \delta |\lambda|$ with $|\lambda| > \lambda_1$, the perturbed resolvent can instead be bounded using and the identity $$\mR(\lambda +i\gamma) = \mR_0(\lambda + i\gamma) - \mR_0(\lambda+i\gamma)V\mR_0(\lambda + i\gamma)
+ \mR_0(\lambda+i\gamma)V\mR(\lambda+i\gamma) V\mR_0(\lambda + i\gamma).$$ The first two terms are bounded operators on $L^2$ with norms less than $\frac{1}{|\gamma|}$ and $\frac{\|V\|}{|\gamma|^2}$ respectively. In the third term we use the decay of $V$ to map $L^2$ to $L^2_\sigma$, or from $L^2_{-\sigma}$ to $L^2$, then the composition is again a bounded operator on $L^2$ with norm no greater than $\frac{C\|V\|^2}{|\gamma|^2}$.
Put together, the perturbed Dirac resolvent $\mR(\lambda + i\gamma)$ exists as a bounded operator on $L^2$ whenever $|\gamma| > \|V\|$ or when $|\lambda| > \max(\lambda_1, \frac{\|V\|}{\delta})$ and $\gamma \not= 0$.
[99]{}
Abramowitz, M. and Stegun, I. A. [*Handbook of mathematical functions with formulas, graphs, and mathematical tables.*]{} National Bureau of Standards Applied Mathematics Series, 55. For sale by the Superintendent of Documents, U.S. Government Printing Office, Washington, D.C. 1964
Agmon, S. [*Spectral properties of Schrödinger operators and scattering theory.*]{} Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4) 2 (1975), no. 2, 151–218.
Arai, M. and Yamada, O. *Essential selfadjointness and invariance of the essential spectrum for Dirac operators,* Publ. Res. Inst. Math. Sci. 18 (1982), no. 3, 973–985.
Balslev, E. and Helffer, B. [*Limiting absorption principle and resonances for the Dirac operator*]{}, Advances in Advanced Mathematics 13 (1992), 186–215.
Bejenaru, I. and Herr, S. *The cubic Dirac equation: small initial data in $H^{1/2}(\R^3)$,* Commun. Math. Phys. 335 (2015), 43–82.
Bejenaru, I. and Herr, S. *The cubic Dirac equation: small initial data in $H^{1/2}(\R^2)$,* Commun. Math. Phys. 343 (2016), 515–562.
Berthier, A. and Georgescu, V.*On the point spectrum of Dirac operators,* J. Funct. Anal. 71 (1987), no. 2, 309–338.
Bouclet, J.-M. and Tzvetkov, N. [*On global Strichartz estimates for non trapping metrics*]{}, J. Funct. Anal. 254 (2008), no. 6, 1661–1682.
Boussaid, N. *Stable directions for small nonlinear Dirac standing waves,* Comm. Math. Phys. 268 (2006), no. 3, 757–817.
Boussaid, N. and Comech, A. *On spectral stability of the nonlinear Dirac equation,* J. Funct. Anal. 271 (2016), 1462–1524.
Boussaid, N. and Comech, A. *Spectral stability of small amplitude solitary waves of the Dirac equation with the Soler-type nonlinearity,* SIAM J. Math. Anal. 49 (2017), 2527–2572.
Boussaid, N., D’Ancona, P., and Fanelli, L. *Virial identiy and weak dispersion for the magnetic Dirac equation,* J. Math. Pures Appl. 95 (2011), 137–150.
Boussaid, N. and Golenia, S. *Limiting absorption principle for some long range perturbations of Dirac systems at threshold energies,* Comm. Math. Phys. 299 (2010), no. 3, 677–708.
Cacciafesta, F. *Virial identity and dispersive estimates for the n-dimensional Dirac equation*, J. Math. Sci. Univ. Tokyo 18 (2011), 1–23.
Carey, A., Gesztesy, F., Kaad, J., Levitina, G., Nichols, R., Potapov, D., and Sukochev, F., *On the Global Limiting Absorption Principle for Massless Dirac Operators*. Ann. Henri Poincaré (2018). https://doi.org/10.1007/s00023-018-0675-5
Christ, M. and Kiselev, A. [*Maximal functions associated with filtrations,*]{} J. Funct. Anal. 179 (2001), 409-425.
Comech, A., Phan, T., and Stefanov, A. *Asymptotic stability of solitary waves in generalized Gross-Neveu model*, Ann. Inst. H. Poincaré Anal. Non Linéaire 34 (2017), 157–196.
D’Ancona, P. and Fanelli, L., [*Strichartz and smoothing estimates for dispersive equations with magnetic potentials*]{}, Comm. Partial Differential Equations 33 (2008), no. 4–6, 1082–1112.
D’Ancona, P. and Fanelli, L. *Decay estimates for the wave and Dirac equations with a magnetic potential*, Comm. Pure Appl. Math. 60 (2007), no. 3, 357–392.
D’Ancona, P., Fanelli, L., Vega, L., and Visciglia, N. [*Endpoint Strichartz estimates for the magnetic Schrödinger equation,*]{} J. Funct. Anal. 258 (2010), no. 10, 3227–3240.
Erdoğan, M. B., Goldberg, M., and Schlag, W. [*Strichartz and smoothing estimates for Schrödinger operators with large magnetic potentials in $\R^3$*]{}, J. Eur. Math. Soc. (JEMS) 10 (2008), no. 2, 507–531.
Erdoğan, M. B., Goldberg, M., and Schlag, W. [*Strichartz and smoothing estimates for Schrödinger operators with almost critical magnetic potentials in three and higher dimensions*]{}, Forum Math. 21 (2009), 687–722.
Erdoğan, M. B. and Green, W. R., [*The Dirac equation in two dimensions: Dispersive estimates and classification of threshold obstructions*]{}, Comm. Math. Phys. 352 (2017), no. 2, 719–757.
Erdoan, M. B., Green, W. R. and Toprak, E. *Dispersive estimates for Dirac operators in dimension three with obstructions at threshold energies,*, Amer. J. Math., to appear. arXiv:1609.05164
Fanelli, L. and Vega, L. *Magnetic virial identities, weak dispersion and Strichartz inequalities,* Math. Ann. 344 (2009), no. 2, 249–278.
Fefferman, C. L. and Weinstein, M. I. [*Wave packets in honeycomb structures and two-dimensional Dirac equations,*]{} Comm. Math. Phys. 326 (2014), no. 1, 251–286.
Georgescu, V. and Mantoiu, M. *On the spectral theory of singular Dirac type Hamiltonians,* J. Operator Theory 46 (2001), no. 2, 289–321.
Georgiev, V., Stefanov, A., and Tarulli, M. [*Smoothing-Strichartz estimates for the Schrödinger equation with small magnetic potential*]{}, Discrete Contin. Dyn. Syst. 17 (2007), no. 4, 771–786.
Ginibre, J. and Velo, G. *Generalized Strichartz inequalities for the wave equation*, J. Funct. Anal. 133 (1995), 50–68.
Goldberg, M. and Schlag, W. [*A Limiting Absorption Principle for the Three-Dimensional Schrödinger Equation with $L^p$ Potentials*]{}, Intl. Math. Res. Not. 2004:75 (2004), 4049–4071.
Hörmander, L. [*[The Analysis of Linear Partial Differential Operators,]{}*]{} Grundlehren der Mathematischen Wissenschaften, Springer–Verlag, Berlin (1985).
Kalf, H. and Yamada, O. [*Essential self-adjointness of n-dimensional Dirac operators with a variable mass term,*]{} J. Math. Phys. 42 (2001), no. 6, 2667–2676.
Keel, M. and Tao, T. [*Endpoint Strichartz estimates*]{}, Amer. J. Math. 120 (1998), no. 5, 955-980.
Kato, T. [*Wave operators and similarity for some non-selfadjoint operators,*]{} Math. Ann. 162 (1965/1966), 258–279.
Machihara, S., Nakamura, M., Nakanishi, K., and Ozawa, T. [*Endpoint Strichartz estimates and global solutions for the nonlinear Dirac equation*]{}, J. Funct. Anal.219 (2005), pp. 1–20.
Marzuola, J., Metcalfe, J., and Tataru, D. [*Strichartz estimates and local smoothing estimates for asymptotically flat Schrödinger equations*]{}, J. Funct. Anal. 255, (2008), no. 6, 1497–1553.
Reed, M. and Simon, B. [*Methods of modern mathematical physics. IV. Analysis of operators.*]{} Academic Press \[Harcourt Brace Jovanovich, Publishers\], New York-London, 1978.
Rodnianski, I. and Schlag, W. [*Time decay for solutions of Schrödinger equations with rough and time-dependent potentials,*]{} Invent. Math. 155 (2004), no. 3, 451–513.
Rodnianski, I. and Tao, T. [*Effective Limiting Absorption Principles, and Applications,*]{} Comm. Math. Phys. (2015) 333: 1. doi:10.1007/s00220-014-2177-8
S. N. Roze, *On the spectrum of the Dirac operator,* Theoret. and Math. Phys. 2 (1970), no. 3, 377–382.
Stefanov, A. [*Strichartz estimates for the magnetic Schrödinger equation*]{}, Adv. Math. 210 (2007), no. 1, 246–303.
Thaller, B. *The Dirac equation.* Texts and Monographs in Physics. Springer-Verlag, Berlin, 1992.
Vogelsang, V. *Absolutely continuous spectrum of Dirac operators for long-range potentials,* J. Funct. Anal. 76 (1988), no. 1, 67–86.
Yamada, O. *A remark on the limiting absorption method for Dirac operators,* Proc. Japan Acad. Ser. A Math. Sci. 69 (1993), no. 7, 243–246.
[^1]: The first author was partially supported by NSF grant DMS-1501041. The second author is supported by Simons Foundation Grant 281057. The third author is supported by Simons Foundation Grant 511825 and acknowledges the support of a Rose-Hulman summer professional development grant
[^2]: Here and throughout the paper, scalar operators such as $-\Delta+m^2-\lambda^2$ are understood as $(-\Delta+m^2-\lambda^2)\mathbbm 1_{\mathbb C^{2^{N}}}$.
[^3]: Constants $C_n$ are allowed to change from line to line.
| {
"pile_set_name": "ArXiv"
} |
---
author:
-
bibliography:
- 'ResDenseNet.bib'
title: |
Lightweight Residual Densely Connected\
Convolutional Neural Network
---
Introduction {#intro}
============
the last decades, Convolutional Neural Networks (CNNs) have changed the landscape of visual recognition tasks such as image classification [@he2016deep; @he2016identity; @hu2018squeeze] and semantic segmentation [@naseer2018indoor; @long2015fully; @fooladgar2019survey]. These models need large training datasets with high-end GPU devices to learn a large number of parameters via a high number of computational operations. However, the most important issues in CNN models are the hardware and the burden of computational cost. Yet, complex CNNs [@hu2018squeeze] with high resource demands have been proposed to increase the accuracy. In addition, the ultra-deep CNN models have further increased the depth of networks from 8 layers (AlexNet) [@krizhevsky2012imagenet] to more than one thousand layers (ResNet) [@he2016deep]. In fact, the general idea of CNNs has proceeded through deeper and more complex networks to boost the performance in terms of model’s accuracy. But, the efficiency of networks in terms of model’s size, inference speedup, and computational costs have been rarely inspected. For mobile and embedded devices (such as robotics, drones, and smartphones) the model size, run-time, memory requirement, inference time, and computational cost are very important. As such, a serious question which arises is whether it is really necessary to have these complex models with a huge number of parameters and computational operations.
The amount of redundancy among the parameterization of CNNs facilitates some novel pruning, quantization, and factorization methods to reduce the model size. Denil et al. [@denil2013predicting] demonstrated this significant redundancy in the weights of neural networks. The efficiency of networks in the viewpoint of special purpose applications (like robotics, augmented realities, and self-driving cars) established some novel model structures with impressive architecture designs. ResNet [@he2016deep; @he2016identity] and DenseNet [@huang2017densely] proposed two novel architectures which reduced the computational cost by a factor of 5$\times$ and 10$\times$ alongside boosting the accuracy when compared to the VGG [@simonyan2014very] on Imagenet dataset[@krizhevsky2012imagenet]. Meanwhile, the MobileNet [@howard2017mobilenets], ShuffleNet [@zhang2018shufflenet], and CondenseNet [@huang2018condensenet] reduced the computational cost approximately by a factor of 25$\times$, 25$\times$, and 20$\times$, respectively, while obtaining a comparable accuracy to the VGG on ImageNet.
The general notion of shortcut connection and residual connection have been proposed by the “highway networks” [@srivastava2015highway; @srivastava2015training] and ResNet model [@he2016deep; @he2016identity], respectively. Thereafter, DenseNet [@huang2017densely] presented the idea of densely connected layers where each layer obtains concatenated feature maps of preceding layers. Subsequently, these two main ideas have been applied on different applications and network architectures. For instance, the skip connections have been exploited on encode-decoder models for dense prediction tasks (such as semantic segmentation and depth estimation) [@fooladgar20193m2rnet; @eigen2015predicting]. The local and global residual connections have been utilized for image super-resolution and restoration via exploiting the hierarchical features from all preceding convolutional layers [@zhang2018residual; @zhang2018residualrestore]. These different models and applications have a share characteristic which create short paths from early layers to later ones via these two main concepts while they differ in network topology, training procedure and challenges.
In this paper, the idea of feature reuse ability of the DenseNet [@huang2017densely] with the residual connection idea of the ResNet [@he2016deep] are exploited to achieve an ultra slimmed deep network. The proposed architecture, called Efficient Residual Densely Connected CNN (RDenseCNN), is illustrated in Figure \[fig\_1\]. In the proposed method, an extremely small-size CNN model is presented to reduce the computational cost, memory requirement, and inference time without utilizing any new operation or hardware-software equipment. The performance of the proposed model is extensivly evaluated on the ImageNet [@krizhevsky2012imagenet], MNIST [@deng2012mnist], Fashion MNIST [@xiao2017/online], SVHN [@netzer2011reading], CIFAR-10, and CIFAR-100 [@krizhevsky2009learning] datasets. The obtained results show that its performance is comparable with most available networks in terms of the accuracy, number of parameters, and Floating Point Operations (FLOPs). Our proposed method has obtained superior performance on MNIST, Fashion MNIST and CIFAR datasets. Moreover, the model has achieved a higher accuracy with 46$\times$ smaller model size when compared to the AlexNet with $\%4.8$ higher in accuracy on ImageNet. It has also attained $2\times$ lower number of parameters at only $\%3$ drops in top-1 accuracy compared to the MobileNet [@howard2017mobilenets].
{width="1\linewidth"}
Main contributions of this paper are as follows:
- Presenting an extremely small CNN model with feature reuse abilities, deep supervision, and efficient gradient flows by residual densely connected blocks.
- Reducing the computational cost and memory requirements with a feasible accuracy.
- Presenting a CNN model with basic common operations without requiring additional software or hardware equipment.
- Analyzing CNN models in terms of accuracy versus the number of parameters.
- Representing the effect of residual connections alongside the densely connected blocks (this can be an important result for those who want to combine these two ideas).
This paper is organized as follows. In Section \[sec:2\], the related work is explained in four categories. The proposed efficient residual densely connected CNN is introduced in Section \[sec:3\]. Experimental results on six most import image classification datasets are reported in Section \[sec:4\]. In Section \[sec:5\] and \[sec:6\], concluding remarks are discussed.
Related Work {#sec:2}
============
The aim of designing efficient CNN architectures is to develop these models for limited resource devices. There are three main limitations for running CNN models on embedded devices; namely, the model size, the run-time memory requirement, and the required number of operations. Some methods have been proposed to overcome these constraints by compressing the model size in different ways. These can be classified into four main categories of: (i) quantizing and factorizing, (ii) pruning redundant connections, (iii) learning efficient architectures, and (iv) designing efficient architectures. These are explained next.\
**Quantizing and factorizing:** To compress large CNN models, Denton et al. in [@denton2014exploiting] proposed to approximate the fully connected weights of CNNs by singular value decomposition methods. They obtained approximately 3$\times$ compression in the model size. Weight quantization is another method to compress the model size. HashNet [@chen2015compressing], Binarized Neural Networks [@courbariaux2016binarized], and Xnor-net [@rastegari2016xnor] (binary or ternary weight quantizations) can also compress the model size, save run-time memory, and speed up the inference time, but to the cost of reduced accuracy.\
**Pruning and sparsifying:** These methods can be performed at different granularity levels (weight, kernel, channel, and layer levels). In [@han2015learning], unimportant connections with small weights were pruned to compress the model size and to save the memory. The authors of [@li2016pruning] and [@zhou2016less] proposed pruning at kernel and neuron levels, respectively. Wen et al. [@wen2016learning] proposed to apply the sparsity at different levels of CNNs’ structure (kernels, channels, and layers). All of these methods decreased the run-time memory and time elapse at a moderate accuracy loss. But, in some cases, they need special libraries to apply the pruning and sparsifying modules.
From the fine-to-coarse grains of pruning level, flexibility, generality, and compression ratio decrease. Meanwhile, special hardware/software requirements are needed for almost fine grain levels.\
**Designing efficient architectures:** Recent CNN structures (such as the ResNet [@he2016deep], DenseNet [@huang2017densely], and Xception [@chollet2017xception]) have more efficient models than early CNNs (like AlexNet [@krizhevsky2012imagenet] and VGG [@simonyan2014very]). In addition, specific purpose CNN models [@huang2018condensenet; @iandola2016squeezenet; @howard2017mobilenets; @zhang2018shufflenet] have been designed to decrease the computational cost and memory requirement with the speedup in inference time. To reduce the model size (number of parameters) and computational cost (number of Add-Multiplication operations), these CNN models utilized the idea of group convolution and depth-wise separable convolution. The extreme version of inception modules (called Xception) was proposed by Chollet [@chollet2017xception] to boost the performance of the Inception V3 by substituting Inception modules with depth-wise separable convolutions. The MobileNet [@howard2017mobilenets] proposed an efficient small architecture that applied the depth-wise convolution followed by the point-wise convolution as the depth-wise separable convolution to reduce the computational cost and model size. Howard et al. [@howard2017mobilenets] proposed two hyper parameters as the width and resolution multipliers that controlled the input width of a layer and the input image resolution, respectively. The Channel shuffle was proposed to generalize the cascaded group convolutions. In other words, the ShuffleNet [@zhang2018shufflenet] reformed the group and depth-wise separable convolutions by shuffling the outputs of the point-wise group convolution fed to the depth-wise separable convolutions. The authors of CondenseNet [@huang2018condensenet] proposed to learn these groups at the training phase by a novel module called learned group convolution. At the half of training iterations, their network eliminated filters with small magnitude weights; hence the structure of convolution layers sparsified. Consequently, the kernel pruning performed as the condensation procedure by the condensation factor. In the second half of the training phase, fixed filters were trained. Then, at each layer, the index layer was proposed to select and rearrange the orders of input feature maps.\
**Learning efficient architectures:** Some new methods proposed to learn the structure of networks, automatically. Baker et al. [@baker2016designing] proposed to learn the structure of CNNs by reinforcement learning. Network slimming [@liu2017learning] was another approach which tried to learn a scale factor for each layer to eliminate channels with small scaling factors. The authors of [@liu2017learning] considered the loss function with the L1-norm penalty on the scaling factors to apply sparsity. These approaches succeeded in dealing with all mentioned limitations, but they need iterative training procedures to train the slimmed network.
Proposed Efficient Residual Densely Connected CNN {#sec:3}
=================================================
Novel CNN models are constructed as sequentially cascaded layers or basic bocks. Each layer $l$ or basic block $k$ can be denoted as a function $F_l(.)$ or $H_k(.)$, respectively. The output of the $l^{th}$ layer or the $k^{th}$ block is determined as $x_l$ or $y_k$, accordingly. The $F_l(.)$ function can be made from the compositions of some linear or non-linear functions; such as Convolution(Conv), Batch Normalization (BN), Rectified Linear Units (Relu), or Pooling (max or average pooling). The $H_k(.)$ function can be formed by the combinations of functions that each of them belongs to each layer of the $k^{th}$ block.
The proposed Residual Dense CNN is built upon the idea of residual and densely connected blocks. The common intention of DenseNet and ResNet is to reuse feature maps of preceding layers in the upper layers. Therefore, these two ideas are first elucidated. Then, the building block is clarified in detail.\
**Residual Blocks.** Common CNN models have a plane structure in which the input of each intermediate layer comes directly from the output of the previous layer as $x_l = F(x_{l-1})$. But, in the residual block, the input goes through two different functions and the final output is the summation of their outputs, given by $$\label{eq:1}
x_l = F(x_{l-1}) + G(x_{l-1}).$$
![Three building blocks: (a) dense block, (b) operations in each layer of dense block, (c) operations in transition block.[]{data-label="fig:denseblock"}](denseblock.png){width="\linewidth"}
![Three building blocks: (a) dense block, (b) operations in each layer of dense block, (c) operations in transition block.[]{data-label="fig:denseblock"}](layer.png){width="\linewidth"}
![Three building blocks: (a) dense block, (b) operations in each layer of dense block, (c) operations in transition block.[]{data-label="fig:denseblock"}](transition.png){width="\linewidth"}
The ResNet [@he2016deep] proposed that the function $G$ is an identity mapping $x_l = F(x_{l-1}) + x_{l-1}$. Therefore, it bypasses the non-linear transformations to resolve the vanishing gradient problem.\
**Densely Connected Blocks.** In feed-forward networks, the output feature maps of each layer can be fed to all subsequent layers (as their input) $$\label{eq:2}
x_l = H([x_{0},x_{1},...,x_{l-1}]).$$ Then, all outputs of $l$ preceding layers construct an input of the $l^{th}$ layer, where $[.]$ denotes the concatenation of all previous feature maps as proposed by Huang et al. [@huang2017densely]. The equal size of feature maps is an essential factor for the concatenation operation. The DenseNet [@huang2017densely] proposed densely connected blocks where the size of feature maps was preserved within each dense block. Therefore, the idea of densely connected CNNs can be applied locally within the range of layers in each block.
[|c|c|c|c|c|c|]{} Layers& Output size & RDenseCNN-k-132&RDenseCNN-k-160&RDenseCNN-k-200\
Convolution & 128 $\times$ 128&\
Avg. Pooling & 64 $\times$ 64&\
Dense Block-1 & 64 $\times$ 64& $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3\ Conv. \end{bmatrix} \times 16$ & $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3\ Conv. \end{bmatrix} \times 20$ & $ \begin{bmatrix} 1 \times 1\ Conv.\\ 3 \times 3\ Conv. \end{bmatrix} \times 24$\
Trans Block-1 & 32 $\times$ 32&\
Dense Block-2 & 32 $\times$ 32& $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3\ Conv. \end{bmatrix} \times 16$ & $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3\ Conv. \end{bmatrix} \times 20$ & $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3\ Conv. \end{bmatrix} \times 24$\
Trans Block-2 & 16 $\times$ 16&\
Dense Block-3 & 16 $\times$ 16& $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3 \ Conv. \end{bmatrix} \times 16$ & $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3\ Conv. \end{bmatrix} \times 20$ & $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3 \ Conv. \end{bmatrix} \times 24$\
Trans Block-3 & 8 $\times$ 8&\
Dense Block-4 & 8 $\times$ 8& $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3\ Conv. \end{bmatrix} \times 16$ & $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3\ Conv. \end{bmatrix} \times 20$ & $ \begin{bmatrix} 1 \times 1 \ Conv. \\ 3 \times 3\ Conv. \end{bmatrix} \times 24$\
Classification &1 $\times$ 1&\
Details of Proposed Residual Densely Connected Model
----------------------------------------------------
The proposed method consists of the dense and transition blocks with one skip connection (see Figure \[fig\_1\]). These combinations of non-linear transformations, performed in two consequent blocks, are bypassed by a skip connection. Suppose $x_l$ is the input to the residual block, then the output of this residual block can be defined as $$\label{eq:3}
x_o = H(x_l) + F(x_l)$$ where $F(x_l)$ is utilized as a skip connection to ensure an unimpeded information flow through the network. The $H(x_l)$ function is defined as $$\label{eq:4}
H(x_l): H_T(H_D([x_l,x_{l+1},...,x_{l+m}]))$$ where $x_{o-1} = H_D([x_l,x_{l+1},...,x_{l+m}]$ denotes the non-linear transformations performed on outputs of $m$ sequential densely connected layers of the dense block and $x_{o-1}$ is an output of this dense block. The $H_T$ indicates the non-linear functions performed in the transition block.\
**Composite function $H_D(.)$**. This function is a sequential combination of the BN, Relu, $1\times1$ Conv , BN, Relu, and $3\times3$ Conv defined as one layer of the dense block, as depicted in Figure \[fig:denseblock\]. The overall structure of a dense block with 4 layers is illustrated in Figure \[fig:denseblock\].(a), where operations of each layer are given in Figure \[fig:denseblock\].(b).\
**Composite function $H_T(.)$**. This function is a sequential combination of the BN, Relu, $1\times1$ Conv , and $2\times2$ average pooling performed in the transition block depicted in Figure \[fig:denseblock\].(c). Hence, the spatial size of the output diminishes by a factor of 2 in each transition. Therefore, the element-wise summation of Equation (\[eq:3\]) imposes a further down-sampling operation in the skip connection. To resolve the inconsistency of the size of the feature maps, the $F$ in Equation (\[eq:3\]) is considered to be an average pooling function instead of an identity map. However, $F$ is a non-linear function and does not lead to loss of information about the original state of the image. Therefore, the input to the next block contains the down-sampled version of the original data; where it has been altered by the non-linearity function denoted by $H$.\
**Growth rate.** Motivated by [@huang2017densely], each layer of the dense block produces $k$ feature maps. Each dense block has $m$ layers. Therefore, the output of the dense block has $(m-1)\times k + k_0$ feature maps, where $k_0$ is the number of input feature maps. The growth rate determines the amounts of new feature maps that each $3 \times 3$ convolution layer can exploit. Since the number of layers in each dense block is large, it is possible to limit the growth rate. Consequently, the width of the proposed method can be managed and controlled.
Architecture Design
-------------------
The proposed architecture is trained by a single end-to-end training procedure in contrast to some methods that need to be trained in two separate sequential training phases [@huang2018condensenet]. All operations of the RDenseCNN model are the basic CNN operations which are performed without requiring any special software or hardware to accelerate them. It has three architecture parameters of: (i) number of dense blocks, (ii) number of layers within each dense block, and (iii) growth rate. The network consists of only three or four dense blocks where each of them has 16, 20, or 24 layers. The growth rate ($k$) can be 12 or 16. The initial convolution layer has $4k$ convolution kernels of size $3 \times 3$ with stride and padding of 1. Hence, the number of convolutional kernels in the $1 \times 1$ convolution layer of the transition block is set to $4k$ to apply the element-wise summation of Equation (\[eq:3\]). In Table \[tab:1\], the more details of three different configurations of RDenseCNN are illustrated.
Experimental Results {#sec:4}
====================
In this section, the effectiveness of the proposed RDenseCNN is evaluated in terms of computational cost, model size, and implementation requirements. It has been evaluated on six main datasets MNIST, Fashion-MNIST, SVHN, CIFAR-10, CIFAR-100, and ImageNet. The proposed model has been trained with different depths and growth rates. To compare the computational efficiency of the method, it has been compared with specific and general-purpose CNN models. It is implemented by using the PyTorch library. The code will be made publicly available on github.
Datasets
--------
The following six important datasets, which are commonly used in the image classification have been employed to evaluate the performance of the method.\
**MNIST.** This database of handwritten digits [@deng2012mnist] contains 60,000 and 10,000 examples as the training and test sets, respectively. The digits have been size-normalized and centered in $28\times28$ images.\
**Fashion MNIST (fMNIST).** It can be seen as similar in flavor to MNIST (e.g., the image size and structure of training and testing splits) [@xiao2017/online]. Each example is a $28\times28$ grayscale image corresponded to one of the 10 class labels.\
**Street View House Numbers (SVHN).** It consists of over 600,000 images obtained from house numbers in Google Street View images [@netzer2011reading]. The digit images are significantly more difficult to process compared to the MNIST examples as they are captured from real-world natural scenes.\
**CIFAR.** It contains 60,000 colored low resolution $32\times32$ images with 10 and 100 classes, determined as CIFAR-10 and CIFAR-100 [@krizhevsky2009learning], accordingly. They contain 50,000 and 10,000 images for training and testing, respectively.\
**ImageNet.** This is the pioneer large scale dataset for image classification [@krizhevsky2012imagenet]. It contains 1.2 million training and 50,000 validation images classified into 1000 classes.
Training Procedures
-------------------
The proposed model was trained from scratch using the Stochastic Gradient Descent (SGD) optimization process. Mini-batch sizes were selected depending on the datasets. The mini-batch size was set to 64 and 30 for ImageNet dataset. It was set to 128 and 256 for five other datasets. The proposed model was trained during 120 and 300 epochs for the ImageNet and five other datasets, respectively. The training process was initialized with the learning rate of 0.1 and was decreased every 30 epochs by a factor of 10. The weight decay was set to $10^{-4}$. Data augmentation methods for training images were utilized as in [@huang2017densely; @he2016deep].
Classification Results on MNSIT and Fashion-MNIST
-------------------------------------------------
One of the well-known benchmark datasets for machine learning algorithms is MNIST which most methods validate their algorithms based on it. The state-of-the-art accuracy in this dataset has been approached to %99.75. Hence, Xiao et al. [@xiao2017/online] introduced Fashion-MNIST as a replacement of MNIST dataset in 2017. It includes 10 classes of “T-shirt, Trouser, Pullover, Dresser, Coat, Sandal, Shirt, Sneaker, Bag, and Ankle boot”. Our network’s inputs for both MNIST and Fashion-MNIST datasets are 28$\times$28 grayscale images. The structure of RDenseCNN model is depicted in Table \[cifar\_arch\]. It contains three dense blocks alongside two transition blocks. The experiments were conducted on the smallest proposed model with $k=12$ and $d=100$.
[|c|c|c|c|c|c|]{} Layers& Output size & RDenseCNN-k-100&RDenseCNN-k-121&RDenseCNN-k-152\
Convolution & 32 $\times$ 32&\
Avg. Pooling & 16 $\times$ 16&\
Dense Block-1 & 16 $\times$ 16& $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3\ Conv. \end{bmatrix} \times 16$ & $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3\ Conv. \end{bmatrix} \times 20$ & $ \begin{bmatrix} 1 \times 1\ Conv.\\ 3 \times 3\ Conv. \end{bmatrix} \times 24$\
Trans Block-1 & 8 $\times$ 8&\
Dense Block-2 & 8 $\times$ 8& $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3\ Conv. \end{bmatrix} \times 16$ & $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3\ Conv. \end{bmatrix} \times 20$ & $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3\ Conv. \end{bmatrix} \times 24$\
Trans Block-2 & 4 $\times$ 4&\
Dense Block-3 & 4 $\times$ 4& $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3 \ Conv. \end{bmatrix} \times 16$ & $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3\ Conv. \end{bmatrix} \times 20$ & $ \begin{bmatrix} 1 \times 1\ Conv. \\ 3 \times 3 \ Conv. \end{bmatrix} \times 24$\
Classification &1 $\times$ 1&\
The performance of proposed method was compared with the most new CNN models which have validated their algorithm on MNIST and Fashion-MNIST datasets. The results are listed in Table \[mnist\]. All available approaches that have published their results on these two datasets have not reported the number of parameters and floating-point operations. Therefore, it was not possible to compare the methods in terms of model size and computational complexity. But in terms of accuracy, the proposed method achieves the state-of-the-are results in Fashion-MNIST. Our proposed method attains %99.3 accuracy on Fashion-MNIST where it is %5.7 and %4 better accuracy than two recent models of CapsNet [@sabour2017dynamic] and DENSER [@assunccao2019denser], respectively.
Model FLOPS Parmas MNIST fMNIST
----------------------------------- ------- -------- ------- --------
WRN-28-10 [@zhong2017random] - 36.5M - 4.1
WRN-28-10 + RE [@zhong2017random] - 36.5M - 3.7
CapsNet [@sabour2017dynamic] - - 0.25 6.4
DENSER [@assunccao2019denser] - - 0.3 4.7
RDenseCNN-12-100 52.4M 0.61M 0.5 0.7
: Comparison of classification error (%) with state-of-the-art CNN models on MNIST and Fashion-MNIST(fMNIST) datasets.[]{data-label="mnist"}
Classification Results on CIFAR and SVHN
----------------------------------------
The network inputs for both CIFAR and SVHN datasets were 32$\times$32 images, with the per-pixel mean subtracted. The overall architectures considered for these two datasets were as those for the MNIST dataset which are depicted in Table \[cifar\_arch\].
The experiments were performed on two different values of growth rate (k=12 and k=16) with three different network depths. They were compared with the most important special-purpose CNN models (such as the pruned version of CNNs and CondenseNet). As the results listed in Table \[cifar10\] shows, the proposed method has surpassed almost all of pruned CNN models. For instance, it obtained a lower classification error with 7$\times$ less number of parameters and approximately 2.3$\times$ fewer FLOPs than the VGG-pruned model [@li2016pruning], more especially in CIFAR-10. The RDenseCNN did not outperform the CondenseNet and pruned DenseNet-40 in terms of model size. But, note that it has lower computational cost with only %0.7 drops in the classification error. It is notable that the most recent efficient CNN models (like MobileNet and ShuffleNet) have not reported their performance on these two datasets.
In Table \[cifars\], the classification error and the number of parameters are compared with the state-of-the-art CNN models. The RDenseCNN-152 has outperformed the VGGNet [@simonyan2014very] and ResNet [@he2016deep] models. ResNet and DenseNet models with a high number of parameters (more than 10M) have obtained a lower classification error in both CIFAR datasets. The proposed model has a comparable performance with DenseNet and ResNet models in terms of error and the number of parameters. CliqueNet [@yang2018convolutional] attained state-of-the-art classification error on SVHN but with 9.45G computational complexity while our proposed method has 121.4M FLOPS, approximately 80$\times$ lower computational complexity than CliqueNet.
Model FLOPs Parmas C-10 C-100
---------------------------------------- -------- -------- ------ -------
VGG-16-pruned [@li2016pruning] 206M 5.40M 6.60 25.3
VGG-19-pruned [@liu2017learning] 250M 5.0M - 26.5
ResNet-56-pruned [@li2016pruning] 90M 0.73M 6.94 -
ResNet-110-pruned [@li2016pruning] 213M 1.68M 6.45 -
ResNet-164-B-pruned [@liu2017learning] 124M 1.2M - 23.9
DenseNet-40-pruned [@liu2017learning] 190M 0.66M 5.19 -
CondenseNet-94 [@huang2018condensenet] 122M 0.33M 5.00 24.1
RDenseCNN-12-100 69.1M 0.61M 6.8 30.3
RDenseCNN-12-121 88.2M 0.78M 6.3 29.2
RDenseCNN-12-152 121.4M 1.1M 5.7 25.9
: Comparison of classification error (%) with efficient state-of-the-art CNNs on CIFAR-10(C-10) and CIFAR-100(C-100) datasets.[]{data-label="cifar10"}
Model FLOPS Parmas C-10 C-100 SVHN
----------------------------------------------- -------- ----------- --------- ---------- ------
CapsNet [@sabour2017dynamic] - - 10.6 - 4.3
SkipNet-ResNet110 [@wang2018skipnet] - - 6.4 28.8 1.9
CMPE-SE-WRN-28-10 [@hu2018competitive] - 36.9M 3.6 18.5 1.59
Wide ResNet-28-10 [@zagoruyko2016wide] - 36.5M 4.0 19.3 -
DenseNet-24-100 [@huang2017densely] - 27.2M 3.7 19.3 1.59
VGGNet[@simonyan2014very] - 20.0M$^+$ 6.3$^+$ 26.7$^+$ -
ResNet-1001 [@he2016identity] - 10.2M 4.6 22.7 -
CliqueNet [@yang2018convolutional] 9.45G 10.14M 5.06 23.14 1.51
ResNet reported by [@huang2016deep] - 1.70M 6.4 27.2 2.0
ResNet-164 [@he2016identity] - 1.70M 5.5 24.3 -
ResNet with Stochastic Depth [@huang2016deep] - 1.70M 5.2 24.6 1.8
RDenseCNN-12-152 121.4M 1.10M 5.7 25.9 2.5
Classification Results on ImageNet
----------------------------------
The proposed method was evaluated on the ImageNet dataset. Figure \[fig:error\] illustrates the behavior of RDenseCNN with different depths and growth rates. It shows top-1 and top-5 validation errors for two different RDenseCNN architectures during 120 epochs. The proposed model achieves lower top-1 and top-5 classification errors when the depth increases. The initial learning rate was set to 0.1 and decreased every 30 epoch.
![Validation error for two RDenseCNN models.[]{data-label="fig:error"}](Error.png){width="0.6\linewidth"}
The comparison results with the state-of-the-art CNN models as well as efficient CNN architectures (designed for specific embedded devices) are shown in Figures \[fig:accvcparam\] and \[fig:accvcparam\_efficient\]. These two charts have been sorted based on top-1 accuracy. It provides a good intuition about the ratio of accuracy versus the model size. The overall architecture of the proposed method is depicted in Table \[tab:1\]. The results presented in Figure \[fig:accvcparam\] reveal that:
- RDenseCNN outperforms AlexNet in terms of both error rate and the number of parameters. It has $26\times$ lower number of parameters with $8$ percent of higher accuracy.
- RDenseCNN has attained reasonable error rate by fewer number of parameters when compared with Inception V1, GoogleNet, and ResNet-18. It has approximately a $3\times$ smaller number of parameters at $\%3.4$ drops in accuracy.
- The best model is Wide-ResNet-50 [@zagoruyko2016wide] which yields a $\%13$ higher accuracy with a $66.6$M additional number of parameters.
![Ratio of top-1 accuracy vs. number of parameters (models are evaluated on ImageNet dataset).[]{data-label="fig:accvcparam"}](accvcparam.png){width="1\linewidth"}
![Ratio of top-1 accuracy vs. number of parameters (efficient models are evaluated on ImageNet dataset).[]{data-label="fig:accvcparam_efficient"}](accvcparam_efficient.png){width="1\linewidth"}
To show more details about these results, Figure \[fig:accvcparam\_efficient\] considers only more efficient models. The proposed RDenseCNN is more accurate than the SqueezNet [@iandola2016squeezenet] even with fewer parameters. The method is also more efficient than some configurations of MobileNet (e.g., 0.25 MobileNet and 0.5 MobileNet). The MobileNet has width and resolution multipliers to design a smaller and faster model. The proposed model has been compared with different configurations of the MobileNet . The results are listed in Table \[tab:mobilenetvsour\]. These different configurations provide a trade-off between the classification error and the model size. At the same level of model size, RDenseCNN has achieved a very competitive classification error.
Model Params Top-1 Error
-------------------- -------- ------------- --
1.0 MobileNet-224 4.2M 29.4
1.0 MobileNet-128 4.2M 35.6
0.75 MobileNet-224 2.6M 31.6
0.5 MobileNet-224 1.3M 36.3
0.5 MobileNet-160 1.3M 39.8
0.25 MobileNet-224 0.5M 49.4
RDenseCNN-12-132 0.8M 41.5
RDenseCNN-16-132 1.3M 38
RDenseCNN-16-200 2.3M 34.7
: Comparison with different configurations of MobileNet [@howard2017mobilenets] on ImageNet dataset.[]{data-label="tab:mobilenetvsour"}
Discussion {#sec:5}
==========
The proposed method evolved based on two well-known CNN models. The main goal of our architecture is to combine two ideas of these two models to achieve a more lightweight and efficient model. In other words, the proposed model is not an extension of DenseNet or ResNet model, it is a novel architecture that utilizes the residual and dense connection ideas with the light CNN structure. Our proposed model has less number of parameters and computational complexity at the level of feasible and comparable accuracy via the residual and dense connection ideas.
The Effect of Residual Connections
----------------------------------
In this section, the importance of skip connection in each residual dense block is investigated. The proposed architecture without the skip connection is considered as “Plane-DenseCNN” (PDenseCNN). In this plane architecture, every dense block is connected in a sequentially cascaded manner. It is notable that this PDenseCNN is our lightweight model without residual connections and it is not a version of the well-known DenseNet model. To demonstrate the effectiveness of skip connections in the proposal residual dense block, two models have been evaluated on three datasets. Figure \[fig:plane\_res\] compares the PDenseCNN and RDenseCNN models in terms of top-1 accuracy during training in both train and validation splits of datasets. The results presented in this figure reveal that the RDenseCNN performs more accurate than the PDenseCNN. It attains approximately a %20 higher top1-accuracy whilst requiring the same number of parameters. This is because the residual connections do not enforce any additional number of parameters in our model. These experiments were performed based on the PDenseCNN-12-100 and RDenseCNN-12-100 models with 0.6M parameters.
{width="\linewidth"}
{width="\linewidth"}
{width="\linewidth"}
Model Analysis
--------------
The efficient flow of gradients, feature reuse ability, and deep supervision in short and long ranges are three main properties of the proposed model. In each layer of the dense block, a small set of feature maps (based on the growth rate) is selected as the nonlinear composition of input feature maps. Then, these feature maps are concatenated with the previous ones and passed to the proceeding layer to produce the new nonlinear composition. In other words, the feature maps of early layers of each dense block can be reused by the last ones in the same dense block. This feature reuse ability preserves the information flows through two consecutive dense and corresponding transition blocks. In other direction, the skip connection utilized in each residual dense block amend the main information flows of the network. Meanwhile, the deep supervision was applied in the network by both dense connections in the residual block (short ranges) and also the skip connection (long ranges). In the densely connected model, each layer receives additional supervision from the loss function via a shorter connection. Hence, feature maps of intermediate layers learn more discriminative features.
The accuracy versus the number of parameters can be more informative for selecting an appropriate model based on requirements of specific applications to analyze the efficiency of CNN models. Therefore, the ratio of top-1 accuracy versus the number of parameters was computed for state-of-the-art CNN models. These are depicted in Figure \[fig:accvcparam\]. As shown in this figure, the RDenseCNN achieves the admissible rank among different state-of-the-art models based on these two evaluation metrics. The VGG-16, Wide ResNet and AlexNet have a significantly large number of parameters.
Conclusion {#sec:6}
==========
One of the main goals of this work was to design an ultra-small model to analyze the amount of complexity of novel CNN architectures which have attained the state-of-the-art performance in terms of classification error. Towards this goal, a small CNN model was proposed based on two main ideas of residual connections and densely connected convolutional layers. This proposed RDenseCNN architecture had a few numbers of parameters with a feasible classification error. It contained densely connected layers in terms of some dense blocks with skip connections to preserve the flow of information in a deep CNN model. The proposed model had 26$\times$ fewer number of parameters than the first widespread CNN model (AlexNet) with a %8 higher classification error. It is worth mentioning that the smallest proposed model had the same level of accuracy (%1 better accuracy) with the AlexNet at 60$\times$ smaller model size. The results revealed that some levels of complexity in CNN models have unacceptable justifications, seriously in early CNN models. Consequently, these unfeasible complex models can be substituted by a significantly smaller and more efficient model, particularly in limited resource applications.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
When animation of a humanoid figure is to be generated at run-time, instead of by replaying pre-composed motion clips, some method is required of specifying the avatar’s movements in a form from which the required motion data can be automatically generated. This form must be of a more abstract nature than raw motion data: ideally, it should be independent of the particular avatar’s proportions, and both writable by hand and suitable for automatic generation from higher-level descriptions of the required actions.
We describe here the development and implementation of such a scripting language for the particular area of sign languages of the deaf, called SiGML (Signing Gesture Markup Language), based on the existing HamNoSys notation for sign languages.
We conclude by suggesting how this work may be extended to more general animation for interactive virtual reality applications.
author:
- |
Richard Kennaway\
School of Computing Sciences,\
University of East Anglia, Norwich NR4 7TJ, U.K.
bibliography:
- 'animgen.bib'
date: 2 November 2006
title: 'Avatar-independent scripting for real-time gesture animation'
---
Introduction
============
In interactive applications of virtual reality, it is necessary to create animations of avatars and other objects at run-time, according to the demands of the moment. Usually, this is done by replaying motion clips selected from a library of pre-composed motions, created by either manual keyframing or motion capture. This method suffers from several disadvantages.
1. It may require a very large motion library to anticipate every possible action that may be demanded.
2. The same movement will be played in exactly the same way every time it is used, reducing the realism of the scene.
3. The motion data are specific to one avatar.
It is possible to warp motion clips, for example, to adapt a motion for one avatar to another avatar, to have a walk cycle for a character follow a specified path through the world, or to smoothly blend together the end of one clip with the start of another. However, this only gives a limited amount of extra flexibility. Inspection of some current video games indicates that at present little run-time warping or blending is done.
Run-time generation of the motion data from some higher-level description would allow for animation to be generated in a much more flexible way, in order to exactly fit the interactive situation at each moment. In this paper we will describe an avatar-independent scripting language for posture and movement in the particular domain of sign languages of the deaf, and how we use it to automatically generate animation data in real time for any avatar. A secondary advantage of the method is that the script for a movement is many times smaller than the motion data generated from it, which greatly reduces storage requirements.
The software [@kennaway:synthetic] was developed in the ViSiCAST [@Ell00; @VerTF:sign:01] and eSign \[[[[@eSignCVHI]]{}]{}; <http://www.sign-lang.uni-hamburg.de/eSIGN>\] projects, as one component of a system which provides signing animations for websites or other interactive multimedia applications. (The software was not deployed until the eSign project; ViSiCAST still relied on motion-captured animations.) A web page using this sytem will include an avatar on the page (implemented as an ActiveX control), and beside each block of text for which signing is available, a button to request that the signing be played. The web page also contains data to drive the avatar when one of these buttons is clicked, not in the form of motion data, but as text in the avatar-independent scripting language. This is translated in real time into frames of motion data (bone rotations and facial morphs). The data frames are passed to the avatar and played at a fixed frame rate. A fuller account of the whole system can be found in [@eSignCVHI]. A publicly available web site employing the technology is at <http://www.gebarennet.nl>[^1].
The scripting language and its translation to motion data are the subject of the present paper. Our concern is with the concrete geometry of the animation, in contrast to the large number of proposals for scripting languages for embodied conversational agents, which focus on much higher-level concepts. Some of these are briefly surveyed in Section \[section:otherwork\].
Related work {#section:otherwork}
============
There have been numerous proposals for scripting languages for embodied conversational agents, including VHML \[[[[@MarS02]]{}]{}; <http://www.vhml.org>\], HumanML \[<http://www.humanmarkup.org>\], EMOTE [@chi:emote], GESTYLE [@NooR03], CML [@Ara-etal02], AML [@Ara-etal02; @PelB03], APML [@CarCBP02], MURML [@KraKW02], MPML [@PreDI04], and STEP [@Huang:STEP:02; @Huang:implementation]. A survey of some of these languages and several others can be found in [@Pel03], and [@Pre:lifelike:04] is a book-length account of various implementations of “social agents”. These languages mainly address a higher level of abstraction than we are concerned with here, typical elements or attributes being *angry*, *alert*, *think*, *emphasize*, and *semiote*. The main purpose of most of these languages is to specify the behaviour of such agents, and especially the emotionally expressive facial and bodily movements that they should make to accompany realistic speech and interact in an engaging way with the user. The specification of the concrete movements which perform these high-level concepts is typically seen as a mere implementation detail. Among the proposals that include explicit description of concrete movement, emphasis has been primarily on facial animation, with body animation either ignored (for talking head applications), undeveloped, as in the most recent (2001) VHML specification, or implemented directly in terms of avatar-dependent joint rotations.
The VHD system [@SanBTT99] for directing real-time virtual actors is based on motion capture and predefined postures, and the VHD++ system [@PonPMTT03] is a high-level virtual reality engine for integrating lower-level technologies.
Dance notations such as Labanotation [@hutchinson:labanotation] have been animated by software such as Don Henderson’s LINTER \[<http://www-staff.mcs.uts.edu.au/~don/pubs/led.html>\], although that system does not attempt naturalistic animation of detailed avatars. Labanotation and other dance notations present particular challenges to computer animation, as transcriptions written to be read by human dancers rely for their interpretation not only on the dancer’s knowledge of dancing, but upon their creative input to flesh out intentionally incomplete transcriptions. Such creative extrapolations are beyond the intended scope of the work we are engaged in.
At the opposite extreme to these are the low-level FAPS (Facial Animation Parameters) and BAPS (Body Animation Parameters) of the MPEG-4 standard [@PandzicF02], and the H-Anim standard \[<http://www.h-anim.org>\]. FAPS are a set of basic facial movements expressed in terms of a standard set of feature points of the face. Amounts of movement of these points are expressed in terms of a set of standard measurements of the face (e.g. distance between the eyes). They are thus avatar-independent: the same stream of FAPS data played through different face models should produce animations that are seen as doing the same thing (as demonstrated in [@EisG98]). BAPS, on the other hand, consist of joint rotations in the form of Euler angles, and are not avatar-independent. A stream of BAPS data which, played through one avatar, causes it to clap its hands, will fail to do so if played through another avatar which is identical except that its shoulders are a couple of inches further apart: the hands will fail to meet by the same distance. H-Anim is an adjunct to the VRML (Virtual Reality Modelling Language) standard defining a standard skeleton and form of motion data for humanoid animation. Like BAPS it expresses motion data in terms of avatar-dependent joint rotations.
We thus perceive a role for a method of specifying physical posture and movement in a manner which is avatar-independent, human-readable and -writable, and which allows unambiguous calculation of the exact avatar-dependent motion data for any given avatar. Such a method can provide the lower level of implementation which must underlie all the higher-level proposals described above. This is the purpose of the work reported here.
In the area of sign language, Vcom3D \[<http://www.vcom3d.com>\] publish a commercial product, SigningAvatar, for performing sign language on web pages. It uses an internal notation for specifying the movements of the avatar, of which details are not available. SignSynth [@grieve:signsynth] is a signing animation system based on the Stokoe notation, producing output in VRML, and also addressing the issue of translating English text to sign language. The GESSYCA system [@lebourque.sylvie:complete; @lebourque.sylvie:high; @gibet:high-level] is concerned with French sign language, and of all the animation systems listed here is closest to the approach of the present paper. It uses a notation system developed for a project called Qualgest (Qualitative Communication Gestures Specification), and uses biomechanical simulation to generate realistic human movements.
HamNoSys and SiGML
==================
The language we have developed is called SiGML (Signing Gesture Markup Language), and was developed from HamNoSys, the Hamburg Notation System. HamNoSys [@prillwitz.leven.ea:hamnosys; @hanke:hamnosys] is a notation for recording sign language gestures, developed by researchers on sign language at the IDGS (Institut für Deutsche Gebärdensprache) at the University of Hamburg. It differs from other signing notations such as Stokoe [@stokoe:dictionary] in that it is not specific to any one sign language, but is intended to cover all signing gestures in all sign languages.
Like most such notations, HamNoSys records signs in terms of hand shape, hand location, and hand movement. It uses a set of special glyphs comprising a HamNoSys font. An example is shown in Figure \[fig:here-hns-avatar\]. The symbols mean:
:
: The hands perform mirror-image actions.
: The hand shape, illustrated in Figure \[fig:modified-handshapes\], is the *finger2* shape of Figure \[fig:handshapes\] with the thumb across the fingers and the index finger bent at the base.
: The axis of the hand points forwards.
: The palm faces down.
: The hands move down a small distance.
: Repeat the movement once from the initial posture.
In the most recent version, HamNoSys has been extended to cover facial movement and expression, and non-manual movements of the upper body such as nodding the head, raising the shoulders, etc.
Principles of HamNoSys
----------------------
A HamNoSys transcription is independent of the person or avatar performing it. It records only those aspects of the gesture which are significant for the correct performance of a sign. Thus for most signs, it records only what the hands do; the rest of the body is assumed to do whatever is natural in order for the hands to perform the gesture. There are a few signs for which some other body part is significant, such as the BSL (British Sign Language) sign for “Scotland”, in which the right elbow is the important part (see Figure \[fig:ScotlandBSL\]).
In such cases, HamNoSys contains provisions for specifying the significant body part.
HamNoSys also includes some non-manual actions of the torso, shoulders, head, and eyes: shrugging the shoulders, tilting or turning the head, etc. There is also a repertoire of facial morphs of two types: visemes (visual equivalents of phonemes, the visible mouth movements used in speech) and other facial movements, such as closing the eyes or puffing the cheeks. These are classified into a number of “tiers”, each tier being the set of movements that affects one part of the body: the shoulders, the torso, the head, the eyes, the facial expression, and the mouth. For each tier a sequence of actions can be given, and actions for different tiers being performed simultaneously. There are only limited capabilities for synchronising different tiers with each other or with the signing actions of the arms and hands. Non-manuals are not as fully developed a part of HamNoSys as manual gesturing, and play a correspondingly limited role in our implementation of HamNoSys.
Positions are specified in terms of named locations. These fall into two classes: locations in “signing space”, the space generally in front of the signer’s upper body, and locations on or close to the surface of the body (see Figures \[fig:signingspace\] and \[fig:locations\]).
-- --
-- --
Signing space is represented in HamNoSys by a three-dimensional grid of discrete points. These are at a number of levels: below the stomach, stomach, chest, shoulders, neck, head, and above the head. From left to right there are five positions: far left, near left, centre, near right, and far right. There are three proximities: near, medium, and far. Points on the body include the palms, the fingertips, the eyebrows, the nose, the chin, and many more; each may be specified with a proximity of medium, near, or touching. In total, there are some hundreds of locations that can be specified. In addition, a position can be specified as the midpoint between two of the above positions.
Hand shapes are classified into the twelve basic types illustrated in Figure \[fig:handshapes\]. Variations such as in Figure \[fig:modified-handshapes\] can be created by specifying the shapes of individual fingers: straight, rounded, curled, etc. Contacts between fingers can also be specified, such as fingers crossed, or thumb between fingers.
[cccccc]{}
[c]{}\
*fist*
&
[c]{}\
*flat*
&
[c]{}\
*finger2*
&
[c]{}\
*finger23*
&
[c]{}\
*finger23spread*
&
[c]{}\
*finger2345*
\
[c]{}\
*pinch12*
&
[c]{}\
*cee12*
&
[c]{}\
*pinchall*
&
[c]{}\
*ceeall*
&
[c]{}\
*pinch12open*
&
[c]{}\
*cee12open*
-- --
-- --
Movements are parallel or sequential combinations of a repertoire of basic movements: straight, curved, and circular. The direction of a movement can be given as a combination of up, down, left, right, forwards, and back. A target location can be given for a movement, instead of, or as well as, the direction of movement.
In many signs, the hands perform mirror image actions. HamNoSys allows transcriptions of such signs to be abbreviated, by specifying the action of the dominant hand, plus a symbol indicating that the nondominant hand mirrors the action. Left-right mirroring is the most common form, but there is also provision for front-back and up-down mirroring, as well as parallel movement of the two hands. The mirroring symbols can be regarded as abbreviations which could be eliminated by expanding the transcription.
For someone practiced in the use of HamNoSys, it takes about a minute to write down a transcription of a sign. This compares very favourably with the time it would take to either animate a sign by hand with animation software such as 3D Studio Max, or to record a sign in a motion capture session.
SiGML
-----
HamNoSys was originally designed as a medium for researchers to communicate with each other, and not with computer processing in mind (other than storing HamNoSys transcriptions in databases). As a result, it had initially no formal syntax, and only a verbal description of the semantics in the form of the HamNoSys manual [@prillwitz.leven.ea:hamnosys]. To separate out the issues of concrete syntax, we obtained from the IDGS a first draft of a formal grammar, which was then further developed jointly. The initial formal grammar was developed *post hoc*, and in some cases was contradicted by actual practice in the existing corpus of HamNoSys transcriptions. Some transcriptions were indeed erroneous (as judged by the designers of HamNoSys), but others had to be accepted as correct and the grammar modified to allow them.
We then developed an XML application language to represent the grammatical structures of HamNoSys more abstractly, called SiGML: Signing Gesture Markup Language. A parser was written to convert HamNoSys into SiGML, and all subsequent processing is performed on the SiGML, which is read and parsed by the freeware *expat* XML parser. Figure \[fig:here-sigml\] gives the SiGML translation of the HamNoSys of Figure \[fig:here-hns-avatar\].
Animating SiGML
===============
HamNoSys and SiGML are signer- and avatar-independent. To generate motion data for a particular avatar, numerical information about the avatar must be supplied, to be combined with the avatar-independent description of the movement, to produce avatar-specific motion data. The information required divides into two classes, information about the geometry of a specific avatar, and information about body language or signing style, which can vary independently of the avatar.
The geometrical information required about the avatar consists of:
1. The dimensions of the avatar’s skeleton: the lengths of its bones, and the positions and orientations of the bones in some standard posture.
2. For each of the non-manuals defined in HamNoSys, the mixture of avatar-specific facial morphs or bone rotations which implements it.
3. The coordinates of every location nameable in HamNoSys, relative to the bone which moves the relevant body part. In principle, a HamNoSys location should be considered as a volume having a certain extent, the choice of which point within that volume to use being dependent on the context, but for implementation we have preferred to model each location by a single fixed point.
The first of these can in principle be read automatically from whatever avatar description file is exported by the modelling software used to create it. The second requires the avatar creator to express each of the HamNoSys facial movements as a combination of the avatar’s morphs. The third requires the avatar creator to specify the coordinates of all the HamNoSys locations on the surface of the body. Some of these locations can be discovered automatically. For example, it is fairly easy to determine surface points on the palmar, dorsal, ulnar, and radial sides of each finger, level with the midpoint of the middle phalanx. Other points are best placed manually: it is easier for the modeller to say where the ears are than to determine them algorithmically from analysis of the surface mesh. Specifying the locations of these surface points should be considered a necessary part of the construction of any avatar intended for synthetic animation. Note that no other knowledge is required of the surface mesh itself, which may be arbitrarily dense, subject to the requirement of rendering it at the desired frame rate.
The “body language” class of information consists of the mapping of HamNoSys categories to numbers. It requires:
1. A numerical definition of the size of a “large”, “medium”, or “small” movement, “near” and “far” proximities, the shape of a “deep” or “shallow” curved arc, and similarly for all the other geometric categories in HamNoSys. Distances are best given as proportions of various measurements of the avatar: for example, “far” from the chest might be defined as a certain proportion of the length of either arm.
2. A numerical specification of the time required for each type of movement, and of how that time should change when HamNoSys specifies the manner of movement as “fast”, “slow”, “tense”, etc.
3. A numerical specification of how the hands should accelerate and decelerate during movements of different types, performed with different manners.
4. The temporal trajectories that the avatar-specific morphs and bone rotations that implement non-manual movements should follow. This takes the form of attack, sustain, and release times, plus a description of how the elements of the movement ramp up from zero to the full value during the attack, and down again to zero during the release.
Body language information is specific, not to a particular avatar body, but to the personality inhabiting it.
Given all of this information, generating motion data requires solving the following problems.
Hand location
-------------
HamNoSys and SiGML always specify the locations of the hands, but to determine exactly what is meant by a piece of notation that is clear to the human reader can be surprisingly non-trivial.
The hand is an extended object: the whole hand cannot be placed at some point in space, only some particular point on the hand. HamNoSys allows a point on the hand to be specified, but does not require it. If it is omitted, an implementation of HamNoSys must guess which point is intended. Generally, when the hands are not in contact with each other or the body, the centre of the palm is a reasonable choice, but this will usually not be suitable when there are contacts or close proxomities. For example, the BSL sign for “me” uses a pointing hand positioned close to or touching the chest, pointing to the signer. If the centre of the palm is placed there, the index finger will penetrate the chest. The significant location on the hand is here the index fingertip.
There is a trade-off here between the effort required of the program and its author in order to correctly guess the intended meaning, and the effort required of the writer of HamNoSys or SiGML in being explicit about such details.
In one significant case there is no way to express in HamNoSys certain information about hand location. Some HamNoSys transcriptions express the spatial relationship of the two hands to each other, and the relationship of the pair of hands considered as a whole to the torso. For the first part of this information, all that HamNoSys can express is the distance between the two hands, and the locations on the hands that are to be that distance apart. The direction from one location to the other cannot be expressed in HamNoSys. There are some heuristic rules that one can formulate for guessing this information, based on inspection of example transcriptions, but there is no way to be sure that the rules will be correct in all cases, because there is no formal definition of what the correct choice is. For making new transcriptions directly in SiGML, it would be easy to extend SiGML with an additional attribute of the relevant XML element specifying the direction. However, this would not help in dealing with legacy HamNoSys transcriptions, such as the IDGS corpus of several thousand signs of German Sign Language.
Arm positioning
---------------
Most HamNoSys transcriptions specify only the hand movements. The arms are assumed to move in whatever way is natural to achieve this. The animation software must therefore have inverse kinematic rules to decide how the elbows and shoulders should move in order to place the hands into the required positions and orientations. The rules that we have implemented were derived from informal observation of signers, experimentation, and by obtaining feedback from signers about the perceived quality of the postures and movement.
1. When a hand is in signing space, the plane formed by the upper and lower arm bones should make some angle $\alpha$ with the vertical plane through the shoulder and wrist joints. This angle should vary from zero when the hand is directly to the right of the body, to some angle $\alpha_{\mbox{\textit{max}}}$ when the hand reaches across to the opposite side.
2. The wrist joint has limited mobility, especially for side-to-side movements. Therefore the angle $\alpha$ chosen by the previous constraint may need to be further modified to prevent excessive wrist bending.
3. When the hand reaches beyond a certain distance from the shoulder, the collarbone should rotate to move the shoulder some distance in the direction of reach.
4. These rules are extended to cover all hand positions and orientations likely to arise in signing, in such as way as to make the resulting arm configuration vary continuously with the hand.
For reaching movements across the body, the animation would be further improved by allowing the upper body to turn and tilt in the direction of reach, but this has not yet been implemented.
Handshapes {#sect:handshapes}
----------
HamNoSys handshape descriptions can be quite complex. To a basic handshape selected from those of Figure \[fig:handshapes\] may be added any or all of the following:
[—]{}
: A finger bending to override the bending of all the extended fingers or thumb of the basic handshape. The “extended fingers” are, for example, all of them for the *flat* handshape, the index finger and thumb for the *pinch12* handshape, etc. There are five possible bendings, beside the default value for the handshape, which can nominally be expressed in terms of the amount of bending of the three joints of the finger, as a proportion of the maximum possible. The joints are listed in order going outwards from the palm. *bent* $= (1,0,0)$, *round* $= (\frac{1}{2},\frac{1}{2},\frac{1}{2})$, *hooked* $= (0,1,1)$, *dblbent* $= (1,1,0)$, *dblhooked* $= (1,1,1)$. The actual joint angles may vary substantially from these nominal values, and depend on the avatar.
[—]{}
: A thumb modifier: splayed, opposing the fingers, or folded across the palm. In the case of the three *cee*-type handshapes (*cee12*, *ceeall*, and *cee12open* in Figure \[fig:handshapes\]), a thumb modifier may also specify that the gap between fingertips and thumb tip should be wider or narrower than normal.
[—]{}
: A specification of which fingers are to be the extended fingers: for example, to specify that the small finger should be extended instead of the index finger in the *finger2* handshape.
[—]{}
: A specification that a particular finger should be given a specific bending.
[—]{}
: A specification that one finger should be crossed over another.
[—]{}
: A specification that the thumb should be placed between two fingers.
Turning a general handshape description into a set of joint angles appropriate to a given avatar would ideally be done by using the dimensions of the avatar’s hands to calculate the joint rotations that would produce the correct relationships between all of its parts. Instead of doing this, we calculated lookup tables for the avatars that we were using, by posing the avatar’s hands in various shapes and reading the joint rotations. This involved less work, given the limited number of different avatars that we had available to animate, but in the longer run it would be preferable to calculate the tables automatically from the hand geometry.
The lookup tables express the basic handshapes and finger bending modifiers, not in terms of joint rotations directly, but in terms of a parameterisation of each joint according to the number of its degrees of freedom. The middle and end joints of the fingers and thumbs are adequately described as hinge joints, and can be parameterised by a single number. The base joints are more complicated.
The base joint of each finger has two degrees of freedom, which can be labelled “bend” and “splay”. In HamNoSys, these are independently specified. Bend is implied by the handshape or explicitly given by a fingerbending modifier. Splay is implied by the handshape: *finger23spread*, *finger2345*, *cee12open*, and *pinch12open* have various degrees of splay of the extended fingers, the other handshapes do not. The problem is then to map any given amounts of bend and splay to the corresponding rotation, taking into account the interaction between the two. The interaction has two effects. Firstly, bend limits splay: when the base joint is fully bent towards the palm, it has almost no mobility about any other axis. So when the value of bend is at its maximum, the rotation should be independent of the value of splay. Secondly, the space of all physiologically possible rotations of the joint is a two-dimensional subset of the three-dimensional space of all mathematically possible rotations. We require a description of that space which makes it easy to calculate the rotation corresponding to a given bend and splay. The third mathematically existing but physiologically infeasible degree of freedom is twist about the axis of the finger.
There are two standard types of mechanical joint having two degrees of rotational freedom: the universal joint and the constant velocity joint[^2]. We found that neither of these, with their natural parameterisations, provides an accurate mapping of bend and splay to joint rotation, no matter how the bend and splay axes are selected. All of them require a significant amount of twist to be added for some combinations of bend and splay in order to reproduce the observed orientation of a real finger. To compute this twist one must either construct a lookup table defining twist as a function of bend and splay, deriving a continuous function by interpolation, or find an algorithmic method of computing the required total rotation of the joint. We adopted the latter approach, and require just three extra measurements of each of the avatar’s fingers. These could be automatically computed from the locations of some of the definable points on the hands, but at present have been measured manually for each avatar. The measurements are of deviations between the actual directions of the proximal and medial phalanges of the fingers in certain handshapes, and the directions they would have in an ideal simplified hand whose fingers in a flat handshape are straight and parallel to the axis of the hand, and which bend along axes exactly transverse to the fingers. See Figure \[fig:fingerangles\].
----------------------------------------------- ------------------------------------------------ ------------------------------------------------
${\theta_{\mbox{\scriptsize \textit{flat}}}}$ ${\theta_{\mbox{\scriptsize \textit{fist1}}}}$ ${\theta_{\mbox{\scriptsize \textit{fist2}}}}$
----------------------------------------------- ------------------------------------------------ ------------------------------------------------
1. In the flat hand (in which the fingers are straight and touch each other all along their length), the angle ${\theta_{\mbox{\scriptsize \textit{flat}}}}$ (a different angle for each finger) between the actual direction of the proximal phalanx of each finger and the ideal direction. Typically, there is some inward tapering of the fingers, depending on their thickness.
2. The same measurement, taken for a fist, giving angles ${\theta_{\mbox{\scriptsize \textit{fist1}}}}$ for each finger.
3. The same measurement, taken for the medial phalanges in a fist, giving angles ${\theta_{\mbox{\scriptsize \textit{fist2}}}}$ for each finger.
Let $\beta$, $\sigma$, and $\lambda$ be the bend, splay, and longitudinal axes of rotation. These are orthogonal: $\lambda$ is the axis of the hand, $\sigma$ is normal to the plane of the hand, and $\beta$ is perpendicular to both. Write $(\alpha,a)$ for a rotation about axis $\alpha$ of angle $a$, and define $\mbox{mid}(a,b,k) = a(1-k) + bk$. Then given angles $b$ and $s$ for bend and splay, and defining $b' = 2b/\pi$, we construct the total rotation as the following composition of four components: $$(\sigma,\mbox{mid}(s,0,b')) \cdot
(\beta,b) \cdot
(\sigma,\mbox{mid}({\theta_{\mbox{\scriptsize \textit{flat}}}},{\theta_{\mbox{\scriptsize \textit{fist1}}}},b')) \cdot
(\lambda,{\theta_{\mbox{\scriptsize \textit{fist2}}}}b')$$ When $b = 0$, this reduces to $(\sigma,s+{\theta_{\mbox{\scriptsize \textit{flat}}}})$ and when $b = \frac{1}{2}\pi$, it reduces to $(\beta,\frac{1}{2}\pi) \cdot (\sigma,{\theta_{\mbox{\scriptsize \textit{fist1}}}}) \cdot (\lambda,{\theta_{\mbox{\scriptsize \textit{fist2}}}})$, which can be verified to produce the finger orientations exhibited in Figure \[fig:fingerangles\] for the flat and fist handshapes. (Some care is necessary about the choice of directions of the axes and signs of the rotation angles.) In the figure, the three angles are all zero for the middle finger. For ${\theta_{\mbox{\scriptsize \textit{flat}}}}$, this is because we define the longitudinal axis of the hand to be the direction of the middle finger in the flat hand. For ${\theta_{\mbox{\scriptsize \textit{fist1}}}}$ and ${\theta_{\mbox{\scriptsize \textit{fist2}}}}$, this is a contingent fact about the particular hand from which the illustrations were made.
Although these angles are typically no more than a few degrees, they make a large difference to the naturalness of the resulting handshapes. Setting them all to zero produces unrealistic claw-like hands. Comparative examples are shown in Figure \[fig:claws\]. (The abnormal shortness of the medial phalanx of the ring finger is a defect of the avatar.)
-- -------- ---------- --------
*flat* *ceeall* *fist*
-- -------- ---------- --------
The carpometacarpal joints of the fourth and fifth fingers (joints which lie inside the palm) have some limited amount of mobility which is used when forming a handshape in which the fingers wrap round an object such as a tennis ball. Although one is typically not aware of the movement, it is significant: without that mobility, one would hardly be able to pick up a tennis ball with the whole hand. These joints can be modelled as hinges with their axes lying obliquely across the palm, and which should receive nonzero rotations when the fourth and fifth fingers have nonzero values for both bend and splay. (Many supposedly fully articulated avatars do not include these joints.)
The base joint of the thumb is more complicated. It appears to have two degrees of freedom, which can be very roughly characterised as the angle between the metacarpal bone and the axis of the hand, and the rotation of the thumb about that axis, but a mapping of two real parameters to the space of rotations that produces realistic articulation of all the required handshapes is difficult to synthesize. We eventually discarded the attempt to describe it in terms of bend, splay, and an algorithmically determined twist, and defined the necessary amount of twist for the basic thumb positions in the lookup tables describing the basic handshapes. Intermediate thumb positions are then defined by interpolation of these three parameters.
HamNoSys can also define a handshape as being intermediate between two given handshapes. (An example is shown in Figure \[fig:me\]; the backslash is the betweenness operator.) In this case we linearly interpolate between the representations of the handshapes in terms of joint parameters, before converting to joint rotations.
Collisions
----------
A certain amount of collision detection and avoidance is required in order to prevent the avatar passing its arms or hands through its body. However, we make the basic assumption that the transcription we are given does not describe something physically impossible, such as a hand movement that begins in contact with the chest and then moves inwards. We have only found it necessary to implement collision avoidance for two situations:
1. When reaching across the body, the upper arm must be prevented from entering the torso, by suitably raising the elbow.
2. Because the front of the torso cannot be assumed to be a vertical plane, a movement that begins in contact with the torso and moves up, down, or across must be modified so as to avoid penetrating the surface of the torso.
To fulfil these constraints, it is not necessary to know the complete surface mesh of the avatar, merely a sufficiency of locations on its surface for interpolation to yield a good enough result. The set of HamNoSys locations is sufficiently dense to achieve this. For this reason, the actual surface mesh of the avatar (which might be arbitrarily complex) does not need to appear in the avatar description used by the synthesis software.
Trajectories
------------
The route by which a hand moves from one location to another is in general explicitly specified by the transcription, and therefore does not need to be decided by the software. The accelerations and decelerations along that path are not specified in any detailed way. HamNoSys distinguishes five modalities of movement: normal, fast, slow, tense[^3], and “sudden stop at the end” (such as in a sign for “punch”). In addition, we distinguish four separate types of “normal” movement in signing: “targetted”, “lax”, “hard contact”, and “linear”. These are not explicitly notated, but are implied by context.
Targetted movement is used for meaningful movement, as opposed to lax movement, which is used for non-meaningful movements. There is a marked visual difference between the two. The deceleration at the end of a targetted movement is much greater and shorter than the acceleration at the beginning, giving a velocity profile that peaks after the midpoint of the movement. Lax movement has a broader velocity peak lying more in the middle of the movement, with a sharper initial acceleration and a longer final deceleration than targetted movement. An example of lax movement is in the return stroke of most repeated movements, such as the upwards stroke of the sign for “here” in Figure \[fig:here-hns-avatar\], and the outwards stroke of the elbow in Figure \[fig:ScotlandBSL\]. HamNoSys can also notate a form of repetition in which the return stroke is as significant as the forward stroke; for a sign using this repetition, both directions would be performed with the targetted trajectory.
The distinction between lax and targetted movement can also be seen in the transitional movement from the end of one sign to the beginning of the next. When the second sign consists of a single static posture (such as the BSL sign for “me”, in which the hand points to the signer), the transition is targetted, but when the second sign contains movement, the movement into its initial posture from the previous sign is lax. (In an early version of the implementation, we used targetted motion for all inter-sign transitions, and received comments from deaf people that those movements looked wrong: they appeared meaningful when they were not. Using lax transitions except before static signs eliminated this criticism.)
Hard contact is targetted movement which ends with the hand contacting a solid object such as the other hand or the torso, and which produces a more sudden stop than a targetted movement ending in mid-air. At the end of the movement, the hand remains in contact with the object for a perceptible time before moving again.
Linear movement, where the distance moved is proportional to elapsed time, is never appropriate for straight line movements in lifelike humanoid animation, but it is applicable when the hand makes a circular motion. Apart from the transitions into and out of the movement, the angular velocity remains close to constant.
To synthesize suitable temporal trajectories for the different manners, we use the semi-abstract biocontrol model illustrated in Figure \[fig:biocontrol\].
(15.5,8)(0,-0.7) (6,7)[(0,-1)[1.5]{}]{} (6.2,6.2)[$x_r =$ reference value of $x$]{} (3,0)[(0,1)[4.75]{}]{} (3,4.75)[(1,0)[1]{}]{} (0.3,2.5)[Actual value of $x$]{} (2.4,-0.5)[$x = \int \dot{x}$]{} (8,4.75)[(1,0)[3]{}]{} (11,4.75)[(0,-1)[0.75]{}]{} (8.2,5)[$\dot{x}_r =$ reference value of $\dot{x}$]{} (7,0)[(0,1)[3.25]{}]{} (7,3.25)[(1,0)[2]{}]{} (4.3,2.5)[Actual value of $\dot{x}$]{} (6.4,-0.5)[$\dot{x} = \int \ddot{x}$]{} (13,3.25)[(1,0)[2]{}]{} (15,3.25)[(0,-1)[1.95]{}]{} (15,1.1) (15,0.9)[(0,-1)[0.9]{}]{} (13.2,3.85) (14.45,-0.5)[$\ddot{x} = \ddot{x}_r$]{} (-0.5,1.5)[*Control system*]{} (-0.5,0.4)[*Physics*]{} (-0.7,0.9)(0.133,0)[125]{}[(1,1)[0.391]{}]{} (4,4)[(4,1.5)[$\dot{x}_r = k(x_r - x)$]{}]{} (9,2.5)[(4,1.5)[$\ddot{x}_r = k'(\dot{x}_r - \dot{x})$]{}]{}
This is a two-level cascade control system built from two proportional controllers. The part below the shaded line is the simulated physics. The controlled variable is $x$, representing the proportion of the distance that has been travelled towards the desired final position, and is initially zero. $x_r$ is the reference or target value, initially 1. The output of the outer controller is $\dot{x}_r = k(x_t - x)$, a demanded rate of change of $x$ which is proportional to the error, the difference between the reference value $x_r$ and the actual value $x$. This is the reference input to the inner controller, which compares it with the actual rate of change $\dot{x}$, and outputs a demanded acceleration $\ddot{x}_r$ proportional to the error $\dot{x}_t - \dot{x}$. The virtual actuator sets $\ddot{x}$ equal to $\ddot{x}_r$, which determines the time evolution of $\dot{x}$ and $x$.
The equation of motion of the resulting system is $$\ddot{x} + k'\dot{x} + kk'x = kk'x_t$$ With constant positive gains in both of the proportional controllers, this is mathematically equivalent to damped simple harmonic motion, critically damped when $k'/k = 4$, underdamped for smaller values of $k'/k$, and overdamped for larger values. The top row of Figure \[fig:trajectories\] illustrates some trajectories of the system. The horizontal axis is scaled so that the major part of the motion (when it has settled to within 99% of the final value) occurs in one unit of time.
[ccc]{} ![Horizontal axes: time; vertical axes: position.[]{data-label="fig:trajectories"}](images/critdamped.pdf "fig:") & ![Horizontal axes: time; vertical axes: position.[]{data-label="fig:trajectories"}](images/underdamped.pdf "fig:") & ![Horizontal axes: time; vertical axes: position.[]{data-label="fig:trajectories"}](images/overdamped.pdf "fig:")\
Critically damped ($k'/k=4$) & Underdamped ($k'/k=2$) & Overdamped ($k'/k=8$)\
\
\
& &\
\
\
![Horizontal axes: time; vertical axes: position.[]{data-label="fig:trajectories"}](images/targetted.pdf "fig:") & ![Horizontal axes: time; vertical axes: position.[]{data-label="fig:trajectories"}](images/lax.pdf "fig:") & ![Horizontal axes: time; vertical axes: position.[]{data-label="fig:trajectories"}](images/suddenstop.pdf "fig:")\
Targetted & Lax & Sudden stop\
We found that all choices of $k'/k$ gave sluggish behaviour, compared with the performance of signers, especially towards the end of the movement. We modified the system by making the stiffness increase as the target is approached (that is, by increasing $k$ and $k'$ as the error $x_r - x$ decreases, while keeping their ratio constant), so as to model the increase in tension of the muscles as the intended location is reached. This resulted in animations that were judged to be satisfactory, and compared well with trajectories observed by frame-by-frame stepping through a signing video, of which an example is given in the middle row of Figure \[fig:trajectories\], for the sign “west” (the flat right hand, facing left, makes a leftwards movement). The movement lasts for 11 frames at 15 frames per second. The “sudden stop” modality uses a greater increase of $k$ and $k'$, sets their ratio to be underdamped, and truncates the trajectory when the final position is reached. Varying the parameters of the system produced suitable trajectories for each of the other HamNoSys modalities, as well as the distinctions between targetted, hard contact, and lax movement. Because HamNoSys manners are a discrete set, we can greatly reduce the computation time by precomputing a trajectory for each one, in the form of a lookup table for a function $f:[0\dots1]\rightarrow[0\dots1]$. When a fraction $k$ of the duration of the movement has elapsed, the hand has moved a fraction $f(k)$ along its path. The bottom row of Figure \[fig:trajectories\] illustrates some of these trajectories. The “targetted” trajectory is the one that would be used to animate the sign that was measured from video.
In our implementation, the time taken for a movement is independent of the size of the movement. Some researchers have adjusted the timing according to Fitts’ law [@Fitts], which states that the time for a reaching movement is a linear function of the log of the distance. However, his observations were for reaching movements to a target of definite size and location, in which subjects were asked to favour accuracy over speed. Signing movements are not of this nature. For example, the movements of the hands in the BSL sign for “here” illustrated in Figure \[fig:here-hns-avatar\] have no particular target location of any definite size, and the sign for “Scotland” (Figure \[fig:ScotlandBSL\]) does not involve hand movements at all. Observation of signers’ movements suggests there is in fact little variation of duration with distance (which, at the level of accuracy and range of distances we are dealing with, is also a reasonable interpretation of the logarithmic relationship expressed by Fitts’ law).
The duration can be observed to vary with the shape of the movement: curved arcs are performed more slowly than straight movements, and circular movements take longer still. The durations used by the implementation were arrived at through feedback from signers.
We use the same temporal trajectories for all the other types of movement: change of hand orientation, change of handshape, non-manual movements, and change of facial morphs.
Interpolation between postures
------------------------------
The trajectories defined in the previous section specify what proportion of the path from one posture to the next the avatar should have moved after any given elapsed time. We now describe the actual interpolation between postures.
The location of a hand at any instant is defined by the spatial path described by the SiGML, the particular point on the hand that must traverse this path, and the proportion of the path that has been traversed. Interpolation is linear with respect to distance travelled along the path.
The path in rotation space from one hand orientation to the next is that defined by spherical linear interpolation.
The location and orientation of the hand at each moment define the location of the wrist joint. From this, the wrist, elbow, shoulder, and clavicle joint rotations are calculated by inverse kinematic routines.
When the hand changes shape, the rotation of each hinge joint of the fingers and thumb is interpolated in the obvious way. The finger base joints interpolate the bend and splay parameters, and the thumb base joints the bend, splay, and twist parameters described in section \[sect:handshapes\].
Non-manual movements and facial morphs are described as real numbers expressing the amount of each one that is to be present at any instant, and therefore linear interpolation applies.
Performance
-----------
Each frame of animation data consists of about 50 bone rotations, the exact number depending on the avatar: 15 to 17 joints in each hand, 4 in each arm, and several joints in the spine, neck and head. The frame also includes an amount of each of the avatar’s facial morphs.
The current system is capable of generating 15,000 frames of animation data per second on a 1.7 GHz desktop PC. This excludes the time taken to read and parse the SiGML input. Although the current system does store SiGML as text, when embedded in a system such as those described in Section \[section:otherwork\] it would be feasible to bypass parsing by directly constructing SiGML data structures.
If the animation is to play at 25 frames per second, which gives acceptable legibility for signing[^4], this means that the synthesis of motion data from SiGML data structures requires only 0.17% of the time budget, even on what is by current standards a modestly powered machine. At the 60 fps required for fast-action videogames, the fraction required of the time budget is still only 0.4% per avatar.
Synthetic animation from a notation such as SiGML also allows for a great reduction in the amount of data to be stored. The SiGML displayed in Figure \[fig:here-sigml\] represents about 0.8 seconds of animation in about 320 bytes. The corresponding 20 frames of 25 fps animation would (uncompressed) occupy about 12kB, assuming 50 rotations per frame, each represented as a triple of floating point numbers. The latter figure can be substantially reduced by various compression schemes, but it is unlikely that any such scheme could achieve the 37-fold reduction obtained by the SiGML representation. The SiGML itself is susceptible of compression: the HamNoSys in Figure \[fig:here-hns-avatar\] only takes 10 bytes.
Automatic generation of SiGML
=============================
SiGML is designed to be human-readable and -writable, but for some applications it is necessary to generate SiGML automatically. One example is its original application, sign language.
Signing is often conventionally transcribed as a sequence of “glosses”: names of the signs performed, such as [*BOOK*]{} [*GIVE*]{}, corresponding to the English sentence “Give me the book”. However, this is misleadingly simplified: to animate such a sequence it is not sufficient to merely replace each gloss by its SiGML transcription and generate animation data. This is because many signs take different forms depending on their context. For example, [*GIVE*]{} will be made with a handshape indicating the type of object that is being given, and the start and end locations of the movement will indicate who is giving to whom: you to me, him to her, me to multiple people, etc. [*GIVE*]{} requires a set of parameters: the above example is better glossed as [*BOOK*]{} [*GIVE*]{}([*you,me,flat-thing*]{}). This variety of instantiations greatly expands the corpus of signs required for interactive discourse beyond what a mere count of the glosses in a dictionary might suggest. For highly inflected signs such as [*GIVE*]{}(…), the SiGML transcriptions of its required forms must be generated automatically. Even for signs which might not appear to require parameterisation, such as signs for ordinary concrete objects like [*BOOK*]{}, the location at which the sign is performed may express some of the meaning. For example, it may be performed at some arbitrarily chosen location in signing space, in order to be able to refer back to the object later in the discourse by pointing to that location.
In the ViSiCAST project, sign language has been generated from English text via the intermediate form of Discourse Representation Structures and Head-Driven Phrase Structure Grammars. The latter form the input to sign language generation based on the Attribute Logic Engine. This pathway has been described in [@Saf01b; @Saf02; @MarS04], and will not be further described here. Sign language represented in the parameterised form described above can then be transformed into SiGML, inserting appropriate parameters such as handshape and location.
In virtual reality applications with autonomous interactive avatars, SiGML (or an extension of SiGML covering a wider repertoire of movements) could be generated from the system’s internal high-level representation of the avatars’ intended actions (e.g. “point to that building”). The SiGML description serves as a low-level, but still avatar-independent, representation of the desired action, which can then be translated into motion data for the particular avatar.
Limitations and extensions of HamNoSys
======================================
The success of our implementation of SiGML demonstrates the practicality of real-time synthesis of human movement in a certain restricted setting. We are interested in expanding this application to wider repertoires of human movement, such as are required by general interactive virtual reality applications. The experience of implementing SiGML has shown some weaknesses of HamNoSys for this application, and suggests ways in which it might be improved.
Limitations
-----------
### Syntax
HamNoSys was originally developed for people to read and write. Its formal grammar was developed in hindsight to fit existing practice. While that practice is easy for those trained in the notation to deal with, the resulting grammar is more complicated than is warranted by the subject matter. This is still true of SiGML. Although it eliminates many of the irrelevant complexities of the concrete syntax of HamNoSys, it is still very much based on HamNoSys. A notation developed specifically for computer animation could be substantially simplified without any loss of expressiveness.
### Semantics
HamNoSys depends to a significant extent on human understanding of signing for both parsing and interpretation. A large part of the task of implementing SiGML was taken up by attempts to automatically fill in information that human transcribers typically omit when transcribing signs, and (often unconsciously) fill in when reading transcriptions.
In the eSign project, among the people who were responsible for creating SiGML transcriptions of the signs required for various applications, those who had less prior experience with HamNoSys sometimes found it easier to produce transcriptions that worked well with the animation system, being less accustomed to depending on human interpretation.
An improved SiGML would have a simpler and more precise semantics, expressed in some form more explicit than merely its implementation.
### Handshapes {#handshapes}
The twelve basic handshapes of HamNoSys can, in principle, be mapped to (avatar-dependent) tables of finger and thumb joint angles, as can each of the fingerbending modifiers. However, there turns out to be some significant context-dependency. For example, the actual bend angle of a finger that is “bent at the base joint” depends on the basic handshape (compare the base joint angles for the two handshapes in Figure \[fig:modified-handshapes\], which both use that finger shape modifier). This requires a significant expansion of the per-avatar lookup tables used to translate handshapes to joint rotations. To handle this in a more flexible way, and to cover the less commonly used features of HamNoSys hand descriptions (such as crossed fingers), we augmented the SiGML [`<handconfig>`]{} element with the possibility to explicitly specify finger joint angles. This is, however, at the expense of avatar-independence; a better solution would be to calculate the required angles from an avatar-independent specification of the significant properties of the handshape. Such properties would include contacts between different parts of the hand, or opposition of the thumb to specified fingers.
We also note that in the *cee* handshapes, modifications expressing a wider or narrower gap between the fingers and thumb are expressed in HamNoSys by a modifying symbol which describes the thumb position. However, inspection of actual performance shows that changes to the finger-thumb distance are accomplished primarily by varying the bend angle at the base joints of the fingers. The HamNoSys notation for these handshapes does not map in a simple way to physical reality.
### Integration of manual and non-manual movements
For historical reasons, HamNoSys makes an artificial distinction between manual and non-manual movements. This separation, enforced by the syntax of HamNoSys and carried over into SiGML, limits the flexibility with which the two classes of movement can be combined and synchronised with each other.
Extensions
----------
### Discrete categories
Several concepts of SiGML are discrete categories: there is a finite set of locations, a finite set of distances, a finite set of directions and sizes of movement, etc. Other applications of scripted animation will require more precision and variability than these provide. For some of these categories HamNoSys allows the specification of a value midway between two built-in values, but this still allows only a limited set of values. Each of these sets of discrete values should be extended to continuous categories by adding the possibility of specifying fractional interpolation. We have already implemented extensions to SiGML for this purpose.
### Timing
We have provided more detailed control of timing than the “fast” and “slow” modalities of HamNoSys, by allowing an absolute duration or a temporal scaling to be applied to any movement or to a whole sign.
### Path specification
A SiGML transcription specifies explicitly the spatial paths of the hands — straight line, curved, circular, etc. The software never has to decide what path to follow to move a hand from one place to another. This should remain true for more general animation. The person or automated process that is composing a SiGML transcription may be assumed to be describing physically possible postures and movements. If the author has provided sufficient detail to uniquely determine the posture and movement, the software need not be burdened with the task of guessing or inventing missing detail. If it is desired to ease the animator’s task by allowing some details to be omitted and automatically synthesized, algorithms for doing so should be implemented either as a separate SiGML-to-SiGML translation or an external library that can be referenced from SiGML, rather than being built into the implementation of SiGML itself.
### Gravitational dynamics
Signing movements are largely independent of gravity. Hands move along their intended paths and the arms do whatever is necessary to achieve that. As a result, sufficiently lifelike accelerations and decelerations can be created from a simple biocontrol model with a small number of parameters, without requiring a detailed physical simulation. In contrast, for lower body movements like walking, or for upper body movements that involve lifting heavy things, gravitational forces are comparable with or greater than muscular forces and play a major role in the dynamics, and hence also the visual appearance. An extension of SiGML to cover these types of movement will require the implementation to be provided with a means of synthesizing them, which will be more complicated than the simple biocontrol model that we have found sufficient for signing animation. Integration with more sophisticated methods such as those surveyed in [@multon.france.ea:computer] would be desirable.
### Expressive movement
HamNoSys was originally devised for recording signing citation forms, and as such it contains only a discrete range of modalities, those necessary to the correct performance of the signs. Thus SiGML does not have to deal with the entire range of expressive movement. Although signing itself can be very expressive, the “newsreader” type applications in the eSign project do not require it. However, it is an essential aspect of general character animation.
We regard expressiveness as being a concept at a higher level than the intended target of SiGML or any extension of it, and outside the intended scope of this work. SiGML is concerned with the avatar-independent description of concrete posture and movement; it is the responsibility of some higher-level system (such as, for example, some of the projects briefly surveyed in section \[section:otherwork\]) to express qualities of expressiveness in concrete terms.
Principles of posture and movement notation
-------------------------------------------
Some of the principles underlying HamNoSys deserve wider application than they in fact receive in that notation.
### Constraint-based posture specification
HamNoSys specifies the positions of the hands either by their relation to each other, and the relationship of the pair of hands considered as a single entity to the body, or by the relationship of each hand to the body separately. This is a useful method of defining posture in an avatar-independent way, but in HamNoSys it is subject to heavy restrictions. The relationship is always one of distance only, not direction (which must be guessed from context). It applies only to position, with hand orientation always defined in absolute terms. We are therefore considering a radically revised language which would specify postures in terms of constraints on the relative positions and orientations of general body parts.
### Intention vs. geometry
There is a large number of geometric features of a posture which a transcription notation might record, and there are many different ways of choosing a set of geometric features whose specification will completely determine the posture. For example, any two of the orientation of a hand, the orientation of a straight index finger, and the angle at the base joint of that finger determine the third. Which properties, therefore, should a notation record?
HamNoSys makes a particular choice of features to record. For example, the direction of the hand is expressed in terms of the “extended finger direction” or e.f.d.: the direction the fingers would be pointing in if they were straight, or equivalently, the direction of the metacarpal bone of the middle finger. However, there is evidence [@crasborn:phonetic] that this property of the hand is not phonologically significant for signing. For a handshape that is pointing with the index finger, what is significant is the direction in which the index finger is pointing, which may be identified with the direction of the distal phalanx. So long as that direction is correct, the angle at the base joint of the index finger and the e.f.d. of the hand can both vary considerably without affecting the meaning of the gesture.
This leads to a tendency in some circumstances for HamNoSys transcribers to record an idea of the sign which differs from its actual performance. An example is the sign for “me”: an inward pointing index finger. This is commonly but incorrectly transcribed as a *finger2* handshape with an inwards e.f.d. That would be a rather awkward posture to perform, and differs from actual performance. The transcription and actual performance are compared in Figure \[fig:me\].
-- -- --
-- -- --
We therefore believe, as argued in [@kennaway:experience], that a gesture transcription notation should record those properties of a gesture which are essential to its significance: the properties which will remain invariant when the same gesture is performed by avatars of different proportions or different gestural styles. In a slogan, *intention is primary, geometry is secondary*. For a pointing handshape, the direction of pointing is significant; this is what should be recorded, and the bending of the pointing finger in Figure \[fig:me\] should be determined from inverse kinematic rules describing the performance of pointing gestures.
Conclusions
===========
The present work has demonstrated the practicality of generating gesture animation in real time from an avatar-independent scripting language. The generation of motion data is done not merely in real time, but in a very small fraction of the time budget, each single frame of data taking on average 1/15 of a millisecond to generate on a typical machine. The animations are of a degree of naturalism which has proved acceptable to deaf users, in terms of both legibility and realism. The SiGML representation of a movement is also many times smaller than the motion data which it generates.
Avatars for use with this method must be furnished with additional data beyond the skeleton topology and dimensions, and repertoire of morphs. The additional data consists primarily of the locations of a standard list of named points on the surface of the body.
The work has suggested directions in which SiGML can be improved and extended in order to apply the method to other applications of real-time, naturalistic, humanoid animation.
Acknowledgements
================
We acknowledge funding from the European Union for our work on the ViSiCAST and eSign projects under the Framework V programme, the support of our colleagues and partners on these projects, and the deaf users of sign language who have provided valuable feedback on the quality of our animations. John Glauert and Ralph Elliott at the University of East Anglia collaborated with the author in designing the SiGML DTD, and Ralph Elliott wrote the HamNoSys parser and the translator to SiGML. Televirtual Ltd., one of our partners in ViSiCAST and eSign, produced the VGuido and Visia avatars, whose images appear in this paper; Figure \[fig:claws\] uses an avatar created by Vincent Jennings at the University of East Anglia.
[^1]: Note that the text of the web site is in Dutch, and the signing is in Dutch Sign Language. The animations will play on Internet Explorer running on Windows (2000 or XP), and require installing some software (available from the web site). There is also a video demo on the site for those without the necessary configuration. The live avatar plays rather more smoothly than the video.
[^2]: The set of rotations of a constant velocity joint is the set of all rotations about all axes passing through a given point and lying in a given plane. The name derives from its use in the front half-axles of a front wheel drive vehicle, where it ensures that when the driving side turns at a constant velocity, so does the driven side.
[^3]: “Tense” here means a slow and steady movement, as if performed with great effort. This is visually quite distinct from a movement which is merely slow, even when the face (which may communicate some of the associated affect) is concealed.
[^4]: A commercially produced CDrom for students of signing, from which the example trajectory in Figure \[fig:trajectories\] was measured, records all video clips at 15 fps.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that the symmetries of image formation by scattering enable graph-theoretic manifold-embedding techniques to extract structural and timing information from simulated and experimental snapshots at extremely low signal. The approach constitutes a physically-based, computationally efficient, and noise-robust route to analyzing the large and varied datasets generated by existing and emerging methods for studying structure and dynamics by scattering. We demonstrate three-dimensional structure recovery from X-ray diffraction and cryo-electron microscope image snapshots of unknown orientation, the latter at 12 times lower dose than currently in use. We also show that ultra-low-signal, random sightings of dynamically evolving systems can be sequenced into high quality movies to reveal their evolution. Our approach offers a route to recovering timing information in time-resolved experiments, and extracting 3D movies from two-dimensional random sightings of dynamic systems.'
author:
- Peter Schwander
- Chun Hong Yoon
- Abbas Ourmazd
- Dimitrios Giannakis
title: 'The symmetries of image formation by scattering. II. Applications'
---
[^1]
[^2]
Introduction {#secIntroduction}
============
In an earlier paper [@GiannakisEtAl11a], hereafter referred to as Paper I, we presented a theoretical framework for analyzing snapshots formed by scattering. In this paper, we demonstrate the power of this approach to reconstruct three-dimensional (3D) models and time-series from random sightings at extremely low signal, with no orientational or timing information. The theoretical framework in Paper I represents the information content of an ensemble of snapshots as a Riemannian manifold, and shows that the properties of operations in space give rise to object-independent symmetries. Purposeful navigation on this manifold is tantamount to reconstructing a 3D model of the sighted system and/or its evolution, in the sense that given any snapshot, any other can be produced on demand. The symmetries of the manifold reveal its natural eigenfunctions, thus allowing physically-based interpretation of graph-theoretic analysis, and enhanced noise discrimination. Simple algorithms then suffice to reach exceptionally low signal-to-noise levels unmatched by other approaches in terms of computational cost, noise robustness, or both. As examples, we demonstrate structure recovery from radiation-sensitive objects at doses at least an order of magnitude below current levels (signal-to-noise ratio (SNR): $ -16\text{ dB} $), and reconstruction of time-series at SNR values as low as $ -21\text{ dB} $. The versatility of the approach is demonstrated in the context of simulated and experimental data from X-ray diffraction, cryo-electron microscopy (cryo-EM), and optical snapshots using a variety of graph-theoretic techniques. These applications demonstrate the generality of the symmetry-based approach, elucidating at the same time the measures needed to deal successfully with experimental data, a key benchmark of the practical utility of any theoretical framework.
This paper is organized as follows. Without claim to be comprehensive, Sec. \[secPreviousWork\] briefly summarizes previous work in the field to provide a context for the applications discussed in this paper. For the convenience of non-mathematical readers, Sec. \[secConceptualSummary\] provides a conceptual outline of the theoretical framework developed in Paper I. Sec. \[secSimulatedData\] describes 3D reconstruction from simulated diffraction snapshots of single biomolecules at the signal level expected from single molecules in upcoming experiments utilizing the new generation of X-ray Free Electron Lasers (XFELs) [@ShneersonEtAl08; @FungEtAl08; @SchwanderEtAl10a]. Sec. \[appSOD\] establishes, in principle, the applicability of our approach to crystalline samples. Sec. \[appCryo\] addresses structure recovery from simulated and experimental cryo-EM snapshots of single molecules. In this case, essential experimental issues such as defocus variation must be faced and incorporated into the theoretical formalism. Sec. \[appPirouettePasDeDeux\] demonstrates reconstruction of time-series (movies) from random sequences of ultralow-signal optical snapshots. The paper concludes in Sec. \[secConclusions\] with a summary of our key findings and their implications. Detailed points of a technical nature are elucidated in appendices, and movies provided as supplementary online material EPAPS [@EPAPS].
Previous work {#secPreviousWork}
=============
As described in Paper I, we are concerned with constructing a model from random sightings of a system viewed in some projection, i.e., by accessing a limited number of variables describing the state of the system. A 3D model of an object and its evolution, for example, can be constructed from an ensemble of low-signal 2D snapshots without orientational information [@Frank02; @ScheresEtAl07; @FungEtAl08; @LohElser09; @SchwanderEtAl10a; @FischerEtAl10]. Modern graph-theoretic algorithms can now be used to discover low-dimensional manifolds representing the information content of datasets in some high-dimensional space determined by the measurement apparatus [@TenenbaumEtAl00; @RoweisSaul00; @BelkinNiyogi03; @DonohoEtAl03; @CoifmanEtAl05; @CoifmanLafon06; @CoifmanEtAl10; @BishopEtAl98]. The power of these methods stems from their generality, in the sense that no assumptions are made as to the nature of the data or their internal correlations. This brings with it four major challenges: (1) Interpretation of the analysis results (“what physical variables do the manifold dimensions represent?”); (2) Computational cost and scaling behavior on moving from simulated (“toy”) datasets to experimental measurements; (3) Robustness against noise, particularly of non-additive, non-Gaussian types; and (4) Incorporation of inevitable and/or desirable experimental factors (“utility of the theoretical framework in practice”.)
These issues can be brought to focus in the context of the much-discussed problem of recovering the orientation of cryo-EM snapshots of faint biological objects. Direct graph-theoretic attempts to determine the orientation of snapshots from a synthetic object were abandoned at a signal-to-noise ratio (SNR) of $ \sim 2 $ dB, even though only additive Gaussian noise was included [@CoifmanEtAl08]. Noting that graph-theoretic analyses often “fail to solve the cryo-EM problem, because the reduced coordinate system that each of them obtains does not agree with the projection directions” [@CoifmanEtAl10], properties specific to cryo-EM images were used to extract information from the snapshots. Graphs were then constructed using this information in order to assign physical meaning to the outcome of the analysis. Orienting low-signal cryo-EM snapshots by utilizing so-called (straight) common-lines identified primarily in simulated data with additive Gaussian noise has reached remarkably low SNR values [@SingerEtAl10]. However, such assumptions, while justified under some circumstances, are not generally valid. Common-lines, for example, are present only when elastic single-scattering dominates, are straight only when the wavelength of the incident radiation is so short that the Ewald sphere can be replaced by a plane, and are compromised by defocus variations essential for reliable structure recovery by cryo-EM [@Frank06].
Symmetry-based assignment of physical meaning to the outcome of graph-theoretic analysis of scattering data and its favorable computational consequences were addressed in Paper I. Here, we are concerned with ability of this theoretical framework to deal with noise and other important factors encountered in experimental datasets. This determines the practical utility of an approach as much as theoretical elegance and computational efficiency. Below, we demonstrate the utility of our symmetry-based approach by applying a number of manifold-embedding techniques to a variety of simulated and experimental datasets (see Table \[tabApplication\]).
Conceptual summary of theoretical framework {#secConceptualSummary}
===========================================
A snapshot formed on a 2D detector by scattered radiation from an object can be represented by a vector, with the intensity values recorded at the $ n $ detector pixels as components (Paper I, Fig. 1). Object motion and/or evolution (dynamics) change the pixel intensities, causing “the vector tips” representing the ensemble of snapshots to trace out a surface — a manifold — in the $ n $-dimensional data space. The number of degrees of freedom available to the object determines the dimensionality of the manifold traced out. Rotations of a rigid object in 3D, for example, result in a 3D manifold.
The data manifold represents the totality of information about the object gathered by the detector in the course of an observation. Learning is tantamount to understanding the properties of this manifold sufficiently to “navigate” on it. Learning the manifold generated by object rotations, for example, is equivalent to constructing a 3D model of the object, because, starting from any 2D projection (point on the manifold) any other 2D projection can be found by navigation, with the shortest route corresponding to a geodesic.
As we are initially concerned with constructing 3D models from 2D snapshots, we consider a formulation of scattering by a single object in Fourier space so as to concentrate on the effect of rotations. This circumvents issues such as rigid shifts, which would otherwise have to be corrected or incorporated as additional manifold dimensions. Rotation operations do not commute. One must therefore consider the order in which they are performed. This leads to a distinction between so-called left and right “translations,” where a rotation operator $ \mathsf{ T } $ is placed to the left or right of another rotation operator $ \mathsf{ R } $, i.e., $ \mathsf{ T } \mathsf{ R } $ vs. $ \mathsf{ R } \mathsf{ T } $. A left translation can be thought of as an active rotation in 3D space of the incident beam-detector arrangement (frame rotation) after $ \mathsf{ R } $. Similarly, a right translation corresponds to an active rotation of the object (object rotation) before $ \mathsf{ R } $. As $ \mathsf{ T } \mathsf{ R } \neq \mathsf{ R } \mathsf{ T } $, left and right translations must be considered separately. Each forms an SO(3) group, and the total set of possible operations to be considered corresponds to $ \text{SO(3)} \times \text{SO(3)} $.
A key question is this: Which, if any, of these operations leave the distances on the manifold unchanged, i.e., which operations are “invisible” to an ant crawling on the manifold? These operations would represent symmetries — more precisely isometries — of the manifold. For a detector with circular symmetry, the distances on the manifold are invariant under beam-detector rotations about the beam axis. This is obvious; a frame rotation about the beam axis rotates all the snapshots by the same amount about that axis without changing them. This leaves the distances on the manifold unchanged. The process of image formation on a circularly-symmetric detector at right angles to the illuminating beam thus has $ \text{SO(2)} $ isometry, i.e., of all possible $ \text{SO(3)} $ frame operations, the $ \text{SO(2)} $ subset of rotations about the beam direction leave the distances on the manifold unchanged. This is related to the projection of a 3D object on the 2D detector, which is equivalent to a “central slice” through the diffraction volume in reciprocal space.
Consider next the $ \text{SO(3)} $ set of operations corresponding to object rotations. It turns out that the metric measuring distance on the manifold can be decomposed into a homogeneous part, which varies uniformly with object rotation, plus a residual term, which acts as a fingerprint of the object (see Paper I Sec. III C). Considering the homogeneous part only, the total set of symmetries, is then $ \text{SO(2)} \times \text{SO(3)} $. The same set of symmetries appears in certain models of the universe in general relativity [@Taub51; @Hu73], and is associated with well-known eigenfunctions familiar in the context of spinning tops in classical and quantum mechanics [@BiedenharnLouck81].
The key point here is that the knowledge of the manifold symmetries, which stem from the nature of operations in space, allows one to determine the leading-order properties of the manifold under a very general set of scattering scenarios, including its natural eigenfunctions. Projection of noisy datasets on these eigenfunctions is tantamount to noise discrimination. The components of a data point representing a snapshot can then be directly related to its orientation (see Paper I Sec. III D).
Applications {#secResults}
============
It has long been known that the use of problem-specific constraints can substantially increase computational efficiency [@LeCunEtAl90]. By combining wide applicability with class specificity, symmetries represent a particularly powerful example of such constraints. In Paper I, we used the object-independent symmetries of image formation to recover 3D structure from a large ensemble of simulated, noise-free diffraction snapshots with a computational complexity $10^4\times $ higher than the state of the art. Here we demonstrate the noise robustness stemming from exploiting the symmetries of image formation. Examples include orientation recovery, 3D reconstruction, and movie extraction from ultra-low-signal diffraction or image snapshots of periodic and non-periodic objects and dynamical systems. Each example was selected to highlight an important application area. As shown in Table \[tabApplication\], a variety of manifold-embedding techniques can be used.
.
Data type Observed system Snapshot type Reconstruction Manifold-embedding technique
-------------- -------------------------------- ----------------------- ---------------------- ------------------------------
Simulated Adenylate kinase molecule[^3] Diffraction 3D structure Diffusion Map
Chignolin molecule Diffraction 3D structure Diffusion Map
Experimental Superoxide dismutase-1 crystal Diffraction Orientation recovery Isomap
Chaperonin molecule Cryo-EM images 3D structure GTM
Pirouette Unsorted image frames Time series Diffusion Map
Pas de deux Unsorted image frames Time series Isomap
Structure recovery from simulated diffraction snapshots of non-periodic objects at ultra-low signal {#secSimulatedData}
---------------------------------------------------------------------------------------------------
First, we demonstrate 3D structure recovery from a collection of two million simulated diffraction snapshots of the synthetic protein chignolin (Protein Data Bank (PDB) descriptor: 1UAO, model 1) at $ 4\times 10^{-2} $ scattered photons per Shannon pixel at $1.8 \text{ \AA}$. (A Shannon pixel is of the size needed for appropriate sampling of the intensity distribution as prescribed by the Shannon-Nyquist theorem.) This scattered intensity is expected from a 500 kD protein exposed to a single pulse from an XFEL [@ShneersonEtAl08; @FungEtAl08]. At this signal level, Poisson (shot) noise dominates. The ability to deal with such levels of non-additive noise was previously demonstrated only by Bayesian algorithms [@FungEtAl08; @LohElser09] with extremely unfavorable scaling behavior [@FungEtAl08; @LohElser09; @SchwanderEtAl10a; @GiannakisEtAl11a], restricting the size of amenable objects to eight times the spatial resolution.
Here, we use the symmetry-based approach described in Paper I after modest denoising. The denoising scheme consists of two steps: (1) Convolve the snapshot pixels with a 2D Gaussian filter with a width approximately equal to that of a Shannon pixel; (2) Replace each snapshot vector by an average over its local neighbors. Depending on the SNR, a number iterations of step (2) may be needed, with a stopping criterion based on a least-squares residual determined through the first nine Laplacian eigenfunctions of the dataset (ordered in order of increasing eigenvalue). These eigenfunctions are employed in our scheme to assign an orientation to each snapshot. For details see Appendix \[appendixNoisy\] and Paper I.
To estimate the accuracy of orientation recovery, we use the following measure for root-mean-square (RMS) distance between the deduced and true orientations: $$\label{eqRMSE}
\varepsilon = \left[ \frac{ 1 }{ s ( s - 1 ) } \sum_{i\neq j}( \tilde D_{ij} - D_{ij} )^2 \right]^{1/2},$$ where $ D_{ij} = 2 \arccos( | \tau_i \bm{ \cdot } \tau_j | ) $ and $ \tilde D_{ij} = 2 \arccos( | \tilde \tau_i \bm{ \cdot } \tilde \tau_j | ) $ are the true and estimated internal distances between orientations $ i $ and $ j $, respectively, and $ \bm{ \cdot } $ is the inner product between quaternions. Moreover, to assess the influence of local averaging on the eigenfunctions employed for orientation recovery, we compute the distance $ \gamma $ of the invariant subspace $ \tilde V $ spanned by the leading nine eigenfunctions of the diffusion matrix $ \mathsf{ P }_\epsilon $ in Table \[algP\] from the corresponding invariant subspace $ V $ of the noise-free diffusion matrix [^4]. Note that $ \mathsf{ P }_\epsilon $ has size $ s \times s $, where $ s $ is the number of snapshots in the data set; i.e., $ \tilde V $ and $ V $ are subspaces of $ \mathbb{ R }^s $.
Here, we employ a standard distance measure from matrix perturbation theory [@Stewart73], viz. $$\label{eqDistP}
\gamma = \lVert \tilde \Pi - \Pi \rVert_2,$$ where $ \tilde \Pi $ and $ \Pi $ are orthogonal projectors from $ \mathbb{ R }^s $ to $ \tilde V $ and $ V $, respectively, and $ \lVert \cdot \rVert_2 $ denotes the spectral norm of matrices. With this definition, $ \gamma $ lies in the interval $ [ 0, 1 ] $, and may be interpreted as the sine of an angle characterizing the deviation of $ \tilde V $ from $ V $. For our purposes, Eq. is more appropriate than an error measure based on the difference between the noisy and noise-free diffusion matrices (or their generators), since the latter depends on higher eigenfunctions which are not used in our scheme.
Diffraction snapshots were simulated in $ 2 \times 10^6 $ different orientations to a spatial resolution of $1.8 \text{ \AA}$ using 1 photons. The orientations were sampled approximately uniformly over SO(3), as described in Ref. [@LovisoloDaSilva01]. Cromer-Mann atomic scattering factors [@CromerMann68] were used for the 77 non-hydrogen atoms, and the hydrogen atoms neglected. The detector pixel was the appropriate Shannon pixel [@FungEtAl08], which oversamples the scattered amplitudes by a factor of two, resulting in $ 40 \times 40 = 1600 $ Shannon detector pixels. To model shot noise, diffracted intensities were scaled so that the mean photon count (MPC) per Shannon pixel was 0.04 at $1.8 \text{ \AA}$ resolution. The quantized photon count at each pixel was obtained from a Poisson distribution by the algorithm described in Ref. [@Knuth97].
With no other information, the noisy diffraction patterns were provided to the algorithm in Table \[algNoisy\] (width of Gaussian filter for image smoothing $ \sigma = 0.7 $; number of nearest neighbors in the sparse distance matrix $ d = 220 $; number of nearest neighbors for local averaging $ l = 20 $; number of datapoints for least-squares fitting $ r = 8 \times 10^4 $; number of nearest neighbors for autotuning $ n = 30 $.) As illustrated in Fig. \[figDenoise\], the least-squares residual $ \mathcal{ G }^* $, the subspace distance $ \gamma $, and the RMS orientation recovery error $ \varepsilon $ all decrease monotonically for the first five iterations of local averaging, but exhibit a mild increase at iteration six. At that point the algorithm was terminated in accordance with the stopping criterion described above and in Appendix \[appendixNoisy\]. The minimum $ \varepsilon $ value attained with this choice of parameters at iteration 5 is $ \sim 1.1 \text{ Shannon~angles} $. We measured comparable levels of orientation-recovery accuracy for various combinations of $ l $ and $ n $ parameters in the range 10–50. In all cases, we observed that small values of $ \mathcal{ G }^* $ correlate strongly with small values of $ \epsilon $, indicating that the least-squares residual provides an effective guideline for setting the parameters of the algorithm. This is particularly important, because $ \mathcal{ G }^* $ depends solely on the Laplacian eigenfunctions (see Table \[algAutotuning\]), and, unlike $ \varepsilon $, can be evaluated in an experimental environment where the correct orientations are not known.
![\[figDenoise\] (a) Least-squares residual $ \mathcal{ G }^*$, (b) invariant-subspace distance $ \gamma $, (c) RMS internal distance error $ \varepsilon $, shown as a function of the local-averaging iteration count. In Panel (a), $ \mathcal{ G }^*$ has been normalized by the number of samples $ r = 8 \times 10^4 $ used for least-squares fitting.](fig1)
The quality of orientation recovery was further tested by inverting the reconstructed 3D diffraction volume compiled on a uniform Cartesian grid by an interpolation scheme consistent with the geometry of diffraction [@SchwanderEtAl10b]. The $ R $-factor between the gridded scattering amplitudes $ \tilde F_i $ and those obtained from the Fourier transform of the recovered electron density from phasing, $ F_i $, was defined as $$R = \frac{ \sum_i ( | \tilde F_i |^2 - | F_i|)^2 }{ \sum_i |F_i|^2 }.$$ The 3D electron density obtained by iterative phasing with the <span style="font-variant:small-caps;">superflip</span> algorithm [@PalatinusChapuis07] ($ R=0.20 $) is shown in Fig. \[figChignolin\]. The close agreement with the known structure of chignolin clearly demonstrates sufficient alignment accuracy for reconstruction to $1.8 \text{ \AA}$ resolution. This is on par with the computationally much more expensive Bayesian approaches [@FungEtAl08; @LohElser09; @MothsOurmazd11].
![\[figChignolin\]Three-dimensional electron density of the synthetic protein chignolin, recovered from $ 2 \times 10^6 $ noisy diffraction patterns of unknown orientation at a mean photon count of 0.04 per pixel at $1.8 \text{ \AA}$ resolution. The ball-and-stick model represents the actual structure.](fig2){width="\linewidth"}
Orienting diffraction patterns of crystals {#appSOD}
------------------------------------------
So far we have shown that our symmetry-based approach can be used to orient diffraction patterns from single molecules to high accuracy. We now demonstrate that this approach can also orient diffraction snapshots from crystals. This is important, because recent XFEL-based “diffract-and-destroy” approaches, which use femtosecond X-ray pulses to “outrun radiation damage”, produce diffraction snapshots of nanocrystals of unknown orientation [@ChapmanEtAl11]. As a representative example, we consider a biological crystal of the enzyme superoxide dismutase-1 (SOD1, PDB designation: 1AR4) with $\sim 3\times 10^{3}$ atoms per unit cell, and thus highly complex diffraction patterns. The key issue is whether manifolds produced by diffraction snapshots of crystals are sufficiently homogeneous (possess sufficiently homogeneous metrics) for snapshot orientations to be recovered in a straightforward manner. To demonstrate this point, we intentionally utilized snapshots spanning an orientation range of $ 90^\circ $ so as to produce an open 1D manifold, and analyzed the dataset with the Isomap manifold-embedding method [@TenenbaumEtAl00]. In contrast to Diffusion Map, whose eigenfunctions are insensitive at the boundaries, Isomap maps a 1D open manifold to a straight line segment, and is sensitive to snapshot orientation over the entire range.
Experimental diffraction patterns of a single crystal of superoxide dismutase-1 with a mosaicity of $ 0.8^\circ $ were obtained at the Advanced Photon Source ($\lambda = 0.98 \text{ \AA}$). The crystal was rotated about an arbitrary axis with a step size of $ 0.5^\circ $, and 1800 diffraction patterns recorded over a range of $ 90^\circ $. To compensate for spurious beam intensity fluctuations, the diffraction pattern intensities were normalized. Isomap was used to embed diffracted amplitudes (square-roots of intensities), using two nearest neighbors for calculation of geodesic distances (integrals of the metric). As shown in Fig. \[figSOD\], a one-dimensional and uniformly populated manifold results, with the projection on the first eigenvector linearly proportional to the snapshot orientation to within $ 1^\circ $, compared with the crystal mosaicity of $ 0.8^\circ $. The homogeneity of this manifold (metric) establishes that, in principle, our symmetry-based approach can be used to treat crystalline objects in the same way as non-periodic single particles, provided, of course, object symmetry is appropriately incorporated.
 Typical experimental diffraction pattern of superoxide dismutase-1. (b) The embedding of the geodesic distances in the space of the first three eigenvectors, with each point representing a diffraction pattern. (c) Relation between the correct and the determined orientations.](fig3){width="\linewidth"}
Structure recovery from experimental cryo-electron micrographs {#appCryo}
--------------------------------------------------------------
A well-studied application of graph-theoretic techniques concerns the 3D reconstruction of faint biological objects by single-particle cryo-EM without orientational information. In cryo-EM, the resolution is strongly degraded by radiation damage. As such, the lowest acceptable exposure to electrons and thus SNR must be employed. As described in Sec. \[secPreviousWork\], this has proved a fertile ground for testing new algorithms. By recourse to specific properties of cryo-EM images, impressive results have been obtained, primarily with simulated data. Beyond noise, however, reconstruction by cryo-EM must contend with a range of key issues, chief among them the loss of information due to zero-crossings in the transfer function of the microscope and thus partial loss of information in any single snapshot. The exact position of the zero-crossings depends sensitively on microscope defocus. This offers a means to recoup some of the lost information by insuring that the dataset includes micrographs obtained over a range of defocus values, each with a different set of zero-crossings in the transfer function. The key point is this: the object structure cannot be recovered in full detail from a single defocus, even if the imaging parameters were known exactly. Thus, for a reconstruction algorithm to be of practical use, it must deal with the effect of defocus variations — a test rarely passed by new algorithms. Here, we demonstrate structure recovery from experimental cryo-EM images of the biological molecule chaperonin. Specifically, we incorporate the effect of defocus, use the symmetry-based homogeneity of the manifold metric to deduce orientations, and thence recover the 3D object structure. This demonstration is mitigated by two factors: (1) in order to expedite the calculations, snapshots with only one orientational degree of freedom were selected from a set presorted by a standard orientation algorithm; and (2) to demonstrate structure recovery at ultra-low signal — far below what is normally used — experimental snapshots were preprocessed to simulate such low signal levels.
Randomly oriented single-particle cryo-EM images images of the wild-type group II chaperonin in *methanococcus maripaludis* (Mm-cpn), obtained with a mean incident electron count (MEC) of $ 20/\text{\AA}^2 $ (equivalent to 135 electrons per $2.6 \text{ \AA}$-square snapshot pixel) were kindly provided by Chiu *et al*. [@ZhangEtAl10]. Each snapshot consisted of $ 96 \times 96 $ pixels. A set of 413 side-view snapshots was selected from 5000 images, whose orientations had been previously determined by the <span style="font-variant:small-caps;">eman</span> program [@LudtkeEtAl99], resulting in a dataset with a single orientational degree of freedom about the object symmetry axis.
To investigate the performance of our method at lower dose, a second data set was produced by applying an additional Poisson process to the raw experimental images. The method is based on an approximation valid for low-contrast images with Poisson noise and sufficiently large MEC. The substitution $ I \mapsto I' = \operatorname{Pois}( I^{1/2} ) $ transforms a signal $ I $ to a signal $ I' $, with mean $ \text{MEC}' = \text{MEC}^{1/2} $ and variance $ \operatorname{var}( I' ) = \operatorname{var}( I ) / 4 $. Simulations verified the accurate validity of this approach at MEC = 100, compared with an MEC per snapshot pixel of 135 for our experimental images. Twenty noisy versions of each image were thus generated to form a data set of 8260 images with an effective MEC of 1.7 per $\text{\AA}^2$ .
Since neither the noise-free signal nor the noise variance was known for our experimental cryo-EM images, a method developed by Frank [@Frank06] was used to estimate the SNR directly from the experimental data. This determines the SNR from the cross-correlation coefficient $ C_{ij} $ between two images in the same orientational class using the definition: $$\label{eqSNR_I}
\text{SNR} = 10 \log_{10} \operatorname{mean}( C_{ij} / ( 1 - C_{ij} ) ),$$ where the mean is taken over all classes and all images within each class. Provided two images represent different realizations of noise from an identical object in the same orientation, the above estimate for SNR agrees with the standard definition $ \text{SNR}=10\log_{10} \frac{\operatorname{var}(signal)}{\operatorname{var}(noise)} $ [@FrankAli75]. With the classification obtained from <span style="font-variant:small-caps;">eman</span> [@LudtkeEtAl99], and the assumption that class members differ only in noise, we estimate a SNR of $ -6\text{ dB} $ for the raw experimental snapshots (MEC: $ 20/\text{\AA}^2 $) and a SNR of $ -16\text{ dB} $ for the preprocessed experimental snapshots (MEC: $ 1.7/\text{\AA}^2 $).
As described above, the defocus value and hence transfer function of the electron microscope vary from snapshot to snapshot. To analyze such cryo-EM data, we implemented a modified version of the manifold-embedding algorithm Generative Topographic Mapping (GTM) [@BishopEtAl98; @Svensen98] to explicitly incorporate the effect of the microscope transfer function. GTM defines a manifold in data space by partitioning the noisy dataset into a number of Gaussians each centered around a point (node) on the manifold. The partitioning is based on a nonlinear mapping of a latent space, in this case the space of rotations. GTM is thus, in essence, a manifold-embedding technique, with the symmetries of scattering manifested in the homogeneity of the data manifold, as described in Paper I. However, the generative capability of GTM allows one to construct an image (in essence a model snapshot) at each node on the data manifold. In our approach, this model image extracted from the data corresponds to the aberration-free projected potential of the object. In order to assign an experimental snapshot to a model image, its distance from the model is calculated after convolving the model with the transfer function of the microscope at the defocus corresponding to that of the experimental snapshot. This convolution proceeds efficiently as multiplication in Fourier space, and is not computationally expensive. A similar approach based on more efficient manifold-embedding techniques will be published elsewhere.
The GTM-based approached was first validated with simulated cryo-EM images of chaperonin over a typical experimental defocus range of $10,000 \text{ \AA}$ to $30,000 \text{ \AA}$ (underfocus). The orientations were successfully recovered to within $ 1^\circ $. To reconstruct 3D density maps from experimental snapshots, model aberration-free projected potentials were generated (lifted) at 16 equally-spaced nodes of the data manifold produced by experimental images replicated according to the 8-fold object symmetry. 3D density maps were then reconstructed tomographically using the back-projection algorithm <span style="font-variant:small-caps;">bg cg</span> of the <span style="font-variant:small-caps;">spider</span> software package [@FrankEtAl96]. For comparison, a simulated density map was obtained from 2D snapshots using the known chaperonin atomic coordinates (PDB identifier: 3LOS) under the following imaging conditions: spherical aberration $ C_s = 4.1 \text{ mm} $; defocus $ \Delta f=\text{24,000 \AA} $ (underfocus); electron energy $ E = 300 \text{ keV} $; damping envelope parameter $ B=50 \text{ \AA}^2$; images phase-flipped. The resulting 3D density map was passed through a $ 5 $ Gaussian filter.
Fig. \[figChaperonin\](a) shows a typical experimental snapshot, Fig. \[figChaperonin\](b) the average of the micrographs assigned to an orientation class by the cryo-EM reconstruction software package <span style="font-variant:small-caps;">eman</span> [@LudtkeEtAl99], and Fig. \[figChaperonin\](c) the snapshots oriented by manifold embedding and reconstructed (lifted) from the manifold. (For a movie of the reconstructed tilt series see EPAPS Movie 1). Note that the manifold is able to generate missing images by interpolation. The improved quality of the manifold-generated snapshots compared to the class averages offers the possibility to reconstruct at significantly reduced dose. Fig. \[figChaperonin\](d) is an experimental snapshot preprocessed to approximate snapshots expected from a single chaperonin molecule at a dose $12\times$ lower than commonly used [@ZhangEtAl10] ($\text{SNR}\sim-16 \text{ dB}$, i.e., 10 dB below a typical dose). Fig. \[figChaperonin\](e) is the snapshot lifted from the manifold after orienting an ensemble of 8000 different raw snapshots by manifold embedding. It is clear from this image, the corresponding tilt series (EPAPS Movie 2), and the 3D reconstructions of Fig. \[figChaperonin\](f–h) that snapshots can be successfully oriented by manifold embedding to produce 3D models, even at $ 12\times $ lower signal than in use today. Note that images at this dose could not be oriented by standard cryo-EM approaches [@FrankEtAl96], even when accurately centered prior to analysis, as was performed here. In contrast, our orientation recovery results were similar to those obtained at an MEC of $20/\text{\AA}^2$, indicating that the effect of lower signal levels can be compensated by increasing the number of snapshots.
 Experimental cryo-electron micrograph of a chaperonin molecule at a mean electron count of $20/\text{\AA}^2$ ($ \text{SNR} = -6\text{ dB}$). (b) Image obtained by averaging the members of an orientation class. (c) Image generated (lifted) from the data manifold. (d) Experimental micrograph of chaperonin molecule processed to reflect a mean electron count of $1.7/\text{\AA}^2$ ($\text{SNR} = -16\text{ dB}$). (e) Image lifted from the data manifold. (f) 3D reconstruction with simulated images. (g) 3D reconstruction with lifted images at a snapshot dose of $20/\text{\AA}^2$. (h) 3D reconstruction with lifted images at a snapshot dose of $1.7/\text{\text{\AA}}^2$.](fig4){width="\linewidth"}
Our results thus offer the tantalizing possibility of reducing the snapshot dose in 3D reconstruction techniques using ionizing radiation, in some cases by at least an order of magnitude. This would significantly mitigate the limits set by radiation damage. As a benchmark, the essentially unfulfilled promise of the costly transition of cryo-EM to liquid He temperatures was aimed at improving dose tolerance by a factor of two. The superior signal extraction capability offered by manifold mapping could also be used to obtain images at smaller defocus values in order to reconstruct the object to higher resolution.
Time-series (movies) from ultra-low-signal random-sequence snapshots {#appPirouettePasDeDeux}
--------------------------------------------------------------------
Our knowledge of the precise time at which an experimental snapshot of a dynamic system was obtained is corrupted by inevitable uncertainties, which can substantially exceed the intrinsic time resolution of the observation technique. Modern pump-probe experiments, for example, can now be performed with pulses as short as a few femtoseconds, but their time-resolution is often determined by timing jitter, which can be up to two orders of magnitude larger [@Glownia:10; @Cryan10]. When the state of a system under observation is not synchronized with the observation windows, a sequence of snapshots can represent random sightings of the system during its evolution. It is thus important to develop means for deducing “time stamps” directly from the data, either to reduce jitter-induced uncertainty, or to order a sequence of snapshots according to the intrinsic evolution of the system under observation. Here, we demonstrate this capability at SNRs as low as $-21\text{ dB}$. Specifically, we show that: (1) Randomized movie sequences can be time-ordered, even when the signal is extremely low; and (2) Frames generated (lifted) from the manifold produce movies of superior quality.
Movies of a pirouette and a pas de deux were downloaded from the web. These represent optical snapshots of a conformationally rigid body in rotations, and the evolution of two flexible bodies in interaction, respectively. In order to reduce the SNR, a constant background was added, and shot noise incorporated at each pixel depending on its intensity value, as described in Ref. [@Knuth97]. For the pirouette, a sequence of 16 turns consisting of 268 frames ($ 210\times160 $ pixels each) was replicated 132 times, a background $ 5\times $ the mean intensity added, and shot noise incorporated to produce an effective mean photon count per pixel of 0.08 and a SNR of $-21\text{ dB}$ (see Eq. \[eqSNR\_I\]). For the pas de deux, a sequence of 870 frames ($ 265\times305 $ pixels each) was replicated 12 times with an added background of twice the mean intensity, and shot noise incorporated to produce a mean photon count of 0.8 and a SNR of $ -11\text{ dB} $. For both movies, camera motion was corrected by reference to a stationary marker.
Each random sequence was ordered by a suitable manifold-embedding technique (Diffusion Map or Isomap). Using the generative property of GTM, images were then lifted from the manifold. As described in Ref. [@SchwanderEtAl10a] and demonstrated in Fig. \[figChaperonin\](c,e), this procedure uses the information content of the entire dataset to generate each snapshot, producing images of significantly higher quality than possible by traditional classifying and averaging techniques. It is also more robust against non-uniform sampling and jitter. Table \[algLifting\] summarizes the lifting procedure.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
**Inputs:** Noisy snapshots $ \mathcal{ M }_I = \{ \underline{I}_1,\ldots,\underline{I}_s \} $ Estimated quaternions $ \mathcal{ T } = \{ \tau_1, \ldots, \tau_s \} $ number of nodes $ K $ number of basis functions $ M $ basis function width $ \sigma $ **Outputs:** Manifold-lifted images $ \mathcal{ M } = \{ \underline{a}_1, \ldots, \underline{a}_s \} $ Generate the grid of latent points $ \{ x_1, \ldots, x_K \} $. Generate the grid of basis function centers $ \{ \mu_1, \ldots, \mu_M \} $. Compute the matrix of basis function activations $ \Phi $ such that $$\Phi_{m}( x ) = \exp( - ( x - \mu_m )^2/ 2 \sigma^2 ).$$ Initialize a set of weights $ W $ using principal component analysis. Initialize inverse Gaussian noise variances $\alpha $ and $\beta$. Compute a set of responsibilities $ R $ that assigns each snapshot to a node from the results of manifold embedding. Compute the diagonal matrix $ G $ using $ R $, where $ G_{kk} = \sum_{i=1}^s R_{ki} $. $ W \leftarrow ( \Phi^\mathrm{ T } G \Phi + \lambda I )^{-1} \Phi^\mathrm{ T } R \mathcal{ M }_I $, where the regularization parameter $ \lambda $ may be zero. Compute $ \Delta $, where $ \Delta_{kn} = \lVert \underline{ I }_n - \Phi_k W \rVert^2 $. Compute $ \gamma $ from $ \lambda $ and $ \alpha $. Update $ \alpha $ and $ \beta $ using $ \gamma $, $ R $ and $ \Delta $. $ \mathcal{ M } \leftarrow \Phi W $ $ \mathcal{ M } $
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: \[algLifting\]Manifold-lifting algorithm based on GTM
For the pirouette, Diffusion Map was used to recover the object orientation in each frame during the dance, and hence the sequence order (number of nearest neighbors in the sparse distance matrix $ d = 5896 $; Gaussian kernel bandwidth $ \epsilon = 200 $.) Snapshots were then lifted from the manifold (number of nodes $ K = 28 $; number of basis functions $ M = 14 $, basis function width $ \sigma = 2 $.) Reconstructed images are shown in Fig. \[figPirouette\] together with a sequence of unsorted, unprocessed snapshots. The movie is available as EPAPS Movie 3. The randomized pas de deux sequence was ordered with Isomap (number of nearest neighbors $ d = 33 $.) To compile the movie, images were lifted from an ordered sequence of 870 points (nodes), corresponding to uniform sampling on the Isomap manifold. Reconstructed images are shown in Fig. \[figPasDeDeux\] together with a sequence of unsorted snapshots. The movie is available as EPAPS Movie 4.
Figs. \[figPirouette\] and \[figPasDeDeux\], and the associated movies clearly show that our approach is able to determine the correct frame sequence and generate high quality snapshots at signal levels as low as 0.08 photon/pixel for the pirouette (modulo one revolution), and 0.8 photon/pixel for the pas de deux, both with added background and non-additive noise corresponding to signal-to-noise ratios in the range $-11 $ to $-21\text{ dB} $. These examples demonstrate the capability to determine the time evolution of systems from unsorted random sightings at extremely low signal. They also highlight the potential to correct timing jitter in pump-probe experiments, and reconstruct the evolution of dynamic systems from random sightings of members of a heterogeneous set, each at a different stage of its evolution. These possibilities will be described in detail elsewhere. The general implications for signal extraction and image processing are clear.
![\[figPirouette\]Top row: First five frames of 35,000 randomly-sequenced snapshots of a pirouette preprocessed to reflect a mean photon count of 0.08 per pixel with added background and shot noise, corresponding to a signal-to-noise ratio of $ -21 $ dB. Bottom row: Five evenly-spaced images extracted from the Diffusion Map manifold. (See also EPAPS Movie 3.)](fig5){width="\linewidth"}
![\[figPasDeDeux\]Top row: First four frames of 10,000 randomly-sequenced snapshots of a pas de deux preprocessed to reflect a mean photon count of 0.8 per pixel with added background and shot noise, corresponding to a signal-to-noise ratio of $ -11 $ dB. Bottom row: Four evenly-spaced images extracted from the Isomap manifold. (See also EPAPS Movie 4.)](fig6){width="\linewidth"}
Conclusions {#secConclusions}
===========
We have shown that manifold mapping, as described in Paper I, augmented with modest noise reduction measures, is able to extract structural and timing information from simulated and experimental snapshots at extremely low signal. The ability to orient simulated and experimental diffraction and image snapshots confirms the accessibility of the homogeneous manifold expected from our theoretical framework for a wide range of objects and imaging scenarios, including crystalline samples. The capability to recover 3D structure at extremely low signal is on par with the more expensive Bayesian approaches, but offers greater reach in terms of sample size and resolution, as demonstrated in Paper I. The noise-robustness of our approach substantially exceeds what has been demonstrated with comparable graph-theoretic approaches without restrictive, application-specific assumptions. The manifold itself offers a powerful route to image reconstruction at low signal, because snapshots reconstructed from the manifold achieve higher signal-to-noise ratios than possible by traditional approaches based on classification and averaging. Taken together, these offer a physically-based, computationally efficient, noise-robust route to analyzing the large and varied datasets generated by existing and emerging structure recovery methods. In the longer term, it should be possible to use these approaches to recover or improve timing information in pump-probe experiments, and construct 3D movies (4D maps) from random sightings of members of structurally heterogeneous and dynamically evolving ensembles.
We are grateful for experimental data and advice to W. Chiu and J. Zhang (cryo-EM images) and M. Schmidt (crystallographic data), and acknowledge discussions with G. N. Phillips, Jr., R. Rosner, W. Schr" oter, and members of the UWM Physics Department. We are grateful to J. Frank and H. Y. Liao for assistance with 3D reconstruction by back-projection. One of us (DG) is grateful to the Random Shapes Program held in 2007 at the Institute for Pure & Applied Mathematics. This work was partially supported by the U.S. Department of Energy Office of Science (SC-22, BES) award \#DE-SC0002164 and a UWM Research Growth Initiative award.
Treatment of noise {#appendixNoisy}
==================
In the manifold picture, noise can be described as a perturbation of the noise-free manifold $\mathcal{ M } = \{ \underline{ a }_1, \underline{ a }_2, \ldots, \underline{ a }_{s} \} $, where $ \underline{ a }_i $ is a snapshot vector of measured pixel amplitudes, viz.
$$\begin{gathered}
\label{eqSnapshotNoisy}
\underline{ a }_i \mapsto \underline{ \tilde a }_i = \kappa \underline{ a }_i + \delta \underline{ a }_i, \\
\delta \underline{ a }_i = ( \delta a_{i1}, \delta a_{i2}, \ldots, \delta a_{in} )^\mathrm{T}, \quad \delta a_{ij} = I_{ij}^{1/2} - \kappa a_{ij}\;.
\end{gathered}$$
This causes the observed, noisy, dataset $$\label{eqDataSetANoisy}
\tilde{ \mathcal{ M } } = \{ \underline{ \tilde{ a } }_1, \underline{ \tilde{ a } }_2, \ldots, \underline{ \tilde{ a } }_{s} \}$$ not to lie exactly on the manifold $ \mathcal{M} $ (up to a global scaling by $ \kappa $). One would expect that if the perturbation norm $ \lVert \delta \underline{ a }_i \rVert $ becomes comparable to the kernel bandwidth $ \epsilon^{1/2} $, the computed eigenfunctions $ \underline{ \psi }_k $ are distorted to the point that the embedded manifold no longer has the topology of SO(3) [@CoifmanLafon06; @BalasubramanianSchwartz02]. Indeed, as illustrated in Fig. \[figDenoise\] direct application of the algorithm for noise-free data (see Paper I Table III) to a noisy dataset $ \tilde{ \mathcal{ M } } $ at an $ \text{MPC} = \operatorname{O}( 10^{-2} ) $ results in poor accuracy. In order to be practically useful, the noise-free orientation-recovery scheme must be augmented by a suitable denoising method.
Conceptually, we denoise the data in three steps: (1) Low-pass filtering of each snapshot by convolution of pixel intensities with a 2D Gaussian; (2) Variance-stabilizing transformation (VST); and (3) Local averaging over nearest neighbors in data space prior to embedding. In practice, we combine (1) and (2), known to be effective for shot noise [@ZhangEtAl07; @ZhangEtAl08; @Guan09], into a single step, and follow the iterative procedure described below.
Low-pass Gaussian filtering and variance-stabilizing transformation (VST)
-------------------------------------------------------------------------
Convolution with a low-pass Gaussian filter of bandwidth $ \sigma $ is represented by:
\[eqGaussianFiltering\] $$\underline{ I } \mapsto ( I \ast H_\sigma )( \vec{ r } ) = \int d \vec{ r }' \, I( \vec{ r }' ) H_\sigma( \vec{ r } - \vec{ r }' ),$$ with $$H_\sigma( \vec{ r } ) = \exp( - \lVert \vec{ r } \rVert^2 / 2 \sigma^2 ) / ( 2 \pi \sigma^2 )^{1/2}.$$
VST, proposed by Guan [@Guan09] for low-intensity data, can be written as: $$\label{eqVST1}
\underline{ I } \mapsto \underline{ I }^{1/2} + ( \underline{ I } + 1 )^{1/2}.$$ $ I( \vec{ r } ) $ denotes the (discretely sampled) intensity pattern on the detector plane obtained by “unpacking” the column vector of intensities $ \underline{ I } $.
Eqs. and are combined into a single operation: $$\label{eqVST}
\operatorname{VST}( I; \sigma ) = ( I \ast H_\sigma )^{1/2} + [ ( I \ast H_\sigma ) + 1 ]^{1/2}.$$ Given an intensity-pattern dataset $ \mathcal{ M }_I $ consisting of $ s $ samples and an index matrix of nearest neighbors $ \mathsf{ N } $ of dimensions $ s \times l $, we introduce a combined VST and aggregation operation taking $ \mathcal{ M }_I $ to a dataset $ \tilde{ \mathcal{ M } }= \operatorname{VSTL}( \mathcal{ M }_I; \sigma, \mathsf{ N } ) $ such that $$\label{eqVSTL}
\tilde{ \mathcal{ M } } = \{ \underline{ \tilde{ a } }_1, \underline{ \tilde{ a } }_2, \ldots, \underline{ \tilde{ a } }_{s} \}, \quad \underline{ \tilde{ a } }_i = \operatorname{VST}\left( \sum_{k=1}^{l} \underline{ I }_{N_{ik}}; \sigma \right) .$$ Noise robustness can be further enhanced by a so-called self-tuning Gaussian kernel introduced by Zelnik-Manor and Perona [@ZelnikManorPerona04]. Here, instead of the isotropic Gaussian kernel $ \mathcal{ K }( \underline{ a }_i, \underline{ a }_j ) = \exp( - \lVert \underline{ a }_i - \underline{ a }_j \rVert^2 / \epsilon ) $ of Eq. (B3) in Paper I, one uses an anisotropic Gaussian kernel with local scaling parameters, $ \epsilon_i $, given by $$\label{eqAutotuningKernel}
\mathcal{ K }( \underline{ a }_i, \underline{ a }_j ) = \exp( - \lVert \underline{ a }_i - \underline{ a }_j \rVert^2 / ( \epsilon_i \epsilon_j )^{1/2} ).$$ A canonical choice for the scaling parameters, which we adopt throughout, is $ \epsilon_i = \lVert \underline{ a }_i - \underline{ a }_{N_{il}} \rVert^2 $, where, as usual, $ N_{il} $ denotes the index of the $ l $-th nearest neighbor of datapoint $ i $.
Iterative local averaging
-------------------------
If the true nearest neighbors of a point in data space are known, and noise produces no systematic bias, local averaging approaches the true manifold. To see this, let $ \varepsilon $ be an error tolerance in data space, and consider a reference orientation $ \mathsf{ R } $ with corresponding noise-free snapshot $ \underline{ a } $. For any $ \varepsilon > 0 $ it is possible to find a ball $ B_\varepsilon $ in data space centered at $ \underline{ a } $, such that for any countable set $ \{ \underline{ a }_1, \ldots, \underline{ a_l } \} $ of noise-free snapshots lying in $ B_\varepsilon $ the mean, $ \bar{ \underline{ a } } = \sum_{i=1}^l \underline{ a }_i / l $, has error $ \lVert \bar{ \underline{ a } } - \underline{ a } \rVert < \varepsilon $. In the presence of noise, the $ \underline{ a }_i $ are replaced by the random variables in Eq. ; i.e., $ \underline{ a }_i \mapsto \tilde{ \underline{ a } }_i $, where $ \tilde{ \underline{ a } }_i $ are statistically independent, have expectation value $ \kappa^{1/2} \underline{ a }_i $ proportional to $ \underline{ a }_i $, and finite variance $ \Delta \underline{ a }_i^2 $. Moreover, the sample mean within $ B_\varepsilon $ becomes a random variable $ \hat{ \underline{ a } } = \sum_{i=1}^l \tilde{ \underline{ a } }_i / l $ with expectation value $ \kappa^{1/2} \bar{ \underline{ a } } $ and variance $ \Delta \underline{ a }^ 2 = \sum_{i=1}^l \Delta \underline{ a }_i^2 / l $. By the law of large numbers, in the limit of an infinite data set with infinite snapshots in $ B_\varepsilon $ (i.e., $ l \to \infty $), $ \hat{ \underline{ a } }_i $ is equal to $ \bar{ \underline{ a } }_i $ (up to an unimportant proportionality constant) with probability one. Thus, recovery of the data manifold with error $ \varepsilon $ is possible almost surely.
In practice, noise corrupts the local neighborhood relations, and without *a priori* information, it is not possible to identify which of the snapshots in a noisy data set are associated with the ball $ B_\varepsilon $ of the underlying noise-free system. Therefore, it cannot be guaranteed that local averaging leads to the correct manifold. We therefore exploit our knowledge of the natural eigenfunctions of scattering manifolds to monitor the effect of local averaging, and terminate the procedure before substantial deviations have occurred.
Specifically, we follow the algorithm described in Table \[algNoisy\]. First, we apply the $ \operatorname{VST}$ operation to the intensity data $ \{ \underline{ I }_i \} = \mathcal{ M }_I $, setting the filter width $ \sigma $ to a relatively small value (e.g., in Sec. \[secSimulatedData\], $ \sigma $ is set to $7/10$ of a pixel width). The autotuning version of the orientation-recovery method (Table \[algAutotuning\]) is executed using the $ \operatorname{VST}$-filtered intensities as input data. The nearest-neighbor indices $ \mathsf{ N }_0 $ obtained in the course of the calculation of the sparse distance matrix then become our initial estimate for the true nearest-neighbor indices. We also record the residual of the nonlinear least-squares output, $ \mathcal{ G }^*_{0} $, and choose a value $ l \leq d $ for the number of nearest neighbors for local averaging.
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
**Inputs:** Noisy snapshots $ \mathcal{ M }_I = \{ \underline{ I }_1, \ldots, \underline{ I }_s \} $ Number of retained nearest neighbors $ d $ Number of nearest neighbors for local averaging Number of datapoints in the least-squares fit, $ r $ Number of nearest neighbors for autotuning, $ n $ Gaussian filter bandwidth $ \sigma $ **Outputs:** Estimated quaternions $ \mathcal{ T } = \{ \tau_1, \ldots, \tau_s \} $ Estimated nearest-neighbor index matrix $ \mathsf{ N } $ Least-squares residual $\mathcal{ G }^* $ $ \underline{ \tilde a }_i \leftarrow \operatorname{VST}( \underline{ I }_i; \sigma ) $ $ \tilde{ \mathcal{ M } }_0 \leftarrow \left\{ \underline{ \tilde a }_i \right\} $ Execute the algorithm in Table \[algAutotuning\] with input data $ \tilde{\mathcal{ M }}_{0} $; store the returned nearest-neighbor index matrix as $ \mathsf{ N }_{0} $ and the least squares residual as $ \mathcal{ G }^*_{0} $. $ i \leftarrow 1 $ $ {\mbox{\textit{terminate}}}\leftarrow \text{false} $ $ \tilde{\mathcal{ M }}_{i} \leftarrow \operatorname{VSTL}( \mathcal{ M }_I; \sigma, \mathsf{ N }_{i-1} ) $ Execute the algorithm in Table \[algAutotuning\] with input data $ \tilde{\mathcal{M}}_{i} $; store the outputs as $ \mathcal{ T }_{i} $, $ \mathsf{ N }_{i} $, and $ \mathcal{ G }^*_{i} $. $ {\mbox{\textit{terminate}}}\leftarrow \mathcal{ G }^*_{i} > \mathcal{ G }^*_{i-1} $ $ i \leftarrow i + 1 $ $ \mathcal{ T } \leftarrow \mathcal{T}_{i-1} $ $ \mathcal{ G }^* \leftarrow \mathcal{ G }^*_{i-1} $ $ \mathsf{ N } \leftarrow \mathsf{ N }_{(i-1)} $ $ \mathcal{T} $, $ \mathcal{ G }^* $, $ \mathsf{ N } $.
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: \[algNoisy\]Orientation-recovery for noisy snapshots
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
**Inputs:** Snapshots $ \mathcal{ M } = \{ \underline{ a }_1, \ldots, \underline{ a }_s \} $ Number of retained nearest neighbors $ d $ Number of datapoints in the least-squares fit, $ r $ Number of nearest neighbors for autotuning, $ n $ **Outputs:** Estimated quaternions $ \mathcal{ T } = \{ \tau_1, \ldots, \tau_{s} \} $, Nearest-neighbor index matrix $ \mathsf{ N } $ Least-squares residual $\mathcal{ G }^* $ Compute the $ s \times d $ matrices $ \mathsf{ N } $ and $ \mathsf{ S } $ such that $$\begin{aligned}
N_{ij} &= \text{index of $j$-th nearest neighbor to snapshot $ \underline{ a }_i $},\\
S_{ij} &= \lVert \underline{ a }_i - \underline{ a }_{N_{ij}} \rVert.
\end{aligned}$$ $ \mathsf{ N } $ Rescale the distance data by the $ n $-th nearest neighbors: $$S_{ij} \leftarrow \left( S_{i,N_{i,n }} S_{j,N_{j,n}} \right)^{1/2}.$$ Compute the sparse transition probability matrix $ \mathsf{ P } $ using the algorithm in Table \[algP\] with inputs $ \mathsf{ S } $, $ \mathsf{ N } $, $ \epsilon $, and $ \alpha = 1 $. Solve the sparse eigenvalue problem $ \mathsf{ P } \underline{ \psi }_k = \lambda_k \underline{ \psi }_k $ for $ 0 \leq k \leq 9 $ and $ 1 = \lambda_0 < \lambda_1 \leq \cdots \leq \lambda_9 $. Solve the nonlinear least-squares problem $$\mathcal{ G }( \{ c_{ijk} \} ) = \sum_{l=1}^{ r } \lVert \tilde{ \mathsf{ R } }_l^\mathrm{T} \tilde{ \mathsf{ R } }_l - \mathsf{ I } \rVert^2 + | \det( \tilde{ \mathsf{ R } }_l ) - 1 |^2 \text{ with } [ \tilde{ \mathsf{ R } }_l ]_{ij} = \sum_{k=1}^9 c_{ijk} \psi_{lk}.$$ $ \mathcal{ G }^* $ Compute an approximate SO(3) matrix $ \tilde{ \mathsf{ R } }_i $ for snapshot $ \underline{ a }_i $ Project $ \tilde{ \mathsf{ R } }_i $ to an orthogonal matrix Convert $ \mathsf{ R }_i $ to a unit quaternion $ \tau_i $ $ \tau_i $
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: \[algAutotuning\]Orientation-recovery using a self-tuning kernel
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
**Inputs:** $ s \times d $ distance matrix $ \mathsf{ S } $ $ s \times d $ nearest-neighbor index matrix $ \mathsf{ N } $ Gaussian width $ \epsilon $ Normalization parameter $ \alpha $ **Outputs:** $ s \times s $ sparse transition probability matrix $ \mathsf{ P } $ Construct an $ s \times s $ sparse symmetric weight matrix $ \mathsf{ W } $, such that $$W_{ij} = \begin{cases}
1, & \text{if $ i = j $}, \\
\exp( - S_{ik}^2 / \epsilon ), & \text{if $ j = N_{ik} $}, \\
W_{ji}, & \text{if $ W_{ij} \neq 0$}, \\
0, & \text{otherwise}.
\end{cases}$$ Evaluate the $ s \times s $ diagonal matrix $ \mathsf{ Q } $ with nonzero elements $ Q_{ii} = \sum_{j=1}^s W_{ij} $. Form the anisotropic kernel matrix $ \mathsf{ K } = \mathsf{ Q }^{-\alpha} \mathsf{ W } \mathsf{ Q }^{-\alpha} $. Evaluate the $ s \times s $ diagonal matrix $ \mathsf{ D } $ with nonzero elements $ D_{ii} = \sum_{j=1}^s K_{ij} $. $ \mathsf{ P }_\epsilon = \mathsf{ D }^{-1} \mathsf{ K } $
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: \[algP\]Calculation of the sparse transition probability matrix $ \mathsf{ P }_\epsilon $ in Diffusion Map, reproduced from Paper I for convenience.
Next, we enter an iteration loop, where in the $ i $-th step the dataset $$\label{eqAIterate}
\tilde{ \mathcal{ M } }_i = \operatorname{VSTL}( \mathcal{ M }_I; \sigma, \mathsf{ N }_{i-1} )$$ is computed, and the algorithm in Table \[algAutotuning\] executed using $ \tilde{ \mathcal{ M } }_{i} $ as input data. The resulting quaternion estimates, nearest-neighbor indices, and least-squares residual are respectively designated $ \mathcal{ T }_i $, $ \mathsf{ N }_{i} $, and $ \mathcal{ G }_{i}^* $. Note that the residual $ \mathcal{ G }_{i}^* $ is a measure of the difference between the eigenfunctions obtained by embedding and the natural eigenfunctions (Wigner $ D $-functions) expected on the basis of symmetry (see Paper I).
If, after an iteration $ i $, $ \mathcal{ G }^*_{i} $ is larger than the residual $ \mathcal{ G }^*_{i-1} $ encountered in the previous step, the loop is terminated. Otherwise, the iteration is repeated using $ \mathsf{ N }_{i} $ as an updated estimate of the true nearest-neighbor indices. Our final orientation (quaternion) assignment is the one corresponding to the minimum least-squares residual, reached in the iteration prior to the termination step.
The empirical evidence in Sec. \[secSimulatedData\] clearly shows that, given a sufficiently large number of sample points, the scheme, applied only a handful of times and terminated using the value of $ \mathcal{ G }_i^* $ as a criterion, provides noise reduction sufficient for accurate orientation recovery at $ \text{MPC} = \operatorname{O}( 10^{-2} ) $.
A potentially fruitful way of interpreting mathematically the success of the process (which lies outside the scope of the present paper) would be to explore its connections with mutually reinforcing models (MRMs) for graph filtering [@ChenSafro10]. This type of model involves iteratively replacing vertices of graphs with weighted averages, whereby the vertex itself and its local neighborhood exert an influence on the vertex in the course of iterative updates. In certain applications, the iterative process in an MRM is terminated after only a small number of iterations. Both of these two features are present in the scheme presented here.
Computational resources {#appendixComputationalResources}
=======================
The calculations reported in this work were primarily performed on a Rocks cluster with 30 nodes, each consisting of two 2.5 GHz Quad-Core Intel Xeon CPUs with 16 GB RAM. Algorithms were usually implemented in <span style="font-variant:small-caps;">matlab r2009b</span> with the Parallel Computing Toolbox together with the <span style="font-variant:small-caps;">matlab</span> Distributed Server using up to 120 workers (parallel processes). For less intensive calculations, a Linux workstation with a 2.66 GHz Quad-Core Intel Xeon CPU, 32 GB RAM and/or a Mac Pro $ 2 \times 2.8 $ GHz Quad-Core Intel Xeon CPUs, 10 GB RAM were used. In Diffusion Map, by far the most CPU-intensive calculations are: (1) the determination of the Euclidean distances of snapshots; and (2) setting up of the sparse distance matrix of nearest neighbors. These calculations were performed in parallel using 100 workers. Such a distance calculation involving $ 2 \times 10^6 $ snapshots typically takes 7 hours for chignolin and 48 hours for adenylate kinase (ADK) in Paper I. Other calculations, including the eigenvector determination and the estimation of the orientation matrices were performed on the Linux workstation in about 8 hours altogether. In total, the orientation determination for $ 2 \times 10^6 $ snapshots requires 56 hours for noise-free ADK and 33 hours for noisy chignolin with 5 local-averaging iterations. Compiling a 3D diffraction volume consisting of a uniform Cartesian grid was implemented in parallel code, with an execution time of less than three hours for two million ADK diffraction snapshots using 80 workers. Diffraction patterns and cryo-EM images were simulated on the cluster and on the Mac Pro. The GTM and phasing algorithms were performed on the Linux workstation and/or the Mac Pro. The <span style="font-variant:small-caps;">chimera</span> package [@Chimera] was used to visualize electron density maps.
[47]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [ ()]{}, @noop [****, ()]{} [****, ()](\doibase 10.1038/nphys1129) @noop [****, ()]{} @noop @noop [****, ()]{} [****, ()](\doibase 10.1038/nmeth992) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [**]{} (, ) @noop [****, ()]{} @noop [****, ()]{} @noop [**]{} (, , ) in @noop [**]{} (, , ) p. @noop [**** ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [**]{}, ed., Vol. (, ) @noop [“,” ]{} (), @noop [****, ()]{} [ ()](http://journals.iucr.org/a/issues/2011/05/00/issconts.html), @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} **, @noop [Ph.D. thesis]{}, () @noop [****, ()]{} [****, ()](\doibase 10.1364/OE.18.017620) [****, ()](\doibase 10.1103/PhysRevLett.105.083004) @noop [****, ()]{} in @noop [**]{}, Vol. () p. @noop [****, ()]{} @noop [****, ()]{} in @noop [**]{}, Vol. () p. @noop [ ()]{}, @noop
[^1]: These authors contributed equally.
[^2]: These authors contributed equally.
[^3]: see Paper I Sec. IV A.
[^4]: $ V $ and $ \tilde V $ consist of all linear combinations of the form $ \protect\sum_{k=1}^9 c_k \protect\underline{ \psi }_k $ where $ \protect\underline{ \psi }_k$ are the first nine eigenvectors of $ \mathsf{ P }_\epsilon $ ordered in order of increasing eigenvalue.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'There has been a surge of recent interest in the role of anisotropy in interaction-induced phenomena in two-dimensional (2D) charged carrier systems. A fundamental question is how an anisotropy in the energy-band structure of the carriers at zero magnetic field affects the properties of the interacting particles at high fields, in particular of the composite fermions (CFs) and the fractional quantum Hall states (FQHSs). We demonstrate here tunable anisotropy for holes and hole-flux CFs confined to GaAs quantum wells, via applying *in situ* in-plane strain and measuring their Fermi wavevector anisotropy through commensurability oscillations. For strains on the order of $10^{-4}$ we observe significant deformations of the shapes of the Fermi contours for both holes and CFs. The measured Fermi contour anisotropy for CFs at high magnetic field ($\alpha_\mathrm{CF}$) is less than the anisotropy of their low-field hole (fermion) counterparts ($\alpha_\mathrm{F}$), and closely follows the relation: $\alpha_\mathrm{CF} = \sqrt{\alpha_\mathrm{F}}$. The energy gap measured for the $\nu = 2/3$ FQHS, on the other hand, is nearly unaffected by the Fermi contour anisotropy up to $\alpha_\mathrm{F} \sim 3.3$, the highest anisotropy achieved in our experiments.'
author:
- Insun
- 'K. A. Villegas'
- 'M. A.'
- 'L. N.'
- 'K. W.'
- 'K. W.'
- 'R.'
- Medini
- 'M.'
title: Transference of Fermi Contour Anisotropy to Composite Fermions
---
High-mobility, two-dimensional (2D), charged carriers at high perpendicular magnetic fields $B$ and low temperatures exhibit rich many-body physics driven by Coulomb interaction. Examples include the fractional quantum Hall state (FQHS), Wigner crystal, and stripe phase [@Shayegan.Review.2006; @Jain.Book.2007]. Recently, the role of *anisotropy* has become a focus of new studies [@Balagurov.PRB.2000; @Shayegan.PSSB.2006; @Gokmen.NatPhys.2010; @Fradkin.ARCMP.2010; @Xia.NatPhys.2011; @Koduvayur.PRL.2011; @Liu.PRB.2013; @Haldane.PRL.2011; @Kamburov.PRL.2013; @Kamburov.PRB.2014; @Mueed.PRL.2015a; @Mueed.PRL.2015b; @Mueed.PRB.2016; @Mulligan.PRB.2010; @Samkharadze.NatPhy.2016; @Wang.PRB.2012; @Qiu.PRB.2012; @BYang.PRB.2012; @KYang.PRB.2013; @Balram.PRB.2016; @Johri.NJP.2016]. This interest has been amplified by the recognition that, although the FQHSs at fillings $1/q$ ($q=$ odd integer) are well described by Laughlin’s wave function with a rotational symmetry [@Laughlin.PRL.1983], there is a geometric degree of freedom associated with the anisotropy of the 2D carrier system [@Haldane.PRL.2011].
\[b!\]
The fundamental issue we address here is how the anisotropy of the energy-band structure of the low-field carriers transfers to the interacting particles at high $B$ and, in particular, to the FQHSs and composite fermions (CFs). The latter are electron-flux quasi-particles that form a Fermi sea at a half-filled Landau level [@Jain.Book.2007; @Halperin.PRB.1993], and provide a simple explanation for the nearby FQHSs [@Jain.PRL.1989]. There is no clear theoretical verdict yet. While some theories predict that the CF Fermi contour anisotropy ($\alpha_\mathrm{CF}$) should be the same as the zero-field (fermion) contour anisotropy ($\alpha_\mathrm{F}$) [@Balagurov.PRB.2000; @Balram.PRB.2016], others conclude that $\alpha_\mathrm{CF}$ is noticeably smaller than $\alpha_\mathrm{F}$ [@BYang.PRB.2012; @KYang.PRB.2013; @Johri.NJP.2016]. This question was also addressed in several recent experimental studies. For 2D electrons occupying AlAs conduction-band valleys with an anisotropic effective mass, a pronounced *transport* anisotropy was reported for CFs, but the anisotropy of the CF *Fermi contour* could not be measured because of the insufficient sample quality [@Gokmen.NatPhys.2010]. More recently, experiments probed the Fermi contour anisotropy of low-field carriers (both electrons and holes), and of CFs in GaAs quantum wells by subjecting them to an additional parallel magnetic field ($B_\|$) [@Kamburov.PRL.2013; @Kamburov.PRB.2014; @Mueed.PRL.2015a; @Mueed.PRL.2015b; @Mueed.PRB.2016]. However, the $B_\|$-induced anisotropy is primarily caused by the coupling between the in-plane and out-of-plane motions of the carriers, rendering a theoretical understanding of the data challenging. Furthermore, a strong $B_\|$ can lead to a bilayer-like charge distribution [@Mueed.PRL.2015b].
As highlighted in Fig. 1, we demonstrate a simple yet powerful technique to tune and probe the anisotropy of both low-field carriers and high-field CFs *without* applying $B_\|$. The experiments consist of subjecting the sample, a GaAs 2D hole system (2DHS), to strain [@fnote1; @Shayegan.APL.2003; @Shkolnikov.APL.2004; @Supple] and measuring $\alpha_\mathrm{F}$ and $\alpha_\mathrm{CF}$, via commensurability oscillations measurements. We find that, for a given value of strain, CFs are less anisotropic than their low-field 2D hole counterparts, and the anisotropies are related through a simple empirical relation: $\alpha_\mathrm{CF} = \sqrt{\alpha_\mathrm{F}}$. In contrast, the measured energy gap of the $\nu = 2/3$ FQHS remains almost constant even for $\alpha_\mathrm{F}$ as large as 3.3. Our results allow a direct and quantitative comparison with theoretical predictions.
Figure 1(a) shows the results of numerical calculations for the strain-induced Fermi contour anisotropy of our sample, a 2DHS confined to a 175-[[Å]{}]{}-wide GaAs (001) quantum well [@fnote2; @Kamburov.PRB.2012; @Kamburov.PRL.2012]. The self-consistent calculations are based on an $8\times8$ Kane Hamiltonian [@Supple; @Bir.Book; @Winkler.Book]. Without strain ($\epsilon = 0$), the Fermi contour of holes is four-fold symmetric but is split into two contours because of the spin-orbit interaction [@Winkler.Book]. The minority-spin contour is nearly circular while the majority-spin contour is warped. When tensile strain ($\epsilon > 0$) is applied along $[\bar{1}10]$, the hole Fermi contours become elongated along $[\bar{1}10]$ and shrink along the $[110]$ direction [@Supple; @Bir.Book; @Winkler.Book; @Kolokolov.PRB.1999; @Habib.PRB.2007; @Shabani.PRL.2008]. On the other hand, compressible strain ($\epsilon < 0$) has the opposite effect \[Fig. 1(a)\]. Our experimental setup for applying *in situ* tunable strain to the sample is shown in Figs. 1(b) and 1(c) [@Shayegan.APL.2003]. An L-shaped Hall bar is etched into the GaAs wafer which is thinned to $\sim$120 $\mu$m and glued on one surface of a stacked piezo-actuator. When a voltage $V_{P} > 0$ ($V_{P} < 0$) is applied to the piezo, the sample expands (contracts) along $[\bar{1}10]$. This is monitored using a strain gauge glued to the opposite face of the piezo [@Shayegan.APL.2003; @Shkolnikov.APL.2004; @Supple].
In order to measure the Fermi wavevectors, we fabricate periodic gratings of negative electron-beam resist, with period $a = 200$ nm, on the surface of the L-shaped Hall bar \[Fig. 1(c)\]. The grating induces a periodic strain onto the GaAs surface which in turn results in a small periodic modulation of the 2DHS density via the piezoelectric effect [@Endo.PRB.2000; @Kamburov.PRB.2012; @Kamburov.PRL.2012]. In the presence of $B$, when the cyclotron motion of holes becomes commensurate with $a$, the magnetoresistance shows oscillations whose minima positions are directly related to the carriers’ Fermi wavevector in the direction perpendicular to the current [@Endo.PRB.2000; @Kamburov.PRB.2012; @Kamburov.PRL.2013; @Gunawan.PRL.2004]. Figure 1(c) shows an example when tensile strain is applied along $[\bar{1}10]$; the elongated cyclotron orbits under a finite $B$ are indicated by black curves.
Figure 2(a) shows magnetoresistance traces for $\epsilon = -1.8 \times 10^{-4}$. The red and blue traces are from the patterned regions along the $[110]$ and $[\bar{1}10]$ directions, respectively, while the black trace is for an unpatterned region. The red and blue traces exhibit commensurability features for holes near $B = 0$ \[Fig. 2(b)\] and for CFs near $\nu = 1/2$ \[Fig. 2(c)\]. To analyze the low-field hole data, we use the *electrostatic* commensurability condition [@Weiss.EPL.1989; @Winkler.PRL.1989; @Gerhardts.PRL.1989; @Beenakker.PRL.1989; @Endo.PRB.2000; @Kamburov.PRB.2012] for the minima positions, $2R_{c}/a = i-1/4$ ($i = 1,2,3, \ldots$) where 2$R_{c} = 2\hbar k/eB$ is the cyclotron orbit diameter, $k$ is the 2DHS Fermi wavevector perpendicular to the current, and $a$ is the period of the density modulation. For CFs, we observe commensurability features near $\nu = 1/2$, or $B_{1/2} = 14.5$ T \[Fig. 2(c)\]. The positions of minima around $\nu = 1/2$ yield the Fermi wavevector of CFs ($k^\ast$) according to the *magnetic* commensurability condition [@Smet.PRL.1999; @Zwerschke.PRL.1999; @Kamburov.PRL.2012], $2R^\ast_{c}/a = i + 1/4$, where the CF cyclotron diameter 2$R^\ast_{c} = 2\hbar k^\ast / eB^\ast$ and $B^\ast = B-B_{1/2}$ is the effective field for CFs [@Kamburov.PRL.2014; @fnote3].
In Fig. 2(d) we mark the measured Fermi wavevectors for holes with red and blue dots along $[\bar{1}10]$ and $[110]$. Although theoretical calculations for holes predict two spin subbands with different Fermi wavevectors \[black solid and dashed curves in Fig. 2(d)\], we measure a single $k$ for each direction from the commensurability features. The measured $k_{[\bar{1}10]}$ (red dots) and $k_{[110]}$ (blue dots) are close to the average calculated Fermi wavevectors for the two spin-subbands. Figure 2(e) shows $k^\ast$ measured for CFs with red and blues dots. We depict the Fermi contour as an ellipse because there are no theoretical calculations available for CFs, and also the area of an ellipse spanned by the two measured $k^\ast$ accounts for the density of CFs which are fully spin-polarized at high fields [@Kamburov.PRL.2012]. Note that the CF Fermi contour anisotropy $\alpha_\mathrm{CF} \equiv k^\ast_{[\bar{1}10]} / k^\ast_{[110]}= 0.77$, which is closer to unity than the 2D hole anisotropy $\alpha_\mathrm{F} \equiv k_{[\bar{1}10]} / k_{[110]} = 0.53$. Quantitatively, we find $\alpha_\mathrm{CF} = \sqrt{\alpha_\mathrm{F}}$ to within 5%; see below.
Next we demonstrate the tunability of CF Fermi contour anisotropy via strain. Figures 3(a) and 3(b) show magnetoresistance traces near $\nu = 1/2$, taken along $[\bar{1}10]$ and $[110]$, at different strains. In each panel, the green trace represents the $\epsilon = 0$ case where the Fermi contour is essentially isotropic and $k^\ast = k^\ast_{0} = \sqrt{4 \pi p}$ [@Supple]. The traces shown above the green trace are for tensile strain ($\epsilon>0$) while those below are for compressive strain ($\epsilon<0$). In Fig. 3(a) the positions of resistance minima move towards (away from) $B^\ast = 0$ for $\epsilon>0$ ($\epsilon<0$), while the opposite is true for Fig. 3(b). These observations imply a distortion in the shape of CF cyclotron orbits as depicted in the side panels of Figs. 3(a) and 3(b).
Figure 3(c) summarizes the measured $k^\ast$ along $[\bar{1}10]$ and $[110]$, normalized by $k^\ast_{0}$. Comparing $k^\ast$ values for compressive and tensile cases, the change of $k^\ast$ for $\epsilon > 0$ is larger than for $\epsilon < 0$. This asymmetry reflects the response of the 2DHS Fermi contour to the applied strain \[Fig. 1(a)\]. We also find that the geometric means of $k^\ast / k^\ast_{0}$ along the two perpendicular directions remain close to unity \[Fig. 3(d)\]. This suggests that CF Fermi contours are nearly elliptical, although we cannot exclude a more complex shape.
Figure 4 illustrates the highlight of our study: comparison of strain-induced Fermi contour anisotropy for CFs and holes. The measured anisotropy for CFs, $\alpha_\mathrm{CF} = k^\ast_{[\bar{1}10]} / k^\ast_{[110]}$, is shown by black circles, and the *square-root* of the calculated anisotropy for holes, $\alpha_\mathrm{F} = k_{[\bar{1}10]} / k_{[110]}$, by a purple curve. Here we use, for each $k_{[\bar{1}10]}$ and $k_{[110]}$, the averaged values of $k$ for the spin-subbands, since experiments measure only a single $k$ for each direction. Remarkably, the measured $\alpha_\mathrm{CF}$ for CFs essentially coincides with $\sqrt{\alpha_\mathrm{F}}$ over the entire range of strains applied in the experiments. This is particularly striking because there are no fitting or adjustable parameters.
Lastly, we study the impact of anisotropy on the strength of FQHSs, focusing on the energy gap for the $\nu = 2/3$ state. The sample used for the measurements has $p = 1.3 \times 10^{11}$ cm$^{-2}$, and exhibits commensurability features only for holes along $k_{[\bar{1}10]}$. Moreover, using a different cool-down procedure [@Supple], we achieved larger strain values ($\epsilon$ up to $5.5 \times 10^{-4}$), and anisotropy ($\alpha_\mathrm{F}$ as large as 3.3) as shown in Fig. 5. The measured energy gap $\Delta$, determined from the expression $R(T) \sim e^{-\Delta / 2 T}$, is 2.1 K for $\epsilon = 0$, and it decreases only to 2.0 K even for a large anisotropy $\alpha_\mathrm{F} = 3.3$. The small decrease of $\Delta$ is consistent with recent theoretical predictions [@Balram.PRB.2016], suggesting that the FQHSs in the lowest Landau level are quite robust against anisotropy.
Returning to the Fermi contour anisotropy, our measurements (Fig. 4) provide quantitative evidence for a simple relation between the anisotropy of low-field fermions and high-field CFs: $\alpha_\mathrm{CF} = \sqrt{\alpha_\mathrm{F}}$. This appears to contradict some of the theories which predict that $\alpha_\mathrm{F}$ and $\alpha_\mathrm{CF}$ should be the same [@Balagurov.PRB.2000; @Balram.PRB.2016]. One can, however, qualitatively justify the square-root relation [@Gokmen.NatPhys.2010]. In an ideal, isotropic 2D system, the Coulomb interaction $V_\mathrm{C}$ ($\propto 1/\sqrt{x^2 + y^2}$) determines the physical parameters of CFs, including their effective mass $m^\ast$ which is linearly proportional to $V_\mathrm{C}$ [@Jain.Book.2007; @Halperin.PRB.1993]. At a given filling factor, $V_\mathrm{C}$ is quantified solely by the magnetic length $l_{B} = \sqrt{\hbar/eB}$ [@Jain.Book.2007; @Halperin.PRB.1993]. Now, a system with an anisotropic dispersion $\alpha_\mathrm{F} \ne 1$ at $B = 0$ can be mapped to a system with an isotropic Fermi contour and an anisotropic $V_\mathrm{C}$ ($\propto 1/\sqrt{x^{2}\alpha_\mathrm{F} + y^{2}/\alpha_\mathrm{F}}$) using the coordinate transformations $x \rightarrow x/\sqrt{\alpha_\mathrm{F}}$ and $y \rightarrow y \sqrt{\alpha_\mathrm{F}}$. In such a system the strength of $V_\mathrm{C}$ at high fields thus depends not only on $l_B$ but also on the direction, i.e., $V_\mathrm{C}$ anisotropy is $\alpha_\mathrm{F}$. If one assumes that CFs have a parabolic dispersion and an anisotropic $m^\ast $ whose anisotropy follows linearly the anisotropy of $V_\mathrm{C}$, then the *mass* anisotropy of CFs is given by $\alpha_\mathrm{F}$, implying that their Fermi wavevector anisotropy is proportional to $\sqrt{\alpha_\mathrm{F}}$.
In conclusion, our results provide direct and quantitative evidence for the inheritance of Fermi contour anisotropy by CFs from their low-field fermion counterparts through a simple relation: $\alpha_\mathrm{CF} = \sqrt{\alpha_\mathrm{F}}$. While the discussion in the preceding paragraph serves as a plausibility argument for this relation, there is also some very recent rigorous theoretical justification. In their numerical calculations for anisotropic fermions with a parabolic band, Ippoliti *et al.* [@Ippoliti.2017] find that the relation $\alpha_\mathrm{CF} = \sqrt{\alpha_\mathrm{F}}$ is indeed empirically obeyed [@fnote4]. It remains to be seen, both experimentally and theoretically, if the relation holds when the fermions’ band deviates significantly from parabolic [@fnote5; @Mueed.PRL.2015a].
Supplemental Material: Transference of Fermi Contour Anisotropy to Composite Fermions
=====================================================================================
S I. Inclusion of strain in Fermi contour calculations
======================================================
The in-plane components of the strain created by the piezo-actuator are transfered to the Hall bar, giving rise to the strain components ${\epsilon}_{{[\bar{1}10]}}$ and ${\epsilon}_{{[110]}}$. The effect of strain on the electronic structure is usually expressed in a conventional coordinate system with $\hat{\bm{x}} \parallel [100]$, $\hat{\bm{y}} \parallel [010]$, and $\hat{\bm{z}} \parallel [001]$ (e.g., Refs.[@Bir.Book; @Winkler.Book]). Here the nonzero components ${\epsilon}^c_{ij}$ of the second-rank strain tensor become
$$\begin{aligned}
{\epsilon}^c_{xx} & = {\epsilon}_{yy}^c = {{\textstyle\frac{1}{2}}} ({\epsilon}_{{[\bar{1}10]}}+ {\epsilon}_{{[110]}}) \\
{\epsilon}^c_{xy} & = {\epsilon}_{yx}^c = {{\textstyle\frac{1}{2}}} ({\epsilon}_{{[\bar{1}10]}}- {\epsilon}_{{[110]}})
\equiv {{\textstyle\frac{1}{2}}} \, {\epsilon}\label{eq:strain:xy} \\
{\epsilon}^c_{zz} & = - \frac{c_{12}} {c_{11}} ({\epsilon}_{{[\bar{1}10]}}+ {\epsilon}_{{[110]}}).
\label{eq:strain:zz}
\end{aligned}$$
Equation (\[eq:strain:xy\]) defines the strain ${\epsilon}$ used in the main paper. It is this shear strain component that is primarily responsible for the Fermi contour anisotropy $\alpha_\mathrm{F} \ne 1$ at $B=0$. Equation (\[eq:strain:zz\]) follows from Hooke’s law, assuming that the Hall bar’s extent in $z$ direction can adjust freely to the in-plane strain. The coefficients $c_{11}$ and $c_{12}$ are the elastic constants of GaAs.
We incorporate the effect of the strain ${\epsilon}^c_{ij}$ into our self-consistent calculations using the Bir-Pikus strain Hamiltonian [@Bir.Book; @Winkler.Book]. Numeric values for the relevant deformation potentials are given in Ref. [@Winkler.Book]. We note that a simple analytical model can be developed that evaluates the effect of strain using perturbation theory in lowest order of the wave vector ${{\bm{k}}}$ and strain ${\epsilon}^c_{ij}$ (similar to the analytical models developed in Ref. [@Winkler.Book]). However, for the 2D hole systems studied here, terms of higher order in ${{\bm{k}}}$ and ${\epsilon}^c_{ij}$ are very important. These terms are fully taken into account in our numerical calculations. The lowest-order analytical model overestimates the 2D hole Fermi contour anisotropy $\alpha_\mathrm{F}$ by about a factor of five.
S II. Determination of strain in experiments
============================================
Because of the different thermal contractions of the piezo-actuator and the sample, a residual, or “built-in" strain ($\epsilon_\mathrm{bi}$) develops when the sample is cooled to low temperatures. The value of $\epsilon_\mathrm{bi}$ depends on the piezo-actuator and the details of the cool-down, and is not fully controllable. For example, when the two leads of the piezo-actuator are left open-circuited during the cool-down, $\epsilon_\mathrm{bi}$ is typically small. This is the case for the data shown in Fig. 3, where $\epsilon_\mathrm{bi} \simeq -0.5\times10^{-4}$. In presenting the data in Figs. 3 and 4, we have corrected for this $\epsilon_\mathrm{bi}$. To determine the value of $\epsilon_\mathrm{bi}$ and make the correction, we first take data at different values of biases ($V_\mathrm{P}$) applied to the piezo-actuator, and find the bias that gives $k_{[\bar{1}10]} =k_{[110]}$. This bias corresponds to $\epsilon=0$. We then determine the applied $\epsilon$ from the relative change of strain as monitored by the strain gauge glued to the piezo-actuator [@Shayegan.APL.2003; @Shkolnikov.APL.2004; @Gunawan.PRL.2004].
For the data of Fig. 5, the sample was cooled down multiple times with a 1 G$\Omega$ resistor across the two leads of the piezo-actuator. This led to large values of $\epsilon_\mathrm{bi}$, up to $5.5 \times 10^{-4}$ [@Shabani.PRL.2008], and $\alpha_\mathrm{F}$ as large as 3.3, as shown in Fig. 5. Unfortunately, we could not achieve such large $\epsilon_\mathrm{bi}$ with the sample of Fig. 3.
We acknowledge support by the DOE BES (DE-FG02-00-ER45841) grant for measurements, and the NSF (Grants DMR 1305691, DMR 1310199, MRSEC DMR 1420541, and ECCS 1508925), the Gordon and Betty Moore Foundation (Grant GBMF4420), and Keck Foundation for sample fabrication and characterization. We thank R. N. Bhatt, S. D. Geraedts, M. Ippoliti, J. K. Jain, and D. Kamburov for illuminating discussions.
[99]{}
M. Shayegan, “Flantland Electrons in High Magnetic Fields", in *High Magnetic Fields: Science and Technology*, Vol. 3, edited by F. Herlach and N. Miura (World Scientific, Singapore, 2006), pp. 31-60 \[condmat/0505520\]. J. K. Jain, *Composite Femions* (Cambridge University Press, Cambridge, 2007).
D. B. Balagurov and Y. E. Lozovik, Phys. Rev. B [**62**]{}, 1481 (2000).
M. Shayegan, E. P. De Poortere, O. Gunawan, Y. P. Shkolnikov, E. Tutuc, and K. Vakili, Phys. Stat. Sol. (b) [**243**]{}, 3629 (2006). T. Gokmen, M. Padmanabhan, and M. Shayegan, Nat. Phys. [**6**]{}, 621 (2010). E. Fradkin, S. A. Kivelson, M. J. Lawler, J. P. Eisenstein, and A. P. Mackenzie, Ann. Rev. Condens. Matter Phys. [**1**]{}, 153 (2010). M. Mulligan, C. Nayak, and S. Kachru, Phys. Rev. B [**82**]{}, 085102 (2010).
F. D. M. Haldane, Phys. Rev. Lett. [**107**]{}, 116801 (2011). J. Xia, J. P. Eisenstein, L. N. Pfeiffer, and K. W. West, Nat. Phys. [**7**]{}, 845 (2011). S. P. Koduvayur, Y. Lyanda-Geller, S. Khlebnikov, G. Csathy, M. J. Manfra, L. N. Pfeiffer, K. W. West, and L. P. Rokhinson, Phys. Rev. Lett. [**106**]{}, 016804 (2011). Y. Liu, S. Hasdemir, M. Shayegan, L. N. Pfeiffer, K. W. West, and K. W. Baldwin, Phys. Rev. B. [**88**]{}, 035307 (2013).
D. Kamburov, Y. Liu, M. Shayegan, L. N. Pfeiffer, K. W. West, and K. W. Baldwin, Phys. Rev. Lett. [**110**]{}, 206801 (2013). D. Kamburov, M. A. Mueed, M. Shayegan, L. N. Pfeiffer, K. W. West, K. W. Baldwin, J. J. D. Lee, and R. Winkler, Phys. Rev. B [**89**]{}, 085304 (2014). M. A. Mueed, D. Kamburov, Y. Liu, M. Shayegan, L. N. Pfeiffer, K. W. West, K. W. Baldwin, and R. Winkler, Phys. Rev. Lett. [**114**]{}, 176805 (2015). M. A. Mueed, D. Kamburov, M. Shayegan, L. N. Pfeiffer, K. W. West, K. W. Baldwin, and R. Winkler, Phys. Rev. Lett. [**114**]{}, 236404 (2015). M. A. Mueed, D. Kamburov, S. Hasdemir, L. N. Pfeiffer, K. W. West, K. W. Baldwin, and M. Shayegan, Phys. Rev. B [**93**]{}, 195436 (2016). N. Samkharadze, K. A. Schreiber, G. C. Gardner, M. J. Manfra, E. Fradkin, and G. A. Csáthy, Nat. Phys. [**12**]{}, 191 (2016).
H. Wang, R. Narayanan, X. Wan, and F. Zhang, Phys. Rev. B [**86**]{}, 035122 (2012). R.-Z. Qiu, F. D. M. Haldane, X. Wan, K. Yang, and S. Yi, Phys. Rev. B [**85**]{}, 115308 (2012). B. Yang, Z. Papic, E. H. Rezayi, R. N. Bhatt, and F. D. M. Haldane, Phys. Rev. B [**85**]{}, 165318 (2012). K. Yang, Phys. Rev. B [**88**]{}, 241105 (2013). S. Johri, Z. Papic, P. Schmitteckert, R. N. Bhatt, and F. D. M. Haldane, New J. Phys. [**18**]{}, 025011 (2016). A. C. Balram and J. K. Jain, Phys. Rev. B [**93**]{}, 075121 (2016). R. B. Laughlin, Phys. Rev. Lett. [**50**]{}, 1395 (1983). B. I. Halperin, P. A. Lee, and N. Read, Phys. Rev. B [**47**]{}, 7312 (1993). J. K. Jain, Phys. Rev. Lett. [**63**]{}, 199 (1989).
Throughout this paper, we denote strain by $\epsilon=\epsilon_{[\bar{1}10]}- \epsilon_{[110]}$ where $\epsilon_{[\bar{1}10]}$ and $\epsilon_{[110]}$ are strains along the $[\bar{1}10]$ and $[110]$ directions and $\epsilon_{[110]}=-0.38 \epsilon_{[\bar{1}10]}$; see Refs. [@Shayegan.APL.2003; @Shkolnikov.APL.2004]. M. Shayegan, K. Karrai, Y. P. Shkolnikov, K. Vakili, E. P. De Poortere, and S. Manus, Appl. Phys. Lett. [**83**]{}, 5235 (2003). Y. P. Shkolnikov, K. Vakili, E. P. De Poortere, and M. Shayegan, Appl. Phys. Lett. [**85**]{}, 3766 (2004).
See Supplemental Material. For details of our sample structure, see Refs. [@Kamburov.PRB.2012; @Kamburov.PRL.2012].
D. Kamburov, H. Shapourian, M. Shayegan, L. N. Pfeiffer, K. W. West, K. W. Baldwin, and R. Winkler, Phys. Rev. B [**85**]{}, 121305 (2012). D. Kamburov, M. Shayegan, L. N. Pfeiffer, K. W. West, and K. W. Baldwin, Phys. Rev. Lett. [**109**]{}, 236401 (2012). G. L. Bir and G. E. Pikus, *Symmetry and Strain-Induced Effects in Semiconductors* (Wiley, New York, 1974). R. Winkler, *Spin-orbit Coupling Effects in Two-Dimensional Electron and Hole Systems* (Springer, Berlin, 2003).
K. I. Kolokolov, A. M. Savin, S. D. Beneslavski, N. Ya. Minina, and O. P. Hansen, Phys. Rev. B [**59**]{}, 7537 (1999). B. Habib, J. Shabani, E. P. De Poortere, M. Shayegan, and R. Winkler, Phys. Rev. B [**75**]{}, 153304 (2007). J. Shabani, M. Shayegan, and R. Winkler, Phys. Rev. Lett. [**100**]{}, 096803 (2008). A. Endo, S. Katsumoto, and Y. Iye, Phys. Rev. B [**62**]{}, 16761 (2000).
O. Gunawan, Y. P. Shkolnikov, E. P. De Poortere, E. Tutuc, and M. Shayegan, Phys. Rev. Lett. [**93**]{}, 246603 (2004).
D. Weiss, K. von Klitzing, K. Ploog, and G. Weimann, Europhys. Lett. [**8**]{}, 179 (1989). R. W. Winkler, J. P. Kotthaus, and K. Ploog, Phys. Rev. Lett. [**62**]{}, 1177 (1989). R. R. Gerhardts, D. Weiss, and K. von Klitzing, Phys. Rev. Lett. [**62**]{}, 1173 (1989). C. W. J. Beenakker, Phys. Rev. Lett. [**62**]{}, 2020 (1989).
J. H. Smet, S. Jobst, K. von Klitzing, D. Weiss, W. Wegscheider, and V. Umansky, Phys. Rev. Lett. [**83**]{}, 2620 (1999). S. D. M. Zwerschke and R. R. Gerhardts, Phys. Rev. Lett. [**83**]{}, 2616 (1999). D. Kamburov, Y. Liu, M. A. Mueed, M. Shayegan, L. N. Pfeiffer, K. W. West, and K. W. Baldwin, Phys. Rev. Lett. [**113**]{}, 196801 (2014). In view of Ref. [@Kamburov.PRL.2014], we take the CF density to equal the minority carrier density in the lowest Landau level, and use the minimum position for $B^\ast > 0$ to extract the $k^\ast$ of CFs.
M. Ippoliti, S. D. Geraedts, and R. N. Bhatt, unpublished.
A similar relation can also be derived from Eq. (15) in Ref. [@KYang.PRB.2013], provided that the range of interaction ($s$) in the postulated Gaussian potential equals $l_{B}$.
Quantifying the 2DHS Fermi contour “warping" by the geometric mean of $k_{[110]}/k_0$ and $k_{[\bar{1}10]}/k_0$, we note that this mean varies between 1.12 and 1.16 in the $\epsilon$ range of Fig. 4. (For a circular or an elliptical Fermi contour, the mean is unity). The equivalently defined parameter for CFs is closer to unity \[see Fig. 3(d)\], implying that CFs’ Fermi contour is less warped than their zero-field counterparts. A similar conclusion was reached in Ref. [@Mueed.PRL.2015a].
| {
"pile_set_name": "ArXiv"
} |
[[ad]{}]{}[ad]{} [[Ad]{}]{}[Ad]{}
\#1
*2[[sl]{}(2, [C]{})]{}*
1\#1
1\#1
Introduction
============
The modifications (or the Hecke transformation) of bundles over complex curves is a correspondence between two bundles $E$ and ${\tilde{E}}$. It is isomorphism in a complement of some divisor. A modification can change the topological type of the original bundle. From the field-theoretical point of view the modification is provided by a gauge transformation of sections, which is singular at the divisor. In [@LOZ1] we apply this procedure to the Higgs bundles. The Higgs bundles are the phase spaces of the Hitchin integrable systems [@Hi]. Modifications acts on the phase space as a symplectic transformation. In this special case we call the modification the Symplectic Hecke Correspondence. For the Higgs bundles over elliptic curves with marked points Symplectic Hecke Correspondence leads to a symplectomorphism between different classical integrable systems such as
- Elliptic Calogero–Moser system $\Leftrightarrow$ Elliptic ${{\rm GL}(N,{\mathbb C})}$ Top, [@LOZ1];
- Calogero–Moser field theory $\Leftrightarrow$ Landau–Lifshitz equation, [@LOZ1; @Kr1];
- Painlevé VI $\Leftrightarrow$ non-autonomous Zhukovsky–Volterra gyrostat, [@LOZ2].
In these examples modifications increase the degree of the underlying bundles on one. In general, modifications act as the Bäcklund transformations of integrable systems. If degree of the bundles (modula rank) is not changed then modifications produce what is called the autoBäcklund transformations. It turned out that the modification in the first example is equivalent to the twist of $R$-matrices [@B; @JMO] that transforms the dynamical $R$-matrices of the IRF models of the ${\rm GL}(N)$ type [@Fe] to the vertex $R$-matrices [@Be] corresponding to the ${\rm GL}(N)$ generalization of the XYZ models.
The modifications are parameterized by vectors $\vec m$ of the weight lattices $P$ of $\SLN$. If $\vec m$ belongs to the root sublattice $Q\subset P$, then the modified bundle ${\tilde{E}}$ has the same degree as $E$. Otherwise, the degree of bundle is changed. The modifications can be described by changing another topological invariant. It is a characteristic class of a bundle. Let the base of $E$ be a Riemann surface ${\Sigma}_g$ of genus $g$. Then the characteristic class of $E$ is an element of $H^2({\Sigma}_g,\mZ_N)\sim\mZ_N$, where $\mZ_N\sim P/Q$ is a center of $\SLN$. Another example of the characteristic classes, is the characteristic class of spin-bundles, that will not considered here, is the Stiefel–Whitney class $H^2({\Sigma}_g,Z_2)$.
Here we discuss a field-theoretical interpretation of modifications. It was established in [@KW] that the modifications are related to the Dirac monopole configurations in a topological version of the $\clN=4$ four-dimensional super-symmetric Yang–Mills theory. If “the space-time” of the topological theory has the form $\mR^2\times{\Sigma}_g$, then the modifications of $E$ over ${\Sigma}_g$ are parameterized by the monopoles charges.
To describe the modification it is sufficient to neglect the “time” dependence and consider $\mR\times{\Sigma}_g$. The condition for fields to preserve the supersymmetry amounts to the Bogomolny equation.
The aims of this paper are
- To define modifications and describe their interrelations with the Bogomolny equation following [@KW]. We consider a special configuration of the space-time $\mR^2\times{\Sigma}_\tau$, where ${\Sigma}_\tau$ is an elliptic curve with the modular parameter $\tau$.
- To find solutions of the Bogomolny equation in the case of line bundles over ${\Sigma}_\tau$. They are generalizations of the Kronecker series [@We]. We give two representations of the solution and prove their equivalence by means of the functional equation generalizing the Kronecker functional equation.
- To describe non-Abelian modifications that are not related directly to solutions of the Bogomolny equation and follows from our previous results.
Characteristic classes of holomorphic bundles\
over complex curves {#section2}
==============================================
We describe holomorphic bundles over complex curves ${\Sigma}_g$ of genus $g$ and define their characteristic classes.
Global description {#section2.1}
------------------
Let $\pi_1({\Sigma}_g)$ be a fundamental group of ${\Sigma}_g$. It has $2g$ generators $\{a_{\alpha},b_{\alpha}\}\,$, corresponding to the fundamental cycles of ${\Sigma}_g$ with the relation $$\begin{gathered}
\label{0.3}
\prod_{{\alpha}=1}^g [a_{{\alpha}},b_{{\alpha}}]=1,\end{gathered}$$ where $[a_{{\alpha}},b_{{\alpha}}]= a_{{\alpha}}b_{{\alpha}}a_{{\alpha}}^{-1}b_{{\alpha}}^{-1}$ is the group commutator.
Let $\rho$ be a representation of $\pi_1$ in $\mC^N$. Consider a holomorphic adjoint ${{\rm GL}(N,{\mathbb C})}$ bundle $E$ over ${\Sigma}_g$. In fact, $E$ is a P${{\rm GL}(N,{\mathbb C})}\sim\PSLN$ bundle, because the center of ${{\rm GL}(N,{\mathbb C})}$ does not act in the adjoint representation. The bundle $E$ can be defined by holomorphic transition matrices of its sections $s\in{\Gamma}(E)$ around the fundamental cycles. Let $z\in {\Sigma}_g$ be a fixed point. Then $$\begin{gathered}
s(a_{\alpha}z)= \rho(a_{\alpha})s(z),\qquad
s(b_{\beta}z)= \rho(b_{\beta})s(z).\end{gathered}$$
Due to (\[0.3\]) we have $$\begin{gathered}
\label{pi} \prod_{{\alpha}=1}^g
[\rho(a_{{\alpha}}),\rho(b_{{\alpha}})]={\rm Id}.\end{gathered}$$
Let $\clK$ be an extension of $\pi_1$ by the cyclic group $\mZ_N\sim\mZ/N\mZ$ $$\begin{gathered}
\label{eseq} 1\to\mZ_N\to\clK\to\pi_1({\Sigma}_g)\to
1. \end{gathered}$$ The group $\clK$ is defined by the relation $$\begin{gathered}
\prod_{{\alpha}=1}^g [a_{{\alpha}},b_{{\alpha}}]={\omega},\qquad {\omega}^N=1.\end{gathered}$$ Let $\hat{\rho}$ be a representation of $\clK$ in ${{\rm GL}(N,{\mathbb C})}$. Then using $\hat{\rho}$ as transition matrices we define a bundle over ${\Sigma}_g$. But now instead of (\[pi\]) we have $$\begin{gathered}
\label{pi1} \prod_{{\alpha}=1}^g
[\hat\rho(a_{{\alpha}}),\hat\rho(b_{{\alpha}})]={\omega}\, {\rm Id}. \end{gathered}$$ Here ${\omega}\,
{\rm Id}$ is the generator of the center $\clZ(\SLN)\sim\mZ_N$ of $\SLN$. It means that $\hat\rho$ can serve as transition matrices only for $\PSLN$ bundles, but not for $\SLN$ or ${{\rm GL}(N,{\mathbb C})}$ bundles. Note, that the fibers of the $\PSLN$-bundles are spaces of representations with highest weights from the root lattice $Q$ (\[rl\]) including the adjoint representation with the highest weight $\varpi_1+\varpi_{N-1}$ (\[fw1\]). For the $\SLN$ representations the highest weights belong to the weight lattice $P$ (\[wl\]). In this way elements from the factor group $P/Q\sim \clZ(\SLN)$ (\[center\]) define an obstruction to lift $\PSLN$ bundles to $\SLN$ bundles.
The obstruction has a cohomological interpretation. Consider the exact sequence following from (\[a1\]) $$\begin{gathered}
\to H^1({\Sigma}_g,\SLN)\to H^1({\Sigma}_g,\PSLN)\to
H^2({\Sigma}_g,\clZ(\SLN))\to\cdots.
\end{gathered}$$ The groups $H^1({\Sigma}_g,\SLN)$, $H^1({\Sigma}_g,\PSLN)$ are the moduli space of $\SLN$ and $\PSLN$ bundles. Then $H^2({\Sigma}_g,\clZ(\SLN))$ defines an obstruction to lift $\PSLN$ bundles to $\SLN$ bundles. We call $\xi\in H^2({\Sigma}_g,\mZ_N)$ *the characteristic class* of a $\PSLN$ bundle. In fact, $H^2({\Sigma}_g,\mZ_N)\sim\mZ_N$ and ${\omega}$ in (\[pi1\]) represents $\xi\in H^2({\Sigma}_g,\mZ_N)$.
This construction can be generalized to any factor-group $G_l=\SLN/\mZ_l$, where $l$ is a nontrivial divisor of $N$, $(N=pl$, $l\neq 1,N)$. Consider an extension $\clK_l$ of $\pi_1({\Sigma}_g)$ by $\mZ_l$ (compare with (\[eseq\])) $$\begin{gathered}
1\to\mZ_l\to\clK_l\to\pi_1({\Sigma}_g)\to 1.\end{gathered}$$ Let $E_l$ be a holomorphic $G_l$-bundle. The fibers of $E_l$ belong to a irreducible representation of $G_l$ with a highest weight $\nu\in{\Gamma}(G_l)$ (\[gl\]). Then the transition matrices representing $\clK_l$ satisfy the relation $$\begin{gathered}
\label{pi2} \prod_{{\alpha}=1}^g
[\hat\rho(a_{{\alpha}}),\hat\rho(b_{{\alpha}})]={\omega}^p\, {\rm Id},\qquad ({\omega}^p)^l=1.\end{gathered}$$ It follows from the exact sequence $$\begin{gathered}
1\to\mZ_l\to\SLN\to G_l\to 1, \end{gathered}$$ that elements from $H^2({\Sigma}_g,\mZ_l)\sim\mZ_l$ are obstructions to lift $G_l$ bundle $E_l$ to a $\SLN$-bundle. The group $\mZ_l$ can be identified with the center of the dual group $^LG_l\sim G_p=\SLN/\mZ_p$ (see (\[dugr\]) and (\[cdg\])). Thus, the obstructions to lift $G_l$ bundles $E_l$ to a $\SLN$ bundles are defined by $H^2({\Sigma}_g,\clZ(^LG_l))$.
On the other hand, since $\mZ_p$ is a center of $G_l$ we have the sequence $$\begin{gathered}
1\to\mZ_p\to G_l\to\PSLN\to 1,
\end{gathered}$$ where $\mZ_p$ is a center of $G_l$. Then elements from $H^2({\Sigma}_g,\clZ(G_l))$ are obstructions to lift a $\PSLN$-bundle to a $G_l$-bundle. Summarizing we have defined two types of the characteristic classes $$\begin{gathered}
H^2({\Sigma}_g,\clZ(G_l)) - {\rm
obstructions~to~lift~a~}\PSLN~{\rm bundle~to~a~}G_l~{\rm bundle},\nonumber\\
\label{obs2}
H^2({\Sigma}_g,\clZ(^LG_l)) - {\rm obstructions~to~lift~a~}G_l~{\rm bundle~to~a~}\SLN~{\rm bundle}.\end{gathered}$$
Though for ${\omega}\neq 1$ $\PSLN$ bundles cannot be lifted to $\SLN$ bundles, they can be lifted to ${{\rm GL}(N,{\mathbb C})}$ bundles. From the exact sequence $$\begin{gathered}
1\to \clO^*\stackrel{\mathrm{det}}\to{{\rm GL}(N,{\mathbb C})}\to {\rm
PGL}(N,\mC)\to 1\end{gathered}$$ we have $$\begin{gathered}
H^1({\Sigma}_g,{{\rm GL}(N,{\mathbb C})})\to
H^1({\Sigma}_g,{\rm PGL}(N,\mC))\to H^2({\Sigma}_g,\clO^*).\end{gathered}$$ The Brauer group $H^2({\Sigma}_g,\clO^*)$ vanishes and therefore, there is no obstruction to lift ${\rm PGL}(N,\mC)\sim{\rm PSL}(N,\mC)$ bundles to ${{\rm GL}(N,{\mathbb C})}$ bundles. We will demonstrate it below.
Holomorphic bundles over elliptic curves {#section2.2}
----------------------------------------
We define an elliptic curve ($g=1$) as the quotient ${\Sigma}_\tau=\mC/(\mZ+\tau\mZ)$. In this case we can construct explicitly the generic transition matrices for $G_l$-bundles.
The curve has two fundamental cycles $a: (z\to z+1)$, $b: (z\to z+\tau)$. We define a trivial bundle $E$ over ${\Sigma}_\tau$ by two commuting matrices $$\begin{gathered}
\label{14}
s(z+1)=\rho_as(z),\qquad s(z+\tau)=\rho_bs(z),\qquad [\rho_a,\rho_b]={\rm Id}.\end{gathered}$$ It is a ${{\rm PGL}(N,{\mathbb C})}$-bundle that can be lifted to $\SLN$ bundles.
Consider a representation of $\hat\rho$ of $\clK$ acting on the sections of $E$ as $$\begin{gathered}
s(z+1)=\hat{\rho}_as(z),\qquad s(z+\tau)=\hat{\rho}_bs(z).\end{gathered}$$ with commutation relation (\[pi1\]) $$\begin{gathered}
[\hat{\rho}_a,\hat{\rho}_b]={\omega}\,{\rm Id}. \end{gathered}$$ One can choose $$\begin{gathered}
\label{qla}
\hat{\rho}_a=\clQ,\qquad \hat{\rho}_b={\Lambda}, \qquad
\clQ={{\rm diag}}\big(1,{\omega},\ldots,{\omega}^{N-1}\big),\qquad
{\Lambda}=\left(
\begin{array}{cccc}
0 & 1 & \ldots & 0 \\
0 & 0 & 1 & 0 \\
\vdots & \vdots &\ddots & 1 \\
1 & 0 & \ldots & 0 \\
\end{array}
\right).\end{gathered}$$ The bundle with these transition functions cannot be lifted to $\SLN$ bundles. Replace $\hat{\rho}_b$ by $$\begin{gathered}
\label{last}
\hat{\rho}'_b=\exp\left(-\frac{2\pi
i}{N}\left(z+\frac{\tau}{2}\right)\right){\Lambda}.\end{gathered}$$ It is a ${{\rm GL}(N,{\mathbb C})}$ bundle since $[\hat{\rho}_a,\hat{\rho}'_b]=Id$ and $\det\hat{\rho}'_b\neq 1$ . It follows from (\[last\]) that a section of the determinant bundle is the theta-function $$\begin{gathered}
\label{theta} \vartheta(z,\tau)=q^{\frac
{1}{8}}\sum_{n\in {\mZ}}(-1)^ne^{\pi i(n(n+1)\tau+2nz)},\qquad q=\exp
2\pi i\tau. \end{gathered}$$ It has a simple pole in the fundamental domain $\mC/(\mZ\oplus\tau\mZ)$. Therefore, the bundle has degree one. It is called the theta-bundle.
To consider a general case [@LZ] represent the rank as the product $N=pl$. Define the transition matrix $$\begin{gathered}
\label{ic}
\hat\rho_a=\clQ , \\
\label{ic1} \hat\rho_b=\bfe(\vec u_l){\Lambda}^p,\end{gathered}$$ where $$\begin{gathered}
\vec u_l={{\rm diag}}(\overbrace{\bfu_p,\ldots,
\bfu_p}^l),\qquad \bfu_p=({\tilde{u}}_1,\ldots,{\tilde{u}}_p). \end{gathered}$$ Since $[\clQ,{\Lambda}^p]={\omega}^p\,{\rm Id}_l$, ${\omega}^p=\exp\frac{2\pi i}{l}$ $$\begin{gathered}
[\hat\rho_a,\hat\rho_b]={\omega}^p\,{\rm Id}_N.\end{gathered}$$ Comparing this relation with (\[pi2\]) we conclude that (\[ic\]) and (\[ic1\]) serve as the transition matrices for a $G_l$-bundle over ${\Sigma}_\tau$. Therefore ${\omega}^p$ represents an element from . It is an obstruction (\[obs2\]).
As in (\[last\]), modify the transition matrix $$\begin{gathered}
\hat\rho_b\to\hat\rho_b'=\exp\left\{-\frac{2\pi
i}{p}\left(z+\frac{\tau}{2}\right)\right\} \hat\rho_b. \end{gathered}$$ We come to the ${{\rm GL}(N,{\mathbb C})}$-bundle of degree $p$ (mod$\,N$).
Local description {#section2.3}
-----------------
There exists another description of a holomorphic bundles over ${\Sigma}_g$. Let $w_0$ be a fixed point on ${\Sigma}_g$ and $D_{w_0}$ ($D^\times_{w_0}$) be a disc (punctured disc) with a center $w_0$ with a local coordinate $z$. A bundle $E$ over ${\Sigma}_g$ can be trivialized over $D$ and over ${\Sigma}_g\setminus w_0$. These two trivializations are related by a ${{\rm GL}(N,{\mathbb C})}$ transformation $g(z)$, holomorphic on $D^\times_{w_0}$. If we consider another trivialization over $D$ then $g$ is multiplied from left by an invertible matrix $h$ on $D$. Likewise, a trivialization over ${\Sigma}_g\setminus w_0$ is determined up to the multiplication on the right $g\to gh$ , where $h\in{{\rm GL}(N,{\mathbb C})}$ is holomorphic on ${\Sigma}_g\setminus w_0$. Thus, the set of isomorphism classes of rank $N$ vector bundles is described as a double-coset[$$\begin{gathered}
{{\rm GL}(N,{\mathbb C})}(D_{w_0})\setminus
{{\rm GL}(N,{\mathbb C})}(D^\times_{w_0}) /{{\rm GL}(N,{\mathbb C})}({\Sigma}_g\setminus w_0), \end{gathered}$$ where ${{\rm GL}(N,{\mathbb C})}(U)$ denote the group of ${{\rm GL}(N,{\mathbb C})}$-valued holomorphic functions on $U$.]{}
Let $\det\,g(z)=1$. If $g(ze^{2\pi i})=g(z)$ then it defines a $\SLN$-bundle over ${\Sigma}_g$. But if the monodromy is nontrivial $$\begin{gathered}
\label{abc11}
g(ze^{2\pi i})={\omega}g(z),\qquad {\omega}^N=1,
\end{gathered}$$ then $g(w)$ is a transition matrix for a $\PSLN$-bundle but not for a $\SLN$-bundle. This relation is similar to (\[pi1\]).
Let us choose a trivialization of $E$ over $D$ by choosing $N$ linear independent holomorphic sections $\vec
s=(s_1,s_2,\ldots,s_N)$. Thereby, the bundle $E$ over $D$ is represented by a sum of $N$ line bundles $\clL_1\oplus\clL_2\oplus\cdots\oplus\clL_N$. The sections over ${\Sigma}_g\setminus w_0$ are obtained by the action of the transition matrix $\vec s\,'=\vec s g$.
Let $\vec m$ belongs to the root lattice $(\vec
m=(m_1,m_2,\ldots,m_N)\in Q)$ (\[rl\]). Transform the restriction of the section $\vec s$ on $D^\times_{w_0}$ as $$\begin{gathered}
\label{mod}
s_j\to z^{-m_j}s_j, \qquad j=1,\ldots,N.\end{gathered}$$ Then the transition matrix is transformed by the diagonal matrix $$\begin{gathered}
\label{mtm}
g(z)\to {{\rm diag}}\big(z^{-m_1},z^{-m_2},\ldots,z^{-m_N}\big)g(z).\end{gathered}$$ It implies the transformation of line bundles over $D$ $$\begin{gathered}
\clL_j\to \clL_j\otimes \clO(m_j). \end{gathered}$$ In this way we come to the new bundle ${\tilde{E}}$ (*the modified bundle*). It is defined by the new transition matrix (\[mtm\]). This transformation of the bundle $E$ to ${\tilde{E}}$ (or more exactly the map of sheaves of its sections) $$\begin{gathered}
{\Gamma}(E)\,\mapright{\Xi(\vec m)} \,{\Gamma}({\tilde{E}}),\qquad \Xi(\vec
m)\sim{{\rm diag}}\big(z^{-m_1},z^{-m_2},\ldots,z^{-m_N}\big),
\end{gathered}$$ is called *the modification or the Hecke transformation* of type $\vec m=(m_1,m_2,\ldots,m_N)$. In field-theoretical terms it corresponds to *the t’Hooft operator,* generating by monopoles (see below).
Let us relax the condition $\sum_{j=1}^N m_j=0$. Then the modification $\Xi(\vec m)$ changes the topology of $E$. We come to a nontrivial bundle of degree ${\rm deg}\,({\tilde{E}})={\rm deg}\,(E)+\sum_{j=1}^N
m_j$. In next section we illustrate this fact.
Now assume that $\vec m$ belongs to the weight lattice $P$ (\[wl\]). Then the modification $\Xi(\vec m)$ changes the characteristic class of a ${{\rm PGL}(N,{\mathbb C})}$-bundle $E$. To prove it let us pass to the basis of the fundamental weights (\[fw1\]) $$\begin{gathered}
\vec
m=\sum_{j=1}^{N}m_je_j=\sum_{k=1}^{N-1}n_k\varpi_k. \end{gathered}$$
It follows from (\[fw1\]) that $m_j$ and $n_k$ are related as $$\begin{gathered}
m_1=\f1{N}((N-1)n_1+(N-2)n_2+\dots+n_{N-1}),\nonumber \\
m_2=\f1{N}(-n_1+(N-2)n_2+\dots+n_{N-1}),\nonumber \\
\cdots\cdots\cdots\cdots\cdots\cdots\cdots \cdots\cdots\cdots\cdots\cdots\cdots\cdots \nonumber\\
m_N=\f1{N}(-n_1-2n_2-\dots-(N-1)n_{N-1}),
\qquad n_k=m_k-m_{k+1}. $$ Rewrite the modification in the form of the product of the diagonal matrices $$\begin{gathered}
\label{gtr} \Xi(\vec
n)\sim \prod_{k=1}^{N-1}{{\rm diag}}\big(z^{-n_k\varpi_k}\big). \end{gathered}$$ It follows from (\[fw1\]) that the monodromy of this matrix around the point $z=0$ is $$\begin{gathered}
\label{mo1} \exp\left(-\frac{2\pi
i}{N}\sum_{k=1}^{N-1}kn_k\right){\rm Id}_N.\end{gathered}$$ Therefore, the characteristic class of the adjoint bundle is unchanged if $$\begin{gathered}
\sum_{k=1}^{N-1}kn_k=N\sum_{j=1}^{N-1}m_j=0,
\qquad ({\rm mod}\,N).\end{gathered}$$ In this case the weight vector $\vec m$ belongs to the root lattice $Q$. Otherwise, we come to the non-trivial monodromy (\[mo1\]). It is an obstruction to lift the ${{\rm PGL}(N,{\mathbb C})}$-bundle to a $\SLN$-bundle. This element can be identified with the monodromy (\[pi1\]) and in this way with an element from $H^2({\Sigma},\mZ_N)$. As it was mentioned above, the modified bundle ${\tilde{E}}$ can be lifted to a ${{\rm GL}(N,{\mathbb C})}$ bundle. Let us act on the modified sections (\[mod\]) by the scalar matrix $$\begin{gathered}
h=z^{\frac{2\pi i}{N}\sum\limits_{k=1}^{N-1}kn_k}\,{\rm Id}_N.\end{gathered}$$ It is a ${{\rm GL}(N,{\mathbb C})}$ gauge transformation. The monodromy of the new transition matrix is trivial. Therefore, we come to the ${{\rm GL}(N,{\mathbb C})}$ bundle. The bundle is topologically nontrivial – it has degree $$\begin{gathered}
\label{edgr}
p=\sum_{k=1}^{N-1}kn_k=N\sum_{j=1}^{N-1}m_j.\end{gathered}$$ It follows from (\[mo1\]) that the characteristic class $\xi$ and the degree $p$ are related as $$\begin{gathered}
\xi=\exp\frac{2\pi i}{N}p.\end{gathered}$$ The set of modifications that changes the degree on $p$ is defined as solutions of (\[edgr\]) in integers $n_k$.
Assume that the bundle $E$ is equipped with a holomorphic connection. On $D^\times_{w_0}$ it takes the form $({\partial}_z+A_z) dz$ and can be considered as an element of the affine Lie coalgebra $\widehat{\rm gl}^*(N,\mC)(D^\times_{w_0})$ The gauge transformation (\[mtm\]) acts on $A_w$ acts as the coadjoint action $$\begin{gathered}
\label{a}
(A_z)_{jk}dz \to \big(z^{m_k-m_j}(A_z)_{jk}(1-{\delta}_{jk})-m_jz^{-1}{\delta}_{jk}\big)dz.\end{gathered}$$ Let $\vec m\in P$. Then the first term in the r.h.s. is well defined, since $m_k-m_j$ is integer. The last term represents the shift action (\[shi\]) of the affine group $\bar W_a$ (\[q25\]) on the connection. The topology of $E$ is not changed if $\vec m\in
Q$ and we come to description of the characteristic class as elements from factor group $\bar W_a/W_a$ (\[fwe\]). We come again to this point in Section \[section4\].
Let $N=pl$ with $l\neq 1,N$ and $G_l=\SLN/\mZ_l$ (\[gg\]). Consider the gauge transformation (\[gtr\]) with $\vec m$ $(\vec\varpi)\in{\Gamma}(^LG)$ (\[gg\]). For example, we can take $\vec\varpi=(p,0,\ldots,0)$. Then the monodromy (\[mo1\]) belongs to the group $\mZ_l$. It means that the modified bundle ${\tilde{E}}$ is the $G_l$-bundle that cannot be lifted to the $\SLN$-bundle (see (\[obs2\])).
The modification can be performed in an arbitrary number of points $w_{a}$, $(a=1,\ldots,n)$. To this end define the isomorphism classes of vector bundles as the quotient $$\begin{gathered}
\prod_{a=1}^n{{\rm GL}(N,{\mathbb C})}(D_{w_a}) \setminus
\prod_{a=1}^n{{\rm GL}(N,{\mathbb C})}(D^\times_{w_a}) /{{\rm GL}(N,{\mathbb C})}({\Sigma}_g\setminus
(w_1,\ldots,w_a)). \end{gathered}$$ We have $n$ transition matrices $g_a(z_a)$ representing an element of the quotient, where $z_a$ is a local coordinate. Let $\Xi(\vec m_a)$ denotes the modification of $E$ at $w_{a}$ and $\Xi=\prod_{a=1}^n\Xi(\vec m_a)$. The order of modifications in the product is irrelevant, since they commute. To calculate the monodromy of $\Xi$ we choose the same orientation in all points $w_a$. The characteristic class of $\xi$ of modified bundle ${\tilde{E}}$ corresponds to $$\begin{gathered}
\prod_{a=1}^n\exp\left(-\frac{2\pi
i}{N}\sum_{k=1}^{N-1}kn_k^a\right). \end{gathered}$$
Bogomolny equation {#section3}
==================
**Definition.** Let $W=\mR\times{\Sigma}_g$. Consider a bundle $V$ over $W$ equipped with the curvature $F$. Let $\phi$ be a zero form on $W$ taking value in sections of the adjoint bundle $\phi\in{\Omega}^0(W,{\rm End}\, V)$. It is the so-called Higgs field.
The Bogomolny equation on $W$ takes the form $$\begin{gathered}
\label{q1} F=*D\phi. \end{gathered}$$ Here $*$ is the Hodge operator on $W$ with respect to the metric $ds^2$ on $W$. In local coordinates $(z, {\bar{z}})$ on $\Sigma_g$ and $y$ on the real line $ds^2=g|dz|^2+dy^2$, where $g(z,{\bar{z}})|dz|^2$ is a metric on ${\Sigma}_g$. Then the Hodge operator is defined as $$\begin{gathered}
\star dy=\tfrac 12 i gdz\wedge d{\bar{z}},\qquad
\star dz= -i dz\wedge dy, \qquad \star d{\bar{z}}= i d{\bar{z}}\wedge dy,\end{gathered}$$ and (\[q1\]) becomes
\[q2\] $$\begin{gathered}
\label{q2-1}
{\partial}_{z} A_{{\bar{z}}}-{\partial}_{{\bar{z}}} A_z+[A_z,A_{{\bar{z}}}]=\frac{ig(z,{\bar{z}})}{2}\left({\partial}_y\phi+[A_y,\phi]\right),
\\
\label{q2-2} {\partial}_y A_z-{\partial}_{z} A_y+[A_y,A_z]=i({\partial}_{z}\phi+[A_z,\phi]),
\\
\label{q2-3} {\partial}_yA_{{\bar{z}}}-{\partial}_{{\bar{z}}}
A_y+[A_y,A_{{\bar{z}}}]=-i({\partial}_{{\bar{z}}}\phi+[A_{{\bar{z}}},\phi]).\end{gathered}$$
In what follows we will consider only $\PSLN$-bundles.
A monopole solution of this equation is defined in the following way. Let ${\tilde{W}}=(W\setminus \vec{x}^0=(y=0,z=z_0))$. The Bianchi identity $ DF=0$ on ${\tilde{W}}$ implies that $\phi$ can be identified with the Green function for the operator $\star D\star D$ $$\begin{gathered}
\label{GF}
\star D\star D\phi=M{\delta}(\vec{x}-\vec{x}^0),\\
\label{mm}
M={{\rm diag}}(m_1,m_2,\ldots,m_N)\in{{\rm gl}(N, {\mathbb C})},\qquad \vec m=(m_1,m_2,\ldots,m_N)\in
P \quad \eqref{wl},\end{gathered}$$ and $(m_1,m_2,\ldots,m_N)$ are the monopole charges. We explain below this choice of $M$. This equation means that $\phi$ is singular at $\vec{x}^0$.
**Boundary conditions and gauge symmetry.** In what follows except Section \[section3.1\] we assume that ${\partial}_y\phi$ vanishes when $y\to \pm\infty$. It is the Neumann boundary conditions for the Higgs field, while the gauge fields are unspecified. Let $V_\pm$ be restrictions of $V$ to the bundles over ${\Sigma}_g$ on the “left end” and “right end” of $W: y\to \pm\infty$. These bundles are flat. It follows from (\[q5-1\]), where the gauge $A_y=0$ is assumed. It was proved in [@KW] that in absence of the source $M=0$ in (\[GF\]) the only solutions of (\[q1\]) with these boundary conditions are $F=0$, $\phi=0$. Note that these boundary conditions differ from ones chosen in [@KW].
The Bogomolny equation defines a transformation $V_-\to V_+$. ($E$ and ${\tilde{E}}$ in our notations in Introduction.) We will see in next sections that in general the characteristic classes of bundles are changed under these transformations. It depends on the monopole charges $\vec m=(m_1,m_2,\ldots,m_N)$.
The system (\[q2\]) is invariant with respect to the gauge group $\clG$ action: $$\begin{gathered}
A_z\rightarrow hA_zh^{-1}+{\partial}_{z} hh^{-1} , \qquad A_{{\bar{z}}}\rightarrow
hA_{{\bar{z}}} h^{-1}+{\partial}_{{\bar{z}}} hh^{-1}, \nonumber\\
A_y\rightarrow hA_yh^{-1}+{\partial}_y hh^{-1}, \qquad
\phi\rightarrow h\phi h^{-1},\label{q3}\end{gathered}$$ where $h\in\clG$ is a smooth map $W\to{{\rm GL}(N,{\mathbb C})}$. To preserve the r.h.s. in (\[GF\]) it should satisfy the condition $[h(\vec{x}^0),M]=0$.
Assume for simplicity that $V$ is an adjoint bundle. Since the gauge fields for $y=\pm\infty$ are unspecified and only flat we can act on them by boundary values of the gauge group . Then $\clM_\pm=\{V_\pm\}/\clG_\pm$ are the moduli spaces of flat bundles.
**Relations to integrable systems.** The moduli spaces of flat bundles are phase spaces of non-autonomous Hamiltonian systems related to the isomonodromy problems over ${\Sigma}_g$. The isomonodromy problem takes the form $$\begin{gathered}
\label{ip}
[{\partial}_{z}+A_z,\Psi]=0 ,\qquad
[\partial_{{\bar{z}}}+A_{{\bar{z}}},\Psi]=0.
\end{gathered}$$ Here $\Psi\in{\Omega}^0({\Sigma}_g,{\rm Aut}\, V)$ is the Baker–Akhiezer function. These system is compatible for any degree of bundle, because it is defined in the adjoint representation. One example of these systems we have mentioned in Introduction ($V_-\to$ Painlevé VI) and ($V_+\to$ Zhukovsky–Volterra gyrostat).
It is known, that the moduli space of flat bundles are deformation (the Whitham deformation) of the phase spaces of the Hitchin integrable systems – the moduli spaces of the Higgs bundles. To consider this limit one should replace a holomorphic connection by the $\kappa$-connection $\kappa{\partial}_{z}+A_z$ introduced by P. Deligne and take a limit $\kappa\to 0$. It is a quasi-classical limit in the linear problem (\[ip\]). Details can be found in [@FN; @Kr; @LO; @Ar]. In this way a monopole solution put in a correspondence (symplectic Hecke correspondence) two Hitchin systems (the first and the last examples in Introduction). But Bogomolny equation tells us more. It describes an evolution from one type of system to another.
It is possible to generalize (\[GF\]) and consider multi-monopole sources $\sum_aM_a{\delta}(\vec{x}-\vec{x}_a^0)$ in the r.h.s. This generalization will correspond to modifications in a few points of ${\Sigma}_g$ described at the end of previous section.
It is interesting that in some particular cases this situation was discussed in the frameworks of a supersymmetric Yang–Mills theory [@SY; @Po][^1]. It was observed there that a monopole configuration corresponds to a soliton type evolution along $y$. Therefore, it can be suggested that the system (\[q2\]) is integrable. We did not succeed to prove this fact, but propose a linear problem related to the Bogomolny equation. An associated linear problem allows one in principal to apply the methods of the Inverse Scattering Problem or the Whitham approximation to find solutions [@ZMNP]. Assume that the metric $g$ on ${\Sigma}_g$ is a constant. Then the system (\[q2\]) is the compatibility condition for the linear system $$\begin{gathered}
\left({\partial}_z+A_z+\tfrac 12 {\lambda}^{-1}g({\partial}_y+A_y+i\phi)\right)\psi=0,\nonumber\\
\left({\partial}_{{\bar{z}}}+A_{{\bar{z}}}+\tfrac 12 {\lambda}g({\partial}_y+A_y-i\phi)\right)\psi=0,$$ where ${\lambda}\in\mC P^1$ is a spectral parameter. It can be suggested that monopole solution of (\[q2\]) corresponds to a soliton solution of this system. We will not develop here this approach[^2].
**Gauge fixing.** Choose a gauge fixing conditions as: $A_{{\bar{z}}}=0$. Holomorphic functions $h=h(y,z)$ preserve this gauge. Then $$\begin{gathered}
-{\partial}_{{\bar{z}}} A_z=\frac{ig}{2}\left({\partial}_y\phi+[A_y,\phi]\right),\nonumber
\\
{\partial}_y A_z-{\partial}_{z} A_y+[A_z,A_y]=i({\partial}_{z}\phi+[A_z,\phi]),\nonumber
\\
{\partial}_{{\bar{z}}} A_y=i{\partial}_{{\bar{z}}}\phi.$$ The last equation means that $A_y-i\phi$ is holomorphic. It follows from (\[q3\]) that the gauge transformation of this function is $$\begin{gathered}
A_y-i\phi\rightarrow h(A_y-i\phi)h^{-1}+{\partial}_y
hh^{-1}.
\end{gathered}$$ Thus, we can keep $A_y=i\phi$ by using holomorphic and $y$-independent part of the gauge group $({\partial}_yh=0)$. Finally, we come to the system
\[q5\] $$\begin{gathered}
\label{q5-1} {\partial}_{{\bar{z}}} A_z=-\frac{ig}{2}{\partial}_y\phi,
\\
\label{q5-2} {\partial}_y A_z-2i{\partial}_{z} \phi+2i[A_z,\phi]=0,
\\
\label{q5-3} A_y=i\phi,
\\
\label{q5-4} A_{{\bar{z}}}=0.\end{gathered}$$
Two upper equations from (\[q5\]) lead to the Laplace type equation $$\begin{gathered}
\label{q6}
{\partial}^2_y\phi+\frac{4}{g}({\partial}_{z}{\partial}_{{\bar{z}}}\phi+{\partial}_{{\bar{z}}}[A_z,\phi])=0.\end{gathered}$$
In scalar case (\[q6\]) is simplified $$\begin{gathered}
\label{q7}
{\partial}^2_y\phi+\frac{4}{g}{\partial}_{z}{\partial}_{{\bar{z}}}\phi=0. \end{gathered}$$
Rational solution in scalar case {#section3.1}
--------------------------------
In this subsection we replace ${\Sigma}_g$ by $\mC$. The coordinates $z,{\bar{z}}$ on $\mC$ will play the role of local coordinates on ${\Sigma}_g$. Consider (\[q7\]) on ${\tilde{W}}=\mR\times\mC\setminus(0,0,0)$. In this particular case we can choose the boundary conditions in the following form: $$\begin{gathered}
\label{q8}
\phi|_{y=\pm\infty}=0, \\
\label{q9}
A_z|_{y=\pm\infty}=0.\end{gathered}$$ The solution of (\[q7\]) with $g=1$ satisfying (\[q8\]) has the form: $$\begin{gathered}
\label{q10} \phi=c\frac{1}{\sqrt{y^2+z{\bar{z}}}}, \end{gathered}$$ where $c$ is a constant. So in fact we deal here with the Laplace equation on $\mathbb{R}\times\mathbb{C}\setminus(0,0,0)$. It follows from (\[q9\]) and from the equation ${\partial}_{{\bar{z}}} A_z=-\frac{i}{2}{\partial}_y\phi$ (\[q5-1\]) that $$\begin{gathered}
A_z(z,\bar{z},y)=A_z^+(z,\bar{z},y),\qquad y>0\quad \mbox{and}\quad y=0, \quad z\neq 0,\nonumber
\\
A_z(z,\bar{z},y)=A_z^-(z,\bar{z},y),\qquad y<0,\label{q101}\end{gathered}$$ where $$\begin{gathered}
A_z^+(z,\bar{z},y)=-ic\left(\frac{1}{z}\frac{y}{\sqrt{y^2+z{\bar{z}}}}-\frac{1}{z}\right)
+{\rm const},\nonumber
\\
A_z^-(z,\bar{z},y)=-ic\left(\frac{1}{z}\frac{y}{\sqrt{y^2+z{\bar{z}}}}+\frac{1}{z}\right)
+{\rm const},$$ and $A_z(z,\bar{z},y)$ is a connection on the line bundle $\clL$ over ${\tilde{W}}$. The connection has a jump $-2ic\frac{1}{z}$ at $y=0$. To deal with smooth connections we compensate it by a holomorphic gauge transform that locally near $\vec x_0$ has the form $h\sim
z^m$. Here $m$ should be integer, because $h$ is a smooth function. Notice that all holomorphic line bundles over $S^2$ are known to be $\mathcal{O}(m)$-bundles, $m\in\mZ$. Thus, we have $c=i\frac{m}{2}$, $m\in\mathbb{Z}$. This usually referred as a quantization of the monopole charge. In fact the constant $c$ contains factor $4\pi$ (area of a unit sphere) which yields a proper normalization of delta-function and appears in Gauss’s law. The gauge transformation $h$ is the modification (\[mod\]), (\[mtm\]) for line bundles over $\mC P^1$. This is what we mean saying that the described 3-dimensional construction characterizes the modification of the corresponding bundle.
Consider for a moment the general situation $ W=\mR\times{\Sigma}_g$ and let $z,{\bar{z}}$ be local coordinates on ${\Sigma}_g$. Locally near $\vec
x_0=(0,0,0)$ connections corresponding to solutions of (\[q7\]) have the form (\[q101\]). Let $S^2$ be a small sphere surrounding the point $\vec x_0$ in $W$ and ${\Sigma}_{g,\pm}$ be the left and right boundaries of $W$ and $\clL_{\pm}$ are the corresponding restrictions of $\clL$. Then (as it is explained in [@KW] in detail) $$\begin{gathered}
\int_{{\Sigma}_{g,+}}F=\int_{{\Sigma}_{g,-}}F +m,\end{gathered}$$ where $F$ is a curvature of the connection $A$. In other words, the monopole solution with the charge $m$ increases the degree of bundle by $m$ $({\rm deg}\, \clL_+={\rm deg}\, \clL_-+m)$.
Elliptic solution in scalar case {#section3.2}
--------------------------------
The Laplace equation (\[q7\]) on ${\Sigma}_\tau$ has the form $$\begin{gathered}
\label{q111} {\partial}^2_y\phi+4({\mathcal
Im}(\tau))^2{\partial}_{z}{\partial}_{{\bar{z}}}\phi=0,\end{gathered}$$ or $$\begin{gathered}
{\partial}^2_y\phi+(2\pi{\alpha})^2{\partial}_{z}{\partial}_{{\bar{z}}}\phi=0,\qquad {\alpha}^{-1}=\frac{2\pi
i}{\tau-\bar{\tau}},\end{gathered}$$ and ${\mathcal Im}(\tau)$ is the area of parallelogram of periods. We give two representations of the Green function $\phi$ and prove their equivalence using the same technique as for the Kronecker series described in [@We].
A naive elliptic solution of (\[q111\]) on ${\tilde{W}}$ is obtained by averaging (\[q10\]) over the lattice $\Gamma=\mZ\oplus\tau\mZ\subset\mC$:[^3] $$\begin{gathered}
\label{q12}
\phi(z,y)=c\sum\limits_{{\gamma}\in\Gamma}\frac{1}{\sqrt{(\pi{\alpha}y)^2+|z+{\gamma}|^2}}. \end{gathered}$$ However the series diverges. That is why we consider its generalization $$\begin{gathered}
\label{q120} \clR(s,x,z,y)=
c\sum\limits_{{\gamma}\in\Gamma}\frac{\chi({\gamma},x)}{({(\pi{\alpha}y)^2+|z+{\gamma}|^2})^s}, \qquad \clR\left(\tfrac{1}{2},0,z,y\right)=\phi(z,y), \end{gathered}$$ where $$\begin{gathered}
\chi({\gamma},x)=e^{{\alpha}^{-1}({\gamma}\bar{x}-\bar{{\gamma}}x)}\end{gathered}$$ is a character ${\mathbb Z}\times{\mathbb Z}\rightarrow {\mathbb
C}^*$ of the additive group $\Gamma$ and $s$, $x$ are complex parameters. The characters are double-periodic $$\begin{gathered}
\chi({\gamma},x+1)=\chi({\gamma},x),\qquad
\chi({\gamma},x+\tau)=\chi({\gamma},x),\qquad {\gamma}\in\Gamma, \end{gathered}$$ while the series $\clR(s,x,z,y)$ are quasi-periodic $$\begin{gathered}
\clR(s,x,z+1,y)=e^{{\alpha}^{-1}(x-\bar x)}\clR(s,x,z,y), \nonumber\\
\clR(s,x,z+\tau,y)=e^{{\alpha}^{-1}(x\bar\tau-\bar x\tau)}\clR(s,x,z,y).\label{qper1}\end{gathered}$$ The variable $x$ describes behavior of $\clR(s,x,z,y)$ on the lattice $\Gamma$. In other words, $x$ parameterizes the moduli space of line bundles on $\Sigma_\tau$. Note that for $\clR e\,s>1$ the series in the r.h.s. of (\[q120\]) converges. The function $$\begin{gathered}
\label{f1} \clR\left(\tfrac{1}{2},x,z,y\right)=
c\sum\limits_{{\gamma}\in\Gamma}\frac{\chi({\gamma},x)}{({(\pi{\alpha}y)^2+|z+{\gamma}|^2})^{\frac{1}{2}}}=\phi(x,z,y)
\end{gathered}$$ is the formal solution of (\[q111\]) with the quasi-periodicity conditions (\[qper1\]).
Another representation of the Green function can be obtained by the Fourier transform. Define the delta-functions $$\begin{gathered}
\delta(y)=\int\limits_{-\infty}^{+\infty}dp e^{2\pi ipy}, \qquad
\delta^{(2)}(z,{\bar z})=\sum\limits_{{\gamma}\in\Gamma}\chi({\gamma},z).\end{gathered}$$ Then $$\begin{gathered}
\sum\limits_{{\gamma}\in\Gamma}\chi({\gamma}+x,z)=\chi(x,z)\sum\limits_{{\gamma}\in\Gamma}\chi({\gamma},z)=
\chi(x,z)\delta^2(z,{\bar{z}})=\delta^2(z,{\bar{z}}), \end{gathered}$$ Let ${\tilde{\phi}}$ be the Green function with the quasi-periodicity (\[qper1\]) $$\begin{gathered}
{\partial}^2_y{\tilde{\phi}}+(2\pi{\alpha})^2{\partial}_{z}{\partial}_{{\bar{z}}}{\tilde{\phi}}=c\delta(y)\delta^{(2)}(z,{\bar
z}). \end{gathered}$$ Expanding it in the Fourier harmonics we find $$\begin{gathered}
{\tilde{\phi}}=-\frac{c}{4\pi^2}\sum\limits_{{\gamma}\in\Gamma}\int\limits_{-\infty}^{+\infty}dp
\frac{ e^{2\pi ipy}}{p^2+|{\gamma}+z|^2}\chi({\gamma}+x,z).\end{gathered}$$ Integrating over $p$ provides factor $\pi$ and leads to the following expression: $$\begin{gathered}
\label{q1211}
{\tilde{\phi}}(x,z,y)=-\frac{c}{4\pi}\sum\limits_{{\gamma}\in\Gamma}\frac{1}{|{\gamma}+x|}e^{-2\pi|{\gamma}+x||y|}\chi({\gamma}+x,z).\end{gathered}$$ It is worthwhile to note that the solution (\[q1211\]) is well defined. Our goal is to find interrelations between (\[q1211\]) and (\[f1\]).
Consider a generalization of (\[q1211\]) $$\begin{gathered}
\label{t}
I(s,x,z,y)=2c\pi^sy^{s-{\frac{1}{2}}}\sum\limits_{{\gamma}\in\Gamma}
\frac{K_{s-{\frac{1}{2}}}(2\pi |y||{\gamma}+x|)}{|{\gamma}+x|^{s-{\frac{1}{2}}}} \chi({\gamma}+x,z).\end{gathered}$$ Here $K_\nu$ is the Bessel–Macdonald function $$\begin{gathered}
K_{\nu}(2\pi yz)=\frac{{\Gamma}(\nu+{\frac{1}{2}})(z)^\nu}{2(\pi y)^\nu{\Gamma}({\frac{1}{2}})}
\int\limits_{-\infty}^{+\infty}dp\frac{ e^{2\pi ipy}}{(p^2+z^2)^{\nu+{\frac{1}{2}}}}.
\end{gathered}$$ The function $I(s,x,z,y)$ is the Green function for the pseudo-differential operator $$\begin{gathered}
\left({\partial}_y^2+4{\alpha}^2\pi^2{\partial}_{z}{\partial}_{{\bar{z}}}\right)^s \end{gathered}$$ on $\mR\times{\Sigma}_\tau$ with the boundary conditions (\[qper1\]). Since $$\begin{gathered}
K_{{\frac{1}{2}}}(x)=\sqrt{\frac{\pi}{2x}}e^{-x},\end{gathered}$$ we conclude that for $s=1$ $I$ coincides with ${\tilde{\phi}}$ (\[q1211\]) up to constant.
We are going to establish a relation between (\[q120\]) and (\[t\]), and in this way between (\[q12\]) and (\[q1211\]). Let us prove that $$\begin{gathered}
\label{q121} I(s,x,z,y)=
c\sum\limits_{\gamma\in\Gamma} \int\limits_{-\infty}^\infty
dp\int\limits_0^\infty \frac{d{t}}{t}t^s e^{-t(p^2+|{\gamma}+x|^2)+2\pi
ipy}\chi({\gamma}+x,z). \end{gathered}$$ In fact, using the integral representation for the Gamma-function $$\begin{gathered}
\label{gam}
\Gamma(s)=\int\limits_{0}^{\infty}\frac{dt}{t}t^{s}e^{-t} \end{gathered}$$ and taking the integral over $t$ in (\[q121\]) we come to (\[t\]).
The representation (\[q121\]) is universal and can serve to define $\clR$ (\[q120\])
\[lemma3.1\] The function $\clR(s,x,z,y)$ has a representation as the Fourier integral $$\begin{gathered}
\label{fi} \clR(s,x,z,y)=\f1{{\Gamma}(s)\chi({\gamma},z)}
\int_{\infty}^{\infty}dk \, I\left(s,z,x,\frac{k}{\pi{\alpha}}\right)e^{-2\pi iky}.\end{gathered}$$
Substitute in (\[fi\]) $I(s,x,z,y)$ (\[q121\]) and take first integral over $k$. We come to the condition $p=\pi{\alpha}y$. Then using the integral representation for the Gamma-function (\[gam\]) we obtain (\[q120\]).
\[remark3.1\] The series (\[t\]) is a three-dimensional generalization of the Kronecker series (see [@We]) $$\begin{gathered}
K(x,x_0,s)=\sum\limits_{\gamma}\chi({\gamma},x_0)|x+{\gamma}|^{-2s}.\end{gathered}$$ Using the Poisson summation formula Kronecker proved that $$\begin{gathered}
\Gamma(s)K(x,x_0,s)={\alpha}^{1-2s}\Gamma(1-s) K(x_0,x,1-s)\chi(x,x_0)\,.\end{gathered}$$
Our purpose is to generalize this functional equation for the 3-dimensional case $\Sigma_\tau\times \mathbb{R}$. It takes the following form.
\[lemma3.2\] The function $I(s,x,z,y)$ satisfies the functional equation: $$\begin{gathered}
\label{q122}
I(s,x,z,y)=\chi(x,z)\pi^{-\frac{1}{2}}{\alpha}^{-2s+1}\int\limits_{-\infty}^{+\infty}
dk\, I\left(\frac{3}{2}-s,z,x,\frac{k}{\pi{\alpha}}\right)e^{-2\pi i ky}.
\end{gathered}$$
Following [@We] we subdivide integral (\[q121\]) into two parts $$\begin{gathered}
I(s,x,z,y)=c\sum\limits_{\gamma\in\Gamma}\int\limits_{-\infty}^\infty
dp\int\limits_0^T \frac{d{t}}{t}t^s e^{-t(p^2+|{\gamma}+x|^2)+2\pi
ipy}\chi({\gamma}+x,z)\\
\phantom{I(s,x,z,y)=}{}+
c\sum\limits_{\gamma\in\Gamma}\int\limits_{-\infty}^\infty
dp\int\limits_T^\infty \frac{d{t}}{t}t^s e^{-t(p^2+|{\gamma}+x|^2)+2\pi
ipy}\chi({\gamma}+x,z),\qquad T\in \mathbb{R},\quad T>0.\nonumber\end{gathered}$$
The second term is a well defined function for all $s$. Consider the first one. It is well known that for the series $$\begin{gathered}
\Theta(t,x,x_0)=\sum\limits_{\gamma}e^{-t|x+\gamma|^2}\chi({\gamma},x_0)\end{gathered}$$ the following functional equation holds: $$\begin{gathered}
\Theta(t,x,x_0)=({\alpha}t)^{-1}\Theta\big({\alpha}^{-2}t^{-1},x_0,x\big)\chi(x_0,x).\end{gathered}$$
The latter follows from the Poisson summation formula which states that the averaging of function over some lattice equals the averaging of its Fourier transform over the dual lattice. In the above case the functional equation appears after the Fourier transform for the Gauss integral. Then $$\begin{gathered}
\sum\limits_{\gamma\in\Gamma}\int\limits_{-\infty}^\infty
dp\int\limits_0^T \frac{d{t}}{t}t^s e^{-t(p^2+|{\gamma}+x|^2)+2\pi
ipy}\chi({\gamma}+x,z) \nonumber\\
\qquad{} = \sum\limits_{\gamma\in\Gamma}\int\limits_{-\infty}^\infty
dp\int\limits_0^T \frac{d{t}}{t}t^s
e^{-tp^2-{\alpha}^{-2}t^{-1}|{\gamma}+z|^2+2\pi ipy}\chi({\gamma}+z,x)\chi(x,z)({\alpha}t)^{-1}\nonumber\\
\qquad{} \overset{\substack{\text{integrating}\\ \text{over $p$}}}{=}
\sum\limits_{\gamma\in\Gamma}\int\limits_0^T \frac{d{t}}{t}t^s
e^{-\pi^2{y^2}{t}^{-1}-{\alpha}^{-2}t^{-1}|{\gamma}+z|^2}\chi({\gamma}+z,x)\chi(x,z)({\alpha}t)^{-1}\sqrt{\frac{\pi}{t}}\nonumber\\
\qquad{} \overset{\substack{\text{making} \\ \text{substitution} \\ \text{${\alpha}^{-2}t^{-1}\rightarrow t$}}}{=}
\sum\limits_{\gamma\in\Gamma}\int\limits_{{\alpha}^{-2}T^{-1}}^\infty
\frac{d{t}}{t}t^{\frac{3}{2}-s} \sqrt{\pi}{\alpha}^{2-2s}e^{-t((\pi{\alpha}y)^2+|{\gamma}+z|^2)}\chi({\gamma}+z,x)\chi(x,z).\end{gathered}$$ Let $T={\alpha}^{-1}$. Then $$\begin{gathered}
I(s,x,z,y)=
c\sum\limits_{\gamma\in\Gamma}\int\limits_{{\alpha}^{-1}}^\infty
\frac{d{t}}{t}t^{\frac{3}{2}-s} \sqrt{\pi}{\alpha}^{2-2s}e^{-t((\pi{\alpha}y)^2+|{\gamma}+z|^2)}\chi({\gamma}+z,x)\chi(x,z)\nonumber\\
\phantom{I(s,x,z,y)=}{}+
c\sum\limits_{\gamma\in\Gamma} \int\limits_{-\infty}^\infty
dp\int\limits_{{\alpha}^{-1}}^\infty \frac{d{t}}{t}t^s
e^{-t(p^2+|{\gamma}+x|^2)+2\pi ipy}\chi({\gamma}+x,z).\label{q1213}\end{gathered}$$ The proof follows from (\[q1213\]). One should only substitute $I(s,x,z,y)$ from (\[q1213\]), into (\[q122\]). Formula (\[q1213\]) represents $I$ as the sum of two terms. Direct evaluation shows that the first (of two) term from the l.h.s. of (\[q122\]) equals to the second one from the r.h.s. and vice versa.
From Lemmas \[lemma3.1\] and \[lemma3.2\] we come to the main result of this section $$\begin{gathered}
\fbox{$
\clR(\frac{3}{2}-s,x,z,y)=\frac{\sqrt\pi{\alpha}^{2s-1}}{{\Gamma}(\frac{3}{2}-s)}I(s,x,z,y).
$} \end{gathered}$$ Now put $s=1$. Then one can see that well-defined series $$\begin{gathered}
\label{phi}
\pi\sum\limits_{\gamma\in\Gamma}\chi({\gamma}+x,z)\frac{e^{-2\pi
|y||{\gamma}+x|}}{|{\gamma}+x|} \end{gathered}$$ describes the analytic continuation of the divergent series $$\begin{gathered}
\pi\sum\limits_{\gamma\in\Gamma}\chi({\gamma},x)\frac{1}{\sqrt{(\pi{\alpha}y)^2+|{\gamma}+z|^2}}.\end{gathered}$$
We use (\[q1211\]) as the Green function. Then $$\begin{gathered}
\label{cA}
A_z(z,\bar{z},y,x)=-\frac{ic}{4\pi}\frac{1}{\pi^2{\alpha}^2}\hbox{sgn}(y)
\sum\limits_{{\gamma}\in\Gamma}\frac{1}{{\gamma}+x}e^{-2\pi|{\gamma}+x||
y|}\chi({\gamma}+x,z),
\\
\hbox{sgn}(y)=1\quad \hbox{for}\quad y\geq 0,\qquad
\hbox{sgn}(y)=-1\quad \hbox{for}\quad y<0.\nonumber\end{gathered}$$ Notice that the jump of $A$ (while coming through $y=0,\ z=0$) is obviously defined by the jump of $\hbox{sgn}(y)$.
Note that (\[cA\]) is a formal solution of the Bogomolny equation. For $x\neq 0$ it is not a connection of a line bundle over ${\Sigma}_\tau$ due to its monodromies similar to (\[qper1\]). We will use this solution in next section to define a genuine connection for higher ranks bundles.
In order to compare elliptic configuration with the rational we take $x=0$. Then on the line $y=0$ the connection is proportional to $$\begin{gathered}
A_z\sim\sum\limits_{{\gamma}\neq
0}\frac{1}{{\gamma}}\chi({\gamma},z)=E_1(z)-{\alpha}^{-1}(z-\bar{z}),\end{gathered}$$ where $E_1(z)={\partial}\ln{\vartheta}(z)$ is the so-called first Eisenstein series and ${\vartheta}(z)$ is the theta-function (\[theta\]). $E_1(z)$ has a simple pole at $z=0$ with $\hbox{Res}_{z=0}E_1(z)=1$ and the connection $A_z$ is double-periodic. In terms of (\[q3\]) the gauge transformation $h$ compensating the jump of the connection is given by an integer power of theta function $\vartheta^m(z)$, $m\in {\mathbb Z}$. Thus $$\begin{gathered}
{\partial}\log h={\partial}\log \vartheta^m(z)=mE_1(z). \end{gathered}$$
Arbitrary rank case {#section4}
===================
Here we describe modification of vector bundles of an arbitrary rank. First, we repeat arguments of [@KW] and justify the choice $M$ in (\[mm\]). As before, we consider $\PSLN=G_{ad}$-bundles.
Near the singular point $\vec x^0$ the bundle $V$ is splited in a sum of line bundles. Using the solution (\[q10\]) for a line bundle we take the Higgs field near the singularity in the form $$\begin{gathered}
\phi=\frac{i}{2\sqrt{y^2+z{\bar{z}}}}{{\rm diag}}(m_1,\ldots,m_N).\end{gathered}$$ It follows from (\[q101\]) that $A_z$ undergoes a discontinuous jump at $y=0$ $$\begin{gathered}
\label{awga}
A^+_{z}-A^-_{z}=\frac{i}z\,{{\rm diag}}(m_1,m_2,\ldots,m_N). \end{gathered}$$ To get rid of the singularity of $A$ at $z=0$, as in the Abelian case, one can perform the singular gauge transform $\Xi$ that behaves near $z=0$ as (\[gtr\]) $$\begin{gathered}
\Xi={{\rm diag}}\big(z^{-m_1},z^{-m_2},\ldots,z^{-m_N}\big). \end{gathered}$$ Assume that $\vec
m$ belongs to the weight lattice $\vec m\in P$. It means that $\Xi$ is inverse to the cocharacter ${\gamma}_{ad}$ of $\PSLN$ (${\gamma}_{ad}\in
t(G_{ad})=P^\vee\sim P$ (\[tb\])). As it was explained before, the modified bundle $V_+$ can not be lifted to a $\SLN$ bundle. On the other hand, if $\vec m=(m_1,\ldots,m_N)$ belongs to the root lattice $Q$ (\[tb\]), then $\Xi^{-1}=\bar {\gamma}$ and there is no obstruction to lift $V_+$ to an $\SLN$ bundle. Note that (\[awga\]) describes the affine group $\bar W_a$ (\[q25\]) action in the former case and the affine group $W_a$ (\[sh\]) action in the latter case. From field-theoretical point of view it is an action of the t’Hooft operator on $A_z$ (see (\[a\])).
If $N=pl,$ $(l\neq 1,N)$ one can consider the intermediate situation and ${\gamma}_{G_l}$ (\[gl1\]). It means that $\vec m\in t(^LG_l)\sim{\Gamma}(G_p)$. This embedding provides the modification that allows the ${{\rm PGL}(N,{\mathbb C})}$-bundle to lift to the $G_l=\SLN/\mZ_l$-bundle but not to a $\SLN$ bundle. In this way the monopole charges are related to the characteristic classes of bundles.
One can use the maps to the Cartan subgroups of the solution $\phi(z)$ for a line bundle over ${\Sigma}_\tau$ (\[phi\]) with $x=0$ $$\begin{gathered}
\phi\to \phi\cdot{{\rm diag}}(m_1,m_2,\ldots,m_N). \end{gathered}$$ Unfortunately, in this case $V$ being restricted on ${\Sigma}_\tau$ is splitting globally over ${\Sigma}_\tau$ and defines an unstable bundle, though it allows one to describe its modifications.
There exists a map of $\phi(z,y,x)$ and $A_z(z,y,x)$ with $x\neq 0$ to a non-semisimple elements of $\sln$ $$\begin{gathered}
\left(
\begin{array}{cccc}
0 & k_1\phi(z,y,x_1) & \dots & k_{N-1}\phi(z,y,x_{N-1}) \\
0& 0 & \ldots & 0 \\
\vdots & \ldots & \ddots & \vdots \\
0 & \dots & \ldots & 0 \\
\end{array}
\right),
\\
\left(
\begin{array}{cccc}
0 & k_1A_z(z,y,x_1) & \dots & k_{N-1}A_z(z,y,x_{N-1}) \\
0& 0 & \ldots & 0 \\
\vdots & \ldots & \ddots & \vdots \\
0 & \dots & \ldots & 0 \\
\end{array}
\right).\end{gathered}$$ Since these matrices commute they are solutions of the matrix equation (\[q6\]). The connection has a jump at $y=0$. The bundle is characterized by the diagonal monodromy matrices (\[14\]) $$\begin{gathered}
\rho_a={{\rm diag}}(a_1,a_2,\ldots,a_N),\qquad
\rho_b={{\rm diag}}(b_1,b_2,\ldots,b_N),\end{gathered}$$ where $$\begin{gathered}
a_1=\prod_{j=1}^{N-1}{\sigma}_j^{\f1{N}},\qquad a_2=a_1{\sigma}_1^{-1}, \qquad a_N=a_1{\sigma}_{N-1}^{-1},
\\
b_1=\prod_{j=1}^{N-1}\varsigma_j^{\f1{N}},\qquad b_2=a_1\varsigma_1^{-1},\qquad
b_N=a_1\varsigma_{N-1}^{-1},\end{gathered}$$ $\sigma_j=\exp({\alpha}^{-1}(x_j-\bar
x_j))$, $\varsigma_j=\exp({\alpha}^{-1}(x_j\bar\tau-\bar x_j\tau
))$. Note that they are $y$-independent. Moreover, the singular gauge transform, leading to a continues solution of the Bogomolny equation, belongs to the upper nilpotent subgroup and in this way does not change the topological type of the bundle.
Now we describe non-diagonal modifications $\Xi$ of a ${{\rm PGL}(N,{\mathbb C})}$-bundles over ${\Sigma}_\tau$. We do not know solutions of the Bogomolny equation in this case, and only can assert that the modification “kill the jump” of $A_z$ at $y=0$: $$\begin{gathered}
\Xi^{-1}{\partial}_z\Xi=A_z^+-\Xi^{-1}A_z^-\Xi.
\end{gathered}$$
We use the global description of a bundle $E$ in terms of the transition matrices $\rho_a$, $\rho_b$ (\[14\]) using the approach of [@LOZ1]. Let $$\begin{gathered}
\label{bc0}
\rho_a={\rm Id}_N,\qquad \rho_b=\bfe\,{-u},\qquad (u={{\rm diag}}(u_1,u_2,\ldots,u_N))
, \qquad {\bf e}(a)=\exp (2\pi i a).
\end{gathered}$$ The group commutator of these matrices is ${\rm Id}_N$. Thereby, $E$ can be lifted to a $\SLN$-bundle.
Define a modification $\Xi$ of $E$ to the bundle ${\tilde{E}}$ with the transition matrices (\[qla\]). Then $\Xi$ should intertwine the transition matrices $$\begin{gathered}
\label{8.12}
\Xi(z+1,\tau)= \clQ\times \Xi(z,\tau), \\
\label{8.13}
\Xi(z+\tau,\tau)=\Lambda(z,\tau)\times \Xi(z,\tau) \times{\rm diag}
({\bf e}(u)).\end{gathered}$$ The matrix $\Xi(z)$ degenerates at $z=0$ and we assume that it has a simple pole. These conditions fix $\Xi(z)$. It can be expressed in terms of the theta-functions with characteristics $$\begin{gathered}
\Xi_{kj}(z, u_1,\ldots,u_N;\tau)=\frac{
\theta{\left[\begin{array}{c}
\frac{k}N-\frac12\\
\frac{N}2
\end{array}
\right]}(z-Nu_j, N\tau ) } { \theta^{\f1{N}}(z, \tau )}, \end{gathered}$$ where $$\begin{gathered}
\theta{\left[\begin{array}{c}
a\\
b
\end{array}
\right]}(z , \tau ) =\sum_{j\in \Bbb Z} \exp\,2\pi
i\left((j+a)^2\frac\tau2+(j+a)(z+b)\right). \end{gathered}$$ The quasi-periodicity properties (\[8.12\]), (\[8.13\]) follow from the properties of the theta-functions $$\begin{gathered}
\theta{\left[\begin{array}{c}
a\\
b
\end{array}
\right]}(z+1 , \tau )={\bf e}(a) \theta{\left[\begin{array}{c}
a\\
b
\end{array}
\right]}(z , \tau ), \\
\theta{\left[\begin{array}{c}
a\\
b
\end{array}
\right]}(z+a'\tau , \tau ) ={\bf e}\left(-{a'}^2\frac\tau2
-a'(z+b)\right) \theta{\left[\begin{array}{c}
a+a'\\
b
\end{array}
\right]}(z , \tau ).\end{gathered}$$
This modification has the type $(\frac{N-1}{N},-\f1{N},\ldots,-\f1{N})$. The modification that allows to lift ${\tilde{E}}$ to ${{\rm GL}(N,{\mathbb C})}$- bundle is $$\begin{gathered}
\Xi_1(z)=h(z)\Xi(z)= \theta{\left[\begin{array}{c}
\frac{k}N-\frac12\\
\frac{N}2
\end{array}
\right]}(z-Nu_j, N\tau ), \end{gathered}$$ where the gauge transformation $h$ is the diagonal matrix $$\begin{gathered}
h(z)=\theta^{\f1{N}}(z, \tau
){\rm Id}_N.\end{gathered}$$ This modification intertwine the boundary conditions (\[bc0\]) with $$\begin{gathered}
\rho_a=\clQ,\qquad \rho_b={\tilde{{\Lambda}}},\qquad {\tilde{{\Lambda}}}=e^{-2\pi
i(\frac{z}N+\frac{\tau}{2N})}{\Lambda}. \end{gathered}$$ The last transformation belongs to ${{\rm GL}(N,{\mathbb C})}$. Moreover, it can be proved that $$\begin{gathered}
\det\left[\frac{\Xi_1(z, u_1,\ldots,u_N;\tau)} {i\eta(\tau)}\right]=
\frac{{\vartheta}(z)}{i\eta(\tau)}\prod\limits_{1\leq k<l\leq N}
\frac{{\vartheta}(u_l-u_k)}{i\eta(\tau)},\end{gathered}$$ where $\eta(\tau)=q^{\frac{1}{24}}\prod_{n>0}(1-q^n)$ is the Dedekind function $(q=\exp 2\pi i\tau)$ and ${\vartheta}(z)$ is the theta-function (\[theta\]). Since ${\vartheta}(z)$ has a simple pole in ${\Sigma}_\tau$ the bundle ${\tilde{E}}$ is a ${{\rm GL}(N,{\mathbb C})}$-bundle of degree one. This modification provides the Symplectic Hecke correspondence between the elliptic Calogero–Moser system and the Elliptic Top.
Now consider the modification of the trivial bundle $E$ with the transition matrices (\[bc0\]) to the ${\tilde{E}}=E_l$ (\[ic\]), (\[ic1\]), where $\bfu_p=({\tilde{u}}_1,{\tilde{u}}_2,\ldots,{\tilde{u}}_p)$ is the moduli of the modified bundle. The modification takes the form $$\begin{gathered}
\Xi_{kj}(z, \tau) =
\frac{ \theta{\left[\begin{array}{c}
\frac{k}l-\frac12\\
\frac{l}2
\end{array}
\right]}(z-l{\tilde{u}}_{i}), l\tau ) } { \theta^{\f1{l}}(z, \tau
)},\qquad (j=mp+i,\ m=0,\ldots,l-1).\end{gathered}$$ As it was explained in Section \[section2\] the modified bundle can be lifted to $G_l=\SLN/\mZ_l$-bundle, but not to $\SLN$-bundle.
$\boldsymbol{\SLN}$ and $\boldsymbol{\PSLN}$ [@OV; @Bo]
=======================================================
The group $\SLN$ is an universal covering of $\PSLN$ with the center $\mZ_N=\mZ/N\mZ$ $$\begin{gathered}
\label{a1} {\rm Id} \to\mZ_N\to\SLN\to\PSLN\to {\rm Id}. \end{gathered}$$ Therefore $\pi_1(\PSLN)=\mZ_N$. The both groups have the same Lie algebra $\gG$.
**Roots and weights.** The Cartan subalgebra $\gH\subset\gG$ is a hyperplane in $\mC^N$ $$\begin{gathered}
\gH=\left\{\bfx=(x_1,\ldots,x_N)\in\mC^N\,|\,\sum_{j=1}^Nx_j=0\right\}.\end{gathered}$$ The simple roots $\Pi=\{{\alpha}_k\}$ $$\begin{gathered}
{\alpha}_1=e_1-e_2,\quad \ldots,\quad {\alpha}_{N-1}=e_{N-1}-e_N \end{gathered}$$ form a basis in the dual space $\gH^*$. Here $\{e_j\}$ $j=1,\ldots,N$ is a canonical basis in $\mC^N$. They generate the set of roots of type $A_{N-1}$ $$\begin{gathered}
R=\{(e_j-e_k),~j\neq k\}.\end{gathered}$$ The root lattice $Q\subset\gH^*$ takes the form $$\begin{gathered}
\label{rl}
Q=\left\{\sum m_je_j\,|\,m_j\in \mZ,~\sum m_j=0\right\}. \end{gathered}$$
We identify $\gH^*$ and $\gH$ by means of the standard metric on $\mC^N$. Then the coroot system $$\begin{gathered}
R^\vee=\left\{{\alpha}^\vee(R)=
\frac{2({\alpha}^\vee,{\beta})}{({\beta},{\beta})}\in\mZ~{\rm for~any~}{\beta}\in R\right\}\end{gathered}$$ coincides with $R$, and the coroot lattice $Q^\vee$ coincides with $Q$.
The fundamental weights $\varpi_k$, $(k=1,\ldots,N-1)$ are dual to the basis of simple coroots $\Pi^\vee\sim\Pi$ $(\varpi_k({\alpha}^\vee_k)={\delta}_{kj})$ $$\begin{gathered}
\label{fw1}
\varpi_j=e_1+\dots+e_j-\frac{j}{N}\sum_{l=1}^Ne_l,\\
\varpi_1=\left(\frac{N-1}N,-\f1{N},\ldots,-\f1{N}\right),\quad
\varpi_2=\left(\frac{N-2}N,\frac{N-2}N,\ldots,-\frac{2}{N}\right), \quad
\ldots,\nonumber\\
\varpi_{N-1}=\left(\frac{1}N,\frac{1}N,\ldots,\frac{1-N}{N}\right) . \nonumber\end{gathered}$$ In the basis of simple roots the fundamental weights are $$\begin{gathered}
\varpi_k= \f1{N}[(N-k){\alpha}_1+2(N-k){\alpha}_2+\dots+ (k-1)(N-k){\alpha}_{k-1}\\
\phantom{\varpi_k=}{} +k(N-k){\alpha}_k+k(N-k-1){\alpha}_{k+1}+\dots+k{\alpha}_{N-1}].\end{gathered}$$ The fundamental weights generate the weights lattice $$\begin{gathered}
\label{wl}
P\subset\gH^*, \qquad P=\left\{\sum_ln_l\varpi_l\,|\,n_l\in\mZ\right\},
\\
P=\sum_{j=1}^Nm_je_j, \qquad m_j\in \f1{N}\mZ, \qquad m_j-m_k\in\mZ.\nonumber\end{gathered}$$ The weight lattice is generated by $Q$ and the vector $$\begin{gathered}
\varpi_1=e_1-\f1{N}\sum_{j=1}^Ne_j. \end{gathered}$$ The weight lattice $P$ defines representations of $\SLN$, while $Q$ define representations of $\PSLN$.
The factor-group $P^\vee/Q^\vee\,$ $(P^\vee\sim P)$ is the center $\mZ_N$ of $\SLN$. On the other hand it can be identified with the cyclic group symmetry $e_j\to e_{j+1}$ mod$(N)$ of the extended Dynkin graph $\Pi\cup ({\alpha}_0=e_N-e_1)$.
**Characters and cocharacters.** Let $\bar\clT$ $(\clT_{ad})$ be a Cartan torus in $\SLN$ ($\PSLN$). Define the groups of characters[^4] $$\begin{gathered}
\bar{\Gamma}=\{\bar\chi(x)\}=\{\bar\clT\to\mC^*\}, \qquad
{\Gamma}_{ad}=\{\chi_{ad}(x)\}=\{\clT_{ad}\to\mC^*\}.\end{gathered}$$ They can be identified with lattice groups in $\gH^*$ as follows. Let $\varpi_k$ be a basic weight and $\phi=(\phi_1,\phi_2,\ldots,\phi_N)$, $\phi_k=\f1{2\pi i}\ln
x_k$. The functions $$\begin{gathered}
\exp 2\pi i
(\varpi_k\phi),\qquad k=1,\ldots,N-1\end{gathered}$$ generate a basis in $\bar{\Gamma}$. Similarly, for ${\alpha}_k\in \Pi$ $$\begin{gathered}
\exp 2\pi i
({\alpha}_k,\phi)
,\qquad k=1,\ldots,N-1\end{gathered}$$ is a basis in ${\Gamma}_{ad}$. Thereby, we have $$\begin{gathered}
\bar{\Gamma}=P,\qquad {\Gamma}_{ad}=Q.\end{gathered}$$
Define the dual groups of cocharacters $t(\bar G)=\bar{\Gamma}^*$ and $t(G_{ad})={\Gamma}_{ad}^*$ as the maps $$\begin{gathered}
t(\bar
G)=\{\bar{\gamma}=\mC^*\to\bar\clT\}, \qquad
t(G_{ad})=\{{\gamma}_{ad}=\mC^*\to\clT_{ad}\}.\end{gathered}$$ In another way $$\begin{gathered}
t(\bar G)=\{\phi\in \gH\,|\,\bar\chi(e^{2\pi
i\phi})=1\},\qquad t(G_{ad})=\{\phi\in \gH\,|\,\bar\chi_{ad} (e^{2\pi
i\phi})=1\}. \end{gathered}$$ These groups are the groups of the coweight and coroot lattices $$\begin{gathered}
\label{tb} t(\bar G)=Q^\vee\sim
Q,\qquad t(G_{ad})=P^\vee\sim P.\end{gathered}$$ The center of $\bar{\Gamma}=\SLN$ belongs to $\bar \clT$ and is identified with the factor-group $$\begin{gathered}
\label{center} \clZ(\SLN)=P^\vee/t(\bar G)\sim P^\vee/Q^\vee\sim \\
\phantom{\clZ(\SLN)=}{}\sim \pi_1(\PSLN)\sim t(G_{ad})/Q^\vee\sim P/Q=\mZ_N.\nonumber\end{gathered}$$
Let $N=pl$, $(l\neq 1,N)$ and $\mZ_l\subset\mZ_N$ be a subgroup of $Z_N$. Define the factor-group $$\begin{gathered}
G_l=\SLN/\mZ_l.\end{gathered}$$ Then the center $\clZ(G_l)$ of $G_l$ is $\mZ_p$ and $\pi_1(G_l)=\mZ_l$. Consider the groups of characters and cocharacters of $G_l$ $$\begin{gathered}
\label{gl} {\Gamma}(G_l)=\{\chi_{G_l}\,:\,\clT(G_l)\to
\mC^*\}, \\
\label{gl1}
t(G_l)=\{{\gamma}_{G_l}\,:\,\mC^*\to\clT(G_l)\},\end{gathered}$$ $({\Gamma}^*(G_l)=t(G_l))$. They are lattices in $\gH^*$ and $\gH$ $Q\subset{\Gamma}(G_l)\subset P$, $Q^\vee\subset t(G_l)\subset P^\vee$. The lattice ${\Gamma}(G_l)$ is generated by the root lattice $Q$ and the vector $l\varpi_1$, while the lattice $t(G_l)$ is generated by the root lattice $Q$ and the vector $p\varpi_1$ $$\begin{gathered}
\label{gg}
{\Gamma}(G_l)=l\varpi_1\cup Q,\qquad t(G_l)=p\varpi_1\cup Q. \end{gathered}$$ The group ${\Gamma}(G_l)$ is the weight lattice of $G_l$ because highest weights of irreducible finite-dimensional representations of $G_l$ belong to ${\Gamma}(G_l)$.
In terms of lattices the center $\clZ(G_l)$ and $\pi_1(G_l)$ take the form $$\begin{gathered}
\clZ(G_l)\sim P^\vee/t(G_l)\sim
{\Gamma}(G_l)/Q\sim\mZ_p, \\
\pi_1(G_l)\sim
t(G_l)/Q^\vee\sim P/{\Gamma}(G_l)\sim\mZ_l.\end{gathered}$$ A subgroup $^LG_l\subset
\SLN$ is the Langlands dual to $G_l$ if $$\begin{gathered}
t(^LG_l)\sim{\Gamma}(G_l)\qquad ({\Gamma}(^LG_l)\sim t(G_l)). \end{gathered}$$ It implies that $$\begin{gathered}
\label{cdg} \clZ(^LG_l)\sim \mZ_l, \\
\pi_1(^LG_l)\sim \mZ_p. \nonumber\end{gathered}$$ Therefore the dual group is $$\begin{gathered}
\label{dugr} ^LG_l=G_p.\end{gathered}$$ In particular, $^L\SLN={{\rm PGL}(N,{\mathbb C})}$.
**Affine Weil group.** The affine Weyl group $W_a$ is a semidirect product $Q^\vee\rtimes
W$ of the Weyl group $W$ and the group $Q^\vee$. It acts on $\gH$ as $$\begin{gathered}
\label{sh} x\to
x-\frac{2({\alpha},x)}{({\alpha},{\alpha})}{\alpha}^\vee+k{\alpha}^\vee,\qquad k\in\mZ. \end{gathered}$$ Consider a semidirect product $$\begin{gathered}
\label{q25}
\bar W_a=P^\vee\rtimes W.
\end{gathered}$$ In particular, the shift operator $$\begin{gathered}
\label{shi}
x\to x+\vec m,\qquad \vec m\in P^\vee
\end{gathered}$$ is an element from $\bar W_a$. It follows from this construction that the factor group $$\begin{gathered}
\label{fwe}
\bar W_a /W_a\sim P^\vee/Q^\vee\sim\clZ(\SLN).
\end{gathered}$$ Let again $N=pl$ and define a subgroup $W_a(G_l)$ of $\bar W_a$, generated by shifts from $t(G_l)$ $$\begin{gathered}
W_a(G_l)=t(G_l)\rtimes W.
\end{gathered}$$ The factor group $W_a(G_l) /W_a$ is isomorphic to $t(G)/Q^\vee$ and in this way to $\clZ(^LG_l)$.
Acknowledgements {#acknowledgements .unnumbered}
----------------
The work was supported by grants RFBR-09-02-00393, RFBR-09-01-92437-$KE_a$ and NSh-3036.2008.2. The work of A.Z. was also supported by the “Dynasty” fund, the President fund MK-5242.2008.2, NWO-RFBR-05-01-80006 and RFBR-09-01-93106-$NCNIL_a$. The authors are grateful for hospitality to the Max Planck Institute of Mathematics, Bonn, where the most part of this work was done.
[99]{}
Levin A., Olshanetsky M., Zotov A., Hitchin systems – symplectic hecke correspondence and two-dimensional version, [*Comm. Math. Phys.*]{} [**236**]{} (2003), 93–133, [nlin.SI/0110045](http://arxiv.org/abs/nlin.SI/0110045).
Hitchin N., Stable bundles and integrable systems, [*Duke Math. J.*]{} [**54**]{} (1987), 91–114.
Levin A., Olshanetsky M., Zotov A., Painlevé VI, rigid tops and reflection equation, [*Comm. Math. Phys.*]{} [**268**]{} (2006), 67–103, [math.QA/0508058](http://arxiv.org/abs/math.QA/0508058).
Krichever I., Vector bundles and Lax equations on algebraic curves, [*Comm. Math. Phys.*]{} [**229**]{} (2002), 229–269, [hep-th/0108110](http://arxiv.org/abs/hep-th/0108110).
Baxter R., Eight-vertex model in lattice statistics and one-dimensional anisotropic Heisenberg chain. I, [*Ann. Physics*]{} [**76**]{} (1973), 48–71.
Date E., Jimbo M., Miwa T., Okado M., Fusion of the eight vertex SOS model, [*Lett. Math. Phys.*]{} [**12**]{} (1986), 209–215.
Felder G., Conformal field theory and integrable systems associated to elliptic curves, in Proceedings of the International Congress of Mathematicians, Vols. 1, 2 (Zurich, 1994), Birkhauser, Basel, 1995, 1247–1255, [hep-th/9609153](http://arxiv.org/abs/hep-th/9609153).
Belavin A., Dynamical symmetry of integrable system, [*Nuclear Phys. B*]{} [**180**]{} (1981), 189–200.
Kapustin A., Witten E., Electric-magnetic duality and the geometric Langlands program, [*Commun. Number Theory Phys.*]{} [**1**]{} (2007), 1–236, [hep-th/0604151](http://arxiv.org/abs/hep-th/0604151).
Levin A., Zotov A., Integrable systems of interacting elliptic tops, [*Teoret. Mat. Fiz.*]{} [**146**]{} (2006), 55–64 (English transl.: [*Theoret. and Math. Phys.*]{} [**146**]{} (2006), 45–52).
Flaschka H., Newell A.C., Monodromy- and spectrum-preserving deformations. I, [*Comm. Math. Phys.*]{} [**76**]{} (1980), 65–116.
Krichever I., The $\tau$-function of the universal Whitham hierarchy, matrix models and topological field theories, [*Comm. Pure Appl. Math.*]{} [**47**]{} (1994), 437–475, [hep-th/9205110](http://arxiv.org/abs/hep-th/9205110).
Levin A., Olshanetsky M., Hierarchies of isomonodromic deformations and Hitchin systems, in Moscow Seminar in Mathematical Physics, [*Amer. Math. Soc. Transl. Ser. 2*]{}, Vol. 191, Amer. Math. Soc., Providence, RI, 1999, 223–262.
Arinkin D., On $\lambda$-connections on a curve where $\lambda$ is a formal parameter, [*Math. Res. Lett.*]{} [**12**]{} (2005), 551–565.
Tong D., Quantum vortex strings: a review, [arXiv:0809.5060](http://arxiv.org/abs/0809.5060).\
Shifman M., Yung A., Supersymmetric solitons and how they help us understand non-Abelian gauge theories, [hep-th/0703267](http://arxiv.org/abs/hep-th/0703267).\
Gorsky A., Shifman M., Yung A., ${\mathcal N}=1$ supersymmetric quantum chromodynamics: how confined non-Abelian monopoles emerge from quark condensation, [*Phys. Rev. D*]{} [**75**]{} (2007), 065032, 16 pages, .
Popov A., Bounces/dyons in the plane wave matrix model and ${\rm SU}(N)$ Yang–Mills theory, [arXiv:0804.3845](http://arxiv.org/abs/0804.3845).
Weyl A., Elliptic functions according to Eisenstein and Kronecker, [*Ergebnisse der Mathematik und ihrer Grenzgebiete*]{}, Band 88, Springer-Verlag, Berlin – New York, 1976.
Zakharov V.E., Manakov S.V., Novikov S.P., Pitaevsky L.P., Theory of solitons. The method of the inverse scattering problem, Nauka, Moscow, 1980 (in Russian).
Ward R., Integrable systems and twistors, in Integrable Systems (Oxford, 1997), [*Oxf. Grad. Texts Math.*]{}, Vol. 4, Oxford Univ. Press, New York, 1999, 121–134.
Vinberg E.B., Onishchik A.L., Seminar on Lie groups and algebraic groups, Moscow, 1988 (English transl.: [*Springer Series in Soviet Mathematics*]{}, Springer-Verlag, Berlin, 1990).
Bourbaki N., Lie groups and Lie algebras, Chapters 4–6, [*Elements of Mathematics (Berlin)*]{}, Springer-Verlag, Berlin, 2002.
[^1]: We are grateful to A. Gorsky who bring our attention to this point.
[^2]: The SU(2) case and $W=\mR^3$ was analyzed in [@Wa] for different boundary conditions.
[^3]: We omit here and in what follows the ${\bar{z}}$ dependence.
[^4]: The holomorphic maps of the tori to $\mC^*$ such that $\chi(xy)=\chi(x)\chi(y)$ for $x,y\in\clT$.
| {
"pile_set_name": "ArXiv"
} |
---
bibliography:
- 'Thesis.bib'
---
People talk fundamentals and superlatives and then make some changes of detail.
Equation (1.2-9) is a second order, nonlinear, vector, differential equation which has defied solution in its present form. It is here therefore we depart from the realities of nature to make some simplifying assumptions...
What is the “Beta-induced Alfvén Eigenmode?”
Les théories ont causé plus d’expériences que les expériences n’ont causé de théories.
On doit exiger de moi que je cherche la vérité, mais non que je la trouve.
| {
"pile_set_name": "ArXiv"
} |
---
address: 'École Normale Supérieure, Paris *and* Mathematisches Institut, Georg-August Universität zu Göttingen'
author:
- Laurent Bartholdi
date: 22 December 2015
title: Growth of groups and wreath products
---
[^1]
Introduction {#introduction .unnumbered}
============
These notes are an expanded version of a mini-course given at “Le Louverain”, June 24-27, 2014. Its main objective was to gather together useful facts about wreath products, and especially their geometry, in its application to problems and questions about growth of groups. The wreath product is a fundamental construction in group theory, and I hope to help make the reader more familiar with it.
It has proven very useful, in the recent years, in better understanding asymptotics of the the word growth function on groups, namely the function assigning to $R\in\N$ the number of group elements that may be obtained by multiplying at most $R$ generators. The papers [@bartholdi-erschler:permutational; @bartholdi-erschler:givengrowth; @bartholdi-erschler:orderongroups; @bartholdi-erschler:imbeddings; @bartholdi-erschler:distortion] may be hard to read, and contain many repetitions as well as references to outer literature; so that, by providing a unified treatment of these articles, I may provide the reader with easier access to the results and methods.
I have also attempted to define all notions in their most natural generality, while restricting the statements to the most important or fundamental cases. In this manner, I would like the underlying ideas to appear more clearly, with fewer details that obscure the line of sight. I avoided as much as possible reference to literature, taking the occasion of reproving some important results along the way.
I have also allowed myself, exceptionally, to cheat. I do so only under three conditions: (1) I clearly mark where there is a cheat; (2) the complete result appears elsewhere for the curious reader; (3) the correct version would be long and uninformative.
I have attempted to make the text suitable for a short course. In doing so, I have included a few exercises, some of which are hopefully stimulating, and a section on open problems. What follows is a brief tour of the highlights of the text.
Wreath products {#wreath-products .unnumbered}
---------------
The wreath product construction, described in §\[ss:wreath\], is an essential operation, building a new group $W$ out of a group $H$ and a group $G$ acting on a set $X$. Assuming[^2] that $G$ is a group of permutations of $X$, the wreath product is the group $W=H{\operatorname{\wr\wr}}_X G$ of $H$-decorated permutations in $G$: if elements of $G$ are written in the arrow notation, with elements of $X$ lined in two identical rows above each other and an arrow from each $x\in X$ to its image, then an element of $H{\operatorname{\wr\wr}}_X G$ is an arrow diagram with an element of $H$ attached to each arrow, e.g. $$\begin{tikzpicture}
\foreach \i/\j/\pos in {1/2/0.3,2/1/0.3,3/4/0.3,4/5/0.7,5/3/0.5}
\draw[->] (1.7*\i,1) -- node[pos=\pos] {\contour{white}{$h_\i$}} (1.7*\j,0);
\end{tikzpicture}$$
One of the early uses of wreath products is as a classifier for extensions, as discovered by Kaloujnine, see Theorem \[thm:kk\]: there is a bijective correspondence between group extensions with kernel $H$ and quotient $G$ on the one hand, and appropriate subgroups of the wreath product $H{\operatorname{\wr\wr}}G$, with $G$ seen as a permutation group acting on itself by multiplication. We extend this result to permutational wreath products:
Let $G,H$ be groups, and let $G$ act on the right on a set $X$. Denote by $\pi\colon H{\operatorname{\wr\wr}}G\to G$ the natural projection. Then the map $E\mapsto E$ defines a bijection between $$\frac{\left\{\begin{array}{l@{\;}r}E:&E\le H\text{ and the $G$-sets $X$ and $H\backslash E$ are isomorphic}\\&\text{via a homomorphism }E\to G\end{array}\right\}}{\text{isomorphism $E\to E'$ of groups intertwining the actions on $H\backslash E$ and $H\backslash E'$}}$$ and $$\frac{\left\{\begin{array}{l@{\;}l}E\le H{\operatorname{\wr\wr}}G:&\pi(E)\text{ transitive on }X\text{ and }\\&\ker(\pi)\cap E\overset\cong\longrightarrow H\text{ via }f\mapsto f(x)\text{ for all }x\in X\end{array}\right\}}{\text{conjugacy of subgroups of }H{\operatorname{\wr\wr}}G}.$$
The wreath product $H{\operatorname{\wr\wr}}_X G$ is uncountable, if $H\neq1$ and $X$ is infinite. It contains some important subgroups: $H\wr_X G$, defined as those decorated permutations in which all but finitely many labels are trivial; and $H{\wr^{\text{\upshape f.v.}}}_X G$, defined as those decorated permutations in which the labels take finitely many different values. Clearly $H\wr_X
G\subseteq H{\wr^{\text{\upshape f.v.}}}_X G\subseteq H{\operatorname{\wr\wr}}_X G$, and $H\wr_X G$ is countable as soon as $H,G,X$ are countable.
Growth of groups {#growth-of-groups .unnumbered}
----------------
Let us summarise here the main notions; for more details, see §\[ss:growth\]. A choice of generating set $S$ for a group $G$ gives rise to a graph, the *Cayley graph*: its vertex set is $G$, and there is an edge from $g$ to $gs$ for each $g\in G$, $s\in S$. The path metric on this graph defines a metric $d$ on $G$ called the *word metric*. The Cayley graph is invariant under left translation, and so is the word metric.
One of the most naive invariants of this graph is its *growth*, namely the function $v_{G,S}(R)$ measuring the cardinality of a ball of radius $R$ in the Cayley graph. If the graph exhibits some kind of regularity, then it should translate into some regularity of the function $v_{G,S}$.
For example, Klarner [@klarner:growth1; @klarner:growth2] studied the growth of crystals (that expand according to a precise and simple rule) via what turns out to be the growth of an abelian group.
A convenient tool to study various forms of regularity of a function $v_{G,S}$ is the associated generating function $\Gamma_{G,S}(z)=\sum_{R\in\N}(v(R)-v(R-1))z^R$. The regularity of $v_{G,S}$ translates then into a property of $\Gamma_{G,S}$ such as being a rational, algebraic, $D$-finite, … function of $z$.
We may rewrite $\Gamma_{G,S}(z)=\sum_{g\in G}z^{d(1,g)}$; then a richer power series keeps track of more regularity of $G$: $${\widehat\Gamma}_{G,S}(z)=\sum_{g\in G}g z^{d(1,g)}.$$ This is a power series with coëfficients in the group ring $\Z G$, and again we may ask whether ${\widehat\Gamma}_{G,S}$ is rational or algebraic[^3].
If $G$ has an abelian subgroup of finite index [@liardet:phd], or if $G$ is word-hyperbolic [@grigorchuk-n:complete], then ${\widehat\Gamma}_{G,S}$ is a rational function of $z$ for all choices of $S$. We give a sufficient condition for ${\widehat\Gamma}_{G,S}$ to be algebraic:
Let $H=\langle T\rangle$ be a group such that ${\widehat\Gamma}_{H,T}$ is algebraic, and let $F$ be a free group. Consider $G=H\wr F$, generated by $S=T\cup\{\text{a basis of }F\}$. Then ${\widehat\Gamma}_{G,S}$ is algebraic.
We then turn to studying the asymptotics of the growth function $v_{G,S}$. Let us write $v\precsim w$ to mean that $v(R)\le w(CR)$ for some constant $C\in\R_+$ and all $R\ge0$, and $v\sim w$ to mean $v\precsim w\precsim v$. Then the $\sim$-equivalence class of $v_{G,S}$ is independent of the choice of $S$, so we may simply talk about $v_G$.
For “most” examples of groups, either $v_G(R)$ is bounded by a polynomial in $R$ or $v_G(R)$ is exponential in $R$. This is, in particular, the case for soluble, linear and word-hyperbolic groups. There exist, however, examples of groups for which $v_G(R)$ admits an intermediate behaviour between polynomial and exponential; they are called groups of *intermediate growth*. The question of their existence was raised by Milnor [@milnor:5603], was answered positively by Grigorchuk [@grigorchuk:growth], and has motivated much group theory in the second half of the 20th century.
Let $\eta_+\approx 2.46$ be the positive root of $T^3-T^2-2T-4$, and set $\alpha=\log2/\log\eta_+\approx0.76$. We shall show that, for *every* sufficiently regular function $f\colon\R_+\to\R_+$ with $\exp(R^\alpha)\precsim f\precsim\exp(R)$, there exists a group with growth function equivalent to $f$:
Let $f\colon\R_+\to\R_+$ be a function satisfying $$f(2R)\le f(R)^2 \le f(\eta_+R)\text{ for all $R$ large enough}.$$ Then there exists a group $G$ such that $v_G\sim f$.
Thus, groups of intermediate growth abound, and the space of asymptotic growth functions of groups is as rich as the space of functions. Furthermore, we shall show that there is essentially no restriction on the subgroup structure of groups of intermediate growth. Let us call a group $H$ *locally of subexponential growth* if every finitely generated subgroup of $H$ has growth function $\precnsim\exp(R)$. Clearly, if $H$ is a subgroup of a group of intermediate growth then it has locally subexponential growth. We show, conversely:
Let $B$ be a countable group locally of subexponential growth. Then there exists a finitely generated group of subexponential growth in which $B$ imbeds as a subgroup.
Finally, it may happen that a group $G$ has exponential growth, namely that the growth rate $\lim_{R\to\infty}v_{G,S}(R)^{1/R}$ is $>1$ for all $S$, but that the infimum of these growth rates, over all $S$, is $1$. Such a group is called of *non-uniform exponential growth*. The question of their existence was raised by Gromov [@gromov:metriques]\*[Remarque 5.12]{}. Again, soluble, linear and word-hyperbolic groups cannot have non-uniform exponential growth; but, again, it turns out that such groups abound. We shall show:
Every countable group may be imbedded in a group of non-uniform exponential growth.
Furthermore, the group $W$ in which the countable group imbeds may be required to have the following property: there is a constant $K$ such that, for all $R>0$, there exists a generating set $S$ of $W$ with $$v_{W,S}(r) \le \exp(K r^\alpha)\text{ for all }r\in[0,R].$$
(Self-)similar groups and branched groups {#self-similar-groups-and-branched-groups .unnumbered}
-----------------------------------------
All the constructions mentioned in the previous subsection take place in the universe of *(self-)similar groups*. Here is a brief description of these groups; see §\[ss:ss\] for details.
Just as a self-similar set, in geometry, is a set describable in terms of smaller copies of itself, a *self-similar group* is a group describable in terms of “smaller” copies of itself. A *self-similar structure* on a group $G$ is a homomorphism $\phi\colon G\to G\wr_X P$ for a permutation group $P$ of $X$. Thus elements of $G$ may be recursively written in terms of $G$-decorated permutations of $X$. For this description to be useful, of course, the homomorphism $\phi$ must satisfy some non-degeneracy condition (in particular be injective), and $P$ should be manageable, say finite.
The fact that the copies of $G$ in $G\wr_X P$ are “smaller” than the original is expressed as follows: there is a norm on $G$ such that, for $g\in G$ and $\phi(g)$ a permutation with labels $(g_x:x\in X)$, the elements $g_x$ are shorter than $g$, at least as soon as $g$ is long enough. For example, consider $G$ finitely generated, and denote by $\|\cdot\|$ the word norm on $G$. One requires $\|g_x\|<\|g\|$ for all $x\in X$ and all $\|g\|\gg1$; this is equivalent to the existence of $\lambda\in(0,1)$ and $K\ge0$ such that $\|g_x\|\le\lambda\|g\|+K$ for all $x\in X,g\in G$.
Furthermore, in cases that interest us, the map $\phi$ is almost an isomorphism, in that its image $\phi(G)$ has finite index in $G\wr_X
P$. Thus $\phi$ may be thought of as a *virtual isomorphism* between $G$ and $G^X$, namely an isomorphism between finite-index subgroups. When one endows $G^X$ with the $\ell^\infty$ metric $\|(g_x)\|=\max_{x\in X}\|g_x\|$, the condition above requires that this virtual isomorphism be a contraction. On the other hand, endowing $G^X$ with the $\ell^1$ metric $\|(g_x)\|=\sum_{x\in X}\|g_x\|$, the optimal Lipschitz constant of the virtual isomorphism plays a fundamental rôle in estimating the growth of $G$.
*Similar* groups are a natural generalization: one is given a set $\Omega$ and a self-map $\sigma\colon\Omega\righttoleftarrow$; for each $\omega\in\Omega$, a group $G_\omega$ and a permutation group $P_\omega$ of a set $X_\omega$; and homomorphisms $\phi_\omega\colon G_\omega\to
G_{\sigma\omega}\wr_{X_\omega}P_\omega$. Taking for $\Omega$ a singleton recovers the notion of self-similar group. Taking $\Omega=\N$ and $\sigma(n)=n+1$ defines in full generality a similar group $G_0$; but it is often more convenient to consider a larger family of groups in which $(G_n)_{n\in\N}$ imbeds. In particular, one obtains a *topological space* of groups, in such a manner that close groups have close properties (for example, their Cayley graphs coincide on a large ball).
Acknowledgments {#acknowledgments .unnumbered}
---------------
I am very grateful to Yago Antolin, Laura Ciobanu and Alexey Talambutsa for having organised the workshop in Le Louverain where I presented a preliminary version of this text, and to the participants of the workshop for their perspicacious questions.
A large part of the material is taken from articles written in collaboration with Anna Erschler, and I am greatly indebted to her for generously embarking me on her projects. It owes much to her energy and enthusiasm that we were able to finish our joint articles.
Hao Chen, Yves de Cornulier and Pierre de la Harpe helped improve these notes by pointing out a number of mistakes and inconsistencies.
Open problems {#open-problems .unnumbered}
-------------
This text presents a snapshot of what is known on growth of groups in 2014; there remain a large number of open problems. Here are some promising directions for further research.
1. Which groups $G$ are such that, for all generating sets $S$, the complete growth series ${\widehat\Gamma}_{G,S}$ is a rational function of $z$?
This is known to hold for virtually abelian groups, and for word-hyperbolic groups. Conjecturally, this holds for no other group.
The related question of which groups have a rational (classical) growth function is probably more complicated, see §\[ss:formal\].
2. Is the analytic continuation $1/{\widehat\Gamma}_{G,S}(1)$ related to the complete Euler characteristic of $G$, just as $1/\Gamma_{G,S}(1)$ is (under some additional conditions) the Euler characteristic of $G$? See [@stallings:centerless]\*[§1.8]{} for complete Euler characteristic.
3. Does there exist a group $G$ with two generating sets $S_1,S_2$ such that ${\widehat\Gamma}_{G,S_1}$ is rational but ${\widehat\Gamma}_{G,S_2}$ is transcendental?
Such an example could be $G=F_2\times F_2$. Set $S=\{x,y\}^{\pm1}$ a free generating set of $F_2$, and $S_1=S\times\{1\}\sqcup\{1\}\times S$ and $S_2=S_1\sqcup\{(s,s):s\in S\}$.
The same properties probably hold for the usual generating series $\Gamma_{G,S_1}$ and $\Gamma_{G,S_2}$. The radius of convergence of $\Gamma_{G,S_1}$ is $1/3$, but that of $\Gamma_{G,S_2}$ is unknown.
This problem is strongly related to the “Matching subsequence problem”, which asks for the longest length of a common subsequence among two independently and uniformly chosen words of length $n$ over a $k$-letter alphabet; see [@chvatal-sankoff:commonsubsequence]. It is easy to see that, for two uniformly random reduced words of length $n$ in $F_2$, the longest common subword has length $\approx\gamma n$ for some constant $\gamma$, as $n\to\infty$. Thus a pair $(g,h)\in G$ with $\|g\|=\|h\|=n$ has length $2n$ with respect to $S_1$, but approximately $(2-\alpha)n$ with respect to $S_2$. We might call $\gamma$ the *Chvátal-Sankoff constant* of $F_2$.
4. Do there exist infinite simple groups of subexponential growth?
There is no reason for such groups *not* to exist; but the construction methods described in this text yield groups acting on rooted trees, which therefore are as far as possible from being simple.
There is also no reason for finitely presented groups of subexponential growth *not* to exist; again, the obstacle is probably more our mathematical limitations than fundamental mathematical reasons.
The following question, by de la Harpe [@harpe:uniform], is still open at the time of writing: “Do there exist groups with Kazhdan’s property (T) and non-uniform exponential growth?”
Similarly, it is not known whether there exist simple finitely generated groups of non-uniform exponential growth, and whether there exist finitely presented groups of non-uniform exponential growth.
5. Do there exist groups whose growth function lies strictly between polynomials and $\exp(R^{1/2})$?
See the discussion in §\[ss:asymptotic\]. There exists a superpolynomial function $f(R)\succsim R^{(\log R)^{1/100}}$ such that no group has growth strictly between polynomials and $f(R)$. There exists no residually nilpotent group whose growth is strictly between polynomials and $\exp(R^{1/2})$, see Theorem \[thm:growth:hp\].
6. What is the asymptotic growth of the first Grigorchuk group? What is its exact growth, for the generating set $\{a,b,c,d\}$? Does the growth series of the Grigorchuk group exhibit some kind of regularity?
Some experiments indicate that this must be the case. For example, consider the quotient $G_n$ of the first Grigorchuk group that acts on $\{\9,\8\}^n$. It is a finite group of cardinality $2^{5\cdot2^{n-3}+2}$. For $n\le 7$, the diameter $D_n$ of its Cayley graph (for the natural generating set $\{a,b,c,d\}$) is the sequence $1,4,8,24,56,136,344$ and satisfies the recurrence $D_n=D_{n-1}+2D_{n-2}+4D_{n-3}$. If this pattern went on, the growth of the first Grigorchuk group would be asymptotically $\exp(R^{\log2/\log\eta_+})\approx\exp(R^{0.76})$.
If a group has subexponential growth, then its growth series is either rational or transcendental, and if the group has intermediate growth, then the growth series must be transcendental; see \[thm:polya-carlson\] in §\[ss:formal\]. Thus Grigorchuk group’s growth series is transcendental. Does the series satisfy a functional equation? That would make it akin to the classical partition function $\sum_{n\ge0}p(n)z^n=\prod_{n\ge1}(1-z^n)^{-1}$, which (up to scaling and multiplying by $z^{1/24}$) is a *modular function*[^4]. Ghys asked me once: “Is the growth series $\Gamma(z)$ of Grigorchuk’s group modular?”
7. For every $k\in\N$, the *space of marked $k$-generated groups* $\mathscr S_k$ may be defined as the space of normal subgroups of the free group $F_k$, by identifying $G=\langle
s_1,\dots,s_k\rangle$ with the kernel of the natural map $F_k\to G$ sending generator to generator. It is a compact space. What properties does the set $\mathscr I$ of groups of intermediate growth, and the set $\mathscr N$ of groups of non-uniform exponential growth, enjoy in this space? For example,
“Is there an uncountable open subset of $\mathscr S_k$ in which $\mathscr N$, or $\mathscr I$, is dense? Is $\mathscr N$ dense in the complement of groups of polynomial growth?”
Recall that similar groups are families of groups $(G_\omega)_{\omega\in\Omega}$ indexed by a space $\Omega$. If all $G_\omega$ are $k$-generated, we obtain a map $\Omega\to\mathscr S_k$, which under favourable circumstances is continuous. This has been exploited e.g. in [@nekrashevych:cantor] to produce groups of non-uniform exponential growth.
It had actually been doubted, before Grigorchuk’s discovery [@grigorchuk:growth], whether there exist groups of intermediate growth. This text tries to convince the reader that they are abundant. Giving a precise meaning to the above question would quantify, in some manner, the extent to which they are abundant.
Notational conventions {#notational-conventions .unnumbered}
----------------------
I try to adhere to standard group-theoretical notation. In particular, the right action of a group element $g$ on a point $x$ is written $xg$, and a left action would be written ${}^g x$. The stabilizer of $x$ is written $G_x$. The conjugation action of a group on itself is written $g^h=h^{-1}g h$, and the commutator of two elements is $[g,h]=g^{-1}h^{-1}g h=g^{-1}g^h=h^{-g}h$.
I also introduce a minimal amount of new, “fancy” notation to represent elements of wreath products or of self-similar groups, and hope that it helps in achieving clarity and conciseness.
Wreath products {#ss:wreath}
===============
We start by the basic construction. Let $H$ be a group, and let $G$ be a group acting on the right on a set $X$. We construct two groups $$\begin{aligned}
H\wr_X G &= \big({\prod_X}'H\big)\rtimes G\qquad\text{the \emph{restricted} wreath product},\\
H{\operatorname{\wr\wr}}_X G &= \big(\prod_X H\big)\rtimes G\qquad\text{the \emph{unrestricted} wreath product}.\end{aligned}$$
Here the unrestricted product $\prod_X H$ may be viewed as the group of functions $X\to H$, with pointwise composition and with left $G$-action given by pre-composition[^5]: ${}^g\!f$ is the function given by $({}^g\!f)(x)=f(x g)$. The *restricted* product $\prod_X' H$ is then identified with finitely supported functions $X\to H$. In both cases, this product is a subgroup of the wreath product, and is called its *base group*.
In the particular case of $X=G$ with natural right action by multiplication, one calls $H{\operatorname{\wr\wr}}_G G$ the *regular unrestricted wreath product*, and writes it simply $H{\operatorname{\wr\wr}}G$; and similarly for the *regular restricted wreath product* $H\wr_G G=H\wr G$.
Assume that the action of $G$ on $X$ is faithful; so that elements of $G$ may be identified with permutations of $X$. The best way to describe elements of $H{\operatorname{\wr\wr}}_X G$ or its subgroup $H\wr_X G$ is by *decorated permutations*: one writes a permutation of $X$, decorated by elements of $H$, such as
$$\label{eq:permutation}
\begin{tikzpicture}[baseline=5mm]
\foreach \i/\j/\pos in {1/2/0.3,2/1/0.3,3/4/0.3,4/5/0.7,5/3/0.5}
\draw[->] (1.7*\i,1) -- node[pos=\pos] {\contour{white}{$h_\i$}} (1.7*\j,0);
\end{tikzpicture}$$
Permutations are multiplied as usual: by stacking them, and pulling the arrows tight. Likewise, decorated permutations are multiplied by stacking them and multiplying the labels along the composed arrows. We do not write the labels when they are the identity. Here is a graphical computation of a product: $$\begin{tikzpicture}[baseline=0mm]
\foreach \i/\j/\pos in {1/2/0.3,2/1/0.3,3/4/0.3,4/5/0.7,5/3/0.5}
\draw[->] (1.2*\i,0.8) -- node[pos=\pos] {\contour{white}{$h_\i$}} (1.2*\j,0.06);
\foreach \i/\j/\pos in {1/2/0.3,2/3/0.7,3/1/0.5,4/5/0.3,5/4/0.3}
\draw[->] (1.2*\i,-0.06) -- node[pos=\pos] {\contour{white}{$k_\i$}} (1.2*\j,-0.8);
\end{tikzpicture}\quad=\quad
\begin{tikzpicture}[baseline=7mm]
\foreach \i/\j/\k/\pos in {1/2/3/0.2,2/1/2/0.3,3/4/5/0.2,4/5/4/0.7,5/3/1/0.5}
\draw[->] (1.4*\i,1.4) -- node[pos=\pos] {\contour{white}{$h_\i k_\j$}} (1.4*\k,0);
\end{tikzpicture}.$$
When writing formulæ, we must sometimes depart from the graphical notation, in which permutations are written top-to-bottom or left-to-right and thanks to their arrows there is no ambiguity in knowing in which order to compose the labels. We invariably let permutations act on the right on sets, and thus ‘$\sigma\tau$’ means ‘first $\sigma$, then $\tau$’.
In the wreath product $W=\{\pm1\}\wr{\operatorname{Sym}}(2)$, consider the element $$h=\begin{tikzpicture}
\draw[->] (0,1) -- (2,0);
\draw[->] (2,1) -- node [pos=0.3] {\contour{white}{$-1$}} (0,0);
\end{tikzpicture}$$ Show by concatenating the diagram with itself that $h$ has order exactly $4$. The group $W$ has order $8$; which of the order-$8$ groups is it?
Let us consider a wreath product $W=H{\operatorname{\wr\wr}}_X G$. In writing elements $w,w'\in W$ algebraically, we may express them in the form $w=f g$ with $f\colon X\to H$ and $g\in G$, or in the form $w'=g f$. In both cases, the element $g$ and the function $f$ are unique, by definition of the semidirect product. The compositions $(f g)(f'g')$ and $(g f)(g'f')$ are, in all cases, computed using the relation $$g\cdot f = {}^g\!f\cdot g,$$ namely $$\label{eq:wreathmult}
(f g)(f'g')=(f\cdot{{}^g\!f'})(g g')\text{ and }(g f)(g' f')=(g
g')({}^{(g')^{-1}}\!f \cdot f').$$
Let $R$ be a ring, viewed as a group under addition, and let $R G$ denote the group ring of $G$, on which $G$ acts by right multiplication. Show that $R\wr G$ is isomorphic to $R G\rtimes
G$. More generally, let $X$ be a $G$-set; then $RX$ is a $G$-module. Show that $R\wr_X G$ and $RX\rtimes G$ are isomorphic.
Actions
-------
Assume now moreover that $H$ acts from the right on a set $Y$. Then there are two natural sets on which $W=H{\operatorname{\wr\wr}}_X G$ acts:
- There is an action on $Y\times X$, given by $(y,x)\cdot
f g=(y f(x),x g)$ for $(y,x)\in Y\times X$; it is called the *imprimitive* action;
- There is an action on $Y^X$, the set of functions $X\to Y$, given by $(\phi\cdot f g)(x g) = \phi(x)f(x)$ for $\phi\colon X\to Y$; it is called the *primitive* action.
There are natural bijections between the sets $(Z\times Y)\times X=Z\times (Y\times X)$ and $(Z^Y)^X=Z^{Y\times X}$. Assume now that a group $G$ acts on $X$, a group $H$ acts on $Y$ and a group $I$ acts on $Z$. Show that the bijections above give isomorphisms between the groups $(I\wr_Y H)\wr_X G$ and $(I\wr_{Y\times X}(H\wr_X G))$ as permutation groups, both of $Z\times Y\times X$ and of $(Z^Y)^X$.
History
-------
Leo Kaloujnine understood the importance of wreath products in the early 1940’s. It is said that he worked, during the second World War, in a uniform factory and observed a rivet machine; it was made of a rotating ring containing many rotating disks in it. Identifying the movement of the ring with an action of $G$ and each subdisk with an action of $H$, one sees that motions of the machine are described by wreath product elements. In his dissertation (under Élie Cartan, [@kaloujnine:struct]), he studied the Sylow subgroups of symmetric groups, and showed that they were iterated wreath products.
However, this description of maximal $p$-subgroups of symmetric groups already appears in the classical 1870 treatise by Camille Jordan [@jordan:subs]\*[II.I.41]{}, who implicitly defined wreath products there. This is all the more remarkable since Sylow’s theorems were only published two years later [@sylow:subs]!
Kaloujnine returned to Soviet Union after the war, and contributed greatly to the development of mathematics in Ukraine, founding in 1959 the department of algebra and mathematical logic. He is remembered for the following important result classifying group extensions; namely, that the wreath product is a universal object containing all extensions:
\[thm:kk\] Let $G,H$ be groups. Denote by $\pi\colon H{\operatorname{\wr\wr}}G\to G$ the natural projection. Then the map $E\mapsto E$ defines a bijection between $$\frac{\big\{E : 1\to H\to E\to G\to 1\big\}}{\text{isomorphism of extensions}}$$ and $$\frac{\big\{E\le H{\operatorname{\wr\wr}}G:\pi(E)=G\text{ and }\ker(\pi)\cap E\overset\cong\longrightarrow H\text{ via }c\in H^G\mapsto c(1)\big\}}{\text{conjugacy of subgroups of }H{\operatorname{\wr\wr}}G}.$$
Prove Theorem \[thm:kk\].
Hint: given an extension $1\to H\to E\overset\tau\to G\to 1$, choose a set-theoretic section[^6] $g\mapsto\widetilde g$ of $\tau$, and define an imbedding $E\to H{\operatorname{\wr\wr}}G$ by $$e\mapsto\left(g\mapsto \widetilde g e(\widetilde{g\tau(e)})^{-1}\right)\tau(e).$$
Theorem \[thm:kk\] may be interpreted more abstractly as saying that, if $1\to H\to E\to G\to 1$ is an exact sequence and $f_0\colon
H\to K$ is a group homomorphism, then $f_0$ extends naturally to a homomorphism $f\colon E\to K{\operatorname{\wr\wr}}G$. The case $f=\mathsf{id}$ and $H=K$ is exactly the statement of the theorem, and the generalised version is proven in exactly the same manner. The advantage of this formulation is that one need not require that $H$ be normal in $E$; and the more general statement is
\[thm:kkwreath\] Let $H\le E$ be groups, and let $f_0\colon H\to K$ be a homomorphism. Then $f_0$ extends naturally to a homomorphism[^7] $f\colon E\to K{\operatorname{\wr\wr}}_{H\backslash E}E/\operatorname{core}(H)$.
More precisely, let $G,H$ be groups, and let $G$ act on the right on a set $X$. Denote by $\pi\colon H{\operatorname{\wr\wr}}G\to G$ the natural projection. Then the map $E\mapsto E$ defines a bijection between $$\frac{\left\{\begin{array}{ll}E:&H\le E\text{ and the $G$-sets $X$ and $H\backslash E$ are isomorphic}\\&\text{via a homomorphism }E\to G\end{array}\right\}}{\text{isomorphism $E\to E'$ of groups intertwining the actions on $H\backslash E$ and $H\backslash E'$}}$$ and $$\frac{\left\{\begin{array}{ll}E\le H{\operatorname{\wr\wr}}G:&\pi(E)\text{ transitive on }X\text{ and }\\&\ker(\pi)\cap E\overset\cong\longrightarrow H\text{ via }c\in H^X\mapsto c(x)\text{ for all }x\in X\end{array}\right\}}{\text{conjugacy of subgroups of }H{\operatorname{\wr\wr}}G}.$$
Note that we could have written $E$ instead of $E/\operatorname{core}(H)$ everywhere above. We included it to obtain a smaller group acting on the set $X$, thus reinforcing the parallel with Theorem \[thm:kk\], and expressing $E$ as a subgroup of a presumably easier-to-construct group. In particular, if $[E:H]=d<\infty$, then $E/core(H)$ is a group of cardinality between $d$ and $d!$, since it acts faithfully and transitively on the set $H\backslash E$ of cardinality $d$.
We follow the sketched proof of Theorem \[thm:kk\]; rather than a section $G\to E$, we choose a right transversal $T$ of $H$ in $E$; namely, a subset $T\subset E$ such that every $e\in E$ may uniquely be written in the form $ht$ with $h\in H,t\in T$.
For $e\in E$, let $f(e)$ be the following $H$-decorated permutation of $H\backslash E$: the permutation is given by the natural right-multiplication action of $E$ on $H\backslash E$, and on every edge, say from $H t$ to $H u$ if $H t\cdot e=H u$, the label is $t e u^{-1}$.
In formulas, this may be written as follows, although all verifications are easier in the “decorated permutations” form. Denote by $\pi$ the natural map $E\to E/\operatorname{core}(H)$. Set then $f(e)=c\pi(e)$ with $c\colon H\backslash E\to H$ given, for $t\in T$, by $c(Ht)=t e u^{-1}$ for the unique $u\in T$ such that $t e u^{-1}\in H$.
Composing the map $c$ by the homomorphism $f_0\colon H\to K$ gives the first statement.
Wreath products have occupied a prominent place in the theory of permutation groups. For illustration, they appear as fundamental constructions in the O’Nan-Scott theorem classifying maximal subgroups of the symmetric group. We skip subcases *(iii.3)…(iii.viii)* which are too technical:
Let $G$ be a maximal subgroup of ${\operatorname{Sym}}(X)$ for a finite set $X$. Then either
1. the action of $G$ is not transitive[^8]; then $X=Y\sqcup Z$ and $G={\operatorname{Sym}}(Y)\times{\operatorname{Sym}}(Z)$; or
2. the action of $G$ is transitive, but not primitive[^9]; then $X=Y\times Z$ and $G={\operatorname{Sym}}(Y)\wr{\operatorname{Sym}}(Z)$ in its imprimitive action; or
3. the action of $G$ is primitive; and then either
1. $G$ is an affine group over a finite field; or
2. $X=Y^Z$ and $G={\operatorname{Sym}}(Y)\wr{\operatorname{Sym}}(Z)$ in its primitive action; or
3. … (3.viii) …
Generators for wreath products {#ss:genwreath}
------------------------------
We now consider restricted wreath products $W=H\wr_X G$ in more detail. We introduce the following notation: if $X$ be a set and $H$ be a group, then for all $x\in X$ and $h\in H$, $$\text{we define $h@x\colon X\to H$ by }\begin{cases}(h@x)(x) = h,\\ (h@x)(y) = 1\text{ for all }y\neq x.\end{cases}.$$
[r]{}[4cm]{} {width="4cm"}
One has the important formula relating conjugation in $W$ with translation: $$\displaywidth=\parshapelength\numexpr\prevgraf+2\relax
(h@x)^g = h@xg$$ (check it by drawing the decorated permutations!)
Let $S$ be a generating set for $G$, and let $T$ be a generating set for $H$. Let $Y\subseteq X$ be a choice of one representative from each $G$-orbit on $X$. Then $W$ is generated by $$\displaywidth=\parshapelength\numexpr\prevgraf+2\relax
\{t@y:t\in T,y\in Y\}\sqcup S.$$
A good, but also slightly misleading example, is the “lamplighter group”: this is the group $W=\{\pm1\}\wr\Z$. It is best understood by its primitive action (on the space of functions $\{\pm1\}^\Z$): imagine an infinite street with at each integer position a lamppost. The lamp there can be “on” or “off”. Imagine also that the street is viewed from the perspective of the “lamplighter”, namely the person in charge of turning various lamps on and off. The generator $s$ of $\Z$ means “move up the street”, or equivalently “shift all lamps down” relatively to the lamplighter; and the generator $t@1$ means “change the state of the lamp currently in front of the lamplighter”.
This is what is expressed by the cartoon above: relativistically speaking, it makes no difference to think that a train moves on its tracks, or that the tracks move under the train. The lamplighter is on the train, and the lamp configurations are on the track.
The misleading aspect of this example is that one must remember that, when the lamplighter moves in one direction, the lamps actually move in the *opposite* direction. In a regular wreath product, this makes no difference; but in the general case it does.
\[def:schreier\] Let $G=\langle S\rangle$ be a finitely generated group acting on the right on a set $X$. The associated *Schreier graph* has vertex set $X$, and for each $x\in X,s\in S$ an edge from $x$ to $x s$.
In the case $X=G$ of a group $G$ acting on itself by (right or left) multiplication, one obtains the (right or left) Cayley graph of $G$. The group $G$ acts then by graph isometries on its Cayley graph by (left or right) multiplication.
Consider $G={\operatorname{Sym}}(4)$ generated by $\{(1,2),(2,3),(3,4)\}$. Draw the Schreier graphs of (in order of difficulty)
1. the action of $G$ on $\{1,2,3,4\}$;
2. the action on the collection of $2$-element subsets of $\{1,2,3,4\}$;
3. the action of $G$ on itself by conjugation.
It would be a great mistake to think that a path in the permutational wreath product $W=H\wr_X G$ is a path in the Schreier graph of $G$ with decorations in $H$ along the path. We shall see where exactly the decorations appear, but their positions are rather at the *inverses* of path points. To complicate matters, in case $G$ is abelian the map $x\mapsto x^{-1}$ is an automorphism of $G$, whence the above “mistaken” description luckily gives the correct answer.
Let us fix generating sets $G=\langle S\rangle$ and $H=\langle T\rangle$, and assume for simplicity that $G$ acts transitively on $X$. We also choose a base point $x\in X$. Then $W$ is naturally generated by $T@x\sqcup S$.
We consider first the description of elements of $W$ as $f g$ with $f\colon X\to H$ and $g\in G$. This is the “lamplighter moving” version, because the “lamp vector” $f$ is not shifted, but the position at which it is changed varies. By , the right action on $w=f g$ of
— a generator $s\in S$ gives $w s=f(gs)$;
— a generator $t@x\in T@x$ gives $w(t@x)=(f\cdot t@xg^{-1})g$.
Thus if $w=(t_0@x)s_1(t_1@x)\cdots s_\ell(t_\ell@x)=f g$, then the support of $f$ is included in $\{x,x s_1^{-1},\dots,x (s_1\dots
s_\ell)^{-1}\}$.
We may also describe elements of $W$ as $g f$; this is the “earth moving” version, because the “lamp vector” $f$ is shifted, and the lamplighter always changes the lamp at a fixed position. By , the right action on $w=g f$ of
— a generator $s\in S$ gives $w s=(gs)\,{}^{s^{-1}}\!f$;
— a generator $t@x\in T@x$ gives $w(t@x)=g(f\cdot t@x)$.
Thus if $w=(t_0@x)s_1(t_1@x)\cdots s_\ell(t_\ell@x)=g f$, then the support of $f$ is included in $\{x s_1\dots s_\ell,x s_2\dots
s_\ell,\dots,x s_\ell,x\}$.
The semidirect product description of $W=H\wr_X G$ gives a presentation of $W$ by generators and relations; see [@cornulier:fpwreath]. Generating as above $W$ by $T@Y\sqcup
S$, we get $$\begin{aligned}
W=\big\langle T@Y\sqcup S\,\big|\,&\text{relations of $G$, relations of $H$, and }\\
&\forall t,t'\in T,\,\forall y\neq y'\in Y:[t@y,t'@y'],\text{ and }\\
&\forall t\in T,\,\forall y\in Y,\,\forall g\in G_y:[t@y,g],\text{ and }\\
&\forall t,t'\in T,\,\forall y\in Y,\,\forall g\in (G_y\backslash G)\setminus\{G_y\}:[t@y,(t'@y)^g]\big\rangle.\end{aligned}$$ The exact criterion, assuming $X\neq\emptyset$ and $H\neq1$, is:
\[thm:cornulier\] The wreath product $W$ is finitely presented if and only if both $G,H$ are finitely presented, $G$ acts on $X$ with finitely generated stabilizers, and $G$ acts diagonally on $X\times X$ with finitely many orbits.
Diestel-Leader graphs {#diestel-leader-graphs .unnumbered}
---------------------
We present a particularly intuitive description of the Cayley graph of “lamplighter groups” $W=F\wr\Z$, for a finite group $F$ of cardinality $q$, see [@woess:dl].
Let $\mathscr T$ denote the $(q+1)$-regular tree, choose a basepoint $o$, and a geodesic ray $\omega\colon\N\to\mathscr T$ starting at $o$. Imagine $\mathscr T$ as “hanging from $\omega$”: orient the edges on $\omega$ from $\omega(n)$ to $\omega(n+1)$, and orient every other edge from its furthest point to $\omega$ to its closest. Define then $h\colon\mathscr T\to\Z$ as follows: for each $x\in\mathscr T$, there is a unique path in $\mathscr T$ from $o$ to $x$, and set $h(x)=($number of edges oriented forward$)-($number of edges oriented backward$)$ on this path[^10].
Consider now the following graph $\mathcal B(q,q)$. Its vertex set is $\{(x,y)\in\mathscr T\times \mathscr T:h(x)+h(y)=0\}$. There is an edge from $(x,y)$ to $(x',y')$ precisely if $\{x,x'\}$ and $\{y,y'\}$ are connected in $\mathscr T$. Here is a portion of $\mathcal B(2,2)$, with one tree pointing up and one tree pointing down:
(x000) at (0,-3,0); (x001) at (0,-2,0); (x010) at (0,-1.4,0); (x011) at (0,-0.4,0); (x100) at (0,0.4,0); (x101) at (0,1.4,0); (x110) at (0,2,0); (x111) at (0,3,0); (0x00) at (1,-2,1); (0x01) at (1,-0.7,1); (0x10) at (1,0.7,1); (0x11) at (1,2,1); (1x00) at (-1,-2,1); (1x01) at (-1,-0.7,1); (1x10) at (-1,0.7,1); (1x11) at (-1,2,1); (00x0) at (2,-1,2); (00x1) at (2,1,2); (01x0) at (0.7,-1,2); (01x1) at (-0.7,1,2); (10x0) at (-0.7,-1,2); (10x1) at (0.7,1,2); (11x0) at (-2,-1,2); (11x1) at (-2,1,2); (000x) at (3,0,3); (001x) at (-0.4,0,3); (010x) at (1.4,0,3); (011x) at (-2,0,3); (100x) at (2,0,3); (101x) at (-1.4,0,3); (110x) at (0.4,0,3); (111x) at (-3,0,3); iin [0,1]{}in [0,1]{}in [0,1]{}łin [0,1]{} [ (xi) – (łxi); (ix) – (łix); (ix) – (łix); (xi) – (0xi) – (00xi) – (000x); (x111) – (x11) – (x1) – (ix); ]{}; (00x0) circle \[radius=2pt\]; (11x1) circle \[radius=2pt\];
Consider a wreath product $G=F\wr\Z$ with a finite group $F$ of cardinality $q$. Denote by $s$ a generator of $\Z$, and consider for $G$ the generating set $S=(F@1)s\sqcup s^{-1}(F@1)$. Then the Cayley graph of $(G,S)$ is $\mathcal B(q,q)$.
Only in this proof, for a subset $A$ of the integers, let us denote by $F^{A}$ the set of finitely-supported functions $A\to F$. The vertex set of $\mathscr T$ may be identified with $\bigsqcup_{n\in\Z}F^{(-\infty,n]}\times\{n\}$, in such a manner that there is an edge from $(\sigma,n)$ to $(\sigma|_{(-\infty,n-1)},n-1)$ for all $n\in\Z$ and all $\sigma\in F^{(\infty,n]}$. Vertices on the ray $\omega$ are those of the form $(1,n)$ with $n\le0$.
Therefore, the vertex set of $\mathcal B(q,q)$ may be identified with $\bigsqcup_{n\in\Z}F^{(-\infty,n]}\times
F^{(-\infty,-n]}\times\{n\}$. This is easily put in correspondence with $F^\Z\times\Z$ via the map $(\sigma,\sigma',n)\mapsto(\tau,n)$ with $\tau\colon\Z\to F$ the finitely supported function given by $\tau(k)=\sigma(k)$ if $k\le n$ and $\tau(k)=\sigma'(1-k)$ if $k>n$.
Thus we put the vertex set of $\mathcal B(q,q)$ is bijection with $G$. It is now routine to check that the generators in $S$ produce the edges of $\mathcal B(q,q)$.
This description of lamplighter groups makes some of their geometric features quite transparent. For example, let us consider a finitely generated group $G=\langle S\rangle$, and its Cayley graph. A vertex $v\in G$ is called a *dead end* if all neighbours of $v$ are at least as close to $1$ as $v$. More generally, let us say $v$ is *on a $k$-hill* if all paths from $v$ to an element of norm $\|v\|+1$ has to go through a vertex of norm $\|v\|-k$. In other words, from the top of the hill the only way of going to infinity is to first go down at least $k$ steps.
Consider the two dots on the picture above. Say one of them is the origin $1$. Then the other one is on a $1$-hill. More generally, any element in the lamplighter group that is reached from the origin by going down $k$ steps, up $2k$ steps and down again $k$ steps along a reduced path reaches the top of a $k$-hill.
This is an fact a familiar phenomenon: consider a long street and two remote addresses we want to visit on that street, in whichever order, from our starting point on the street. The shortest way of doing this is to first go to the closest, and then the other one. The *worst* possible place to start is at equal distance from both addresses.
Growth of groups {#ss:growth}
================
This section is not a treatise on growth of groups; for that, see rather [@mann:howgroupsgrow]. We do recall some elementary, basic notions, and provide some motivation for the material to appear later. Let $G$ be a group generated by a finite symmetric set ($S=S^{-1}$). One defines the *word norm* on $G$ by $$\|g\|=\min\{n:g=s_1\cdots s_n,\,s_i\in S\},$$ and a distance[^11] on $G$ that is invariant under left translation[^12]: $$d(g,h)=\|g^{-1}h\|.$$
Thus $G$ is viewed as a normed space, and as a metric space on which $G$ acts by isometries via left translation.
Formal growth {#ss:formal}
-------------
One may be interested in regularity properties of the metric space $G$; these are best studied via the *growth series*, the formal power series $$\Gamma_G(z)=\sum_{g\in G}z^{\|g\|}\in\Z[[z]].$$
The natural questions that arise are: what is the domain of convergence of $\Gamma_G(z)$? What can be said of analytic continuations of $\Gamma_G(z)$? What are its singularities? Is the function $\Gamma_G(z)$ rational (i.e. in $\Q(z)$)? or at least algebraic (i.e. there exists a two-variable polynomial $F(y,z)\in\Z[y,z]$ with $F(\Gamma_G(z),z)\equiv0$)?
The consideration of the power series $\Gamma_G(z)$, and of the above questions, is justified by the answers that have been given:
- $\Gamma_G(z)$ converges in a disk of radius at least $1/\#S$. If $S$ is symmetric, then the convergence radius is it fact at least $1/(\#S-1)$, with equality if and only $G$ is a free product of $\Z$’s and $C_2$’s with its natural generating set, as in Definition \[defn:freelike\] below.
- If groups $G,H$ are respectively generated by $S,T$, then the direct product $G\times H$ is naturally generated by $S\sqcup
T$. One then has $$\Gamma_{G\times H}(z)=\Gamma_G(z)\Gamma_H(z),$$ see Proposition \[prop:CG:x\].
- If groups $G,H$ are respectively generated by $S,T$, then the free product $G*H$ is naturally generated by $S\sqcup T$. One then has $$\frac1{\Gamma_{G*H}(z)}=\frac1{\Gamma_G(z)}+\frac1{\Gamma_H(z)}-1,$$ see Proposition \[prop:CG:\*\].
- If the group $G$ is virtually abelian [@benson:zn], or word-hyperbolic [@gromov:hyperbolic], or the discrete Heisenberg group [@duchin-shapiro:heisenberg] $H_3=\big(\begin{smallmatrix}1&\Z&\Z\\&1&\Z\\&&1\end{smallmatrix}\big)$, then $\Gamma_G$ is a rational function of $z$ for all choices of the finite generating set $S$.
- Wreath products give some examples of power series $\Gamma_G$ that are algebraic functions, as we shall see in Corollary \[cor:parry\] below.
- If $G$ is a $2$-step nilpotent group with cyclic derived subgroup, then there exist generating sets for $G$ such that $\Gamma_G$ is a rational function of $z$. However, if $G$ is the $5$-dimensional Heisenberg group $H_5=\bigg(\begin{smallmatrix}1&\Z&\Z&\Z&\Z\\&1&&&\Z\\&&1&&\Z\\&&&1&\Z\\&&&&1\end{smallmatrix}\bigg)$, then there exist generating sets for which $\Gamma_G(z)$ is transcendental; see [@stoll:heisenberg].
In respect to this last point, note that, for $G$ nilpotent, the growth of $G$ is polynomial so $\Gamma_G(z)$ converges in the unit disk. It is either rational or transcendental, by the Fatou theorem [@fatou:seriesentieres]. More is known:
\[thm:polya-carlson\] Let $A(z)=\sum_{n\ge0} a_n z^n$ be a power series with integer coëfficients. If $A$ is not rational, then $A$ does not extend analytically beyond the unit circle.
Most importantly, the growth series is a convenient object that encodes information on $G$. A quite satisfactory theory of “Euler characteristic” has been developed for groups, see [@chiswell:euler]. Here is a special case: if $G$ is the fundamental group of a cellular complex $\mathscr X$ with contractible universal cover, one declares $\chi(G)$ to be $\chi(\mathscr X)$. More generally, if $G$ has a finite-index subgroup $H$ which is the fundamental group of a space $\mathscr Y$, one sets $\chi(G)=\chi(\mathscr Y)/[G:H]$; this makes sense because if $G$ is the fundamental group of $\mathscr X$ then $H$ is the fundamental group of a $[G:H]$-sheeted covering of $\mathscr X$, whose Euler characteristic is $[G:H]\chi(\mathscr X)$. In particular, if $G$ is finite then $\chi(G)=1/\#G$. This led to the idea that $1/\Gamma_G(z)$ could behave like an Euler characteristic, and that its limit $1/\Gamma_G(1)$ could express $\chi(G)$. This is not always true, but it does hold in some illustrative cases.
Let us compute for instance the growth series of a free group $F_k$ generated by a basis[^13]. It follows from the formula for free products, or by direct counting if one notes, for all $\ell\ge1$, that there are $2k(2k-1)^{\ell-1}$ elements of norm $\ell$ in $F_k$, that $$\label{eq:growthFm0}
\Gamma_{F_k}(z)=\frac{1+z}{1-(2k-1)z}.$$ The value $\Gamma_{F_k}(1)$ is uniquely defined by analytic continuation, and one has $1/\Gamma_{F_k}(1)=1-k$, in agreement with $F_k$ being the fundamental group of a graph with $1$ vertex and $k$ edges. See [@floyd-p:fuchsian; @MR97d:20031a; @smythe:euler; @lewin:graphgroup] for more such examples of the ‘$1/\Gamma_G(1)=\chi(G)$’ phenomenon.
Complete growth series {#ss:completegrowth}
----------------------
There exists a stronger property than having a rational growth series: a group $G=\langle S\rangle$ has a *rational geodesic combing* if there exists a finite directed graph with edge labels in $S$ and a fixed “initial” vertex, such that the set $L\subseteq S^*$ of words read from the initial vertex along paths in the graph has the following property: $L$ maps bijectively to $G$ by the natural evaluation map of words as elements of $G$, and the words in $L$ have minimal length among all words in $S^*$ having the same evaluation in $G$. Kervaire suggested to consider the *complete* growth series $${\widehat\Gamma}_G(z)=\sum_{g\in G}g z^{\|g\|}\in\Z G[[z]].$$ Note that ${\widehat\Gamma}_G$ depends on the choice of generating set $S$, even though we do not mention it explicitly. The series ${\widehat\Gamma}_G(z)$ is a power series with coëfficients in the group ring, and one may again ask whether it is rational or algebraic. Since $\Z G$ need not be commutative, let us define more precisely these notions; we refer to [@salomaa-s:fps] for details. Let $\Lambda\subseteq\overline\Lambda$ be rings. An *algebraic system over $\Lambda$* in variables $X_1,\dots,X_n$ is a non-degenerate[^14] $n$-tuple of polynomials $P_1,\dots,P_n$ in non-commuting indeterminates $X_1,\dots,X_n$ and coëfficients in $\Lambda$. In a *linear system over $\Lambda$*, the polynomials are restricted to have degree $1$ and contain the indeterminate on the right; i.e. the $P_i$ are sums of monomials all belonging to $\Lambda\cup\bigcup_{1\le i\le n}\Lambda X_i$. A *solution* is an $n$-tuple $(f_1,\dots,f_n)\in\overline\Lambda^n$ such that all $P_i(f_1,\dots,f_n)=f_i$ for all $i=1,\dots,n$.
We then say that a power series $F(z)\in\Z G[[z]]$ is *rational*, respectively *algebraic*, if it is the first coördinate of the solution of a linear, respectively algebraic system over the polynomial ring $\Z G[z]$. A more direct definition of the ring of rational functions is that it is the smallest subring of $\Z G[[z]]$ containing $\Z G[z]$ and closed under Kleene’s *quasi-inversion*, the operation $F(z)^*=(1-F(z))^{-1}=1+F(z)+F(z)^2+\cdots$ defined for all $F(z)\in \Z G[[z]]$ with $F(0)=0$.
If $G$ admits a rational geodesic combing, then its growth series is rational.
Hint: define one variable $X_i$ for each vertex $i$ of the graph defining the combing, and encode the edges of the graph into polynomials.
\[exse:quasi-geodesic combing\] If the growth series of $G$ is rational, then $G$ admits a *quasi-geodesic combing*: a language $L\subset S^*$, recognised by a finite graph as above, with the property that, for some constant $C\in\N$, all words $s_1\dots s_\ell\in L$ have the property $\ell\le C\|s_1\cdots s_\ell\|$.
Hint: consider a polynomial system defining ${\widehat\Gamma}_{G,S}(z)$, make sure that every term in $\Z G$ is accompanied by at least one factor $z$. Write the terms in $\Z G$ as linear combinations of words over $S$, yielding a polynomial system over $\Z S^*$. Let $C$ be the maximal length of all these words over $S$ that appear in the polynomial system. A solution to the polynomial system will be a sum of monomials of the form $s_1\dots s_\ell z^n$, where $n=\|s_1\cdots s_\ell\|\ge \ell/C$.
These notions strengthen the ones for the classical growth series: if ${\widehat\Gamma}_G(z)$ is rational or algebraic, then its image under the augmentation map $\Z G\twoheadrightarrow\Z$ is rational or algebraic. On the other hand, statements concerning the complete growth series are usually not much harder to prove than the analogous ones concerning the classical growth series:
\[prop:CG:x\] Let the groups $G,H$ and $G\times H$ be respectively generated by $S,T$ and $S\sqcup T$. One then has $${\widehat\Gamma}_{G\times H}(z)={\widehat\Gamma}_G(z){\widehat\Gamma}_H(z).$$
Every element $(g,h)\in G\times H$ satisfies $\|(g,h)\|=\|g\|+\|h\|$; so $${\widehat\Gamma}_{G\times H}(z) = \sum_{(g,h)\in G\times H}g h z^{\|g\|+\|h\|}
= \sum_{g\in G}g z^{\|g\|}\sum_{h\in H}h z^{\|h\|} = {\widehat\Gamma}_G(z){\widehat\Gamma}_H(z).\qedhere$$
\[prop:CG:\*\] Let the groups $G,H$ and $G*H$ be respectively generated by $S,T$ and $S\sqcup T$. One then has $$\frac1{{\widehat\Gamma}_{G*H}(z)}=\frac1{{\widehat\Gamma}_G(z)}+\frac1{{\widehat\Gamma}_H(z)}-1.$$
Every element of $w\in G*H$ may be uniquely written in the form $w=h_0 g_1 h_1\cdots g_\ell$ with $h_0\in H$, $h_1,\dots,h_{\ell-1}\in H\setminus\{1\}$, $g_1,\dots,g_{\ell-1}\in
G\setminus\{1\}$, $g_\ell\in G$. Thus $$\begin{aligned}
{\widehat\Gamma}_{G*H}(z) &= \sum_{\ell\ge0}{\widehat\Gamma}_H(z)(({\widehat\Gamma}_G(z)-1)({\widehat\Gamma}_H(z)-1))^\ell{\widehat\Gamma}_G(z)\\
&= {\widehat\Gamma}_H(z)\frac1{1-({\widehat\Gamma}_G(z)-1)({\widehat\Gamma}_H(z)-1)}{\widehat\Gamma}_G(z);
\intertext{so}
\frac1{{\widehat\Gamma}_{G*H}(z)} &= \frac1{{\widehat\Gamma}_G(z)}({\widehat\Gamma}_G(z)+{\widehat\Gamma}_H(z)-{\widehat\Gamma}_G(z){\widehat\Gamma}_H(z))\frac1{{\widehat\Gamma}_H(z)}\\
& = \frac1{{\widehat\Gamma}_H(z)} + \frac1{{\widehat\Gamma}_G(z)} -1.\qedhere
\end{aligned}$$
These results are generalized to *graph products* in [@allen+:growthproducts]: given a graph $\Gamma$ with vertex set $V$ and a group $G_v$ for each $v\in V$, the *graph product* of the $G_v$ is $$G_\Gamma:=\bigast_{v\in V}G_v\Big/\big\langle[G_v,G_w]\text{ for each edge }(v,w)\big\rangle.$$ Recall that a *clique* in a graph is a subset of the vertices any two of which are connected by an edge. They show:
Let each group $G_v$ have generating set $S_v$, and consider the generating set $\bigcup_{v\in V}S_v$ of $G_\Gamma$. Then $$\frac1{{\widehat\Gamma}_{G_\Gamma}(z)}=\sum_{\text{clique }W\subseteq V}\prod_{v\in W}\Big(\frac1{{\widehat\Gamma}_{G_v}}-1\Big).$$
There are few classes of groups in which ${\widehat\Gamma}_G(z)$ is rational for all choices of generating set:
If $G$ is virtually abelian then ${\widehat\Gamma}_G(z)$ is rational for all choices of generating set.
If $G$ is word-hyperbolic then ${\widehat\Gamma}_G(z)$ is rational for all choices of generating set.
Note that there is no need to consider rings such as $\Z G$; the definition is more naturally phrased in terms of a *semiring* such as $\N G$. The polynomials $P_i$ are restricted to be sums of products of monomials, and no subtraction is allowed. The notions of $\Z$-rationality and $\N$-rationality differ subtly, see e.g. [@berstel:Nrat].
Let us now compute explicitly the complete growth series of a free group. It simplifies a little the notation to answer a slightly more general question. We denote throughout the text the cyclic group of order $p$ by $C_p$:
\[defn:freelike\] A *free-like* group is a finite free product of $\Z$’s and $C_2$’s.
Say that a free-like group $G$ has $m_1$ factors isomorphic to $C_2$ and $m_2$ factors isomorphic to $\Z$; then it has a symmetric generating set of size $m_1+2m_2$, consisting of one generator for each $C_2$ and a generator and its inverse for each $\Z$. The group $G$ is characterised by the property that its Cayley graph (see after Definition \[def:schreier\]) is an $(m_1+2m_2)$-regular undirected tree. We call such an $S$ a *natural* generating set for $G$.
For instance, the free group $F_{m/2}$ is free-like for $m$ even, and $\text{\LARGE$\ast$}^m C_2$ is free-like.
Let $G$ be free-like, and let $S$ denote a natural generating set of $G$ with cardinality $m$. We shall see that $G$ has rational complete growth series. For ease of notation, we write $\overline s$ for $s^{-1}$. We identify elements of $G$ with *reduced* words over $S$, i.e. words not containing consecutive $s\overline s$.
For all $s\in S$, define $F_s\in\Z G[[z]]$ by $$F_s=\sum_{\substack{w\in G,\text{ not}\\\text{starting with
}\overline s}} w z^{\|w\|};$$ so ${\widehat\Gamma}_G(z)=1+\sum_{s\in S}s z F_s$. We have the linear system in non-commutative unknowns $F_s$ $$F_s = 1 + \sum_{t\in S,t\neq \overline s}t z F_t,$$ with solution $F_s=(1-\overline s z)(1-\sum_{t\in S}t
z+(m-1)z^2)^{-1}$, so finally $$\label{eq:growthFm}
{\widehat\Gamma}_G(z)=\frac{1-z^2}{1-\sum_{s\in S}s z+(m-1)z^2},$$ compare with .
Asymptotic growth {#ss:asymptotic}
-----------------
We return to a group $G$ with generating set $S$, and view it as a metric space for the word metric. We consider the volume growth of balls in the metric space $(G,d)$; this is the *growth function* $v_G\colon\R_+\to\N$ given by $$v_G(R)=\#\{g\in G: \|g\|\le R\}.$$
It is naturally related to the formal power series $\Gamma(z)$: indeed $v_G(R)$ is the sum of all coëfficients of $\Gamma(z)$ of degree $\le R$; equivalently, for $R\in\N$ it is the degree-$R$ coëfficient of $\Gamma(z)/(1-z)$. Thus, by Tauberian and Abelian theorems (see e.g. [@nathanson:density]), asymptotics of $v_G(R)$ as $R\to\infty$ may be related to asymptotics of $\Gamma_G(z)$ as $z\to$ the convergence radius. In particular, the function $v_G(R)$ grows as $R^d$ if and only if $\Gamma_G(z)$ converges in the unit disk and has an order-$d$ pole singularity at $1$.
The norm $\|\cdot\|$ depends on the choice of generating set $S$, but only mildly: different choices of generating sets give equivalent norms, and equivalent metrics. If for $v,w\colon\R_+\to\N$ we write $v\precsim w$ to mean that $v(R)\le w(CR)$ for a constant $C\in\R_+$ and all $R\gg0$, and we write $v\sim w$ to mean $v\precsim w\precsim
v$, then
\[lem:indepgens\] The $\sim$-equivalence class of $v_G$ is independent of the choice of generating set.
Let $S,S'$ be two finite generating sets for $G$, and let us temporarily write $\|g\|_S,\|g\|_{S'}$ and $v_{G,S},v_{G,S'}$ for the norms and growth functions with respect to $S,S'$. There exists then a constant $C\in\N$ such that $\|s'\|_S\le C$ for all $s'\in S'$, and thus $\|g\|_{S'}\le
C\|g\|_S$. This gives $v_{G,S}(R)\le v_{G,S'}(CR)$. The reverse inequality holds by symmetry.
Note, as a consequence, that all exponentially-growing functions are equivalent, and that $R^d$ and $C\cdot R^d$ are equivalent as soon as $d>0$. The exponential growth rate $$\label{eq:lambda}
\lambda_{G,S}=\lim v_{G,S}(R)^{1/R}$$ is nevertheless worthy of consideration, and will be discussed in §\[ss:nug\].
History
-------
Interest in asymptotic growth of groups dates back at least to the early 1950’s, in the works of Krause [@krause:growth], Efremovich [@efremovich:proximity] and Švarc [@svarts:growth]; they were seeking coarse invariants of manifolds based on their fundamental group. Milnor noted in [@milnor:curvature] that, if $G$ is the fundamental group of a compact riemannian manifold $\mathscr M$, then $v_G$ is equivalent to the volume growth of balls in the universal cover of $\mathscr M$.
Here is a schematic of the known equivalence classes of growth functions of groups. Note the two dots for the two groups of order $4$, respectively the two groups $\Z^4$ and the Heisenberg group $H_3$ with quartic growth: $$\begin{tikzpicture}
\node at (0,0.5) {$1$};
\node at (0,0) {$\bullet$};
\node at (0,-0.8) {$1$};
\node (z/2) at (0.3,1) {$\tfrac{\Z}{2\Z}$};
\node at (0.3,0) {$\bullet$};
\draw[dotted] (z/2) -- (0.3,0);
\node at (0.3,-0.8) {$2$};
\node at (0.6,0) {$\bullet$};
\node at (0.6,-0.8) {$3$};
\node at (0.9,0.2) {$\bullet$};
\node at (0.9,-0.2) {$\bullet$};
\node at (0.9,-0.8) {$4$};
\node (z/5) at (1.2,1) {$\tfrac{\Z}{5\Z}$};
\node at (1.2,0) {$\bullet$};
\draw[dotted] (z/5) -- (1.2,0);
\node at (1.2,-0.8) {$5$};
\node at (1.6,0) {$\cdots$};
\node at (1.6,-0.8) {$\cdots$};
\node at (2,0.5) {$\Z$};
\node at (2,0) {$\bullet$};
\node at (2.0,-0.6) {$R$};
\node (z2) at (2.3,1) {$\Z^2$};
\node at (2.3,0) {$\bullet$};
\node (r2) at (2.3,-1.0) {$R^2$};
\draw[dotted] (z2) -- (r2);
\node at (2.6,0) {$\bullet$};
\node at (2.6,-0.6) {$R^3$};
\node at (2.9,1.0) {$\Z^4$};
\node at (2.9,0.5) {$H_3$};
\node at (2.9,0.2) {$\bullet$};
\node at (2.9,-0.2) {$\bullet$};
\node (r4) at (2.9,-1.0) {$R^4$};
\draw[dotted] (2.9,-0.2) -- (r4);
\node at (3.4,0) {$\cdots$};
\node at (3.4,-0.8) {$\cdots$};
\draw (4.6,0) ellipse (10mm and 4mm);
\node at (4.6,0) {gap};
\node at (5.6,-0.8) {$v_{\min}$};
\node at (6.25,0) {$???$};
\node at (7,0.5) {$W_{012}$};
\node at (7,0) {$\bullet$};
\node at (7,-0.8) {$\exp(R^{0.76\dots})$};
\draw[pattern=north west lines] (8.5,0) ellipse (14mm and 4mm);
\node at (10,0.5) {$F_2$};
\node at (10,0) {$\bullet$};
\node at (10,-0.8) {$\exp(R)$};
\end{tikzpicture}$$
The left of the graph is occupied by finite groups; the growth of a finite group is equivalent to the constant function taking value the order of the group. Abelian groups, and more generally virtually[^15] nilpotent groups have polynomial growth of type $R^d$ for an integer $d$. The converse is a deep result by Gromov:
\[thm:gromov\] A finitely generated group has growth function bounded by a polynomial if and only if it is virtually nilpotent.
It follows also from Gromov’s argument that there exists a superpolynomial function $v_{\min}(R)$ such that all groups with growth $\precnsim v_{\min}$ are virtually nilpotent; so there are no functions with growth strictly between polynomial and $v_{\min}$. Explicit estimates in [@shalom-tao:gromov] imply that one may take $v_{\min}(R)=R^{(\log R)^{1/100}}$, although the gap is probably larger. Note also that there need not exist a $\precnsim$-largest function $v_{\min}$. If one restricts oneself to virtually residually[^16] nilpotent groups, then the gap extends at least to $\exp(R^{1/2})$, see [@grigorchuk:hp]; if one restricts oneself to virtually residually solvable groups, then the gap extends at least to $\exp(R^{1/6})$, see [@wilson:gap].
Milnor asked in 1968, in a famous problem in the “American Math Monthly” [@milnor:5603], whether there exist groups whose growth function is neither polynomial nor exponential. He also conjectured in that note that groups of polynomial growth are precisely the virtually nilpotent groups. Milnor and Wolf showed in [@milnor:solvable; @wolf:solvable] that virtually solvable groups have either polynomial or exponential growth, and the *in*existence of groups with growth between polynomial and exponential became known as “Wolf’s conjecture”. Recall the celebrated “Tits’ alternative” [@tits:linear]: a finitely generated subgroup of a linear group in characteristic $0$ either is virtually solvable or contains a non-abelian free subgroup; from this it follows that linear groups always have polynomial or exponential growth (see furthermore §\[ss:nug\]).
However, groups of *intermediate growth* exist, and Grigorchuk [@grigorchuk:growth] gave such an example, known as the *first Grigorchuk group* $G_{012}$; see §\[ss:ss\].
The growth of $G_{012}$ is not known, even up to $\sim$-equivalence; conjecturally, it is the same as the growth of the group $W_{012}(C_2)$, which will be introduced in §\[ss:ssgrowth\]. The hatched region above indicates that, in fact, there are many groups of intermediate growth, and that any “reasonable” function between $\exp(R^{0.76\dots})$ and $\exp(R)$ is equivalent to the growth function of a group, see Theorem \[thm:givengrowth\].
There are at least two arguments for considering asymptotic growth rather than exact growth of groups. Firstly, the asymptotics of the growth function does not depend on the generating set, by Lemma \[lem:indepgens\], so is an invariant of the group itself. Secondly, we expect “most” growth series to be transcendental power series, so that they are probably difficult to describe, manipulate or expand; this happens e.g. for groups of subexponential growth, whose growth series converges in the unit disk so is either rational or transcendental, by Fatou’s theorem (see Theorem \[thm:polya-carlson\]).
Growth of regular wreath products {#ss:growthrw}
=================================
We consider in this section a wreath product $W=H\wr_X G$, and compute its growth series. We assume that generating sets $S,T$ for $G,H$ respectively have been chosen, and that the growth series of $G$ and $H$ are known.
Wreath products over finite sets
--------------------------------
As a first step, let us suppose that the set $X$ is finite, say $X=\{x_1,\dots,x_d\}$. Then, as generating set for $W$, we may take $\{t@x:t\in T,x\in X\}\sqcup S$. For this generating set, we have $$\Gamma_W(z)=\Gamma_H(z)^{\#X}\Gamma_G(z),$$ and by a small abuse of notation the same relation on complete growth series: $${\widehat\Gamma}_W(z)=\Big(\prod_{i=1}^d {\widehat\Gamma}_H(z)@x_i\Big) {\widehat\Gamma}_G(z).$$ Indeed, every element $w\in W$ may uniquely be written in the form $w=(h_1@x_1)\cdots(h_d@x_d)g$ for some $h_1,\dots,h_d\in H,g\in G$, and the growth series of $\{h_i@x_i\colon h_i\in H\}$ naturally coincides with that of $H$. In particular, if $\Gamma_G(z)$ and $\Gamma_H(z)$ are rational, then so is $\Gamma_W(z)$.
Johnson obtained in [@johnson:wreath] the same conclusion for more complicated generating sets of $W$.
Lamplighter groups
------------------
The next case we consider is $G=X=\Z$, and in particular “lamplighter groups”. Since the computations will be generalised in the next section, we content ourselves with a brief description of the growth series, and for simplicity assume that $H$ is a finite group. We consider $W=H\wr\Z$, denote a generator of $\Z$ by $s$, and let $W$ be generated by the set $\{s,s^{-1}\}\sqcup H@1$.
Consider an element $w\in W$. If its image under the natural map $W\to\Z$ is nonnegative, then it may be written minimally in the form $s^{-m}(h_0@1)s(h_1@1)\cdots s(h_p@1)s^{-n}$ with $h_i\in H$, $m,n\ge0$ and $p\ge m+n$, while if its image in $\Z$ is negative, then it may be written minimally in the form $s^m(h_0@1)s^{-1}(h_1@1)\cdots s^{-1}(h_p@1)s^n$ with $p>m+n$. Furthermore, $h_0$ must be non-trivial unless $m=0$, and $h_p$ must be non-trivial unless $n=0$.
All these constraints are local and therefore rational, except the long-range relation between $m,n,p$. However, in terms of computing growth series, the letters in the expression $s^{-m}(h_0@1)s(h_1@1)\cdots s(h_p@1)s^{-n}$ can be permuted at no cost; and the set of expressions of the form $$(h_0@1)s^{-1}s (h_1@1)\cdots s^{-1}s(h_m@1)s(h_{m+1}@1)\cdots
s(h_{p-n}@1)s s^{-1} (h_{p-n+1}@1)\cdots s s^{-1}(h_p@1)$$ is indeed a rational language, so that its growth function is rational.
Compute the growth series of $C_2\wr\Z$ with the standard generators $C_2@1\cup\{s,s^{-1}\}$. Note that the growth function grows exponentially, at the same rate as Fibonacci numbers. Could you have guessed the appearance of Fibonacci numbers without going through the calculations?
Regular wreath products with free groups
----------------------------------------
We compute in this subsection the complete growth series of a wreath product of the form $W=H\wr G$ for $G$ a free group. In fact, we suppose more generally that $G$ is free-like, see Definition \[defn:freelike\], so that its Cayley graph for the generating set $S$ is an $m$-regular tree $\mathscr T$. We keep the convention of writing $\overline s$ for $s^{-1}\in S$. We suppose as usual that $W$ is generated by $T@1\sqcup S$.
Consider $w\in W$, written as $w=(h_0@1)g_1(h_1@1)\cdots g_\ell(h_\ell@1)$ with $g_i\in G$ and $h_i\in H$. Following the arguments in §\[ss:genwreath\], one may write it as $$w=\prod_{i=0}^\ell(h_i@e_i)\cdot g_1\cdots g_\ell,\qquad\text{with }e_i=(g_1\cdots g_i)^{-1}.$$ The *support* of $w$ is the subgraph of $\mathscr T$ traced by inverses of prefixes of the word $g_1\dots g_\ell$; it is the convex hull of $\{e_0,\dots,e_\ell\}$ in $\mathscr T$. We shall count elements of $W$ by examining their possible supports and summing over them.
For each $s\in S$, let $\Theta_s$ denote the set of finite subtrees of $\mathscr T$, containing $1$ and no element of $S$ except possibly $s$. Each $\theta\in\Theta_s$ has *outer* vertices, with at most one neighbour in $\theta$, and *inner* vertices, with at least two neighbours in $\theta$. We introduce non-commutative power series $E_s(x,y,z)$ with coëfficients in $\Z G[z]$, which count the number of Eulerian cycles[^17] in trees $\Theta_s$, weighted by length in $z$; the variables $x,y$ belong to $G$ and in particular are *not* assumed central. The series $E_s(x,y,z)$ are defined by the algebraic system $$\label{eq:parry1}
E_s(x,y,z)=1+s y\overline s z^2+s x\Big(\prod_{t\in S,t\neq\overline s}E_t(x,y,z)-1\Big)\overline s z^2.$$
$$\begin{tikzpicture}
\draw[very thin] (0,0) node[above=2mm] {1} -- +(150:1.5)
(0,0) -- +(180:1.5)
(0,0) -- +(210:1.5)
(0,0) -- (1.5,0) node[above=2mm] {$s$} -- ++(30:1.5) -- ++(0:1) ++(0:-1) -- +(-20:1)
(1.5,0) -- ++(0:1.5) -- ++(20:1) ++(20:-1) -- ++(0:1) ++(0:-1) -- ++(-20:1)
(1.5,0) -- ++(-30:1.5) -- ++(20:1) ++(20:-1) -- ++(0:1);
\draw[thick,rounded corners,->] (0,0.1) -- ++(1.5,0) -- ++(30:1.6) -- ++(-60:0.2) -- ++(30:-1.6) -- ++(-1.6,0);
\end{tikzpicture}$$
The monomials in $E_s$ are in bijection with (Eulerian cycles tracing) trees in $\Theta_s$; if a tree $\theta$ with $p$ edges has inner vertices at $f_1,\dots,f_n$ and non-trivial outer vertices at $f'_1,\dots,f'_{n'}$, then the monomial corresponding to it is the product, in some order, of $z^{2p},x^{f_1},\dots,x^{f_n},y^{f'_1},\dots,y^{f'_{n'}}$. Indeed the equation defining $E_s$ says that a monomial counted by $E_s$ is either the empty tree (counted as $1$), or a single edge from $1$ to $s$ (counted as $s y\overline s z^2$), or an edge from $1$ to $s$, followed by $m-1$ subtrees counted recursively by $E_t$ for all $t\neq\overline s$, which are not all empty. Note that, if $\theta$ has $p$ edges, then a minimal closed path that explores all vertices of $\theta$ has length $2p$.
Let $D_s(z)$ denote the sum of $w z^{\|w\|}$ over all elements $w\in W$ which belong to the base group $H^G$ and whose support is an element of $\Theta_s$. Such elements may be counted as follows: starting from a support $\theta\in\Theta_s$, choose a word $g_1\dots g_\ell$ in $G$ of minimal length that visits all vertices of $\theta$; and, each time a vertex is first visited, insert an element of $H$, which furthermore must be non-trivial if the vertex is outer. Therefore, $$\label{eq:parry2}
D_s(z)=E_s({\widehat\Gamma}_H(z),{\widehat\Gamma}_H(z)-1,z).$$
Next, let $F_s(z)$ denote the sum of $w z^{\|w\|}$ over all elements $w=f g\in W$ with $f\colon G\to H$, $g\in G$ not beginning in $\overline s$, and whose support does not contain $\overline s$. We have a linear system $$\label{eq:parry3}
F_s(z)=\prod_{t\neq\overline s}D_t(z) + \sum_{t\neq\overline s}\Big(\prod_{u\neq t,\overline s}D_u(z)\Big) t z F_t,$$ since in every such element either $g=1$ and the support explores all the neighbours $t$ of $1$ except $\overline s$, or $g$ begins by a generator, say $t$, and then its support explores all neighbours of $1$ except $\overline s,t$, then moves to $t$, and continues by an element not starting by $\overline t$. Finally, $$\label{eq:parry4}
{\widehat\Gamma}_W(z)=\prod_{s\in S}D_s(z) + \sum_{s\in S}\Big(\prod_{t\neq s}D_t(z)\Big)s z F_s(z),$$ for the same reasoning as above. Combining Equations (\[eq:parry1\]–\[eq:parry4\]), we deduce:
\[thm:wreathalg\] If $H$ is a finitely generated group whose complete growth series ${\widehat\Gamma}_H(z)$ is algebraic, and $G$ is a free-like group, then the complete growth series of $W$ is also algebraic.
\[cor:parry\] If $H$ is a finitely generated group whose growth series $\Gamma_H(z)$ is algebraic, and $G$ is a free-like group, then the growth series of $W$ is also algebraic.
If furthermore $\Gamma_H(z)$ is rational and $m\le2$, then $\Gamma_W(z)$ is also rational.
On the other hand, if $m\ge3$ then $\Gamma_W(z)$ does not belong to the field generated by $z$ and $\Gamma_H(z)$.
Apply the augmentation map $\varpi\colon g\mapsto1$ to Equations (\[eq:parry1\]–\[eq:parry4\]); this gives an algebraic system of degree $\max(m-1,1)$ expressing $\Gamma_W(z)$ in terms of $z$ and $\Gamma_H(z)$. In particular, for $m=2$ it is a linear system.
Conversely, assume $m\ge3$ and let $\rho$ denote the convergence radius of the series of the image of $D_s$ under $\varpi$. Note that $\lim_{z\to\rho^-}\varpi(D_s)(z)$ is finite: if the limit were infinite, convergence to infinity would be order $(m-1)\times$ itself, a contradiction. Therefore $\varpi(D_s)$ has a non-pole singularity, so is not in $\Q(z,\Gamma_H(z))$.
Show that, for the lamplighter group $C_2\wr\Z$, the complete growth series is not rational.
Hint: use Exercise \[exse:quasi-geodesic combing\].
Traveling salesmen
------------------
To glimpse at the limit of what can be computed, consider now the case $G=\Z^2$. No property of $\Gamma_W(z)$ is known, and this is due to the fact that there is no good description of words of minimal norm describing group elements.
In fact, the problem can be quite precisely stated as follows. One is given a point $p_\infty$ and a set $\{p_1,\dots,p_\ell\}$ in $\Z^2$, and is required to find a walk of minimal length on the grid that starts at $(0,0)$, visits all the points $p_1,\dots,p_\ell$ in some order, and ends at $p_\infty$. This is a classical *travelling salesman* problem, and is known to be NP-complete, see [@garey-johnson-stockmeyer:graphpb; @garey-johnson:steinertree]. It is a small step to venture that finding a good description of minimal paths is at least as hard as finding those paths’ length.
Asymptotic growth {#asymptotic-growth}
-----------------
Regular wreath products, in non-degenerate cases, all have exponential growth. This is in stark contrast to the case of permutational wreath products, as we shall see in §\[ss:ssgrowth\].
If $H\neq 1$ and $G$ is infinite, then $W=H\wr G$ has exponential growth.
Choose $h\neq1\in H$, and without loss of generality assume that $h$ is a generator of $H$. Since $G$ is infinite, there exists an infinite word $g_1g_2\dots$ that traces a geodesic in the Cayley graph (see after \[def:schreier\]) of $G$, with $g_1,g_2,\dots$ generators of $G$ and also of $W$. In particular, all $(g_1\cdots g_i)^{-1}$ are distinct. Consider then, for any $\ell\in\N$, the set of elements $$\big\{(h@1)^{\epsilon_0}g_1(h@1)^{\epsilon_1}\cdots g_\ell(h@1)^{\epsilon_\ell}:\epsilon_0,\dots,\epsilon_\ell\in\{0,1\}\big\}.$$ All these elements have norm at most $2\ell+1$, and there are $2^{\ell+1}$ such elements. They are all distinct, since when they are rewritten in the form $f g_1\dots g_\ell$ one has $f((g_1\cdots
g_i)^{-1})=h^{\epsilon_i}$ so that the $\epsilon_i$ can be recovered from the element. Therefore, $v_W(2\ell+1)\ge2^{\ell+1}$.
(Self-)similar groups {#ss:ss}
=====================
We begin by introducing self-similar groups. They are groups with an additional structure:
A group $G$ is *self-similar* if it is endowed with a homomorphism $\phi\colon G\to G\wr_X{\operatorname{Sym}}(X)$ for some set $X$. The map $\phi$ is called the *wreath recursion* of $G$.
In this text, we shall always assume that the set $X$ is finite, and shall (unless stated otherwise) also assume that the homomorphism $\phi$ is injective. A self-similar group is a group $G$ in which elements may be recursively described by a permutation of $X$, decorated by elements of $G$ itself. In case $X=\{0,1,\dots,d-1\}$, we also write elements of $G\wr_X{\operatorname{Sym}}(X)$ in the form $\pair<g_0,\dots,g_{d-1}>\pi$ for group elements $g_0,\dots,g_{d-1}$ and a permutation $\pi\in{\operatorname{Sym}}(X)$.
It is essential to understand that being self-similar is an *attribute* of a group, and not a property. Thus, for example, a topological group is a group endowed with a topology; and every group is a topological group, for the discrete and the coarse topology. In the same vein, every group is self-similar, merely for the reason that it is similar to itself. Taking $X=\{0\}$ and $\phi(g)=\pair<g>$ is uninteresting, but is not illegal.
Finite-state self-similar groups {#ss:fsa}
--------------------------------
We describe two fundamental constructions of self-similar groups.
For the first, start by a well-understood group $F$, such as a free group; and choose a (not necessarily injective!) homomorphism $\tilde\phi\colon F\to F\wr_X{\operatorname{Sym}}(X)$. There exists then a maximal quotient of $F$ on which the map $\tilde\phi$ induces an injective wreath recursion. To wit, one defines an increasing sequence $N_i$ of normal subgroups of $F$ by $$\label{eq:normalF}
N_0=1,\quad N_{i+1}=\tilde\phi^{-1}(N_i^X),$$ and sets $G=F/\bigcup_i N_i$. By construction, the map $\tilde\phi$ induces an injective map $\phi\colon G\to G\wr_X{\operatorname{Sym}}(X)$.
An important example of group defined by this method — and which, essentially, cannot be defined differently — is the *first Grigorchuk group*, introduced in [@grigorchuk:burnside] and based on [@aleshin:burnside]. Consider $$\label{eq:F}
F=\langle a,b,c,d\mid a^2,b^2,c^2,d^2,bcd\rangle\cong C_2*(C_2\times C_2),$$ and define $\tilde\phi\colon F\to F\wr{\operatorname{Sym}}(2)$ by $$\tilde\phi(a)=\tikz[baseline=-2ex]{\draw[->] (0,0) -- (0.5,-0.5); \draw[->](0.5,0) -- (0,-0.5);},\qquad
\tilde\phi(b)=\tikz[baseline=-2ex]{\draw[->] (0,0) -- node[right=-2pt] {\small$a$} (0,-0.5); \draw[->](0.5,0) -- node[right=-2pt] {\small$c$} (0.5,-0.5);},\qquad
\tilde\phi(c)=\tikz[baseline=-2ex]{\draw[->] (0,0) -- node[right=-2pt] {\small$a$} (0,-0.5); \draw[->](0.5,0) -- node[right=-2pt] {\small$d$} (0.5,-0.5);},\qquad
\tilde\phi(d)=\tikz[baseline=-2ex]{\draw[->] (0,0) -- (0,-0.5); \draw[->](0.5,0) -- node[right=-2pt] {\small$b$} (0.5,-0.5);}.$$
It is straightforward to see that $\tilde\phi$ is a homomorphism — just compute the images of the relators. It is, however, remarkable that one may compute efficiently in $G$ just using this description. We use the same letters $a,b,c,d$ for the corresponding generators of $G$. As an illustration, let us check that the relation $(ad)^4$ holds in $G$. Writing the permutation diagrams horizontally, one has $$\phi((ad)^4)=\Big(\tikz[baseline=1.5mm]{\draw[->] (0,0.5) -- +(0.7,-0.5); \draw[->] (0,0) -- +(0.7,0.5);
\draw[->] (0.75,0) -- +(0.7,0); \draw[->] (0.75,0.5) -- node[above=-1pt] {\small$b$} +(0.7,0);
}\Big)^4
=\Big(\tikz[baseline=1.5mm]{\draw[->] (0,0.5) -- +(0.7,-0.5); \draw[->] (0,0) -- +(0.7,0.5);
\draw[->] (0.75,0) -- +(0.7,0); \draw[->] (0.75,0.5) -- node[above=-1pt] {\small$b$} +(0.7,0);
\draw[->] (1.5,0.5) -- +(0.7,-0.5); \draw[->] (1.5,0) -- +(0.7,0.5);
\draw[->] (2.25,0) -- +(0.7,0); \draw[->] (2.25,0.5) -- node[above=-1pt] {\small$b$} +(0.7,0);
}\Big)^2
=\Big(\tikz[baseline=1.5mm]{\draw[->] (0,0) -- node[below=-1pt] {\small$b$} +(0.7,0); \draw[->] (0,0.5) -- node[above=-1pt] {\small$b$} +(0.7,0);}\Big)^2
=\tikz[baseline=1.5mm]{\draw[->] (0,0) -- node[below=-1pt] {\small$b^2$} +(0.7,0); \draw[->] (0,0.5) -- node[above=-1pt] {\small$b^2$} +(0.7,0);}
=1,$$ so $(ad)^4=1$ in $G$ because $\phi$ is injective.
Using similar calculations, compute the exponent of $ab$ and $ac$ in $G$.
Note that, in that example, $G=\langle S\rangle$ for the set $S=\{1,a,b,c,d\}$ which has the property that $\phi(S)$ is contained in $S\times S\times{\operatorname{Sym}}(2)$. More generally,
Let $G$ be a self-similar group. A subset $S\subseteq G$ is *state-closed* if $\phi(S)$ is contained in $S^X\times{\operatorname{Sym}}(X)$.
An element $g\in G$ is *finite-state* if there exists a state-closed subset of $G$ containing $g$. A subset of $G$ is *finite-state* if all its elements are finite-state.
Let $\Z=\langle t\rangle$ be endowed with the self-similar structure $\phi(t)=\tikz[baseline=-2ex]{\draw[->] (0,0) -- node[near
start,left] {\small$t$} (0.5,-0.5); \draw[->] (0.5,0) --
node[near start,right=-1pt] {\small$t^2$} (0,-0.5);}$. Show that only $t^0$ is finite-state.
The product and inverse of finite-state elements is again finite-state.
If $g,h$ are finite-state contained respectively in state-closed sets $S,T$, then $g h^{-1}$ is finite-state, since it belongs to the finite state-closed set $ST^{-1}$.
Therefore, a finitely generated self-similar group $G$ is finite-state if and only if its generators are finite-state, and one may assume that $G$ is generated by a state-closed set.
In that case, the wreath recursion of $G$ may conveniently be represented by an *automaton*, more precisely a *Mealy automaton*. This is a directed graph with vertex set $S$ called its *states*, and with an edge from $s\in S$ to $t\in S$, with label ‘$x|y$’, whenever the decorated permutation $\phi(s)$ maps $x\in X$ to $y\in X$ and has label $t$ on the edge $x\to y$. Thus, in a sense, the graph is the dual of the permutation diagram, with the roles of $X$ and $S$ exchanged. The automaton generating the first Grigorchuk is, with the convention $X=\{\9,\8\}$, $$
\begin{tikzpicture}[baseline,->,>=Stealth,shorten >=1pt,auto,
node distance=2.8cm,semithick,double equal sign distance]
\tikzstyle{every state}=[fill=gray,draw=none,text=white]
\node[state] (b) at (1.4,3) {$b$};
\node[state] (d) at (4.2,3) {$d$};
\node[state] (c) at (2.8,0.7) {$c$};
\node[state] (a) at (0,0) {$a$};
\node[state] (e) at (5.6,0) {$1$};
\path (b) edge node {$\8|\8$} (c) edge node {$\9|\9$} (a)
(c) edge node {$\8|\8$} (d) edge node {$\9|\9$} (a)
(d) edge node[above] {$\8|\8$} (b) edge node {$\9|\9$} (e)
(a) edge [double,bend right=30] node {$\9|\8,\8|\9$} (e)
(e) edge [double,loop right] node {$\9|\9,\8|\8$} (e);
\end{tikzpicture}$$
Assume that the self-similar group was obtained as above as a quotient of a self-similar group $F$ with a map $\tilde\phi\colon F\to
F\wr_X{\operatorname{Sym}}(X)$. Recall that the *word problem* asks, given a word in the generators of a finitely generated group, to determine whether the group element that it defines is trivial. There are groups, even finitely presented, in which the word problem is unsolvable [@novikov:wp]; however,
Let $F$ be a finite-state finitely generated self-similar group with wreath recursion $\tilde\phi$ and solvable word problem, and $G$ be the maximal quotient of $F$ on which the induced wreath recursion $G\to G\wr_X{\operatorname{Sym}}(X)$ is injective. Then $G$ also has solvable word problem.
Assume without loss of generality that $F$ is generated by the finite state-closed set $S$, and denote also by $S$ the corresponding generating set of $G$. Given a word $w\in S^*$ of length $\ell$, it defines a state-closed element of $F$, belonging to the state-closed set $S^\ell$. Consider the corresponding automaton with vertex set $S^\ell$.
Let $U\subseteq S^\ell$ denote the set of states that are reachable from $w\in S^\ell$ by arbitrarily long paths. This set is computable: set $U'_0:=\{w\}$, and for $i\ge0$ set $U'_{i+1}:=U'_i\cup\{$endpoints of edges starting in $U'_i\}$; then the $U'_i$ form an increasing sequence of subsets of $S^\ell$, hence stabilize, say to $U''_0$. Note that $U''_0$ is the set of states reachable from $w$. For all $i\ge0$, let $U''_{i+1}$ denote those endpoints of edges starting in $U''_i$; then the $U''_i$ form a decreasing sequence of subsets of $S^\ell$, hence stabilize, to $U$.
The element of $G$ defined by $w$ is trivial in $G$ if and only if both all the edges starting in $U''_0$ have labels of the form ‘$x|x$’ for some $x\in X$, and all elements of $U$ define trivial elements of $F$ under the evaluation map $S^*\to F$.
More precisely, let $m\in\N$ be minimal such that every element of $U$ may be reached from $w$ by a path of length at most $m$. Then the conditions above imply that $w$ belongs to the normal subgroup $N_m$ of $F$, see .
The above proof amounts to constructing a Mealy automaton for the action of $S^\ell$, and examining it to determine which of its states are trivial in $G$.
Linear groups
-------------
Here is another construction of self-similar groups. Consider a group $G$, a subgroup $H$, and a homomorphism $\phi_0\colon H\to G$. By the “permutational Kaloujnine-Krasner theorem” \[thm:kkwreath\], there exists a natural extension $\phi\colon G\to G\wr_X{\operatorname{Sym}}(X)$, with $X=H\backslash G$, in such a manner that $\phi(h)=\pair<\dots,\phi_0(h),\dots>\dots$ for all $h\in H$, with the ‘$\phi_0(h)$’ in position $H\in H\backslash G$.
Alternatively, this map $\phi$ may be directly constructed as follows: choose a transversal $T$ of $H$ in $G$, namely a subset $T\subseteq G$ such that every $g\in G$ may uniquely be written in the form $ht$ with $h\in H,t\in T$. Identify $X$ with $T$. Let then $\phi(g)$ be the decorated permutation that sends $t\in T$ to $u\in T$ with label $\phi_0(t g u^{-1})$ whenever $t g u^{-1}$ belongs to $H$.
Here is a fundamental example: choose a prime number $p$, and consider $$G=\Gamma_0(p)=\begin{pmatrix}\Z&\Z\\p\Z&\Z\end{pmatrix}\cap{{\operatorname{SL}}}_2(\Z).$$ Consider also the matrix $\Phi=(\begin{smallmatrix}p&0\\0&p^{-1}\end{smallmatrix})\in{{\operatorname{SL}}}_2(\Q)$, and $H=G\cap G^{\Phi^{-1}}$. Set $\phi_0(h)=h^\Phi$. This example generalizes naturally to $G$ any matrix group, such as for instance a congruence subgroup of ${{\operatorname{SL}}}_n(\Z)$ for arbitrary $n$, or even ${{\operatorname{SL}}}_n(\Z)$ itself. This shows that the class of linear groups over $\Z$ is contained in the class of self-similar groups.
In fact, the only essential ingredient of the above construction is the element $\Phi$ in the *commensurator* of $G$. Recall that, for a subgroup $G$ of a group $L$, the commensurator of $G$ is the subgroup of those $x\in L$ such that $G\cap G^x$ has finite index in $G$ and in $G^x$. If $G$ is an irreducible lattice in a Lie group $G$, then $G$ may be called *arithmetic* [@margulis:subgroups] if its commensurator is dense in $L$; e.g. the commensurator of ${{\operatorname{SL}}}_n(\Z)$ in ${{\operatorname{SL}}}_n(\R)$ is ${{\operatorname{SL}}}_n(\Q)$. Then all arithmetic lattices admit self-similar actions on rooted trees [@kapovich:arithmetic].
Rooted trees
------------
Let $X$ be a set, and consider the associated *rooted regular tree* $\mathscr T$: its vertex set is $X^*=\{x_i\dots x_1:x_j\in X\}=\bigsqcup_{i\ge0}X^i$, and it has an edge between $x_{i+1}x_i\dots x_1$ and $x_i\dots x_1$ for all $x_i\in X$. The tree is rooted at the empty word, the unique element of $X^0$; the set of vertices at distance $i$ from the root is identified with $X^i$, and the Cartesian product $X^\infty$ is naturally interpreted as the boundary $\partial\mathscr T$ of the tree, namely the set of infinite rays emanating from the root. Here for illustration is the top of the binary tree: $$\begin{tikzpicture}[level distance=10mm,
every node/.style={inner sep=1pt},
level 1/.style={sibling distance=40mm},
level 2/.style={sibling distance=20mm},
level 3/.style={sibling distance=10mm}]
\node {$\emptyset$}
child {node {$\9$}
child {node {$\9\9$}
child {node {$\9\9\9$}}
child {node {$\8\9\9$}}
}
child {node {$\8\9$}
child {node {$\9\8\9$}}
child {node {$\8\8\9$}}
}
}
child {node {$\8$}
child {node {$\9\8$}
child {node {$\9\9\8$}}
child {node {$\8\9\8$}}
}
child {node {$\8\8$}
child {node {$\9\8\8$}}
child {node {$\8\8\8$}}
}
};
\end{tikzpicture}$$ Let $W$ denote the isometry group of $\mathscr T$, namely the set of bijections of $X^*$ that fix the root $\emptyset$ and preserve the edge structure of $\mathscr T$. Given $g\in W$, let $\sigma\in{\operatorname{Sym}}(X)$ denote the action of $g$ on $X=X^1$, and for all $x\in X$ define an element $g_x\in W$ by $(x_n\cdots x_1 x)g=(x_n\cdots x_1)g_x\,(x^\sigma)$; namely, $g_x$ describes the action of $g$ on the subtree $X^*x$ as is it carried to $X^*x^\sigma$ by $g$.
The map $$\phi\colon\begin{cases}
W &\to W\wr{\operatorname{Sym}}(X)\\
g &\mapsto\pair<g_x:x\in X>\sigma
\end{cases}$$ is a group isomorphism.
Given $g_x\in W$ and $\sigma\in{\operatorname{Sym}}(X)$, an element $g\in W$ may uniquely be defined by $(x_n\cdots x_1 x)g=(x_n\cdots x_1)g_x\,(x^\sigma)$. This proves that $\phi$ is bijective.
To see that $\phi$ is a homomorphism, consider the $\#X$ subtrees below the root. They are permuted according to the permutation part $\sigma$ of $\phi$, and simultaneously acted upon by the decorations $g_x$. Composition of decorated permutations therefore coincides with composition of tree isometries.
Let us now start with a self-similar group $G$. Its wreath recursion $\phi\colon G\to G\wr_X{\operatorname{Sym}}(X)$ then defines an action of $G$ on $X$. Furthermore, the wreath recursion can be “iterated”: one has maps $$\begin{aligned}
G\overset\phi\longrightarrow G\wr_X{\operatorname{Sym}}(X)\overset{\phi^X}\longrightarrow&(G\wr_X{\operatorname{Sym}}(X))\wr_X{\operatorname{Sym}}(X)\\
&=G\wr_{X^2}({\operatorname{Sym}}(X)\wr_X{\operatorname{Sym}}(X))\longrightarrow\cdots\end{aligned}$$ so that $G$ acts on $X^i$ for all $i\in\N$. Furthermore, these actions are compatible with each other, in the sense that the map $(x_{i+1},x_i,\dots,x_1)\mapsto(x_i,\dots,x_1)$ interlaces the actions on $X^{i+1}$ and $X^i$. Taking the inverse limit of $X^i$ under these projection maps gives an action on the Cartesian product $X^\infty$. Note that sequences in $X^\infty$ are infinite on the left, namely are of the form $(\dots,x_{i+1},x_i,\dots,x_1)$.
The compatibility between the actions on $X^i$ and $X^{i+1}$ precisely means that $G$ acts by tree isometries on the rooted regular tree $\mathscr T$ with vertex set $X^*$.
Note that, even if the wreath recursion $\phi$ is injective, the action of the $G$ on the tree $\mathscr T$ need not be faithful. This is, however, the case for the examples of self-similar action of the first Grigorchuk group (see Proposition \[prop:faithful,infinite\]) and of the congruence subgroup $\Gamma_0(p)$.
We introduced the self-similar structure on $\Gamma_0(p)$ and not on ${{\operatorname{SL}}}_2(\Z)$ because the latter does not act on the *rooted $p$-regular tree*, in which each vertex has degree $p+1$ except the root which has degree $p$. If we add an edge upwards from the root, and a rooted $p$-regular tree above it, to the rooted $p$-regular tree, we obtain a $(p+1)$-regular tree on which the action of $\Gamma_0(p)$ extends to an action of ${{\operatorname{SL}}}_2(\Z)$. In fact, this action is already well-known, see [@serre:trees]\*[§II.1]{}: the $(p+1)$-regular tree is the *Bruhat-Tits tree* of ${{\operatorname{SL}}}_2(\Z_p)$. Its vertices are homothety classes of lattices $\cong\Z_p^2$ in $\Q_p$, and there is an edge between classes $\Q_p^\times\Lambda$ an $\Q_p^\times\Lambda'$ if they admit representatives $\alpha\Lambda,\alpha'\Lambda'$ with $\alpha\Lambda\subset\alpha'\Lambda'$ and $[\alpha'\Lambda':\alpha\Lambda]=p$. The group ${{\operatorname{SL}}}_2(\Q_p)$ naturally acts on lattices, and ${{\operatorname{SL}}}_2(\Z_p)$ acts as the stabilizer of the root $\Q_p^\times\Z_p^2$. The congruence subgroup $\Gamma_0(p)$ fixes an edge adjacent to the root, and the rooted $p$-regular tree $\mathscr T$ is spanned by those lattices of the form $\langle (p^n,0),(x_0+x_1p+\cdots+x_{n-1}p^{n-1},1)\rangle$ for $n\in\N$ and $x_0,\dots,x_{n-1}\in\{0,\dots,p-1\}$.
Let us remark in passing that obtaining an action on a rooted tree is not spectacular in itself: every countable residually-$p$ group acts on a rooted $p$-regular tree. Indeed, choose a descending sequence $G=G_0>G_1>\cdots$ of subgroups with $[G:G_i]=p^i$ and $\bigcap G_i=1$. Let the vertices of $\mathscr T$ be the set of right cosets of all $G_i$, with an edge between $G_i g$ and $G_{i+1} g$ for all $i\in\N,g\in G$; and let $G$ act by right multiplication on $\mathscr T$.
This action is in general not self-similar, nor is it “economical”, in the sense that the permutation group acting on $X^i$ may have order comparable to $(\#X)^i$ rather than $(\#X!)^{\#X^i}$.
Finally, let us return to the construction of a self-similar group $G$ as a quotient of a self-similar group $F$ so that the wreath recursion becomes injective. Knowing that the action of $G$ is faithful helps in solving the word problem, in case $G$ is not finite-state:
Let $F$ be a self-similar group with wreath recursion $\tilde\phi$ and solvable word problem, and let $G$ be the maximal quotient of $F$ on which the induced wreath recursion is injective. Assume that the action of $G$ on the tree $\mathscr T$ is faithful. Then $G$ also has solvable word problem.
Let $S$ be a generating set for $F$, and consider $w\in S^*$. We start two semi-algorithms in parallel; the first one will stop if $w$ is non-trivial in $G$, and the second one will stop if $w$ is trivial in $G$.
If $w$ is non-trivial in $G$, then it will act non-trivially on some vertex of $\mathscr T$, and this vertex may be found by enumerating all vertices of $\mathscr T$ and computing the action of $w$ on it by applying $\tilde\phi$.
If $w$ is trivial, then it belongs to one of the normal subgroups $N_i$. Going through all $i=0,1,\dots$ in sequence, and iterating $i$ times $\tilde\phi$ on $w$ yields $\#X^i$ elements of $F$. If all of them are trivial in $F$, then $w$ is trivial; otherwise continue with the next $i$.
Here are two fundamental examples of self-similar groups. Let $\mathscr T=X^*$ be a rooted regular tree, and let $Q$ be a group acting transitively on $X$. Consider the iterated wreath products $Q_i=Q\wr_X Q\wr_X\cdots\wr_X Q$, with $i$ factors; these groups act naturally on $X^i$ by the imprimitive action.
On the one hand, there is a natural map $Q_{i+1}\twoheadrightarrow Q_i$, given by deleting the leftmost factor, i.e. naturally mapping $Q_{i+1}=Q\wr_X Q_i$ to $Q_i\cong1\wr_X Q_i$. Set then $\overline G=\varprojlim Q_i$, the projective limit being taken along these epimorphisms. The self-similarity structure $\phi\colon \overline G\to \overline G\wr_X Q$ is induced by the identity map $Q_i\overset\cong\longrightarrow Q_{i-1}\wr_X Q$. Since it ‘peels off’ the rightmost factor, it is compatible with the inverse limit. It defines a profinite self-similar group $\overline G$.
On the other hand, there is a natural map $Q_i\hookrightarrow
Q_{i+1}$, given by inserting a trivial leftmost factor, i.e.naturally mapping $Q_i\cong1\wr_X Q_i$ to $Q_{i+1}=Q\wr_X Q_i$. Set then $L=\varinjlim Q_i$, the union (= injective limit) being taken along these monomorphisms. The self-similarity structure $\phi\colon L\to
L\wr_X Q$ is induced by the identity map $Q_i\overset\cong\longrightarrow Q_{i-1}\wr_X Q$. Since it ‘peels off’ the rightmost factor, it is compatible with the union. It defines a locally finite group $L$.
Show that $L$ is a dense subgroup of $\overline G$.
A *law* for a group $G$ is a word $w(x_1,x_2,\dots)$ in variables $x_1,x_2,\dots$ such that, whenever the elements $x_1,x_2,\dots$ are replaced by group elements from $G$, the word evaluates to $1$ in $G$. For example, abelian groups are characterised as those groups satisfying the law $w=[x_1,x_2]$.
\[exse:nolaw\] Show that $L$ satisfies no non-trivial law.
Hint: it suffices to look at the case $X=\{1,2\}$. By Theorem \[thm:kk\], every finite $2$-group imbeds in $Q_i$ for some $i\in\N$, and therefore in $G$. Finally, the free group is residually $2$.
See [@abert:nonfree] for a general result about inexistence of group laws, which covers the group $L$.
Similar families of groups {#ss:ssfam}
--------------------------
The notion of self-similar group may be generalised to a *family* of similar groups.
Let $\Omega$ be a set, and let $\sigma\colon\Omega\righttoleftarrow$ be a map. A *similar family of groups* over $\Omega$ is a family $(G_\omega)_{\omega\in\Omega}$ of groups and a family of homomorphisms $$\phi_\omega\colon G_\omega\to G_{\sigma\omega}\wr_{X_\omega}{\operatorname{Sym}}(X_\omega),$$ for a family of sets $(X_\omega)_{\omega\in\Omega}$.
Just as before, each group $G_\omega$ acts on a tree $\mathscr
T_\omega$ with vertex set $\bigsqcup_{i\ge0}X_{\sigma^{i-1}\omega}\times\cdots\times
X_\omega$. This rooted tree is now not anymore regular, but it is still *spherically homogeneous*, in that its isometry group is transitive on the set of vertices at given distance from the root.
As before, there are two fundamental examples of similar families of groups. Let $(X_\omega)_{\omega\in\Omega}$ be a family of sets, and let $(Q_\omega)_{\omega\in\Omega}$ be a family of groups, with $Q_\omega$ acting on $X_\omega$, say transitively for simplicity.
Consider the iterated wreath products $Q_{\omega,i}=Q_{\sigma^{i-1}\omega}\wr_{X_{\sigma^{i-2}\omega}}\cdots\wr_{X_\omega}
Q_\omega$, with $i$ factors; these groups act naturally on $X_{\sigma^{i-1}\omega}\times\cdots\times X_\omega$ by the imprimitive action.
On the one hand, there is a natural map $Q_{\omega,i+1}\twoheadrightarrow Q_{\omega,i}$, given by deleting the leftmost factor, i.e. naturally mapping $Q_{\omega,i+1}=Q_{\sigma^i\omega}\wr_{X_{\sigma^{i-1}\omega}}
Q_{\omega,i}$ to $Q_{\omega,i}\cong1\wr_{X_{\sigma^{i-1}\omega}}
Q_{\omega,i}$. Set then $\overline G_\omega=\varprojlim Q_{\omega,i}$, the projective limit being taken along these epimorphisms. The self-similarity structure $\phi_\omega\colon \overline G_\omega\to
\overline G_{\sigma\omega}\wr_{X_\omega} Q_\omega$ is induced by the identity map $Q_{\sigma,i}\overset\cong\longrightarrow
Q_{\sigma\omega,i-1}\wr_{X_\omega} Q_\omega$. Since it ‘peels off’ the rightmost factor, it is compatible with the inverse limit. It defines a profinite self-similar group $\overline G_\omega$.
On the other hand, there is a natural map $Q_{\omega,i}\hookrightarrow
Q_{\omega,i+1}$, given by inserting a trivial leftmost factor, i.e.naturally mapping $Q_{\omega,i}\cong1\wr_{X_{\sigma^{i-1}\omega}}
Q_{\omega,i}$ to $Q_{\omega,i+1}=Q_{\sigma^i\omega}\wr_{X_{\sigma^{i-1}\omega}}
Q_{\omega,i}$. Set then $L_\omega=\varinjlim Q_{\omega,i}$, the injective limit being taken along these monomorphisms. The self-similarity structure $\phi\omega\colon L_\omega\to
L_{\sigma\omega}\wr_{X_\omega} Q_\omega$ is induced by the identity map $Q_{\sigma,i}\overset\cong\longrightarrow
Q_{\sigma\omega,i-1}\wr_{X_\omega} Q_\omega$. Since it ‘peels off’ the rightmost factor, it is compatible with the union. It defines a locally finite self-similar group $L_\omega$.
The Grigorchuk family $G_\omega$ {#ss:grigorchuk}
--------------------------------
We shall concentrate particularly on one specific example. Write $\{0,1,2\}$ for the three non-trivial homomorphisms $C_2\times C_2\to
C_2$, identified for definiteness as follows. We view the source $C_2\times C_2=\{1,b,c,d\}$ as a subgroup of the group $F$ given in , and the range $C_2=\{1,a\}$ in that same group $F$. The three homomorphisms are then uniquely defined by $\ker(0)=\langle b\rangle$ and $\ker(1)=\langle c\rangle$ and $\ker(2)=\langle d\rangle$. Set then $$\Omega=\{0,1,2\}^\infty,\qquad\sigma(\omega_0\omega_1\omega_2\dots)=\omega_1\omega_2\dots.$$
We start by the similar family $(F)_{\omega\in\Omega}$ with maps $\phi_\omega\colon F\to F\wr{\operatorname{Sym}}(2)$ given by $$\tilde\phi_\omega(a)=\tikz[baseline=-2ex]{\draw[->] (0,0) -- (0.5,-0.5); \draw[->](0.5,0) -- (0,-0.5);},\quad\text{and for all }x\in\{b,c,d\}:\quad
\tilde\phi_\omega(x)=\tikz[baseline=-2ex]{\draw[->] (0,0) -- node[right=-2pt] {\small$\omega_0(x)$} +(0,-0.5); \draw[->](1.0,0) -- node[right=-2pt] {\small$x$} +(0,-0.5);},$$ we define normal subgroups $(N_{\omega,i})_{i\in\N,\omega\in\Omega}$ of $F$ by $N_{\omega,0}=1$ and $N_{\omega,i+1}=\tilde\phi_\omega^{-1}(N_{\omega,i}^2)$, and set $G_\omega=F/\bigcup_{i\in\N}N_{\omega,i}$. This is the same construction as above, and computes $G_\omega$ as the maximal quotient of $F$ such that the maps $\tilde\phi_\omega\colon F\to F\wr{\operatorname{Sym}}(2)$ descend to injective maps $\phi_\omega\colon G_\omega\to
G_{\sigma\omega}\wr{\operatorname{Sym}}(2)$.
In particular, letting $G$ denote the Grigorchuk group defined in §\[ss:fsa\], we have isomorphisms $$\begin{aligned}
&\begin{tikzpicture}
\path [use as bounding box,red] (-2,0) rectangle (8,0);
\node (g) at (0,0) {$G$};
\node (g012) at (2,0) {$G_{(012)^\infty}$};
\node (g120) at (4,0) {$G_{(120)^\infty}$};
\node (g201) at (6,0) {$G_{(201)^\infty}$};
\draw[double equal sign distance] (g) -- node[above] {$\sim$} (g012);
\draw[double equal sign distance] (g012) -- node[above] {$\sim$} (g120);
\draw[double equal sign distance] (g120) -- node[above] {$\sim$} (g201);
\end{tikzpicture}
\intertext{identifying the generating sets as follows:}
&\begin{tikzpicture}
\path [use as bounding box,red] (-2,0) rectangle (8,0);
\node (s) at (0,0) {$\{a,b,c,d\}$};
\node (s012) at (2,0) {$\{a,d,c,b\}$};
\node (s120) at (4,0) {$\{a,b,d,c\}$};
\node (s201) at (6,0) {$\{a,c,b,d\}.$};
\end{tikzpicture}\end{aligned}$$
The groups $G_\omega$ all act on the binary rooted tree, and it is easy to see that the orbit of the ray $\8^\infty$ is dense. Therefore, the groups $G_\omega$ could just as well have been defined by their actions on their respective orbit $\8^\infty G_\omega$. These are naturally graphs, called *Schreier graphs*, with vertex set $\8^\infty G_\omega$, and with an edge from $\8^\infty g$ to $\8^\infty gs$ for each generator $s\in\{a,b,c,d\}$, see Definition \[def:schreier\].
The graph $\8^\infty G_\omega$ is isometric to the half-infinite line $\N$ with multiple edges and loops. Under this identification with $\N$, the action of $G_\omega$ is given by
[2]{} a(2j)&=2j+1, & a(2j+1)&=2j,\
x{b,c,d},x(0)&=0, & x(2\^i(2j+1))&=2\^i(2\_i(j)+1).
Consider the infinite dihedral group $D=\langle a,x\mid
a^2,x^2\rangle$, and the wreath recursion $\phi\colon D\to
D\wr{\operatorname{Sym}}(2)$ defined by $$\phi(a)=\tikz[baseline=-2ex]{\draw[->] (0,0) -- (0.5,-0.5); \draw[->](0.5,0) -- (0,-0.5);},\qquad
\tilde\phi(x)=\tikz[baseline=-2ex]{\draw[->] (0,0) --
node[right=-2pt] {\small$a$} (0,-0.5); \draw[->](0.5,0) --
node[right=-2pt] {\small$x$} (0.5,-0.5);}.$$ It defines a faithful action of $D$ on the binary rooted tree $\mathscr T$ with vertex set $\{\9,\8\}^*$, and the action on the ray $\8^\infty$ is isomorphic to the action on the set of cosets $\langle x\rangle\backslash D$, since the stabilizer in $D$ of $\8^\infty$ is $\langle x\rangle$. We abbreviate $\8^\infty=:\overline\8$. The Schreier graph of the latter is a half-infinite line $$\begin{tikzpicture}
\footnotesize
\path (0,0) edge[loop left] node {$x$} ();
\foreach\i/\l in {0/\overline\8,1/\overline\8\9,2/\overline\8\9\9,3/\overline\8\9\8,4/\overline\8\9\9\8,5/\overline\8\9\9\9,6/\overline\8\9\8\9,7/\overline\8\9\8\8,8/\overline\8\9\9\8\8,9/\overline\8\9\9\8\9,10/\overline\8\9\9\9\9} \fill (\i,0) circle [radius=2pt] node[rotate=-45,right=1mm] {$\l$};
\foreach\i in {0,2,...,8} \path (\i,0) edge node[above] {$a$} (\i+1,0);
\foreach\i in {1,3,...,9} \path (\i,0) edge node[above] {$x$} (\i+1,0);
\path (10,0) edge[densely dotted] (10.5,0);
\end{tikzpicture}$$ Since the action of generators of $D$ change only a single symbol on sequences in $\{\9,\8\}^\infty$, the identification of the Schreier graph’s vertices with $\N$ is explicit: it is the “Gray code” [@gray:pulse] enumeration starting from the left-infinite word $\overline\8$. Thus the sequence $\dots\8\8 x_i\dots x_1$ is identified with the integer $\sum_{j=1}^i(1-x_j)2^{j-1}$, reading the number in base $2$ with $0$’s and $1$’s switched.
Now, to obtain the Schreier graph of $G_\omega$, one replaces each ‘$x$’ edge by a pair of edges labeled by two letters out of $\{b,c,d\}$, and puts loops at the extremities of the edge labeled by the remaining letter. The choice of which letter becomes a loop is determined by the position of the edge on the graph and the sequence $\omega$.
For example, here is the Schreier graph of the action of the first Grigorchuk group $G_{012}=\langle a,b,c,d\rangle$ on $\8^\infty G$: $$\label{eq:grigorbit}
\begin{tikzpicture}[xscale=0.68,baseline]
\footnotesize
\path (0,0) edge[loop above] node {$b$} ();
\path (0,0) edge[loop left] node {$c$} ();
\path (0,0) edge[loop below] node {$d$} ();
\foreach\i in {0,2,...,14} \path (\i,0) edge node[above] {$a$} (\i+1,0);
\foreach\i/\g in {1/b,3/d,5/b,7/c,9/b,11/d,13/b,15/b} \path (\i,0) edge[bend left] node[above] {$\g$} (\i+1,0);
\foreach\i/\g in {1/c,3/b,5/c,7/d,9/c,11/b,13/c,15/c} \path (\i,0) edge[bend right] node[below] {$\g$} (\i+1,0);
\foreach\i/\g in {1/d,2/d,3/c,4/c,5/d,6/d,7/b,8/b,9/d,10/d,11/c,12/c,13/d,14/d,15/d,16/d} \path (\i,0) edge[loop above] node[above] {$\g$} ();
\path (16,0) edge[densely dotted] (16.5,0);
\end{tikzpicture}$$
Growth estimates for self-similar groups {#ss:ssgrowth}
========================================
One of the purposes of this section is to reprove the following result. Let $\Omega'$ denote the subset of $\Omega$ consisting of sequences containing infinitely many of each of the symbols $0,1,2$.
\[thm:grigorchuk\] If $\omega\in\Omega'$, namely if $\omega$ contains infinitely many of each of the symbols $0,1,2$, then $G_\omega$ has intermediate word growth.
We shall in fact prove much more, in preparation for the construction, in §\[ss:perm\], of groups with prescribed growth. We mainly follow [@bartholdi-erschler:givengrowth].
A lower bound via algebras {#ss:algebralb}
--------------------------
We begin by a general lower bound on growth, coming from the theory of Hopf algebras. Recall that the lower central series of a group $G$ is defined by $\gamma_1(G)=G$ and $\gamma_{n+1}(G)=[\gamma_n(G),G]$ for all $n\ge1$.
\[thm:growth:hp\] Let $G$ be a finitely generated group, and assume that there is a subgroup $H<G$ such that $\gamma_n(H)\neq\gamma_{n+1}(H)$ for all $n\in\N$. Then $G$’s growth function satisfies $$\gamma_G\succsim \exp(\sqrt R).$$
In particular, if $G$ is residually virtually nilpotent, then either $G$ is virtually nilpotent (in which case $\gamma_G$ is polynomial) or $\gamma_G\succsim \exp(\sqrt R)$.
Before embarking on the proof, let us set up some algebraic notions. Let $\mathbb K$ be a field, and let $G$ be a group. The group ring $\mathscr A=\mathbb K G$ is the $\mathbb K$-vector space with basis $G$, and multiplication extended linearly. It is a *Hopf algebra*: it admits a *coproduct*, which is an algebra homomorphism $\Delta\colon\mathscr A\to\mathscr A\otimes\mathscr A$ defined on the basis $G$ by $g\mapsto g\otimes g$, a *counit*, which is an algebra homomorphism $\varepsilon\colon\mathscr A\twoheadrightarrow\mathbb K$ defined on the basis by $g\mapsto1$; and an *antipode*, which is an antihomomorphism $\sigma\colon\mathscr A\to\mathscr A$ defined on the basis by $g\mapsto g^{-1}$. Various axioms are satisfied, in particular the coproduct is *coassociative*: $(1\otimes\Delta)\circ\Delta=(\Delta\otimes1)\circ\Delta\colon\mathscr
A\to(\mathscr A)^{\otimes3}$, and *cocommutative*: $\Delta=\tau\circ\Delta$, for $\tau\colon\mathscr A^{\otimes2}\to\mathscr A^{\otimes2}$ the map $x\otimes y\mapsto y\otimes x$ flipping both factors. See [@sweedler:ha] for details.
Denote by $\varpi$ the kernel of $\varepsilon$, called the *augmentation ideal*. The *associated graded* of $\mathscr
A$ is the vector space $$\overline{\mathscr A}=\bigoplus_{n\ge0}\varpi^n/\varpi^{n+1}.$$
The associated graded $\overline{\mathscr A}$ is a graded, cocommutative Hopf algebra.
We have $\Delta(\varpi)\le\mathscr A\otimes\varpi+\varpi\otimes\mathscr A$; so $\Delta(\varpi^n)\le\sum_{i=0}^n\varpi^i\otimes\varpi^{n-i}$. Now given $\overline x\in\varpi^n/\varpi^{n+1}$, choose $x\in\varpi^n$ representing it; write $\Delta(x)=\sum\sum_{i=0}^n y_i\otimes z_{n-i}$ with $y_i,z_i\in\varpi^i$; and set $\Delta(\overline x)=\sum\sum_{i=0}^n
\overline{y_i}\otimes\overline{z_{n-i}}$ where $\overline{y_i},\overline{z_i}\in\varpi^i/\varpi^{i+1}$ are the images of their respective representatives. It is easy to show that this definition does not depend on the choices of $x,y_i,z_i$; and it defines a coassociative and cocommutative coproduct since $\mathscr A$’s coproduct was already coassociative and cocommutative.
Let $H$ be a Hopf algebra. An element $x\in H$ is called *primitive* if $\Delta(x)=x\otimes 1+1\otimes x$. The set of primitive elements in a Hopf algebra forms a Lie subalgebra of $H$ for the usual bracket $[x,y]=x y-y x$. Conversely, if $L$ is a Lie algebra, then its universal enveloping algebra is a Hopf algebra whose primitive elements are $L$. Note that, in characteristic $p$, one should consider *restricted* Lie algebras.
\[prop:mm\] Let $H$ be a cocommutative, primitively generated, graded Hopf algebra. Then it is the universal enveloping algebra of its primitive elements.
Let $P$ denote the Lie algebra of primitive elements in $H$. By the universal property of $U(P)$, there is a map $f:U(P)\to H$ which is graded, and surjective because $P$ generates $H$. We show that $f$ is also injective. Consider a homogeneous element $x\in U(P)$, say of degree $n$. If $n=0$, then $x\in\ker(f)$ if and only if $x=0$. Assume then that $f$ is injective on elements of degree $<n$.
We have $\Delta(x)=1\otimes x+x\otimes 1+y$ for some $y\in
U(P)_{<n}\otimes U(P)_{<n}$. If $f(x)=0$, then $f(y)=0$; but we had assumed $f$ to be injective on elements of degree $<n$, so $y=0$ and $x\in P$. By assumption, $f$ is injective on $P$, so $x=0$ and therefore $f$ is injective on elements of degree $n$ as well.
We apply these considerations to $\mathscr A=\mathbb K G$. First, we identify the primitive elements in $\overline{\mathbb K G}$. Let us define the series $\gamma^{\mathbb K}_n(G)=\{g\in G\mid
g-1\in\varpi^n\}$ of normal subgroups of $G$.
The space of primitive elements in $\overline{\mathbb K G}$ is $$\label{eq:liealg}
\mathscr L^{\mathbb K}(G)=\bigoplus_{n\ge1}(\gamma^{\mathbb K}_n(G)/\gamma^{\mathbb K}_{n+1}(G))\otimes\mathbb K.$$
The natural map $g\mapsto g-1$ from $\gamma^{\mathbb K}_n(G)$ to $\varpi^n/\varpi^{n+1}$ extends to a Lie algebra isomorphism from $\gamma^{\mathbb K}_n(G)$ onto primitive elements of $\overline{\mathbb K G}$.
In fact, $\gamma^{\mathbb K}_n(G)$ only depends on the characteristic $p$ of $\mathbb K$ and, up to extension of scalars, $\mathscr L^{\mathbb K}(G)$ depends only on $p$. Furthermore, it may be identified directly within $G$, and is a variant of the lower central series [@jennings:gpring; @jennings:gpringnilp]. Indeed one has $\gamma^p_1(G)=G$, and if $p=0$ then $$\gamma_{n+1}^0(G)=\langle g\in G\mid g^t\in[G,\gamma_n^0(G)]\text{ for some }t>0\rangle,$$ while if $p>0$ then $$\gamma_{n+1}^p(G)=\langle[G,\gamma_n^p(G)]\{x^p\mid x\in\gamma_{\lceil n/p\rceil}(G)\}.$$
\[prop:alg-gp\] Let $S$ be such that $S\cup S^{-1}$ generates $G$, and let $v_{G,S}$ denote the corresponding growth function. Let $w(n)=\dim_{\mathbb K}\varpi^n/\varpi^{n+1}$ denote the growth function of $\overline{\mathbb K G}$. Then $$v_{G,S}(n)\ge w(0)+w(1)+\dots+w(n)\text{ for all }n\in\N.$$
Consider first an element $x=1-g\in\varpi$, and write $g=s_1^{\epsilon_1}\cdots s_\ell^{\epsilon_\ell}$ as a product of generators and inverses. Using the identities $$\begin{aligned}
1-g h &=(1-g)+(1-h)-(1-g)(1-h),\\
1-g^{-1}&=-(1-g)+(1-g)(1-g^{-1}),
\end{aligned}$$ we get $x\equiv\sum_{i=1}^\ell\epsilon_i(1-s_i)$ modulo $\varpi^2$.
The ideal $\varpi^n$ is generated, qua ideal, by all $x=(1-g_1)\cdots(1-g_n)$ with $g_i\in G$. By the above, $\varpi^n/\varpi^{n+1}$ is generated, again qua ideal, by all $x=(1-s_1)\cdots(1-s_n)$ with $s_i\in S$. Now $\varpi^n/\varpi^{n+1}$ has trivial multiplication, so the $(1-s_1)\cdots(1-s_n)$ also generate $\varpi^n/\varpi^{n+1}$ as a vector space.
This generating set is contained in the linear span of $B_{G,S}(n)\subseteq\mathbb K G$; so for $0\le i\le n$ we have $w(i)\le\dim(\varpi^i\cap\mathbb
KB_{G,S}(n))-\dim(\varpi^{i+1}\cap\mathbb KB_{G,S}(n))$ and the claim follows.
Consider a subgroup $H$ such that $\gamma_n(H)\neq\gamma_{n+1}(H)$ for all $n\in\N$. Without loss of generality, suppose $H$ is finitely generated. For $n\in\N$, let $\mathcal P(n)$ be the set of prime numbers $p$ such that $(\gamma_n(H)/\gamma_{n+1}(H))\otimes\mathbb F_p\neq0$. Each $\mathcal P(n)$ is non-empty because $\gamma_n(H)/\gamma_{n+1}(H)$ is a finitely generated abelian group, and $\mathcal P(n+1)\subseteq\mathcal P(n)$ because the commutator map $(\gamma_n(H)/\gamma_{n+1}(H))\times
H\to\gamma_{n+1}(H)/\gamma_{n+2}(H)$ is onto. There exists therefore a prime number $p$ such that $\gamma_n(H)/\gamma_{n+1}(H)\otimes\mathbb F_p\neq0$ for all $n\in\N$. In particular, $\gamma_n^p(H)\neq\gamma_{n+1}^p(H)$ for all $n\in\N$.
Let $\mathbb K$ be a field of characteristic $p$. It follows that $\overline{\mathbb K H}$ contains a primitive element $x_n$ of degree $n$ for all $n\in\N$, namely $x_n=g_n-1$ for some $g_n\in\gamma_n^p(H)\setminus\gamma_{n+1}^p(H)$; so $\overline{\mathbb K H}$ contains $\pi(n)$ linearly independent elements of degree $n$, where $\pi(n)$ denotes the partition function. Indeed to every partition $n=i_1+\dots+i_k$ with $i_1\le\cdots\le i_k$ associate the element $x_{i_1}\cdots x_{i_k}$; these elements are linearly independent by Proposition \[prop:mm\]. Therefore, the growth function of $\overline{\mathbb K H}$ is at least $\pi(n)$.
It now follows from Proposition \[prop:alg-gp\] that $v_{H,S}(R)\ge \pi(R)$ holds for any generating set $S$ of $H$. Classical results on partitions [@hardy-r:asymptotic] tell us that $\pi(n)\propto \exp(\sqrt n)$; so *a fortiori* $v_G(R)\succsim \exp(\sqrt R)$.
Note that, for Grigorchuk’s group $G=G_{012}$, the quotients $\gamma_n(G)/\gamma_{n+1}(G)$ have bounded rank, in fact $1$ or $2$, see [@rozhkov:lcs]; so that no improvement on the lower bound can be obtained using Proposition \[prop:alg-gp\].
Metrics on $G_\omega$
---------------------
We already saw that the groups $G_\omega$ are contracting, namely if $\phi(g)=\pair<g_0,g_1>\pi$ then $g_0$ and $g_1$ are shorter than $g$. We shall need a strengthening of this property: we assign norms $\|\cdot\|_\omega$ to the groups $G_\omega$ to obtain relations of the form $$\label{eq:supercontract}
\|g_0\|_{\sigma\omega}+\|g_1\|_{\sigma\omega}\le\frac2{\eta_\omega}\big(\|g\|_\omega+\|a\|_\omega\big)$$ with $\eta_\omega>2$ as large as possible.
We do this by assigning norms $\in\R_+$ to the generators $a,b,c,d$ of $G_\omega$, and extend $\|\cdot\|_\omega$ to $G_\omega$ by the triangular inequality: $$\|g\|_\omega=\min\{\|s_1\|+\cdots+\|s_n\|:g=s_1\cdots s_n,\;s_i\in S\}.$$ For this purpose, consider the open $2$-simplex $$\Delta=\{(\beta,\gamma,\delta)\in\R^3:\max\{\beta,\gamma,\delta\}<\frac12,\beta+\gamma+\delta=1\}.$$ Its extremal points are $(\frac12,\frac12,0)$ and its permutations. A choice of $p_\omega=(\beta,\gamma,\delta)\in\Delta$ defines the following norm on the generators of $G_\omega$: $$\|a\|_\omega=1-2\max\{\beta,\gamma,\delta\},\qquad\|b\|_\omega=\beta-\|a\|_\omega,\;\|c\|_\omega=\gamma-\|a\|_\omega,\;\|d\|_\omega=\delta-\|a\|_\omega.$$ In particular, note that the triangular inequality $\|x\|_\omega+\|y\|_\omega\le\|xy\|_\omega$ is sharp for $x,y\in\{b,c,d\}$. $$\begin{tikzpicture}
\coordinate (A0) at (9,4);
\coordinate (A1) at (3,-4);
\coordinate (A2) at (-3,4);
\coordinate (M01) at ($(A0)!0.5!(A1)$);
\coordinate (M02) at ($(A0)!0.5!(A2)$);
\coordinate (M12) at ($(A1)!0.5!(A2)$);
\coordinate (C012) at ($(A0)!0.6667!(M12)$);
\draw[thick] (M12) node[below left] {$(0,\frac12,\frac12)$} --
(M01) node[below right] {$(\frac12,\frac12,0)$} --
(M02) node[above] {$(\frac12,0,\frac12)$} -- cycle;
\node[text width=3.5cm] at (8,2) {The simplex $\Delta$, and a 3-cycle preserved by $\overline M_2\overline M_1\overline M_0$};
\draw[thin] (M01) -- (C012);
\draw[thin] (M02) -- (C012);
\draw[thin] (M12) -- (C012);
\node at (4.0,1.4) {$\overline M_0(\Delta)$};
\node at (2.8,0.4) {$\overline M_1(\Delta)$};
\node at (2.2,1.8) {$\overline M_2(\Delta)$};
\node at ($0.4053*(A0)+0.3285*(A2)+0.2662*(A1)$) {$\bullet$};
\node at ($0.4053*(A1)+0.3285*(A0)+0.2662*(A2)$) {$\bullet$};
\node at ($0.4053*(A2)+0.3285*(A1)+0.2662*(A0)$) {$\bullet$};
\end{tikzpicture}$$
We extend the family of groups $(G_\omega)_{\omega\in\{0,1,2\}^\N}$ to a family $(G_\omega)_{\omega\in\{0,1,2\}^\Z}$; namely, the parameter space is now $\Omega=\{0,1,2\}^\Z$ with the two-sided shift map $\sigma\colon\Omega\righttoleftarrow$. The group $G_\omega$ itself only depends on the restriction of $\omega$ to $\N$, but the norm on $G_\omega$ depends on the restriction of $\omega$ to $-\N$.
Analogously to before, we denote by $\Omega'$ the subset of $\Omega$ consisting of sequences that contain infinitely many $0,1,2$ in both directions. For $\omega\in\Omega'$, we construct the point $p_\omega\in\Delta$ as follows. Consider the matrices $$M_0=\begin{pmatrix}1 & 1 & 1\\ 0 & 2 & 0\\ 0 & 0 & 2\end{pmatrix},\qquad
M_1=\begin{pmatrix}2 & 0 & 0\\ 1 & 1 & 1\\ 0 & 0 & 2\end{pmatrix},\qquad
M_2=\begin{pmatrix}2 & 0 & 0\\ 0 & 2 & 0\\ 1 & 1 & 1\end{pmatrix}.$$ Define then $$\eta\colon\Delta\times\{0,1,2\}\to(2,3],\qquad\overline M_\lambda\colon\Delta\righttoleftarrow,\qquad\mu\colon\Delta\to(0,\tfrac13)$$ by setting, for all $\lambda\in\{0,1,2\}$ and all $p\in\Delta$, $$\begin{aligned}
\eta(p,\lambda)&=(1\;1\;1)\cdot M_\lambda(p)&&\text{the $\ell^1$-norm of }M_\lambda(p),\\
\overline M_\lambda(p)&=\frac{M_\lambda(p)}{\eta(p,\lambda)}&&\text{the projection of }M_\lambda(p)\text{ to }\Delta,\\
\mu(p)&=\min\{\beta,\gamma,\delta\}&&\text{the minimal distance of $p$ to a vertex of }\Delta.\end{aligned}$$
Endow $\Delta$ with the Hilbert metric $d_\Delta(V_1,V_2)=\log(V_1,V_2;V_-,V_+)$, computed using the cross-ratio of the points $V_1,V_2$ and the intersections $V_-,V_+$ of the line containing $V_1,V_2$ with the boundary of $\Delta$. The transformations $\overline M_\lambda$ are projective transformations of $\Delta$, and are therefore contracting:
\[lem:birkhoff\] Let $K$ be a convex subset of affine space, and let $A\colon
K\righttoleftarrow$ be a projective map. Then $A$ contracts the Hilbert metric.
If furthermore $A(K)$ contains no lines from $K$ (that is, $A(K)\cap
\ell\neq K\cap \ell$ for every line $\ell$ intersecting $K$), then $A$ is strictly contracting.
Since $A$ is projective, it preserves the cross-ratio on lines, so we have $d_{A(K)}(A(V_1),A(V_2))=d_K(V_1,V_2)$ for all $V_1,V_2\in K$. Furthermore, on the line $\ell$ through $V_1,V_2$, the intersection points $\ell\cap\partial A(K)$ are not further from $V_1,V_2$ than $\{V_+,V_-\}=\ell\cap\partial K$; the Hilbert metric decreases as $V_\pm$ are moved further apart from $V_1,V_2$, and this gives strict contraction under the condition $A(K)\cap \ell\neq K\cap \ell$.
We are ready to define the points $p_\omega\in\Delta$, and therefore the metrics $\|\cdot\|_\omega$. Choose an arbitrary point $p\in\Delta$, and set $$\label{eq:pomega}
p_\omega=\lim_{n\to\infty} \overline M_{\omega_{-1}}\circ\overline
M_{\omega_{-2}}\circ\cdots\circ \overline M_{\omega_{-n}}(p).$$ Note that, since the transformations $\overline M_\lambda$ are contracting, the limit $p_\omega$ is independent of the choice of $p$. Note also that, by our assumption that the negative part of $\omega$ contains infinitely many $0$, $1$ and $2$, the limit $p_\omega$ does not belong to the boundary of $\Delta$. From now on, we write $$\eta_\omega:=\eta(p_\omega,\omega_0).$$
\[lem:supercontract\] For all $g\in G_\omega$ with $\phi(g)=\pair<g_0,g_1>\pi$, we have the inequality $$\|g_0\|_{\sigma\omega}+\|g_1\|_{\sigma\omega}\le\frac2{\eta_\omega}\big(\|g\|_\omega+\|a\|_\omega\big);$$ and furthermore, if $g\not\in\{b,c,d\}$ then up to replacing $g$ with a conjugate we have $$\|g_0\|_{\sigma\omega}+\|g_1\|_{\sigma\omega}\le\frac2{\eta_\omega}\|g\|_\omega.$$
Without loss of generality, we suppose $\omega_0=0$, and write $p_\omega=(\beta,\gamma,\delta)$ and $p_{\sigma\omega}=\overline
M_{\omega_0}p_\omega=(\beta',\gamma',\delta')$. We have $$\eta_\omega=3-2\beta,\quad(\beta',\gamma',\delta')=\frac{(\beta+\gamma+\delta,2\gamma,2\delta)}{\eta_\omega}=\frac{(1,2\gamma,2\delta)}{\eta_\omega}.$$ Thus $$\begin{aligned}
\|b\|_{\sigma\omega}+\|\omega_0(a)\|_{\sigma\omega} &= \|b\|_{\sigma\omega} = \beta'-\|a\|_{\sigma\omega} = \beta'-(1-2\beta')=3\beta'-1\\
& =\frac3{\eta_\omega}-1=2\frac\beta{\eta_\omega}=\frac2{\eta_\omega}\|ab\|_\omega,\\
\|c\|_{\sigma\omega}+\|\omega_0(a)\|_{\sigma\omega} &= \frac{2\gamma}{\eta_\omega}=\frac2{\eta_\omega}\|ac\|_\omega,\\
\|d\|_{\sigma\omega}+\|\omega_0(a)\|_{\sigma\omega} &= \frac{2\delta}{\eta_\omega}=\frac2{\eta_\omega}\|ad\|_\omega.
\end{aligned}$$ Now given $g\in G_\omega$, write it as a word of minimal norm as $g=a^?x_1\cdots a x_\ell a^?$, with $x_i\in\{b,c,d\}$, and the $a^?$ mean that the initial and final ‘$a$’ may be present or absent. Thus $\|g\|_\omega\ge\|ax_1\|_\omega+\cdots+\|ax_\ell\|_\omega-\|a\|_\omega$. On the other hand, each ‘$a$’ in the expression of $g$ contributes nothing to $g_0$ and $g_1$, while each ‘$x_i$’ contributes an ‘$x_i$’ and a ‘$\omega_0(x_i)$’ to $g_0$ and $g_1$, in some order. Summing together the inequalities above gives the claimed .
The second claim follows, because the extra ‘$\|a\|_\omega$’ term occurs only if $g$ both starts and ends with a letter in $\{b,c,d\}$, and this case can be prevented by conjugating $g$ by its last letter.
Lemma \[lem:supercontract\] can be used to prove statements on $G_\omega$ by induction. For example,
\[prop:faithful,infinite\] If $\omega\in\Omega'$, then the action of $G_\omega$ on the tree $\mathscr T=X^*$ is faithful, and it is transitive on each orbit $X^i$. In particular, $G_\omega$ is infinite.
We first use induction on $\|\cdot\|_\omega$, simultaneously on all $\omega\in\Omega'$, to show that if $g\in G_\omega$ acts trivially on $\mathscr T$ then $g=1$.
The induction starts by noting that the generators $b,c,d$ act non-trivially by our assumption that $\omega$ contains infinitely many $0,1,2$. Indeed, without loss of generality consider $b$; let $k\in\N$ be minimal such that $\omega_k(b)\neq1$; then $b$ acts non-trivially on $X^{k+1}$.
Consider then $g\in G_\omega$ acting trivially on $\mathscr T$. In particular, $g$ fixes $X$, so $\phi_\omega(g)=\pair<g_0,g_1>$. By Lemma \[lem:supercontract\], both $g_0$ and $g_1$ are shorter, so by induction they are trivial; thus $g=1$ because $\phi_\omega$ is injective.
To check that $G_\omega$ acts transitively on $X^i$, it suffices to show that the stabilizer $H$ of $\8$ acts transitively on $X^{i-1}\8$; because then $X^{i-1}\8 G_\omega=X^{i-1}\8\langle
a\rangle=X^i$. Now $H$ contains $b,c,d,y^a$ for a letter $y\in\{b,c,d\}$ such that $\omega(y)=a$; and the action of $b,c,d,y^a$ on $w\8$ is $w b\8,w c\8,w d\8,w a\8$ respectively, so that the $H$-orbit of $w\8$ is $W\8$ for a $G_{\sigma\omega}$-orbit $W\subseteq X^{i-1}$. Again we are done by induction.
Note that the action of $G_\omega$ is still faithful if $\omega$ only contains infinitely many of two symbols; it is *not* faithful if $\omega$ contains finitely many of two symbols.
The $G_\omega$ are infinite torsion groups
------------------------------------------
One of Burnside’s questions [@burnside:question] asks whether there exist infinite, finitely generated groups in which every element has finite order. The first such examples were constructed by Golod [@golod:nil]; here, we show that the groups $G_\omega$ are other examples:
\[thm:torsion\] If $\omega\in\Omega'$, then $G_\omega$ is an infinite torsion $2$-group.
The group $G_\omega$ is infinite by Proposition \[prop:faithful,infinite\]. We prove the claim by induction on $\|g\|$. It is easy to check that the generators $a,b,c,d$ all have order $2$. Consider $g\in G_\omega$, with $\phi_\omega(g)=\pair<g_0,g_1>\pi$. If $\pi=()$, then $g_0,g_1$ are shorter than $g$ by Lemma \[lem:supercontract\] so have finite order, say $2^{n_0},2^{n_1}$ respectively. Then $g$ has order $2^{\max\{n_0,n_1\}}$. If $\pi=(\9,\8)$, then $g^2=\pair<g_0g_1,g_1g_0>$, and $g_0g_1$ is shorter than $g$ again by Lemma \[lem:supercontract\], so has finite order, say $2^n$. Then $g$ has order $2^{n+1}$.
Note that, if $\omega$ contains finitely many copies of a symbol, then $G_\omega$ is not a torsion group anymore. In fact, suppose that $\omega$ contains no $0$; then $\omega_i(b)=a$ for all $i$, and the element $ab$ has infinite order, since $\phi_\omega((ab)^{2n})=\pair<(b a)^n,(ab)^n>$ for all $n\in\Z$.
We also note that the groups $G_\omega$ resemble very much infinitely iterated wreath products; namely the map $\phi_\omega\colon
G_\omega\to G_{\sigma\omega}\wr{\operatorname{Sym}}(2)$ is almost an isomorphism:
Let $(G_\omega)_{\omega\in\Omega}$ be a similar sequence of groups. It is called *branched* if every $G_\omega$ has a finite-index subgroup $K_\omega$ such that $$K_{\sigma\omega}^{X_\omega}\le\phi_\omega(K_\omega).$$ If the subgroups $K_\omega$ are merely required to be non-trivial, then $(G_\omega)$ is called *weakly branched*.
Every group in a (weakly) branched family of groups is also called *(weakly) branched*.
\[prop:Gbranched\] The groups $G_\omega$ are branched for all $\omega\in\Omega'$.
For each $\omega\in\Omega'$, let $x_\omega\in\{b,c,d\}$ be such that $\omega(x_\omega)=1$, and set $K_\omega=\langle[x_\omega,a]\rangle^{G_\omega}$. Choose also $y_\omega\in\{b,c,d\}\setminus\{x_\omega\}$. Then $1\times
K_{\sigma\omega}$ is normally generated by $\pair<1,[x_{\sigma\omega},a]>$, and $$\pair<1,[x_{\sigma\omega},a]>=\phi_\omega([x_\omega,y_\omega^a])=\phi_\omega([x_\omega,a][x_\omega,a]^{a y_\omega})\in \phi_\omega(K_\omega);$$ the same computation holds for $K_{\sigma\omega}\times1$.
We now show that the groups $K_\omega$ have finite index. Consider first the quotient $G_\omega/\langle
x_\omega\rangle^{G_\omega}$. This group is generated by two involutions $a$ and $y_\omega$, so is a finite dihedral group, because $G_\omega$ is torsion. It follows that $\langle
x_\omega\rangle^{G_\omega}$ has finite index in $G_\omega$. Then $\langle x_\omega\rangle^{G_\omega}/K_\omega=\langle x_\omega\rangle
K_\omega$ has order $2$, so $K_\omega$ also has finite index.
\[prop:branched=>fullwreath\] If $G$ is a $p$-torsion weakly branched group, then it contains $L=\bigcup\wr^i C_p$ as a subgroup.
Let $(G_\omega)_{\omega\in\Omega}$ be a similar family of groups, with $G=G_\omega$, and let $K_\omega\le G_\omega$ be the subgroups given by the condition that $(G_\omega)_{\omega\in\Omega}$ is branched. For each $\omega\in\Omega$, let $g_\omega\in K_\omega$ be an element of order $p$, and let $n(\omega)\in\N$ be such that $g_\omega$ acts non-trivially on $X_{\sigma^{n(\omega)-1}\omega}\times\cdots\times X_\omega$; let $v_\omega\in X_{\sigma^{n(\omega)-1}\omega}\times\cdots\times
X_\omega$ be a point on a non-trivial orbit.
Define then simultaneously and recursively $L_\omega=\langle
L_{\sigma^{n(\omega)}\omega}@v_\omega, g_\omega\rangle$. It contains an element $g_\omega$ permuting $p$ copies of $L_{\sigma^{n(\omega)}\omega}$, so is isomorphic to $L$.
If $G$ contains no torsion, or torsion of different primes, analogous (but harder-to-state) results hold. In particular, by Exercise \[exse:nolaw\], branched groups satisfy no law. This recovers a result by Abért [@abert:nonfree].
Lower growth estimates for $G_\omega$
-------------------------------------
The proof of Theorem \[thm:grigorchuk\] requires upper and lower bounds on the growth function of $G_\omega$. A lower bound is easily provided by Theorem \[thm:growth:hp\]: the group $G_\omega$ is infinite (Proposition \[prop:faithful,infinite\]) and torsion (Theorem \[thm:torsion\]), so cannot be virtually nilpotent, since nilpotent groups with finite-order generators are finite. However, in some cases a direct argument also gives the bound $v_{G_\omega}(R)\succsim\exp(\sqrt R)$.
Define indeed maps $\tilde\theta_\omega\colon F\to F$ by $$\label{eq:theta}
\begin{cases}
\tilde\theta_\omega(a) &=a y a\text{ if }\omega_0=\cdots=\omega_{k-1}\neq\omega_k,\text{ and }\omega_0(y)=a,\;\omega_k(y)=1,\\
\tilde\theta_\omega(x)&=x\text{ for all }x\in\{b,c,d\}.
\end{cases}$$ A direct calculation shows that $\tilde\theta_\omega$ induces a map $\theta_\omega\colon G_{\sigma\omega}\to G_\omega$, with $$\phi_\omega(\theta_\omega(g))=\tikz[baseline=-2ex]{\draw[->] (0,0) --
node[right=-2pt] {\small$*$} (0,-0.5); \draw[->](0.5,0) --
node[right=-2pt] {\small$g$} (0.5,-0.5);}$$ for some $*\in\langle a,y\rangle$ using the notation introduced in . Furthermore, $\langle a,y\rangle$ is a dihedral group of order $2^{k+2}$.
Prove that $\theta_\omega$ is a group homomorphism.
Let us now **cheat**, and assume that the element $*$ is always trivial, rather than an element of a finite group of order $2^{k+2}$. If $k$ is bounded, this is unimportant; however, if $k$ is unbounded then an additional argument is really required.
Denote by $B_\omega(R)$ the ball of radius $R$ in $G_\omega$, and abbreviate $v_\omega(R)=v_{G_\omega}(R)$. Consider the map $G_{\sigma\omega}^2\to G_\omega$ given by $(g_0,g_1)\mapsto\theta_\omega(g_0)^a\cdot\theta_\omega(g_1)$. For $R_0,R_1$ even, it defines an injective map $B_{\sigma\omega}(R_0)\times B_{\sigma\omega}(R_1)\to
B_\omega(2(R_0+R_1))$, hence $v_{\sigma\omega}(R)^2\le v_\omega(4R)$ for all $R$ even. We conclude $v_\omega(2\cdot4^k)\ge v_{\sigma^k\omega}(2)^{2^k}\ge5^{2^k}$, so $v_\omega(R)\ge \exp(\frac12\log5\sqrt R)$.
Upper growth estimates for $G_\omega$
-------------------------------------
We are ready to give an upper bound on the growth of $G_\omega$. As before, we abbreviate $v_\omega(R)=v_{G_\omega}(R)$ for the growth function of $G_\omega$ We start by a
\[lem:gamma(Amu)\] For every $A>0$ there exists $B\in\N$ such that, for all $\omega\in\Omega'$, we have $v_\omega(A\mu_\omega)\le B$.
It suffices to bound the number of elements of $G_\omega$ whose minimal expression has length $\le A\mu_\omega$ and has the form $g=a x_1\cdots a x_\ell$ for some $x_i\in\{b,c,d\}$, since there are at most $8$ times more elements of norm $\le A\mu_\omega$, namely all $\{1,b,c,d\}g\{1,a\}$. Now by definition $g$ has norm at least $\ell\mu_\omega$, and there are at most $3^\ell$ such $g$’s, so one may take $B=8\cdot3^A$.
\[lem:etamu\] For all $\omega\in\Omega'$ we have $\eta_\omega\mu_{\sigma\omega}\ge\mu_\omega+\|a\|_\omega$.
Write $p_\omega=(\beta,\gamma,\delta)$, and assume without loss of generality $\omega_0=0$ so $\mu_{\sigma\omega}=2/\eta_\omega\,\min\{\gamma,\delta\}$. Consider the six orderings $\beta<\gamma<\delta$ etc. in turn to check the inequalities $$\eta_\omega\mu_{\sigma\omega}=2\min\{\gamma,\delta\}\ge\min\{\beta,\gamma,\delta\}+(1-\max\{\beta,\gamma,\delta\})=\mu_\omega+\|a\|_\omega.\qedhere$$
\[lem:concavehull\] Let $f$ be a positive sublinear function, namely $f(n)/n\to 0$ as $n\to\infty$. Then $f$ is bounded from above by a concave sublinear function.
For every $\theta\in(0,1)$, let $n_\theta$ be such that $f(n)-\theta
n$ is maximal. Given $n\in\R$, let $\zeta<\theta$ be such that $n\in[n_\theta,n_\zeta]$ with maximal $\zeta$ and minimal $\theta$, and define $\overline f(n)$ on $[n_\theta,n_\zeta]$ by linear interpolation between $(n_\theta,f(n_\theta))$ and $(n_\zeta,f(n_\zeta))$. Clearly $\overline f\ge f$, and $\overline
f(n)/n$ is decreasing and coincides infinitely often with $f(n)/n$, so it converges to $0$.
\[prop:upperG\] There is an absolute constant $B$ such that, for all $\omega\in\Omega'$ and all $k\in\N$, $$v_\omega(\eta_\omega\cdots\eta_{\sigma^{k-1}\omega}\mu_{\sigma^k\omega})\le B^{2^k}.$$
For all $i\in\N$, set $\alpha_i(R)=18(R+2)v_{\sigma^i\omega}(R+\mu_{\sigma^i\omega})$. We **cheat**, in assuming that the functions $\alpha_i$ are log-concave, i.e. satisfy $\alpha_i(R_0)\alpha_i(R_1)\le\alpha_i((R_0+R_1)/2)^2$. This assumption is in fact harmless, since each function $\alpha_i$ can be replaced by its log-concave majorand: the smallest log-concave function that is pointwise larger than $\alpha_i$, given by Lemma \[lem:concavehull\]. For details, see [@bartholdi-erschler:givengrowth]\*[§3.1]{}.
The proposition will follow from the inequalities $\alpha_i(R)\le
\alpha_{i+1}(R/\eta_{\sigma^i\omega})^2$ for all $i,R$; because then $$\begin{aligned}
v_\omega(\eta_\omega\cdots\eta_{\sigma^{k-1}\omega}\mu_{\sigma^k\omega}) &\le \alpha_0(\eta_\omega\cdots\eta_{\sigma^{k-1}\omega}\mu_{\sigma^k\omega})\\
&\le\alpha_1(\eta_{\sigma\omega}\cdots\eta_{\sigma^{k-1}\omega}\mu_{\sigma^k\omega})^2\\
\le\cdots &\le\alpha_k(\mu_k)^{2^k}\le B^{2^k}\text{ by Lemma~\ref{lem:gamma(Amu)}}.
\end{aligned}$$
To simplify notation, we consider only the case $i=0$, since all cases are the same. We have $$\begin{aligned}
\alpha_0(R) = 18(R+2)v_\omega(R+\mu_\omega) &\le
18(R+2)\hspace{-2em minus 1fil}\sum_{R_0+R_1\le \tfrac2{\eta_\omega}
(R+\mu_\omega+\|a\|_\omega)}\hspace{-2em minus 1fil}2v_{\sigma\omega}(R_0)v_{\sigma\omega}(R_1)\\
&\le 36(R+2)^2 \hspace{-2em minus 1fil}\max_{R_0+R_1\le \tfrac2{\eta_\omega}
(R+\mu_\omega+\|a\|_\omega)}\hspace{-2em minus 1fil}v_{\sigma\omega}(R_0)v_{\sigma\omega}(R_1)\\
&\le
6^2(3/\eta_\omega)^2(R+2)^2v_{\sigma\omega}((R+\mu_\omega+\|a\|_\omega)/\eta_\omega)^2\\
&\le \big(18(R+2)/\eta_\omega\,v_{\sigma\omega}(R/\eta_\omega+\mu_{\sigma\omega})\big)^2\\
&\le\alpha_1(R/\eta_\omega)^2.\qedhere
\end{aligned}$$
We are now ready to conclude the proof of Theorem \[thm:grigorchuk\], by showing that the groups $G_\omega$ have subexponential growth. There are in fact different methods for this. Let $\lambda_\omega$ be the exponential growth rate of $G_\omega$, as in ; we are to show $\lambda_\omega=1$.
Since by assumption $\omega$ contains infinitely many $0,1,2$, there are infinitely many positions $k\in\N$ with $\omega_k=0$ and that are separated by $1$ and $2$ in $\omega$. For these $k$, the point $p_{\sigma^k\omega}\in\Delta$ belongs to the subsimplex $\{\beta<\gamma\wedge\beta<\delta\}$. Thus $\beta<\tfrac13$ on that subsimplex, $\eta_{\sigma^k\omega}=3-2\beta>7/3$ is uniformly bounded away from $2$, and $\gamma,\delta>\tfrac16$ so $\mu_{\sigma^{k+1}\omega}>\frac1{18}$. Thus by Proposition \[prop:upperG\] $$\log\lambda_\omega=\lim\frac{\log v_\omega(R)}{R}\le\liminf\frac{2^k\log B}{\eta_\omega\cdots\eta_{\sigma^{k-1}\omega}\mu_{\sigma^k\omega}}=0$$ since on a subsequence the $\mu_{\sigma^k\omega}$ are bounded away from $0$, all terms $2/\eta_{\sigma^i\omega}$ are bounded by $1$, and infinitely many of them are bounded by $6/7$.
A more “abstract” proof may be obtained by noting that the map $\omega\mapsto\lambda_\omega$ is continuous and bounded by $3$, and that the proof of Proposition \[prop:upperG\] gives $\log\lambda_\omega\le2/\eta_\omega\log\lambda_{\sigma\omega}$. Since the action of $\sigma$ on $\Omega'$ is ergodic, we must have $\log\lambda_\omega=0$ for all $\omega\in\Omega'$.
Let us compute more precisely an upper bound for the growth of the first Grigorchuk group $G_{(012)^\infty}$. Since the sequence $\omega=(012)^\infty$ is $3$-periodic, we can find $p_\omega\in\Delta$ explicitly. The calculation is made even simpler by noting that $p_{\sigma\omega}$ and $p_{\sigma^2\omega}$ are cyclic permutations of $p_\omega$; thus $p_\omega$ is the normalised eigenvector of $M_0\cdot(\begin{smallmatrix}0&0&1\\1&0&0\\0&1&0\end{smallmatrix})$, and its spectral radius is $\eta_+\approx2.46$, the positive root of the characteristic polynomial $T^3-T^2-2T-4$. Thus for the first Grigorchuk group we get $v_\omega(\eta_+^k\mu_\omega)\le
B^{2^k}$ for all $k$, and therefore $$v_\omega(R)\precsim\exp(R^{\log 2/\log\eta_+})\approx\exp(R^{0.76}).$$
Growth of permutational wreath products {#ss:perm}
=======================================
The upper and lower bounds on the growth of $G_\omega$ are both of intermediate type $\exp(R^\alpha)$, but do not match. We consider, in this section, permutational wreath products based on the groups $G_\omega$.
Choose a sequence $\omega\in\Omega'$ and a ray $\xi\in\{\9,\8\}^\infty$, and consider the ray’s orbit $X=\xi
G_\omega$. Choose a group $H$. Set then $$W_\omega(H):=H\wr_X G_\omega.$$ (Even though the notation does not make it clear, the group $W_\omega(H)$ depends on $\xi$.) We shall show, in this section:
\[thm:wreathgrowth\] If $H$ has subexponential growth, then so does $W_\omega(H)$.
\[thm:givengrowth\] Let $\eta_+\approx2.46$ be the positive root of $T^3-T^2-2T-4$. Let $f\colon\R_+\to\R_+$ be a function satisfying $$\label{eq:gdoubling}
f(2R)\le f(R)^2 \le f(\eta_+R)\text{ for all $R$ large enough}.$$ Then there exists $\omega\in\Omega'$ such that $v_{W_\omega(C_2)}\sim f$.
We shall give more illustrations of the growth functions that may occur in §\[ss:illustrations\]. We content ourselves with the following:
\[thm:grigWgrowth\] For any finite group $H$, the group $W_{012}(H)$ has growth $$v_{W_{012}(H)}\sim\exp(R^{\log2/\log\eta_+}).$$
The proofs of Theorems \[thm:wreathgrowth\] and \[thm:givengrowth\] rely on estimates of the *support* $\subset X$ of an element of $W_\omega(H)$ of norm $\le R$. Recall that every element of $W_\omega(H)$ may be written in the form $cg$ with $c\colon X\to H$ and $g\in G_\omega$; its support is $\{x\in X:c(x)\neq1\}$. To better understand the support of elements of $W_\omega(H)$, let us introduce the following
\[def:invorbit\] Let $G$ be a group acting on the right on a set $X$ with basepoint $\xi$. For a word $w=w_1\dots w_\ell\in G^*$, its *inverted orbit* is the set $$\mathcal O(w)=\{\xi w_{i+1}\cdots w_\ell:0\le i\le \ell\}.$$ If furthermore $G$ is given with a metric $\|\cdot\|$, then its *inverted orbit growth* is the function $\Delta\colon\R_+\to\N$ given by $$\Delta(R)=\max\{\#\mathcal O(w):\|w\|\le R\}.\qedhere$$
We write $\mathcal O_\omega(w)$ and $\Delta_\omega(R)$ in the case of $G=G_\omega$ with its metric $\|\cdot\|_\omega$. Thus, for example, taking $\xi=\8^\infty$, the inverted orbit of $acadab$ is $$\mathcal O(acadab)=\{\xi acadab,\xi cadab,\xi adab,\xi dab,\xi ab,\xi
b,\xi\}=\{\overline\8\9\8\9,\overline\8,\overline\8\9\9\},$$ see the Schreier graph at the end of §\[ss:ss\].
Note that a basepoint $\xi$ is implicit in the definitions; yet,
\[exse:invgrowth\] Assume that $G$ acts transitively on $X$. Show that $\Delta(R)$ depends only mildly of the choice of $\xi$, in the following sense: if $\xi,\xi'\in X$ are two choices of a basepoint and $\Delta(R),\Delta'(R)$ are the corresponding inverted orbit growth functions, then there exists a constant $C\in\R$ such that $\Delta(R)\le\Delta'(R+C)$ and $\Delta'(R)\le\Delta(R+C)$ for all $R$.
\[prop:invgrowth\] There exists a universal constant $C$ such that, for all $\omega\in\Omega'$ and all $k\in\N$, $$2^k\le\Delta_\omega(\eta_\omega\cdots\eta_{\sigma^{k-1}\omega}\mu_{\sigma^k\omega})\le C\,2^k.$$
For the upper bound, we note that Lemma \[lem:supercontract\] applies just as well to the group $G_\omega$ as to the monoid $$R:=S^*/\{b^2=c^2=d^2=1,b c=d,c d=b,db=c\},$$ see Equation . Indeed, in a minimal-length representative of an element of $R$, the number of ‘$a$’ is at least the number of $b,c,d$-letters minus one, and this is the only property required for Lemma \[lem:supercontract\]. Now given a word $w\in S^*$, its inverted orbit may be read from the image of $w$ in $R$. Every element of $R$ has a unique *reduced form*: the reduced form of a word $w\in S^*$ is the word $\overline w\in S^*$ obtained by replacing every subword equal to a left-hand side of a relation by the corresponding right-hand side.
Without loss of generality and merely at the cost of increasing the constant $C$, we may suppose $\xi=\8^\infty$. We claim that the inverted orbit of a word $w\in S^*$ coincides with the inverted orbit of its reduction $\overline w$. To see this, consider $w=w_1\dots w_\ell\in S^*$, a subword $w_j w_{j+1}$ equal to a left-hand side of a relation, and the word $w'$ obtained by replacing $w_j w_{j+1}$ by the right-hand side of the relation. All terms $\xi w_{i+1}\dots w_\ell$ with $i\neq j$ clearly appear both in $\mathcal O(w)$ and $\mathcal O(w')$. For the remaining term in $\mathcal O(w)$, we have $\xi w_{i+1}\dots w_\ell=\xi w_{i+2}\dots w_\ell$ because $w_{i+1}$ fixes $\xi$, so this term also belongs to $\mathcal O(w')$.
If $w\in F$ satisfies $\tilde\phi_\omega(w)=\pair<w_0,w_1>\pi$ and $\xi=\xi'\9$, then $$\mathcal O_\omega(w)\subseteq\mathcal
O_{\sigma\omega}(w_0)\9\sqcup\mathcal O_{\sigma\omega}(w_1)\8,$$ where the inverted orbits $\mathcal O_{\sigma\omega}$ are computed with respect to the basepoint $\xi'$; and similarly if $\xi=\xi'\8$. We therefore get $$\Delta_\omega(R)\le\max_{R_0+R_1\le2/\eta_\omega(R+\|a\|_\omega)}\big(\Delta_{\sigma\omega}(R_0)+\Delta_{\sigma\omega}(R_1)\big).$$ The same argument as in Proposition \[prop:upperG\] finishes the proof of the upper bound.
For the lower bound, it suffices to exhibit for all $k\in\N$ a word of length at most $\eta_\omega\cdots\eta_{\sigma^{k-1}\omega}\mu_{\sigma^k\omega}$ and inverted orbit of size at least $2^k$. For that purpose, define self-substitutions $\zeta_x$ of $\{ab,ac,ad\}^*$, for $x\in\{0,1,2\}$, by $$\begin{array}{rlll}
\zeta_0:&ab\mapsto adabac,&ac\mapsto acac,&ad\mapsto adad,\\
\zeta_1:&ab\mapsto abab,&ac\mapsto abacad,&ad\mapsto adad,\\
\zeta_2:&ab\mapsto abab,&ac\mapsto acac,&ad\mapsto acadab,
\end{array}$$ and note that for any word $w\in\{ab,ac,ad\}^*$ representing an element of $F$ we have $$\tilde\phi_\omega(\zeta_{\omega_0}(w))=\begin{cases}
\tikz[baseline=1.5mm]{\draw[->] (0,0) -- node[above=-1pt] {\small$w$} +(1.0,0); \draw[->] (0,0.5) -- node[above=-1pt] {\small$a w a$} +(1.0,0);} & \text{ if $\zeta_{\omega_0}(w)$ contains an even number of `$a$',}\\
\tikz[baseline=1.5mm]{\draw[->] (0,0) -- node[near start,below] {\small$w$} node[near end,above=-1pt] {\small$a$} +(1.0,0.5); \draw[->] (0,0.5) -- node[near start,above=-1pt] {\small$a$} node[near end,below] {\small$w$} +(1.0,-0.5);} & \text{ if $\zeta_{\omega_0}(w)$ contains an odd number of `$a$'}.
\end{cases}$$ In particular, $\zeta_{\omega_0}$ induces a homomorphism $G_{\sigma\omega}\to G_\omega$.
By induction, we see that for any non-trivial $w\in\{ab,ac,ad\}^*$ (representing an element of $G_{\sigma^k\omega}$) we have $$\Delta_\omega(\zeta_{\omega_{0}}\cdots\zeta_{\omega_{k-1}}(w))\ge2^k.$$
Note then that, if $Z\in\N^3$ count the numbers of $ab,ac,ad$ respectively in $w$, then $M_x^{\mathrm t}Z$ counts the numbers of $ab,ac,ad$ respectively in $\zeta_x(w)$. Indeed without loss of generality consider $x=0$; then every $ab$ in $w$ contributes one each of $ab,ac,ad$ to $\zeta_x(w)$, while every $ac$ and $ad$ in $w$ contributes two copies of itself to $\zeta_x(w)$.
Let $as\in\{ab,ac,ad\}$ be such that $\|as\|_{\sigma^k\omega}$ is minimal — recall the notation $p_\omega$ from : if $p_{\sigma^k\omega}=(\beta,\gamma,\delta)$ and $\beta\le\gamma,\delta$ then $s=b$, etc. Let $W$ be the basis vector in $\R^3$ with a ‘$1$’ at the position which $as$ has in $\{ab,ac,ad\}$: if $\beta\le\gamma,\delta$ then $W=(1,0,0)^{\mathrm
t}$, etc. Set $w=\zeta_{\omega_{0}}\cdots\zeta_{\omega_{k-1}}(as)$. We have $\Delta_\omega(w)\ge2^k$, and $$\begin{aligned}
\|w\|_\omega&= Z^t p_\omega = W^{\mathrm t}M_{\omega_{k-1}}\cdots M_{\omega_{0}}p_\omega\\
&= \eta_\omega\cdots\eta_{\sigma^{k-1}\omega}W^{\mathrm t}p_{\sigma^k\omega} = \eta_\omega\cdots\eta_{\sigma^{k-1}\omega}\mu_{\sigma^k\omega}.\qedhere
\end{aligned}$$
Finally, let us introduce the “choice of inverted orbits growth” function, first generally for a group $G$, with given metric $\|\cdot\|$, acting on a set $X$ with basepoint $\xi$: $$\Sigma(R):=\#\{\mathcal O(w):\|w\|\le R\}.$$ This function counts the number of subsets that may occur as inverted orbit of a word of length at most $R$. Since $\mathcal O(w)$ is a subset of cardinality at most $R+1$ of the Schreier graph $X$, and furthermore lies in the ball of radius $R$ about $\xi$ in $X$, we get the crude estimate $\Sigma(R)\le\binom{v_{X,\xi}(R)+R}{R}$ based on the growth function $v_{X,\xi}$ of balls centered at $\xi$ in the graph $X$. However, in the particular case of the groups $G_\omega$, we can do better:
\[prop:invchoices\] There is an absolute constant $D$ such that for all $\omega\in\Omega'$ and all $k\in\N$ we have $$\Sigma_\omega(\eta_\omega\cdots\eta_{\sigma^{k-1}\omega}\mu_{\sigma^k\omega})\le D^{2^k}.$$
Consider $w\in G_\omega^*$ with $\tilde\phi_\omega(w)=\pair<w_0,w_1>$. The inverted orbit of $w$ is determined by the inverted orbits of $w_0$ and $w_1$, two words of total $\sigma\omega$-length at most $2/\eta_\omega(\|w\|_\omega+\|a\|_\omega)$ by Lemma \[lem:supercontract\]. Therefore, $$\Sigma_\omega(\eta_\omega R)\le\sum_{R_0+R_1\le \tfrac2{\eta_\omega}
(R+\|a\|_\omega)}\Sigma_{\sigma\omega}(R_0)\Sigma_{\sigma\omega}(R_1),$$ and the same argument as in Proposition \[prop:upperG\] applies.
The growth of $W_\omega(H)$
---------------------------
We start by general estimates on the growth of a permutational wreath product:
\[prop:growthW\] Let $H$ be a group with growth function $v_H$, and suppose that $v_H$ is log-concave.
Let $G$ be a group acting transitively on a set $X$ with basepoint $\xi$, and let $v_G$ denote the growth function of $G$. Denote the inverted orbit growth of $G$ on $(X,\xi)$ by $\Delta$, and denote its inverted orbit choice growth by $\Sigma$.
Consider the wreath product $W=H\wr_X G$, generated by $S\cup T@\xi$ for the generating sets $S,T$ of $G,H$ respectively. Then $$\begin{aligned}
& v_G(R) v_H(R/\Delta(R))^{\Delta(R)}\le v_W(3R),\\
v_W(R) \le & v_G(R) v_H(R/\Delta(R))^{\Delta(R)}(2R)^{\Delta(R)}\Sigma(R),\\
v_W(R) \le & v_G(R) (\#H)^{\Delta(R)}\Sigma(R)\text{ if $H$ is finite}.
\end{aligned}$$
Note that the assumption that $v_H$ be log-concave is mild, thanks to Lemma \[lem:concavehull\].
We begin by the lower bound. For every $R\in\N$, consider a word $w\in G^*$ of norm $\le R$ realizing the maximum $\Delta(R)$; write $\mathcal O(w)=\{x_1,\dots,x_k\}$ for $k=\Delta(R)$. Choose then $k$ elements $a_1,\dots,a_k$ of norm $\le R/k$ in $A$. Define $f\in
\sum_X H$ by $f(x_i)=a_i$, all unspecified values being $1$. Then $w f\in W$ may be expressed as a word of norm $R+|a_1|+\dots+|a_k|\le
2R$ in the standard generators of $W$, by inserting $a_1@\xi,\dots,a_k@\xi$ appropriately into the word $w$.
Furthermore, different choices of $a_i$ yield different elements of $W$. Finally multiplying $w f$ with an arbitrary $g\in G$ of length at most $R$, we obtain $v_G(R) v_H(R/k)^k$ elements in the ball of radius $3R$ in $W$.
For the upper bound, consider a word $w$ of norm $R$ in $W$, and let $f\in\sum_X H$ denote its value in the base of the wreath product. The support of $f$ has cardinality at most $\Delta(R)$, and may take at most $\Sigma(R)$ values.
Write then $\sup(f)=\{x_1,\dots,x_k\}$ for some $k\le\Delta(R)$, and let $a_1,\dots,a_k\in H$ be the values of $w$ at its support; write $\ell_i=\|a_i\|$.
Since $\sum\ell_i\le R$, the norms of the different elements on the support of $f$ define a composition of a number not greater than $R$ into at most $k$ summands; such a composition is determined by $k$ “marked positions” among $R+k$, so there are at most $\binom{R+k-1}{k}$ possibilities, which we bound crudely by $(2R)^k$. Each of the $a_i$ is then chosen among $v_H(\ell_i)$ elements, and (by the assumption that $v_H$ is log-concave) there are $\prod v_H(\ell_i)\le v_H(R/k)^k$ total choices for the elements in $H$.
We have now decomposed $w$ into data that specify it uniquely, and we multiply the different possibilities for each of the pieces of data. Counting the possibilities for the value of $w$ in $G$, the possibilities for its support in $X$, and the possibilities for the elements in $H$ on its support, we get $$v_W(R)\precsim v_G(R)v_A(R/k)^k(2R)^k\Sigma(R),$$ which is maximised by $k=\Delta(R)$.
Finally, if $H$ is finite then we may more simply bound the possible values of $f$ by $(\#H)^k$.
\[cor:growthW\] Let $G,H$ be groups of subexponential growth. Let $G$ act transitively on a set $X$ with basepoint $\xi$, with sublinear inverted orbit growth and subexponential inverted orbit choice growth. Then the wreath product $W=H\wr_X G$ has subexponential growth.
Assume that $H$ has subexponential growth; then by Lemma \[lem:concavehull\] there exists a log-concave subexponentially growing function $v_H$ bounding the growth of $H$ from above.
By Propositions \[prop:invgrowth\] and \[prop:invgrowth\], the function $\Delta$ is sublinear and the function $\Sigma$ is subexponential. By Proposition \[prop:upperG\], the growth of $G_\omega$ is subexponential. Corollary \[cor:growthW\] then shows that $W_\omega(H)$ has subexponential growth.
In the special case of $H$ finite and $G=G_\omega$, Proposition \[prop:growthW\] gives the
\[cor:global\] Let $H$ be a non-trivial finite group. There are then two absolute constants $F,E>1$ such that the growth function $v$ of $W_\omega(H)=H\wr_X G_\omega$ satisfies $$E^{2^k}\le v(\eta_\omega\cdots\eta_{\sigma^{k-1}\omega}\mu_{\sigma^k\omega})\le F^{2^k}.$$
Take together the upper bound on the growth of $G_\omega$ from Proposition \[prop:upperG\], the bounds on the inverted orbit growth from Proposition \[prop:invgrowth\], and the choices for the inverted orbits from Proposition \[prop:invchoices\]. The conclusion follows from Proposition \[prop:growthW\].
This follows directly from Corollary \[cor:global\], using the fact that $\eta_{\sigma^i\omega}=\eta_+$ for all $i\in\N$.
Proof of Theorem \[thm:givengrowth\]
------------------------------------
Our approach will be construct, out of the function $f$ satisfying , a sequence $\omega\in\{0,012\}^\infty$ with long stretches of $0$ when $f$ grows fast, and long stretches of $012$ when $f$ grows slowly.
We start by introducing some shorthand notation. For a finite sequence $\omega=\omega_0\dots\omega_{n-1}\in\{0,1,2\}^n$ and $p\in\Delta$, we write by extension $$\overline M_\omega=\overline M_{\omega_0}\cdots\overline M_{\omega_{n-1}}\colon\Delta\righttoleftarrow$$ and $$\eta(p,\omega_0\dots\omega_{n-1})=\eta(p,\omega_0)\eta(\overline M_{\omega_0}p,\omega_1)\cdots\eta(\overline M_{\omega_0\dots\omega_{n-2}}p,\omega_{n-1}).$$
For $\omega\in\{0,1,2\}^\Z$, recall the construction of $p_\omega\in\Delta$ from , and $\eta_\omega=\eta(p_\omega,\omega_0)$ and $\mu_\omega=\mu(p_\omega)$. Assume that a sequence $\omega\in\{0,1,2\}^\Z$ is under construction, and that there exists $k\in\N$ such that $\omega_i$ has been determined for all $i\le
k$. Then $p_\omega$, $\eta_\omega$ and $\mu_\omega$ are determined, and so are $p_{\sigma^i\omega},\eta_{\sigma^i\omega},\mu_{\sigma^i\omega}$ for all $i\le k$. We abbreviate $$p_i=p_{\sigma^i\omega},\qquad \eta_i=\eta_{\sigma^i\omega},\qquad\mu_i=mu_{\sigma^i\omega}.$$
\[lem:mu\_k bounded\] If the restriction to $\N$ of the sequence $\omega$ has the form $$\label{eq:omega}
\omega=(012)^{i_1}2^{j_1}(012)^{i_2}2^{j_2}(012)^{i_3}2^{j_3}\dots,$$ with $i_1,j_1,i_2,j_2,\dots\ge1$, then the $\mu_k$ are all bounded away from $0$.
In fact, the image of $\overline M_{012}$ is the open triangle spanned by $(\frac13,\frac13,\frac13)$, $(\frac27,\frac27,\frac37)$ and $(\frac4{17},\frac6{17},\frac7{17})$, so after each $012$ the $\mu_k$ belongs to $(\frac4{17},\frac13)$.
The image of that triangle under $\overline M_{2^i}$ is contained in the convex quadrilateral spanned by $(\frac13,\frac13,\frac13)$, $(\frac4{17},\frac6{17},\frac7{17})$, $(\frac14,\frac14,\frac12)$ and $(\frac15,\frac3{10},\frac12)$, so $\mu_k\in(\frac15,\frac13)$ for all $k$.
The heart of the argument is the following lemma, which shows that $\eta$ approaches very quickly its limiting values $2$ and $\eta_+$ as the sequence $\omega$ contains long segments of $2$ or of $012$:
\[lem:expconv\] There exist constants $A'\le1$, $B'\ge1$ such that,
1. \[lem:expconv1\] For all $p\in\Delta$ and all $n\in\N$, $$\eta(p,(012)^n)\ge\eta_+^{3n}A';$$
2. \[lem:expconv2\] For all $p\in\Delta$ and all $n\in\N$, $$\eta(p,2^n)\le2^n B'.$$
Let $\mathcal U$ denote the image of $\overline M_{012}$, and let $p_+\in\mathcal U$ denote the fixed point of $\overline
M_{012}$. Note first that $\eta(-,012)$ is differentiable at $p_+$, and that $\overline M_{012}$ is uniformly contracting on $\mathcal
U$; let $\rho<1$ be such that $\overline M_{012}$ is $\rho$-Lipschitz on $\mathcal U$, and let $D$ be an upper bound for the derivative of $\log\eta(-,012)$ on $\mathcal U$. Recall that $\eta_+^3=\eta(p_+,012)$.
For all $k\in\N$, write $p_k=\overline M_{(012)^k}(p)$. For $k\ge1$ we have $\|p_k-p_+\|\le\rho^{k-1}$, so $|\log\eta(p_k,012)-3\log\eta_+|<D\rho^{k-1}$, while for $k=0$ we write $|\log\eta(p,012)-3\log\eta_+|<3\log3$. Therefore, $$|\log\eta(p,(012)^n)-3n\log\eta_+|\le
3\log3+\sum_{k=1}^{n-1}|\log\eta(p_k,012)-3\log\eta_+| \le
3\log3+D/(1-\rho)$$ is bounded over all $n$ and $p$. The estimate follows, with $A'=3^3\exp(D/(1-\rho))$.
For the second part, consider $p_k=\overline M_{2^k}(p)$, and note that $p_k$ converges at exponential (Euclidean) speed to a point $p_\infty$ on the side $\{\delta=\frac12\}$ of $\partial\Delta$, since $\partial\overline
M_2(\beta,\gamma,\delta)/\partial\delta(*,*,\frac12)=\frac12$; so we have $\|p_k-p_\infty\|<\rho^{k-1}$ for some $\rho<1$. As above, $\eta(-,2)$ is differentiable in a neighbourhood $\mathcal U$ of $\{\delta=\frac12\}$, and the derivative of $\log\eta(-,2)$ is bounded on $\mathcal U$, say by $D$. Recall that $\eta(p,2)=2$ for all $p\in\partial\Delta$. The same computation as above yields , with $B'=3\exp(D/(1-\rho))$.
We now reformulate the statement of Theorem \[thm:givengrowth\] as follows. Set $g(R)=\log f(R)$, so that we have $$\label{eq:g}
g(2R)\le 2g(R)\le g(\eta_+R)$$ for all $R$ large enough. For simplicity (since growth is only an asymptotic property) we assume that holds for all $R$. Without loss of generality (since we are allowed to replace $g$ by an equivalent function), we also assume that $g$ is increasing and satisfies $g(1)=1$.
We are ready to construct $\omega$. Fix arbitrarily the value of $\omega$ on its negative part, say $\omega_{|-\N}=(012)^{-\infty}$. This determines an initial metric $p_0\in\Delta$. Out of the function $g$, we will construct a sequence $\omega$ such that, for constants $A,B$, we have $$\label{eq:AB}
A\le \frac{g(\eta(p_0,\omega_0\dots\omega_{k-1}))}{2^k}\le B\text{ for
all }k;$$ in fact, it will suffice to obtain this inequality for a set of values $k_0,k_1,\dots$ of $k$ such that $\sup_i(k_{i+1}-k_i)<\infty$. Indeed, the orbit $p_i$ of $p_0$ in $\Delta'$ will remain bounded, so we will have $\mu(p_i)\in[C,1]$ for some $C>0$. By Corollary \[cor:global\], $$\begin{aligned}
E^{2^k}\le v(\eta(p_0,\omega_0\dots\omega_{k-1})\mu_k)\le &v(B g^{-1}(2^k)),\\
&v(A C g^{-1}(2^k))\le v(\eta(p_0,\omega_0\dots\omega_{k-1})\mu_k)\le F^{2^k}\end{aligned}$$ and therefore $v(R)\sim \exp(g(R))=f(R)$.
We will extend the sequence $\omega$ to be, on the positive integers, of the form , $$\omega=(012)^{i_1}2^{j_1}(012)^{i_2}2^{j_2}(012)^{i_3}2^{j_3}\dots,$$ with $i_1,j_1,i_2,j_2,\dots\ge1$. The $\mu_k$ are bounded away from $0$ by Lemma \[lem:mu\_k bounded\].
Assuming by induction that $\omega'=\omega_0\dots\omega_{k-1}$ has been constructed, we repeat the following:
- while $g(\eta(p_0,\omega'))<2^k$, we append $012$ to $\omega'$;
- while $g(\eta(p_0,\omega'))>2^k$, we append $2$ to $\omega'$.
For our induction hypothesis, we assume that the stronger condition $$\frac122^k\le g(\eta(p_0,\omega'))\le 2^k$$ holds for each $k$ of the form $i_1+j_1+\dots+i_m+j_m$, and that $$2^k\le g(\eta(p_0,\omega'))\le3^32^k$$ holds for each $k$ of the form $i_1+j_1+\dots+i_m$; these conditions apply whenever $\omega$ is a product of ‘syllables’ $(012)^{i_t}$ and $2^{j_t}$.
Consider first the case $\frac122^k\le g(\eta(p_0,\omega'))\le2^k$; and let $n$ be minimal such that $g(\eta(p_0,\omega'(012)^n))>2^{k+3n}$. Then, for all $i\in\{1,\dots,n\}$, Lemma \[lem:expconv\] gives $\eta(p_0,\omega'(012)^i)\ge\eta(p_0,\omega')\eta_+^{3i}A'$. Let $u\in\N$ be minimal such that $A'\ge\eta_+^{-u}$; this, combined with $g(\eta_+R)\ge2g(R)$, gives $$g(\eta(p_0,\omega'(012)^i))\ge g(\eta(p_0,\omega')\eta_+^{3i-u})\ge
2^{-1-u}2^{k+3i}.$$ By minimality of $n$, we have $g(\eta(p_0,\omega'(012)^{n-1}))\le2^{k+3(n-1)}$; since $g$ is sublinear and $\eta\le3$, we get $$g(\eta(p_0,\omega'(012)^n))\le3^32^{k+3n}.$$
Consider then the case $2^k\le g(\eta(p_0,\omega'))\le3^32^k$, which is similar, and let $n$ be minimal such that $g(\eta(p_0,\omega'2^n))<2^{k+n}$. Then, for all $i\in\{1,\dots,n\}$, Lemma \[lem:expconv\] gives $\eta(p_0,\omega'2^i)\le\eta(p_0,\omega')2^i B'$; this, combined with $g(2R)\le2g(R)$, gives $$g(\eta(p_0,\omega'2^i))\le3^32^{k+i}B'.$$ By minimality of $n$, we have $g(\eta(p_0,\omega'2^{n-1}))\ge2^{k+n-1}$; since $g$ is increasing, we get $$g(\eta(p_0,\omega'2^n))\ge\frac122^{k+n}.$$ We have proved the claim , with $A=2^{-1-u}$ and $B=3^3B'$.
\[rem:solvablewp\] The construction of $\omega$ from $f$ is algorithmic, in the following sense. The initial point $p_0$ may be computed to arbitrary precision by an algorithm. If there exists an algorithm that computes values of $f$, then there exists an algorithm that, with $k\in\N$ as input, computes $g(\eta(p_0,\omega_0\dots\omega_{k-1}))$ to arbitrary precision; so there exists an algorithm that computes the digits of $\omega$.
It then follows via Theorem \[thm:cornulier\] that the groups $G_\omega$ and $W_\omega(H)=H\wr_X G_\omega$ are recursively presented, for recursively presented $H$.
Illustrations {#ss:illustrations}
-------------
We now consider illustrations of Theorem \[thm:givengrowth\], and examples of growth functions that may occur for groups $W_\omega(C_2)$. In fact, the examples can be constructed in both directions: either choose a “nice” function $f$ that satisfies , or choose a “nice” sequence $\omega$ and estimate the corresponding growth using Corollary \[cor:global\], Lemma \[lem:mu\_k bounded\] and Lemma \[lem:expconv\]. We follow both approaches.
- For every $\alpha\in[\log2/\log\eta_+,1]$, there exists a group of growth $\sim\exp(R^\alpha)$. Furthermore, for a dense set of $\alpha$ in that interval, there exists a periodic sequence $\omega$ such that $W_\omega(C_2)$ has growth $\sim\exp(R^\alpha)$.
For every $\alpha\le\beta\in[\log2/\log\eta_+,1)$, one may construct a function $f$ satisfying that coincides, on arbitrarily large intervals, sometimes with the function $\exp(R^\alpha)$ and sometimes with the function $\exp(R^\beta)$. Therefore, there exists a group whose growth function accumulates both at $\exp(R^\alpha)$ and at $\exp(R^\beta)$. This recovers a result by Brieussel, see [@brieussel:growth].
- there exists groups of growth $\sim\exp(R/\log R)$, of growth $\sim\exp(R/\log\log R)$, of growth $\exp(R/\log\cdots\log R)$.
- Consider conversely the sequence $\omega=(012)2^1(012)2^2(012)2^3(012)2^4\dots$. Among the first $k$ entries, approximately $\sqrt k$ instances of $012$ will have been seen; therefore $\eta(p_0,\omega_0\dots\omega_{k-1})\approx 2^{k+\mathcal O(1)\sqrt
k}$. This gives a growth function of the order of $$\exp\big(R/\exp(\mathcal O(1)\sqrt{\log R})\big).$$
Consider next the sequence $\omega=(012)2^1(012)2^2(012)2^4(012)2^8\dots$. Among the first $k$ entries, approximately $\log k$ instances of $012$ will have been seen; therefore $\eta(p_0,\omega_0\dots\omega_{k-1})\approx
2^{k+\mathcal O(1)\log k}$. This gives a growth function of the order of $$\exp\big(R/(\log R)^{\mathcal O(1)}\big).$$
Consider further the sequence $\omega=(012)2^{2^1}(012)2^{2^2}(012)2^{2^4}(012)2^{2^8}\dots$. Among the first $k$ entries, approximately $\log\log k$ instances of $012$ will have been seen; therefore $\eta(p_0,\omega_0\dots\omega_{k-1})\approx 2^{k+\mathcal
O(1)\log\log k}$. This gives a growth function of the order of $$\exp\big(R/(\log\log R)^{\mathcal O(1)}\big).$$
These constructions generalise easily to give “nice” sequences $\omega$ such that $W_\omega(C_2)$ has growth of the order of $\exp(R/(\log\cdots\log R)^{\mathcal O(1)})$.
- Consider the Ackermann function $$A(m,n)=\begin{cases}
n+1 & \text{ if }m=0,\\
A(m-1,1) & \text{ if $m>0$ and }n=0,\\
A(m-1,A(m,n-1)) & \text{ if $m>0$ and }n>0,
\end{cases}$$ and set $B(n)=A(n,n)$. Then there exists a group of growth $\sim\exp(R/B(R)^{-1})$; this last growth function is faster than any subexponential primitive-recursive function.
Considering $\omega=(012)2^{A(0,0)}(012)2^{A(1,1)}(012)2^{A(2,2)}\dots$ gives a group $W_\omega$ whose growth function is subexponential, but at least as fast as $\exp(R/B(R)^{-1})$;
Imbeddings and subgroups {#ss:imbed}
========================
A classical result by Higman, Neumann and Neumann [@higman-n-n:embed] states that every countable group imbeds in a finitely generated group. It was then shown that many properties of the group can be inherited by the imbedding: in particular, solvability (Neumann-Neumann [@neumann-n:embed]), torsion (Phillips [@phillips:embedding]), residual finiteness (Wilson [@wilson:embeddingrf]), and amenability (Olshansky-Osin [@olshanskii-osin:qiembedding]).
Seen the other way round, these results show that there is little restriction, apart from being countable, on the subgroups of a finitely generated group; even if that group is furthermore assumed to be residually finite, amenable, or solvable.
On the other hand, very strong restrictions exist on the subgroups of a virtually nilpotent group: they are all finitely generated, for example.
Since by Gromov’s Theorem \[thm:gromov\] the finitely generated virtually nilpotent groups are precisely the groups of polynomial growth, we naturally ask what conditions are imposed on the subgroups of a group of subexponential word growth. As we shall see, there are essentially none.
In particular, there are torsion branched groups of subexponential growth such as the group $G_\omega$. They contain iterated wreath products, by Proposition \[prop:branched=>fullwreath\], so contain infinitely generated subgroups.
Let us say that a group has *locally subexponential growth* if all of its finitely generated subgroups have subexponential growth. Clearly, if $G$ has subexponential growth then all its subgroups have locally subexponential growth. This is the only restriction, and the objective of this section is to prove the following result:
\[thm:imbed\] Let $B$ be a group. Then there exists a finitely generated group of subexponential growth in which $B$ imbeds as a subgroup if and only if $B$ is countable and locally of subexponential growth.
For example, this implies that there exists a group of subexponential growth containing $\Q$ as a subgroup. I do not know any explicit such example of group.
Neumann’s proof
---------------
By way of motivation, we start with the classical result by Higman and the Neumanns:
\[thm:hnn\] Every countable group imbeds in a finitely generated group.
We shall not follow the original proof (which proceeds by a sequence of “HNN extensions”), but rather that by the two Neumanns [@neumann-n:embed], which uses wreath products. It follows immediately from combining the following two propositions:
\[prop:neumann-n:1\] Every countable group $B$ imbeds in the commutator subgroup $[G,G]$ of a countable group $G$.
\[prop:neumann-n:2\] For every countable group $G$, there exists a $2$-generated group $W$ such that $[G,G]$ imbeds in $[W,W]$.
Consider the following subgroup $G$ of the unrestricted wreath product $B{\operatorname{\wr\wr}}\Z$. The group $G$ is generated by $\Z$ and, for all $b\in B$, the function $f_b\colon\Z\to B$ defined by $f_b(m)=b^m$. Denoting by $t$ the generator of $\Z$, we see that $[t,f_b]$ is the constant function $b$; so $B$ is in fact imbedded in $[t,G]$.
Could we also have defined $f_b(m)=b$ for $m\ge0$ and $f_b(m)=1$ for $m<0$? What would be the advantages and disadvantages of this alternative construction?
Consider the following subgroup $W$ of the unrestricted wreath product $G{\operatorname{\wr\wr}}\Z$. Denote by $u$ a generator of $\Z$, and by $\{b_1,b_2,\dots\}$ a generating set of $G$. Choose a sparse-enough sequence of elements $x_1,x_2,\dots$ of $\Z$, and define $f\colon
\Z\to G$ by $f(x_i)=b_i$, all other values being trivial. The group $W$ is then $W=\langle f,u\rangle$.
Let us spell out below what it means to be “sparse enough”. We write $\Z$ additively. Since in fact by Proposition \[prop:neumann-n:1\] we only need to imbed $[t,G]$ in $W$, we may set $f(0)=t$ and a sufficient condition on the $x_i\in\Z$ is that $x_i\neq0$ for all $i$; all $x_i$ are distinct; and $x_i+x_j\not\in\{0,x_k\}$ for all $i,j,k\in\N$. One then sees that $[f,f^{u^{-x_i}}]$ is a function supported only at $0$, with value $[t,b_i]$ there. This defines the imbedding of $[t,G]$ in $W$.
Note that the group $W$ contains the standard wreath product $B\wr\Z$, so always has exponential growth.
The above construction shows that every countable solvable group imbeds in a $2$-generated solvable group.
Show first that *not* every countable nilpotent group imbeds in a $2$-generated nilpotent group.
Show then that every finitely generated nilpotent group imbeds in a $2$-generated nilpotent group.
Here is a useful, small improvement on Theorem \[thm:hnn\]:
\[thm:hnntorsion\] Let $p\ge5$ be an integer. Then every countable group imbeds in a $2$-generated group both of whose generators have order $p$.
Piggybacking on Propositions \[prop:neumann-n:1\] and \[prop:neumann-n:2\], it suffices to consider a $2$-generated group $G=\langle x,y\rangle$, and to imbed $[G,G]$ into a $2$-generated group $W$ with generators of order $p$. Write $C_p=\langle t|t^p\rangle$, and define $f\colon C_p\to G$ by $$f(1)=x,\qquad f(t^{-1})=y,\qquad f(t^2)=y^{-1}x^{-1},\qquad f(g)=1\text{ for all other }g\in C_p.$$ Consider then the group $W=\langle t,ft\rangle$. It is clearly generated by two elements of order $p$, since in $(ft)^p$ all the coördinates contain some cyclic permutation of the product $x\cdot
y\cdot y^{-1}x^{-1}$. Now $W$ contains $f$, so it also contains $[f,f^t]$, which is the function $C_p\to G$ taking value $[x,y]$ at $1$ and $1$ elsewhere. Furthermore conjugating $[f,f^t]$ by an arbitrary word in $f$ and $f^t$, one obtains the function taking value an arbitrary conjugate of $[x,y]$ at $1$; so $W$ contains $[G,G]@1$.
Where have we used the assumption that $p\ge5$? Can you improve the above result to arbitrary $p\ge3$? Can you imbed any countable group into a group generated by an involution and an element of order $p$?
Finite-valued permutational wreath products
-------------------------------------------
Our goal is, starting from a countable group $B$ locally of subexponential growth, to construct a finitely generated group $W$ of subexponential growth containing $B$. We take inspiration from Neumann’s proof given above, with two modifications: first, we consider permutational wreath products rather than standard ones; secondly, we consider *finite-valued permutational wreath products*:
Let $H$ be a group acting on a set $X$, and let $H$ be a group. Their *finite-valued permutational wreath product* is the group $H{\wr^{\text{\upshape f.v.}}}_X G$, defined as the extension of functions $X\to
H$ with finite image by $G$: $$H{\wr^{\text{\upshape f.v.}}}_X G=\{(\phi,g)\in H^X\times G\colon\#\phi(X)<\infty\}.$$
Note that it is a subgroup, because if $(\phi,g)^{-1}(\phi',g')=(\phi'',g^{-1}g')$ then $\phi''(X)\subseteq\phi(X)^{-1}\phi'(X)$ is finite. Clearly, we have $$H\wr_X G\le H{\wr^{\text{\upshape f.v.}}}_X G\le H{\operatorname{\wr\wr}}G.$$ We also introduce a condition on imbeddings that guarantees control on growth:
Let $B$ be a group. A group $G$ is called *hyper-$B$* if it is a directed union of finite extensions of finite powers of $B$.
Clearly, if a group $B$ is locally of subexponential growth and a group $G$ is hyper-$B$, then $G$ is also locally of subexponential growth. Indeed, for every finite subset $S$ of $G$ there exists a finite extension of a finite power of $B$ that contains $S$.
\[lem:hyperhyper\] Let $G$ be a hyper-$B$ group, and let $H$ be a hyper-$G$ group. Then $H$ is hyper-$B$.
Consider $h\in H$; then $h$ belongs to a finite extension of a finite power of $G$, which may be assumed of the form $G\wr F$ for a finite group $F$. Let us write $h=\phi f$ with $\phi\colon F\to G$ and $f\in F$; then $\phi(f)$ belongs for all $f\in F$ to a finite extension of a finite power of $B$, which can be assumed to be the same for all $f$. This extension may be assumed to be of the form $B\wr E$ for a finite group $E$. It follows that $h$ belongs to $B\wr_{E\times F}(E\wr F)$, a finite extension of a finite power of $B$; so $H$ is hyper-$B$.
\[lem:hyperB\] If $H$ is a hyper-$B$ group and $U$ is locally finite, then $H{\wr^{\text{\upshape f.v.}}}U$ is a hyper-$B$ group.
We first show that $H{\wr^{\text{\upshape f.v.}}}U$ is hyper-$H$. By hypothesis, $U$ is a directed union of finite subgroups $E$. The partitions $\mathscr
P_0$ of $U$ into finitely many parts also form a directed poset; and for every such partition $\mathscr P_0$ and every finite subgroup $E\le U$ there exists a finite partition $\mathscr P$ of $U$ that is invariant under $E$ and refines $\mathscr P_0$, namely the wedge (= least upper bound) of all $E$-images of $\mathscr P_0$.
Consider now the directed poset of pairs $(E,\mathscr P)$ consisting of finite subgroups $E\le U$ and $E$-invariant partitions of $U$. Consider the corresponding subgroups $H^{\mathscr P}\rtimes E$ of $H{\wr^{\text{\upshape f.v.}}}U$. If $(E,\mathscr P)\le(E',\mathscr P')$ then $H^{\mathscr
P}\rtimes E$ is naturally contained in $H^{\mathscr P'}\rtimes
E'$, so these subgroups of $H{\wr^{\text{\upshape f.v.}}}U$ form a directed poset, which exhausts $H{\wr^{\text{\upshape f.v.}}}U$.
It follows that $H{\wr^{\text{\upshape f.v.}}}U$ is a hyper-$H$ group, and we are done by Lemma \[lem:hyperhyper\].
Imbedding in the derived subgroup {#ss:imbedG'}
---------------------------------
Our main goal, in this section, is to prove the following proposition, which replaces Proposition \[prop:neumann-n:1\].
\[prop:imbed’\] Let $B$ be a group. Then there exists a hyper-$B$ group $G$ such that $[G,G]$ contains $B$ as a subgroup.
If $B$ is infinite, then $G$ may furthermore be supposed to have the same cardinality as $B$.
\[lem:CB\] Let $B$ be a group. Then there exists a subgroup $C$ of $B$, containing $[B,B]$, such that $B/C$ is torsion and $C/[B,B]$ is free abelian.
$B/[B,B]\otimes_\Z\Q$ is a $\Q$-vector space, hence has a basis, call it $X$. It generates a free abelian group $\Z X$ within $B/[B,B]$, whose full preimage in $B$ we call $C$. Then $B/C\otimes_\Z\Q=0$ so $B/C$ is torsion.
We set up the following notation for the proof of Proposition \[prop:imbed’\]. We choose a subgroup $C\le B$ as in Lemma \[lem:CB\] and write $T:=B/C$. We choose a basis $X$ of $C/[B,B]$, for every $x\in X$ we choose an element $b_x\in C$ representing it, and we define a homomorphism $\theta_x\colon
C\to\langle b_x\rangle\subseteq B$, trivial on $[B,B]$, by $\theta_x(b_x)=b_x^{-1}$ and $\theta_x(b_y)=1$ for all $y\neq x\in
X$. In particular, we have for all $b\in C$ $$b\cdot\prod_{x\in X}\theta_x(b)\in[B,B]$$ and the product is finite.
We write $\pi\colon B\to T$ the natural projection, and define a section $\sigma$ of $\pi$ with the following property, which we single out as a lemma:
\[lem:section\] There exists a set-theoretic section $\sigma\colon T\to B$ such that, for every $t\in T$, the subset $\{\sigma(t u)\sigma(u)^{-1}:u\in T\}[B,B]$ of $B/[B,B]$ is finite.
Since every abelian torsion group is the direct sum of its $p$-subgroups, we may first define the section on each of $T$’s $p$-components, and then extend it to $T$ multiplicatively (in any order).
We therefore suppose that $T$ is a $p$-group. Recall the notation $\Omega_n(T)=\{t\in T:t^{p^n}=1\}$. Each quotient $V_n:=\Omega_n(T)/\Omega_{n-1}(T)$ is a vector space over $\mathbb F_p$, and the homomorphism $t\mapsto t^p$ induces an injective linear map $V_n\to V_{n-1}$. Choose inductively subsets $X_1,X_2,\dots$ of $T$ such that $X_n$ maps to a basis of $V_n$ and such that $t\mapsto
t^p$ induces an injective map $X_n\to X_{n-1}$. Set $X=\bigcup_n
X_n$, and give an arbitrary total order to $X$. Choose for each $x\in X$ a $\pi$-preimage $\sigma(x)\in B$.
Since $T$ is torsion, every element $t\in T$ belongs to $\Omega_n(T)$ for some $n\in\N$, so can be uniquely written as a product $t=x_1^{\alpha_1}\cdots x_n^{\alpha_n}$ with ordered $x_i\in X$ and $0<\alpha_i<p$ for all $i$. Extend then $\sigma$ by $\sigma(t)=\sigma(x_1)^{\alpha_1}\cdots \sigma(x_n)^{\alpha_n}$.
Consider now $t\in T$, and write it in the form $t=x_1^{\alpha_1}\cdots x_n^{\alpha_n}$ as above. Extend $\{x_1,\dots,x_n\}$ to a finite subset $Y=\{x_1,\dots,x_n,\dots,x_s\}$ of $X$, by adding all $p^j$-th powers of all $x_i$ to it. The set $$T':=\{x_1^{\gamma_1}\cdots x_s^{\gamma_s}:0\le\gamma_i<p\}$$ is then a finite subgroup of $T$. Consider next $u\in T$, and note that it may be written uniquely in the form $u=y_1^{\beta_1}\cdots
y_m^{\beta_m}z$ with $y_i\in X\setminus Y$ and $z\in T'$. Then $t u=y_1^{\beta_1}\cdots y_m^{\beta_m}(z t)$ and this representation is unique; so $\sigma(t u)\sigma(u)^{-1}[B,B]$ belongs to the finite set $\{\sigma(t')[B,B]:t'\in T'\}$.
Rephrase Lemma \[lem:section\] in terms of the cohomology of $T$ with coëfficients in $\Z X$.
Let $F$ be a locally finite group of cardinality $>\#X$, and fix an imbedding of $X$ in $F\setminus\{1\}$. As a first step, we consider the group $G_0=B{\wr^{\text{\upshape f.v.}}}(T\times F)$, and define a map $\Phi_0\colon B\to
G_0$ as follows: $$\label{eq:Phi_0}
\Phi_0(b)=(\phi,\pi(b),1)\text{ with }\phi(t,f)=\begin{cases}
b & \text{ if }f=1,\\
\theta_f(\sigma(t)b\sigma(t\pi(b))^{-1}) & \text{ if }f\in X,\\
1 & \text{ otherwise.}
\end{cases}$$
\[lem:Phi\_0\] The map $\Phi_0$ is well-defined and is an injective homomorphism.
To see that $\Phi_0$ is well-defined, note that the argument $\sigma(t)b\sigma(t\pi(b))^{-1}$ belongs to $\ker(\pi)=C$, so that $\theta_f$ may be applied to it. Note then that, by Lemma \[lem:section\], the expression $\sigma(t)b\sigma(t\pi(b))^{-1}$ takes finitely many values in $C/[B,B]$, so that $\phi(t,f)$ takes finitely many values for varying $t$ and fixed $f$. Finally, $\theta_f(\sigma(t)b\sigma(t\pi(b))^{-1})=1$ except for finitely many values of $f\in X$. In summary, the function $\phi\in
B^{T\times F}$ is such that $\phi(t,f)$ takes only finitely many values.
It is clear that $\Phi_0$ is injective: if $b\neq1$ and $\Phi_0(b)=(\phi,\pi(b),1)$ then $\phi(1,1)=b\neq1$. It is a homomorphism because all $\theta_f$ are homomorphisms.
\[lem:C imbeds\] We have $\Phi_0(C)\le[G_0,G_0]$.
If $b\in[B,B]$ then clearly $\Phi_0(b)\in[G_0,G_0]$. Since $C$ is generated by $[B,B]\cup\{b_x\}_{x\in X}$, it suffices to consider $b=b_x$.
We define $g\in G_0$ by $$g=(\psi,1,1)\text{ with }\psi(t,f)=\begin{cases}b_x & \text{ if }f=1,\\ 1 & \text{ otherwise}.\end{cases}$$ Then $\Phi_0(b_x)=(\phi,1,1)$ with $\phi(t,1)=b_x$ and $\phi(t,x)=b_x^{-1}$, all other values being trivial, according to ; so, as was to be shown, $$\Phi_0(b_x)=(\phi,1,1)=({}^x\!\psi^{-1}\cdot\psi,1,1)=[(1,1,x^{-1}),g]\in[G_0,G_0].\qedhere$$
We finally define $$G=G_0{\wr^{\text{\upshape f.v.}}}(\Q/\Z)$$ and a map $\Phi\colon B\to G$ by $$\Phi(b)=(\phi,0)\text{ with }\phi(r)=\Phi_0(b)\text{ for all }r\in\Q/\Z.$$
\[lem:B imbeds\] The map $\Phi$ is an injective homomorphism, and $\Phi(B)\le[G,G]$.
Clearly $\Phi$ is an injective homomorphism, since $\Phi_0$ is an injective homomorphism by Lemma \[lem:Phi\_0\].
We identify $\Q/\Z$ with $\Q\cap[0,1)$. For every $n\in\N$, consider the map $\Psi_n\colon B\to G$ defined by $$\Psi_n(b)=(\phi,0)\text{ with }\phi(r)=\begin{cases}
\Phi_0(b) & \text{ if }r\in[0,1/n),\\
1 & \text{ otherwise;}
\end{cases}$$ so $\Phi=\Psi_1$. We know from Lemma \[lem:C imbeds\] that $\Psi_n(C)$ is contained in $[G,G]$.
Consider now $b\in B$. Since $B/C$ is torsion, there exists $n\in\N$ such that $b^n\in C$. We define $g\in G$ by $$g=(\psi,0)\text{ with }\psi(r)=\Phi_0(b)^{\lfloor r n\rfloor}\text{
for }r\in[0,1)\cap\Q.$$ Let us write $h=\Phi_0(b)$, and consider the element $[(1,1/n),g]\cdot\Psi_n(b^n)=(\phi,0)$. If $r\in[0,1/n)$ then $\phi(r)=\psi(r-1/n)^{-1}\psi(r)h^n=h$, while if $r\in[1/n,1)$ then $\phi(r)=\psi(r-1/n)^{-1}\psi(r)=h$; therefore $$\Phi(b)=[(1,1/n),g]\cdot\Psi_n(b^n)\in[G,G].\qedhere$$
The first assertion is simply Lemma \[lem:B imbeds\]. For the last one: if $B$ is infinite, we wish to find a subgroup $H$ of $G$ with the same cardinality as $B$, such that $\Phi$ maps into $[H,H]$. For each $b\in B$, choose a finite subset $S_b$ of $G$ such that $\Phi(b)\in[\langle S_b\rangle,\langle S_b\rangle]$, and a subgroup $G_b$, containing $S_b$, that is virtually a finite power of $B$. Consider the group $H$ generated by the union of all the $G_b$. As soon as $B$ is infinite, all $G_b$ have the same cardinality as $B$, and so does $H$.
Spreading, stabilizing, rectifiable sequences
---------------------------------------------
We also extend Proposition \[prop:neumann-n:2\] by replacing the wreath product $G{\operatorname{\wr\wr}}\Z$ by permutational wreath products of the form $G{\operatorname{\wr\wr}}_X P$ for a group $P$ acting on a set $X$, and considering subgroups of the form $W=\langle f,P\rangle$ for a function $f\colon X\to G$.
We already encountered in Proposition \[prop:growthW\] sufficient conditions for such a group $W$ to be of subexponential growth, in case $f$ is finitely generated. As we shall see, the group $W$ may also have subexponential growth if $f$ is *infinitely* supported, but its support is sufficiently sparse, in a sense that we describe now.
In this section, we assume a finitely generated group $P=\langle S\rangle$ acting on the right on a set $X$ has been fixed. We use the same notation for $X$ as a set and as a *Schreier graph*, namely as the graph with vertex set $X$, and with for all $x\in X,s\in S$ an edge labelled $s$ from $x$ to $x s$, see Definition \[def:schreier\]. We denote by $d$ the path metric on this graph.
\[def:spreading\] A sequence $(x_0,x_1,\dots)$ in $X$ is *spreading* if for all $R$ there exists $N$ such that if $i,j\ge N$ and $i\neq j$ then $d(x_i,x_j)\ge R$.
If all $x_i$ lie in order on a geodesic ray starting from $x_0$ (for example if $X$ itself is a ray starting from $x_0$) and for all $i$ we have $d(x_0,x_{i+1})\ge 2 d(x_0,x_i)$, then $(x_i)$ is spreading.
Show that a sequence $(x_0,x_1,\dots)$ in $X$ is spreading if and only if for all $R$ there exists $N$ such that if $i\neq j$ and $i\ge N$ then $d(x_i,x_j)\ge R$.
\[def:locally stabilizing\] A sequence $(x_i)$ in $X$ *locally stabilises* if for all $R$ there exists $N$ such that if $i,j\ge N$ then the $S$-labelled radius-$R$ balls centered at $x_i$ and $x_j$ in $X$ are isomorphic as labelled graphs.
\[def:rectifiable\] A sequence of points $(x_i)$ in $X$ is *rectifiable* if for all $i,j$ there exists $g \in P$ with $x_i g =x_j$ and $x_k g\ne x_\ell$ for all $k\notin\{i,\ell\}$.
For example, if $X=\Z$ and $P=\Z$ acting by translations, then $\Sigma=\{2^i:i\in\N\}$ is rectifiable, since $2^j-2^i=2^\ell-2^k$ only has trivial solutions $i=k,j=\ell$ and $i=j,k=\ell$. It is also spreading and locally stabilizing.
Show that the sequence $\Sigma=(x_i)\subseteq X$ is rectifiable if and only if for all $i,j$ there exists $g\in P$ with $x_i g = x_j$ and $\Sigma\cap\Sigma
g\subseteq\{x_j\}\cup\operatorname{fixed.\!points}(g)$.
Fix a point $z\in X$. A sequence $(g_i)$ in $P$ is *parallelogram-free at $z$* if, for all $i,j,k,\ell$ with $i\neq
j$ and $j\neq k$ and $k\neq\ell$ and $\ell\neq i$ one has $z g_i^{-1}g_j g_k^{-1} g_\ell\ne z$.
$$\begin{tikzpicture}[decoration={markings,mark=at position 0.5 with {\arrow{>}}}]
\node[inner sep=0] (z) at (0,0) [label=right:$z$] {$\bullet$};
\node[inner sep=0] (zi) at (-3,-1) [label=left:$z g_i^{-1}$] {$\bullet$};
\node[inner sep=0] (zij) at (-4.5,0.5) {$\bullet$};
\node[inner sep=0] (zl) at (-1.5,1.5) {$\bullet$};
\node[inner sep=0] (zx) at (-0.2,0.2) {$\bullet$};
\draw[postaction={decorate}] (zi) -- node[below] {$i$} (z);
\draw[postaction={decorate}] (zi) -- node[below left] {$j$} (zij);
\draw[postaction={decorate}] (zl) -- node[above] {$k$} (zij);
\draw[postaction={decorate}] (zl) -- node[above right] {$\ell$} (zx);
\end{tikzpicture}$$
\[lem:parallelogram\] If $z\in X$ and $(g_i)$ is parallelogram-free at $z$, then $(z
g_i^{-1})$ is a rectifiable sequence in $X$.
Set $x_i=z g_i^{-1}$ for all $i\in\N$. Given $i,j\in\N$, consider $g=g_i g_j^{-1}$, so $x_i g = x_j$. If furthermore we have $x_k
g=x_\ell$, then we have $z g_k^{-1} g_i g_j^{-1} g_\ell = z$, so either $k=i$, or $i=j$ which implies $k=\ell$, or $j=\ell$ which implies $k=i$, or $\ell=k$. In all cases $k\in\{i,\ell\}$ as was to be shown.
It is clear that, if $P$ is finitely generated and $X$ is infinite, then it admits spreading and locally stabilizing sequences. Indeed every sequence contains a spreading subsequence, and every sequence contains a stabilizing subsequence, and a subsequence of a spreading or locally stabilizing sequence is again spreading, respectively locally stabilizing.
We will content ourselves with the following rectifiable sequences, based on the Grigorchuk groups $G_\omega=\langle a,b,c,d\rangle$, with $\omega\in\Omega'$, for example the first Grigorchuk group $G_{012}$. Recall the description of $G_\omega$ from §\[ss:grigorchuk\], and in particular the Schreier graph of its action on $\8^\infty$ in . We construct explicitly a spreading, locally stabilizing, rectifiable sequence for the action of $G_\omega$ on $X$: for all $i\in\N$, let us define $$x_i=\8^\infty\9^i,$$ the point at distance $2^i$ from the origin on the Schreier graph.
\[lem:Gspreading\] For all $\omega\in\Omega$ containing infinitely many $0,1,2$’s, and for all $i,j\in\N$,
1. the marked balls of radius $2^{\min(i,j)}$ in $X$ around $x_i$ and $x_j$ coincide;
2. the distance $d(x_i,x_j)$ is $|2^i-2^j|$;
3. there exists $g_{i,j}\in G_\omega$ of length $|2^i-2^j|$ with $x_i g_{i,j} = x_j$ and $x_k g_{i,j}\neq x_\ell$ for all $(k,\ell)\neq (i,j)$.
(1), (2) Consider the map $\theta_\omega$ from Equation . It maps the stabilizer of $\8^\infty$ in $G_{\sigma\omega}$ to the stabilizer of $\8^\infty$ in $G_\omega$, and therefore defines a self-map of $X$ by sending $\8^\infty g$ to $\8^\infty\theta_\omega(g)$. A direct calculation shows that it sends $x\in X$ to $x\9$.
Since $\theta_\omega$ is $2$-Lipschitz on words of even length in $\{a,b,c,d\}$, it maps the ball of radius $n$ around $x$ to the ball of radius $2n$ around $x\9$. Its image is in fact a net in the ball of radius $2n$: two points at distance $1$ in the ball of radius $n$ around $x$ will be mapped to points at distance $1$ or $3$ in the image, connected either by a segment or by a path for some $\{x,y\}\subset\{b,c,d\}$. In particular, the $2^n$-neighbourhoods of the balls about the $x_m$ coincide for all $m\ge n$.
\(3) Note, first, that there exists $g_{i,j}$ with $x_i g_{i,j} =
x_j$, because the rays ending in $\8^\infty$ form a single orbit. Note, also, that we have $x_k g_{i,j} = x_\ell$ for either finitely many $(k,\ell)\neq(i,j)$ or for all but finitely many $(k,\ell)$, because there is a level $N$ at which the decomposition of $g_{i,j}$ consists entirely of generators; if the entry at $\9^N$ of $g_{i,j}$ is trivial or ‘$d$’ then all but finitely many of the $x_k$ are fixed; while otherwise (up to increasing $N$ by at most one) we may assume it is an ‘$a$’; then $\9^{N+1} g_{i,j}=\9^N\8$, so $x_k\neq x_\ell$ for all $k>N+1$.
We use the following property of the Grigorchuk groups $G_\omega$: for every finite sequence $u\in\{\9,\8\}^*$ there exists an element $h_u\in G_\omega$ whose fixed points are precisely those sequences in $\{\9,\8\}^\infty$ that do not start with $u$. One may take for $h_u$ the element $[a,b]$, $[a,c]$ or $[a,d]$ inside the copy of the branching subgroup of $G_\omega$ that acts on $\{\9,\8\}^\infty u$, see Proposition \[prop:Gbranched\].
If the entry at $\9^N$ of $g_{i,j}$ is trivial, then we multiply $g_{i,j}$ with $h_{\9^M}$ for some $M>\max(N,i)$, so as to fall back to the second case.
Then, for each pair $(k,\ell)\neq(i,j)$ with $x_k g_{i,j}=x_\ell$, we multiply $g_{i,j}$ with $h_{\8\9^\ell}$, so as to destroy the relation $x_k g_{i,j}=x_\ell$.
The resulting element $g_{i,j}$ satisfies the required conditions.
Subexponential growth of wreath products
----------------------------------------
The next step in the proof is an argument controlling the growth of a subgroup of the form $W=\langle P,f\rangle\le G{\operatorname{\wr\wr}}_X P$, for a function $f\colon X\to G$ with sparse-enough (but infinite!) support.
We select a finitely generated group $P$ of subexponential growth acting on a set $X$ with sublinear inverted orbit growth (see Definition \[def:invorbit\]; recall, from Exercise \[exse:invgrowth\], that the property of having sublinear orbit growth is independent of a choice of basepoint) and subexponential inverted orbit choice growth. These are the hypotheses for Corollary \[cor:growthW\], which guarantee that $G\wr_X P$ has subexponential growth as soon as $G$ has subexponential growth. The main example we have in mind is a Grigorchuk group $P=G_\omega$ acting on $X=\8^\infty P$, for $\omega\in\{0,1,2\}^\infty$ containing infinitely many times each symbol, see Lemma \[lem:Gspreading\].
We also assume that there are rectifiable sequences in $X$, and (using the results of the previous section) we fix a rectifiable, spreading, locally stabilizing sequence $(x_0,x_1,\dots)$ of elements of $X$.
Finally, we fix a countable group $G$, and a finite or infinite sequence of elements $(b_1,b_2,\dots)$ generating $G$.
For an increasing finite or infinite sequence $0\le n(1) < n(2)
<\dots$ of integers, define $f\colon X\to G$ by $$f(x_{n(1)})=b_1,\qquad f(x_{n(2)})=b_2,\qquad \dots,\qquad f(x)=1\text{ for other }x.$$ The group $W_{(n)}=W_{n(1),n(2),\dots}$ is then defined as the subgroup $\langle P,f\rangle$ of the unrestricted wreath product $G{\operatorname{\wr\wr}}_X P$.
If all but finitely many $b_i$ are trivial, then $W_{(n)}$ has subexponential growth since then $G$ is finitely generated and $W_{(n)}$ is a subgroup of $G\wr_X P$. However, if $(n(i))$ is sparse enough then $W_{(n)}$ may have subexponential growth even if $f$ has infinite support:
\[prop:Wsubexp\] If $G$ has locally subexponential growth, then there exists an infinite sequence $(n(i))$ such that the group $W_{(n)}$ has subexponential growth.
The proof of Proposition \[prop:Wsubexp\] follows from a stronger, and independently interesting, statement: arbitrarily large balls in $W_{(n)}$ are approximable by groups of the form $W_{(n(1),\dots,n(i))}$, which have subexponential growth by the remark above:
\[prop:Wconvergence\] Assume that the sequence $(x_i)$ in $X$ is spreading and locally stabilizing, and that all elements $b_i$ have the same order.
Then for every increasing sequence $(m(i))$ there exists an increasing sequence $(n(i))$ with the following property: the ball of radius $m(i)$ in $W_{(n)}$ coincides with the ball of radius $m(i)$ in $W_{n(1),n(2),\dots,n(i)}$, via the natural identification $P\leftrightarrow P,f\leftrightarrow f$ between $W_{(n)}$ and $W_{n(1),\dots,n(i)}$.
Furthermore, the term $n(i)$ depends only on the previous terms $n(1),\dots,n(i-1)$, on the initial terms $m(1),\dots,m(i)$, and on the ball of radius $m(i)$ in the subgroup $\langle b_1,\dots,b_{i-1}\rangle$ of $G$.
Choose $n(i)$ such that $d(x_j,x_k)\ge m(i)$ for all $j\neq k$ with $k\ge n(i)$, and such that the balls of radius $m(i)$ around $x_{n(i)}$ and $x_j$ coincide for all $j>n(i)$.
Consider then an element $h\in W_{(n)}$ in the ball of radius $m(i)$, and write it in the form $h=(c,g)$ with $c\colon X\to G$ and $g\in P$. The function $c$ is a product of conjugates of $f$ by words of length $<m(i)$. Its support is therefore contained in the union of balls of radius $m(i)-1$ around the $x_j$, with $j$ either $\ge n(i)$ or of the form $n(k)$ for $k<i$. In particular, the entries of $c$ are in $\langle
b_1,\dots,b_{i-1}\rangle\cup\bigcup_{j\ge i}\langle b_j\rangle$. For $j>n(i)$, the restriction of $c$ to the ball around $x_j$ is determined by the restriction of $c$ to the ball around $x_{n(i)}$, via the identification $b_i\mapsto b_j$, because the neighbourhoods in $X$ coincide and all cyclic groups $\langle b_j\rangle$ are isomorphic.
It follows that the element $h\in W$ is uniquely determined by the corresponding element in $W_{n(1),\dots,n(i)}$.
Let $Z=\langle z\rangle$ be a cyclic group whose order (possibly $\infty$) is divisible by the order of all the $b_i$’s. We replace $G$ by $G\times Z$ and each $b_i$ by $b_i z$, so as to guarantee that all generators in $G$ have the same order.
Let $\epsilon_i$ be a decreasing sequence tending to $1$. We now construct a sequence $m(i)$ inductively, and obtain the sequence $n(i)$ by Proposition \[prop:Wconvergence\], making always sure that $m(i)$ depends only on $m(j),n(j)$ with $j<i$.
Denote by $v_i$ the growth function of the group $W_{n(1),\dots,n(i)}$. Since the group $W_{n(1),\dots,n(i)}$ is contained in $G\wr_X P$, it has subexponential growth. Therefore, there exists $m(i)$ be such that $$v_i(m(i))\le \epsilon_i^{m(i)}.$$ By Proposition \[prop:Wconvergence\], the terms $n(i+1),n(i+2),\dots$ can be chosen in such a manner that the balls of radius $m(i)$ coincide in $W_{(n)}$ and $W_{n(1),\dots,n(i)}$.
Denote now by $w$ the growth function of $W_{(n)}$. We then have $w(m(i))\le\epsilon_i^{m(i)}$ for all $i\in\N$. Therefore, $$w(R)\le\epsilon_i^{R+m(i)}\text{ for all }R>m(i),$$ so $\limsup\sqrt[R]{w(R)}\le\epsilon_i$ for all $i\in\N$. Thus the growth of $W_{(n)}$ is subexponential.
Finally, the rectifiability of the sequence $(x_i)$ guarantees that functions with singleton support and arbitrary values in $[G,G]$ belong to $W_{(n)}$ for all sequences $n$:
\[lem:contains \[G,G\]\] If the sequence $(x_i)$ is rectifiable, then $[W_{(n)},W_{(n)}]$ contains $[G,G]$ as a subgroup for all choices of $n=n(1)<n(2)<\cdots$.
We denote by $\iota\colon G\to G^X\rtimes P$ the imbedding of $G$ mapping the element $b\in G$ to the function $X\to G$ with value $b$ at $x_0$ and $1$ elsewhere. We abbreviate $W=W_{(n)}$. We shall show that $[W,W]$ contains $\iota([G,G])$. For this, denote by $H$ the subgroup $\iota([G,G])\cap[W,W]$.
We first consider an elementary commutator $g=[b_i,b_j]$. Let $g_i,g_j\in P$ respectively map $x_i,x_j$ to $x_0$, and be such that $g_i g_j^{-1}$ maps no $x_k$ to $x_\ell$ with $k\neq\ell$, except for $x_i g_i g_j^{-1}=x_j$. Consider $[f^{g_i},f^{g_j}]\in [W,W]$; it belongs to $G^X$, and has value $[b_i,b_j]$ at $x_0$ and is trivial elsewhere, so equals $\iota(g)$ and therefore $\iota(g)\in
H$.
We next show that $H$ is normal in $G^X$. For this, consider $h\in
H$. It suffices to show that $h^{\iota(b_i)}$ belongs to $H$ for all $i$. Now $h^{\iota(b_i)}=h^{f^{g_i}}$ belongs to $H$, and we are done.
We are now ready to complete the proof of Theorem \[thm:imbed\]. By Proposition \[prop:imbed’\], the countable, locally subexponentially growing group $B$ imbeds in $[G,G]$ for a countable, locally subexponentially growing group $G$. Let $(b_1,b_2,\dots)$ be a generating set for $G$. By Proposition \[prop:Wsubexp\], there exists an increasing sequence $(n(i))$ such that the group $W=W_{(n)}$ has subexponential growth. By Lemma \[lem:contains \[G,G\]\], $[G,G]$ imbeds in $[W,W]$, so $B$ imbeds in $[W,W]$ and we are done.
Give examples of sequences $(x_i)$ that are only spreading, or only stabilizing, and such that $W_{(n)}$ has exponential growth, even when the sequence $(n(i))$ grows arbitrarily fast.
Groups of non-uniform exponential growth {#ss:nug}
========================================
Recall that $v_{G,S}(R)$ denotes the growth function of a group $G$ generated by a finite set $S$. The *volume entropy* of $(G,S)$ is $$\lambda_{G,S}:=\lim_{R\to\infty}\frac{\log v_{G,S}(R)}R.$$ The limit exists because the function $v_{G,S}$ is submultiplicative (namely, $v_{G,S}(R_1+R_2)\le v_{G,S}(R_1) v_{G,S}(R_2)$.) Indeed, apply the following lemma to the sequence $(\log v_{G,S}(n))_{n\ge1}$:
Let $(a_n)$ be a subadditive sequence: $a_{n+m}\le a_n+a_m$ for all $m,n\ge1$. Then $\lim a_n/n$ exists and equals $\inf a_n/n$.
Set $A=\inf a_n/n$ and consider any $B>A$. Choose $k\ge1$ such that $a_k/k<B$. Every $n\in\N$ may be written in the form $n=r k+s$ with $r\in\N$ and $s\in\{0,\dots,k-1\}$. Thus $$\frac{a_n}n=\frac{a_{r k+s}}n\le\frac{r a_k+a_s}n=\frac{a_k}k\frac{n-s}n+\frac{a_s}n,$$ so $\limsup_{n\to\infty}a_n/n\le a_k/k\le B$. Since $B>A$ was arbitrary, $\limsup a_n/n\le L$ and we are done.
Furthermore, the following are equivalent: $G$ has exponential word growth; $\lambda_{G,S}>0$ for some generating set $S$; and $\lambda_{G,S}>0$ for all generating sets $S$.
Let $G$ be a group of exponential growth. Note that, even though $\lambda_{G,S}>0$ for all $S$, one might have $\lambda_{G,S_i}\to0$ along a sequence of generating sets $S_i$. It is easy to see that this cannot happen for $G$ a free group of rank $k\ge2$; indeed then each $S_i$ contains a subset of cardinality $k$ generating a free subgroup, so $\lambda_{G,S_i}\ge\log(2k-1)>0$ for all $i$. Let us say that a finitely generated group $G$ has *uniform exponential growth* if $\inf_S\lambda_{G,S}>0$, and *non-uniform exponential growth* if $\lambda_{G,S}>0$ for all $S$ yet $\inf_S\lambda_{G,S}=0$.
The existence of groups of non-uniform exponential growth is asked by Gromov in [@gromov:metriques]\*[Remarque 5.12]{}; see [@harpe:uniform] for a survey. There have been quite a few positive results: Osin showed in [@osin:entropy] that virtually solvable groups have uniform exponential growth unless they are virtually nilpotent; Eskin, Mozes and Oh obtained the same result in [@eskin-m-o:uniform] for finitely generated linear groups in characteristic $0$; Koubi showed in [@koubi:unifexpo] that word-hyperbolic groups have exponential growth unless they are virtually cyclic.
A *Golod-Shafarevich group* is a residually-$p$ group such that the associated Hopf algebra $\bigoplus_{n\ge0}\varpi^n/\varpi^{n+1}$ from §\[ss:algebralb\] has exponential growth; equivalently, the Lie algebra $\bigoplus_{n\ge1}\gamma_n(G)/\gamma_{n+1}(G)$ from has exponential growth. Since by Proposition \[prop:alg-gp\] the growth of $\overline{\Bbbk G}$ is always a lower bound for the growth of $G$, such groups have uniformly exponential growth.
Among groups $G$ of uniform exponential growth, one may ask whether the infimal entropy $\inf_{\langle S\rangle=G}\lambda_{G,S}$ is realised. Recall that a group $G$ is *Hopfian* if it is not isomorphic to a proper quotient of itself. Sambusetti proves in [@MR2000j:20056] that if $G$ is a free product $G=G_1*G_2$ with $G_1$ non-Hopfian and $G_2$ non-trivial, then $\inf_{\langle
S\rangle=G}\lambda_{G,S}$ is not attained.
It was widely suspected, since the appearance of groups of intermediate growth, that examples of groups of non-uniform exponential growth should exist. The first examples of groups of non-uniform exponential growth were exhibited by Wilson, see [@wilson:ueg]. We now give a simple construction showing that such groups abound:
\[thm:nueg\] Every countable group may be imbedded in a group of non-uniform exponential growth.
Furthermore, let $\eta_+\approx2.46$ denote the positive root of the polynomial $T^3-T^2-2T-4$. Then the group $W$ in which the countable group imbeds may be required to have the following property: there is a constant $K$ such that, for all $R>0$, there exists a generating set $S$ of $W$ with $$v_{W,S}(r) \le \exp(K r^{\log2/\log\eta_+})\text{ for all }r\in[0,R].$$
Let $B$ be a countable group. By Theorem \[thm:hnntorsion\], one may imbed $B$ into a group $G$ generated by two elements $s,t$ of order $5$. Without loss of generality, assume that $G$ has exponential growth (if needed, replace first $B$ by $B\times F_2$).
The group $W$ in which $G$ imbeds is the wreath product $G\wr_X G_{012}$ of $G$ with the first Grigorchuk group. We also consider $A=C_5\times C_5=\langle s',t'\rangle$, and the wreath product $W'=A\wr_X G_{012}$.
We consider the points $x_0=\9^\infty$ and $x_i=\9^\infty\8\9^i$ for all $i\ge1$ in the Schreier graph $X$, and the generating sets $S_i=\{a,b,c,d,s@x_0,t@x_i\}$ of $W$ and $S'=\{a,b,c,d,s'@x_0,t'@x_0\}$ of $W'$.
Note that the sequence $(x_i)$ is spreading and locally stabilizing (see Definitions \[def:spreading\] and \[def:locally stabilizing\]); better, for all $R$ the radius-$R$ balls around $x_0$ and $x_i$ in $X$ are isomorphic as labelled graphs for all $i$ large enough, because the action on a sequence $\in\{\9,\8\}^\infty$ of an element of $G_{012}$ of length $R$ depends only on the last $\lceil\log_2(R)\rceil$ symbols of the sequence.
We now claim that, for all $R\in\N$, there exists $i$ such that the balls of radius $R$ in the Cayley graphs of $(W,S_i)$ and $(W',S')$ coincide. By Theorem \[thm:grigWgrowth\], there exists a constant $K$ such that, for all $R\in\N$, the ball of radius $R$ in the Cayley graph of $(W',S')$ has cardinality $v_{W',S'}(R)\le\exp(K R^{\log2/\log\eta_+})$. Assuming the claim, we get $v_{W,S_i}(R)\le\exp(K R^{\log2/\log\eta_+})$, from which the second claim of the theorem follows.
For every $\epsilon>0$ there exists $R$ such that $K R^{\log2/\log\eta_+}<\epsilon R$, so $v_{W,S_i}(R)\le\exp(\epsilon R)$ and $\lambda_{W,S_i}\le\epsilon$ for some $i$. It follows then that $W$ has non-uniform exponential growth.
It remains to prove the claim. Given $R\in\N$, let $i$ be large enough so that the distance between $x_0$ and $x_i$ in $X$ is at least $2R$. Consider a word $w$ in $S_i$ of length $\le R$, and let $w'$ be the corresponding word in $S'$ obtained by replacing $s@x_0,t@x_i$ respectively by $s'@x_0,t'@x_0$. We show that $w$ represents the identity in $W$ if and only if $w'$ represents the identity in $W'$.
Write $w=(c,g)$ in $W$, with $c\colon X\to G$ and $g\in G_{012}$. Similarly, write $w'=(c',g)$ in $W'$, with $c'\colon X\to A$ and the same $g\in G_{012}$. Note that the support of $c$ is contained in the union of the balls of radius $R$ around $x_0$ and $x_i$, and these balls are isomorphic and disjoint. Therefore, $c$ can be written in the form $c=c_1c_2$ with $c_1\colon X\to\langle s\rangle$ and $c_2\colon X\to\langle t\rangle$, and $c_1,c_2$ have disjoint support so they commute. The function $c'$ may correspondingly be written as $c'=c'_1c'_2$ with $c'_1\colon X\to\langle s'\rangle$ obtained by composing $c_1$ with the isomorphism $\langle s\rangle\to\langle s'\rangle$, and $c'_2\colon X\to\langle t'\rangle$ obtained by composing the isomorphism from the radius-$R$ ball around $x_0$ to the radius-$R$ ball around $x_i$, the map $c_2$, and the isomorphism $\langle t\rangle\to\langle t'\rangle$. Therefore, $c'=1$ if and only if $c=1$, so the balls in $W$ and $W'$ are isomorphic.
=wncyr8
[^1]: Partially supported by ANR grant ANR-14-ACHN-0018-01 and DFG grant BA4197/6-1
[^2]: There is no need to require the action of $G$ on $X$ to be faithful; this is merely a visual aid. See §\[ss:wreath\] for the complete definition.
[^3]: The *rational subring* of $\Z G[[t]]$ is the smallest subring of $\Z G[[t]]$ containing $\Z G[t]$ and closed under Kleene’s star operation $A^*=1+A+A^2+\cdots$, for all $A(z)$ with $A(0)=0$.
The *algebraic subring* of $\Z G[[t]]$ is the set of power series that may be expressed as the solution $A_1$ of a non-trivial system of non-commutative polynomial equations $\{P_1(A_1,\dots,A_n)=0,\cdots,P_n(A_1,\dots,A_n)=0\}$ with coëfficients in $\Z G[t]$. The solution is actually rational if furthermore the $P_i$ are of the form $c_{i,0}+\sum_{j=1}^n c_{i,j}A_j$ with $c_{i,j}\in\Z G[t]$.
[^4]: i.e. a function $A(z)=A(\exp(2\pi i\tau))$ such that the corresponding function $\tau\mapsto A(\exp(2\pi i\tau))$ on the upper half plane is invariant under a finite-index subgroup of ${{\operatorname{SL}}}_2(\Z)$.
[^5]: Note that the side of the action changes! It is best to always use the appropriate side, so as to avoid inverses. Recall however that every left action can be converted into a right action by setting $f^g:={{}^{g^{-1}}\!f}$ and vice versa.
[^6]: namely, a map satisfying $\tau(\widetilde g)=g$ for all $g\in G$
[^7]: recall that the core of the subgroup $H$ is the intersection of its conjugates. $E/\operatorname{core}(H)$ is the natural permutation group acting faithfully on the left coset space $H\backslash E$.
[^8]: i.e. there are at least two orbits on $X$
[^9]: i.e. there is a non-trivial $G$-invariant equivalence relation on $X$
[^10]: This is usually called a *Busemann function*
[^11]: it is only here that we use the fact that $S$ is symmetric, to obtain $d(g,h)=d(h,g)$; in fact, it suffices in all that follows to assume that $S$ generates $G$ as a monoid.
[^12]: i.e. $d(t g,th)=d(g,h)$
[^13]: i.e. $S=\{x_1,x_1^{-1},\dots,x_k,x_k^{-1}\}$ and $F_k$ may be identified with reduced words over $S$.
[^14]: Let us not detail this too much; suffice it to say that the system must have a unique solution once its initial terms $f_1(0),\dots,f_n(0)$ have been fixed.
[^15]: A property is said to *virtually* hold if it holds for a finite-index subgroup.
[^16]: A property is said to hold *residually* if for every non-trivial element there exists a quotient in which this element remains non-trivial and the property holds.
[^17]: i.e. cycles that traverse each edge once
| {
"pile_set_name": "ArXiv"
} |
****
Juan Orendain
We study the problem of existence of essential indecomposable submodules of direct sums of copies of the ring of quotients of Ore domains. We provide, for each Ore domain $D$, with at least three non-associate central irreducibles, and with no non-zero infinitely divisible elements, a lower bound for the supremum of all cardinals $\kappa$ such that the direct sum of $\kappa$ copies of the ring of quotients of $D$ contains indecomposable essential $D$-submodules, and a lower bound for the number of times this bound is attained up to isomorphisms. We provide explicit examples illustrating these results.
Introduction
============
Solutions to the problem of existence of essential indecomposable submodules of direct sums of copies of the ring of quotients of the integers, that is, solutions to the problem of existence of indecomposable torsion-free abelian groups of given ranks, can be traced back to [@F4] or even [@K]. We study, using essentially ideas appearing in [@F4], the more general problem of existence and uniqueness of essential indecomposable submodules of direct sums of copies of the ring of quotients of Ore domains. For a given Ore domain $D$, admitting sets of at least three non-associate central irreducibles, and not admitting non-zero infinitely divisible elements, we provide a lower bound for the supremum of all cardinals $\kappa$ such that the direct sum of $\kappa$ copies of the ring of quotients of $D$ admits essential indecomposable $D$-submodules, and a lower bound for the number of times this bound is attained up to isomorphisms. We provide explicit examples of these types of calculations.
In what follows the letter $D$ will, unless otherwise stated, always stand for a left Ore domain, the letter $E$ will always denote the left ring of quotients of $D$, the word *module* will mean a left unital $D$-module, and the word *morphism* will always mean a morphism in $D$-Mod. The choice of side on which the Ore condition is required on $D$ is arbitrary, and it is made for convenience, it should be evident nevertheless that all statements and proofs work for right Ore domains just as well. We will say that an element $a\in D$ is irreducible in $D$ if the equation $a=bc$ in $D$ implies that either $b$ or $c$ is a unit in $D$. Given central irreducibles $a,b\in D$, we will say that $a$ and $b$ are associates in $D$ if there exists a unit $u\in D$ such that $a=ub$. We will say that a non-zero element $a\in R$ is infinitely left divisible in $R$ if there exists $b\in R\setminus U(R)$ such that for every $n\geq 1$, there exists $x_n\in R$ such that $a=b^nx_n$. We will denote by $\mathfrak{S}$ the class of all modules $M$ such that $M$ can be embedded in a direct sum of copies of $E$. Given $M\in\mathfrak{S}$, we will write $rk(M)$ for the minimal cardinal $\kappa$ such that $M$ can be embedded in $E^{(\kappa)}$. Finally, we will denote by $\chi(D)$ the supremum of all cardinals $\kappa$ such that there exists an indecomposable module $M$ such that $rk(M)=\kappa$. $\chi(D)$ will thus take values in the class of all cardinals together with an artificial element $*$ satisfying the convention that $\kappa+*=*$ for every cardinal $\kappa$. Observe that $\chi(D)=1$ if $D$ is a division ring. More generally, in [@K] it is proven that if $D$ is a valuation ring, then $\chi(D)=1$ if and only if $D$ is a maximal discrete valuation ring. In particular, for every prime $p$, the ring of $p$-adic integers, $\mathbb{Z}_p$, satisfies equation
$$\chi(\mathbb{Z}_p)=1$$
We will be interested in statements of the form:
$$\chi(D)\geq \kappa$$
*and this bound is attained at least $\mu$ times up to isomorphisms*. Any such statement should be translated as: *There exist at least $\mu$ non-isomorphic indecomposable modules $M$, in $\mathfrak{S}$, such that $rk(M)=\kappa$*. Or equivalently, as: *There exist at least $\mu$ non-isomorphic essential indecomposable submodules of the direct sum of $\kappa$ copies of $E$*. The following will be the main result of this note
Let $\kappa_1,\kappa_2$ be cardinals. Suppose $D$ admits sets of $2\kappa_1+\kappa_2$ non-associate central irreducibles and no non-zero element of $R$ is infinitely left divisible in $R$. Then
$$\chi(D)\geq 2^{\kappa_1}$$
and this bound is attained at least $2^{\kappa_2}-1$ times up to isomorphisms.
We postpone a proof of Theorem 1.1 until section 3. In section 2 we present explicit examples of Ore domains satisfying the conditions of Theorem 1.1.
Acknowledgements {#acknowledgements .unnumbered}
================
The author would like to dedicate this paper to the late professor Francisco Raggi, without whose attention, guidance, and friendship this paper would have not been possible
Examples
========
In this section we provide explicit examples of Ore domains satisfying the conditions of Theorem 1.1. Observe first that left noetherian domains are examples of left Ore domains that do not admit non-zero infinite left divisible elements. We begin by presenting examples of commutative domains satisfying the conditions of Theorem 1.1.
1. Let $a\in\mathbb{Z}$ be squarefree. Let $\Omega_1,\Omega_2$ be (possibly empty) finite sets such that $\Omega_1\cup\Omega_2$ is algebraically independent over $\mathbb{Z}$. If in this case we make $D$ to be equal to $\mathbb{Z}[\sqrt{a}][\Omega_1][[\Omega_2]]$, then $D$ is a commutative noetherian domain admitting sets of $\aleph_0$ non-associate irreducibles. Thus, by Theorem 1.1 we have
$$\chi(D)\geq 2^{\aleph_0}$$
and this bound is attained at least $2^{\aleph_0}$ times up to isomorphisms. We consider this result as a generalization of results obtained in [@F3].
2. Let $k$ be a field. Let $\Omega_1,\Omega_2$ be finite sets such that $\Omega_1\cup\Omega_2$ is algebraically independent over $k$. Suppose that $\Omega_1\neq\emptyset$. If in this case we make $D$ to be equal to $k[\Omega_1][[\Omega_2]]$, then $D$ is again a commutative noetherian domain, now admitting sets of $\aleph_0\left|k\right|$ non-associate irreducibles. Thus, by Theorem 1.1 we now have
$$\chi(D)\geq 2^{\aleph_0\left|k\right|}$$
and this bound is attained at least $2^{\aleph_0\left|k\right|}$ times up to isomorphisms.
3. Let $R$ be a unique factorization domain. Let $a\in R$. Suppose $a$ admits $n$ ($n\geq 3$) non-associate prime divisors in $R$. If in this case we make $D$ to be equal to the localization $R_{(a)}$, then $D$ is a commutative domain not admitting infinitely divisible elements and admitting exactly $n$ non-associate irreducibles. Thus, by Theorem 1.1 we now have
$$\chi(D)\geq 2^{\left\lfloor n/2\right\rfloor-1}$$
4. Let $k$ be an algebraically closed field. Let $V$ be an algebraic variety over $k$. Let $a\in V$. Suppose $V$ is smooth at $a$. In this case, if we make $D$ to be equal to $\mathcal{O}_{V,a}$, then $D$ is, by the Auslander-Buchsbaum Theorem [@AB], a noetherian UFD. Thus, if $V$ admits sets of $2\kappa_1+\kappa_2$ subvarieties containing $a$, $D$ admits sets of $2\kappa_1+\kappa_2$ non associate irreducibles. In particular, if $V$ is a hyperplane of dimension $\geq 2$, by Theorem 1.1 we have
$$\chi(D)\geq 2^{\left|k\right|}$$
and this bound is attained at least $2^{\left|k\right|}$ times up to isomorphisms.
Observe that example 1 above generalizes the results of [@F3]. Observe also that the bounds obtained in this example are weak compared to those obtained in [@F4]. We regard the following conjecture as a possible generalization of the results of [@F4] and a possible improvement of Theorem 1.1
There exists a model in ZFC in which, for every pair of cardinals $\kappa_1,\kappa_2$, and for every Ore domain $D$, such that $D$ admits infinite sets of non-associate central irreducibles and no non-zero element of $D$ is infinitely left divisible in $R$, we have
$$\chi(D)\geq 2^{\kappa_1}$$
and this bound is attained at least $2^{\kappa_2}-1$ times up to isomorphisms.
We now present examples of non-commutative Ore domains satisfying the conditions of Theorem 1.1. Let $R$ be a commutative noetherian domain, let $\sigma$ be an automorphism of $R$, and let $\delta$ be a $\sigma$-derivation of $R$. In this case, the skew polynomial ring $R[y;\sigma,\delta]$ is a left-right noetherian domain (see [@R]), non-commutative if and only if $\sigma\neq id$ or $\delta\neq 0$. Observe that, in this case, every irreducible in $R$ is irreducible in $R[y;\sigma,\delta]$, as a degree 0 polynomial, and that a central element $a\in R$ is central in $R[y;\sigma,\delta]$, as a degree 0 polynomial, if and only if identities
$$\sigma(a)=a \ \mbox{and} \ \delta(a)=0$$
hold. That is, an element $a\in R$ is central in $R[y;\sigma,\delta]$, as a degree 0 polynomial, if and only if $a$ is central in $R$, $a$ is fixed by $\sigma$, and $a$ is a constant with respect to $\delta$. Moreover, observe that two central irreducibles in $R$ are associates in $R$ if and only if they are associates in $R[y;\sigma,\delta]$. We conclude that, if $R$ is a commutative noetherian domain, $\sigma$ is an automorphism of $R$ and $\delta$ is a $\sigma$-derivation of $R$, $\sigma\neq id$ or $\delta\neq 0$, and $R$ admits sets of $\kappa$ non-associate central irreducibles, fixed by $\sigma$, and constant with respect to $\delta$, then $R[y;\sigma,\delta]$ is a non-commutative noetherian domain, admitting sets of $\kappa$ non-associate central irreducibles. The following are explicit examples of non-commutative noetherian domains, satisfying the conditions of Theorem 1.1, given by the above construction.
1. Let $S$ be a commutative noetherian domain. Let $\Omega$ be a finite nonempty set, algebraically independent over $S$. Let $\omega\in\Omega$. Let $R$ be the ring of ploynomials $S[\Omega]$, let $\sigma$ be equal to $0$ and let $\delta$ be equal to the formal partial derivative $\partial/\partial\omega$. Suppose $S$ admits sets of $\kappa$ non-associate irreducibles. If in this case we make $D$ to be equal to $R[y;\delta]$, then $D$ is a non-commutative noetherian domain admitting sets of $\kappa$ non-associate central irreducibles. In particular, if we make $S$ to be equal to $\mathbb{Z}[\sqrt{a}][\Omega_1][[\Omega_2]]$ as in example 1 above, by Theorem 1.1 we have
$$\chi(D)\geq 2^{\aleph_0}$$
and this bound is attained at least $2^{\aleph_0}$ times up to isomorphisms. Moreover, if we make $S$ to be equal to $k[\Omega_1][[\Omega_2]]$ as in example 2 above, again, by Theorem 1.1 we have
$$\chi(D)\geq 2^{\aleph_0\left|k\right|}$$
and this bound is attained at least $2^{\aleph_0\left|k\right|}$ times up to isomorphisms. Finally, if we make $S$ to be equal to $R_{(a)}$ as in example 3 above, again, by Theorem 1.1 we have
$$\chi(D)\geq 2^{\left\lfloor n/2\right\rfloor}-1$$
2. Let $k$ be a field. Let $\sigma$ be an automorphism of $k$ such that $\sigma\neq id$. Let $\Omega$ be a finite set, algebraically independent over $k$. Let $R$ be equal to $k[\Omega]$. Let $\tilde{\sigma}$ be the only extension of $\sigma$ to $R$ such that $\tilde{\sigma}(\omega)=\omega$ for every $\omega\in\Omega$. Suppose $\sigma$ fixes nonzero elements of $k$. If in this case we make $D$ to be equal to $R[y;\tilde{\sigma}]$, then $D$ is a non-commutative noetherian domain admitting sets of $\aleph_0\left|k\right|$ non-associate central irreducibles. In particular, if we make $k$ to be equal to the algebraic closure, $\overline{\mathbb{F}}_p$, of $\mathbb{F}_p$ for some prime $p$, and we make $\sigma$ to be equal to any power of the Frobenius automorphism of $k$, then, by Theorem 1.1 we have
$$\chi(D)\geq 2^{\aleph_0}$$
and this bound is attained at least $2^{\aleph_0}$ times up to isomorphisms. Moreover, if we now make $k$ to be equal to $\mathbb{C}$ and $\sigma$ to be equal to the complex conjugation in $\mathbb{C}$, then, again by Theorem 1.1 we have
$$\chi(D)\geq 2^\mathfrak{c}$$
and this bound is attained at least $2^\mathfrak{c}$ times up to isomorphisms.
We end this section with the following problem related to examples 1 and 2 presented above.
Compute lower bounds for the following cardinals.
1. $\chi(k[y;\delta])$, where $(k,\delta)$ is any differential field.
2. $\chi(k[x][y;d])$, where $k$ is any field and $d$ denotes the formal derivative, with respect to variable $x$, in $k[x]$.
3. $\chi(\overline{\mathbb{F}}_p[y;\tilde{\sigma}])$, where $\sigma$ denotes the Frobenius automorphism in $\overline{\mathbb{F}}_p$
4. $\chi(\mathbb{C}[y,\tilde{\sigma}])$, where $\sigma$ denotes the complex conjugation on $\mathbb{C}$.
Proof of Theorem 1.1
====================
In this final section we reduce the proof of Theorem 1.1 to the proof of a couple of lemmas. Observe first that the rank, $rk(M)$, of a module $M\in\mathfrak{S}$, is characterized by equation
$$dim_EE\otimes_D M=rk(M)$$
or equivalently, by the fact that $M$ can be essentially embedded in $E^{(rk(M))}.$
Let $\left\{M_\lambda:\lambda\in\Lambda\right\}$ be a collection of modules. Let $\lambda_0\in \Lambda$ and for each $\lambda\in\Lambda$ such that $\lambda\neq \lambda_0$, let $a_\lambda\in M_\lambda$, $b_\lambda\in M_{\lambda_0}$, and $\xi_\lambda\in D\setminus \left\{0\right\}$. Suppose
1. $M_\lambda$ is indecomposable for all $\lambda\in\Lambda$.
2. $Hom(M_\lambda ,M_\mu)=\left\{0\right\}$ for all $\lambda, \mu\in\Lambda$ with $\lambda\neq\mu$.
3. $a_\lambda$ is not divisible by $\xi_\lambda$ in $M_\lambda$ for each $\lambda\in\Lambda\setminus\left\{\lambda_0\right\}$.
4. $b_\lambda$ is not divisible by $\xi_\lambda$ in $M_{\lambda_0}$ for each $\lambda\in\Lambda\setminus\left\{\lambda_0\right\}$.
Then the $D$-submodule
$$M=\bigoplus_{\lambda\in\Lambda}M_\lambda +\sum_{\lambda\in\Lambda\setminus\left\{\lambda_0\right\}}D\xi_\lambda^{-1}(a_\lambda+b_\lambda)$$
of $\bigoplus_{\lambda\in\Lambda}(E\otimes_D M_\lambda)$ is indecomposable. Moreover, if $M_\lambda\in\mathfrak{S}$ for each $\lambda\in \Lambda$, then
$$rk(M)=\left|\Lambda\right|\sum_{\lambda\in \Lambda}rk(M_\lambda)$$
Let $M=N\oplus L$ be a direct sum decomposition of $M$. By condition 2 above, $M_\lambda$ is strongly invariant in $M$ for each $\lambda\in \Lambda$. The identity
$$M_\lambda=(M_\lambda\cap L)\oplus (M_\lambda\cap N)$$
follows for all $\lambda\in \Lambda$. Thus, since by condition 1 above $M_\lambda$ is indecomposable for each $\lambda\in\Lambda$, each $M_\lambda$ is contained in either $N$ or $L$. Suppose $M_{\lambda_0}\leq N$. Suppose now that there exists $\lambda\in\Lambda$ such that $M_\lambda\leq L$. In this case equation
$$\xi_\lambda^{-1}(a_{\lambda}+b_\lambda)=x+y$$
with $x\in N$ and $y\in L$, implies that $\xi_\lambda$ divides $a_\lambda$ in $L$ and $b_\lambda$ in $N$, which in turn implies that $\xi_\lambda$ divides both $a_\lambda$ and $b_\lambda$ in $M$, a contradiction, since by conditions 3 and 4 above, niether $a_\lambda$ nor $b_\lambda$ are divisible by $\xi_\lambda$ in $\bigoplus_{\lambda\in\Lambda}M_\lambda$, and from this and from the definition of $M$ it follows that niether $a_\lambda$ nor $b_\lambda$ are divisible by $\xi_\lambda$ in $M$. It follows that $M_\lambda\leq N$ for all $\lambda\in \Lambda$. Now, since $\bigoplus_{\lambda\in\Lambda}M_\lambda$ is clearly essential in $M$ we conclude that $M$ is indecomposable. The conclusion on the rank of $M$ is trivial.
Let $\kappa_1,\kappa_2$ be cardinals. Suppose $D$ admits sets of $2\kappa_1+\kappa_2$ non-associate central irreducibles and that $D$ does not admit non-zero infinitely left divisible elements. Then, in this case, there exists a collection of modules $\left\{M_\lambda:\lambda\in\Lambda\right\}$, of cardinality $\kappa_1$, an element $\lambda_0\in\Lambda$, and elements $a_\lambda\in M_\lambda$, $b_\lambda\in M_{\lambda_0}$, and $\xi_\lambda\in D\setminus \left\{0\right\}$, $\lambda\in\Lambda\setminus\left\{\lambda_0\right\}$, all satisfying the conditions of Lemma 3.1.
Let $\left\{a_\gamma:\gamma\in \Gamma \right\}, \left\{b_\gamma:\gamma \in\Gamma\right\}$, $\left\{c_\delta :\delta\in\Delta\right\}$ be three disjoint sets consisting of different central irreducibles in $D$ such that no two elements in the union of the three sets are associates in $D$. Suppose $\left|\Gamma\right|=\kappa_1$ and $\left|\Delta\right|=\kappa_2$. Let $\Theta$ be the collection of all sets of the form
$$\left\{d_\gamma: d_\gamma\in \left\{a_\gamma ,b_\gamma\right\} \forall\gamma\in\Gamma\right\}$$
It is immediate that $\left|\Theta\right|=2^{\kappa_1}$. Now, for each $A\in \Theta$, let $M_A$ be the following submodule of $E$
$$\sum_{d_{\gamma}\in A}D[1/d_\gamma]$$
For each $A\in \Theta$, $M_A$ thus defined, being a submodule of the uniform module $E$, is itself uniform, and thus indecomposable, that is, the collection $\left\{M_A:A\in \Theta\right\}$ satisfies condition 1 of Lemma 3.1. Now, let $A,B\in \Theta$ such that $A\neq B$, we prove that $Hom(M_A ,M_B)=\left\{0\right\}$. From the definition of $\Theta$ it is clear that in this case niether $A$ is contained in $B$ nor $B$ is contained in $A$. Let $d_{\gamma}\in A\setminus B$. Let $\Phi\in Hom(M_A ,M_B)$. From the fact that for each $\gamma\in\Gamma$, $a_\gamma$ and $b_\gamma$ are non-associate, central, and irreducible, it is easily seen that no non-zero element in $M_B$ is infinitely divisible by $d_{\gamma}$. From this and from the fact that $1_D$ is infinitely divisible by $d_\gamma$ in $M_A$ it follows that $\Phi(1_D)=0$, and since $D$ is essential in $M_A$, we conclude, from this, that $\Phi=0$. Thus the collection $\left\{M_A:A\in \Theta\right\}$ satisfies condition 2 of Lemma 3.1. Finally, let $\Theta$ play the rôle of $\Lambda$, let $A_0\in \Theta$ play the rôle of $\lambda_0$, and let any sequence formed by elements in $\left\{c_\delta :\delta\in\Delta\right\}$ play the rôle of the sequence $\xi_\lambda$, $\lambda\in\Lambda\setminus\left\{\lambda_0\right\}$ in Lemma 3.1. Finally, let any sequence of elements in $A$ and in any sequence of elements in $A_0$, play the rôle of elements $a_\lambda$ and $b_\lambda$ in Lemma 3.1. It is immediate that in this case collection $\left\{M_A :A\in \Theta\right\}$, together with $A_0$ and corresponding subsets of $\left\{c_\delta :\delta\in\Delta\right\}$, $A$, and $A_0$ satisfy the conditions of Lemma 3.1. This concludes the proof.
With the aid of Lemma 3.1 and Lemma 3.2 we now prove Theorem 1.1.
***Proof of theorem 1.1*** Let $\left\{a_\gamma:\gamma\in\Gamma\right\}$, $\left\{b_\gamma:\gamma\in\Gamma\right\}$, $\left\{c_\delta:\delta\in\Delta\right\}$, and $\Theta$ be as in the proof of lemma 3.2. It follows, from Lemma 3.1 and from the proof of Lemma 3.2 that every sequence with indeces in $\Theta$, formed by elements in any non-empty subset of $\left\{c_\delta:\delta\in\Delta\right\}$, defines an indecomposable module of rank $2^{\kappa_1}$. Let $S,T$ be two different non-empty subsets of $\left\{c_\delta:\delta\in\Delta\right\}$. Let $s,t$ be two sequences with indeces in $\Theta$, formed by elements in $S$ and $T$ respectively. Finally, let $M_s,M_t$ be indecomposable modules of rank $2^{\kappa_1}$, defined, as in the proof of Lemma 3.2, by sequences $s$ and $t$ respectively. Suppose, without any loss of generality, that $S\setminus T\neq \emptyset$. Let $a\in S\setminus T$. Then, by the way they were constructed, $M_s$ admits non-zero elements divisible by $a$, while $M_t$ does not. It follows that $M_s$ and $M_t$ are non-ismorphic. We conclude, from this, from Lemma 3.1, and from Lemma 3.2, that there exist at least $2^{\kappa_2}$ non-ismorphic indecomposable modules of rank $2^{\kappa_1}$. This concludes the proof. $\blacksquare$
[50]{}\[bibliography\]
F. Anderson, K. Fuller, *Rings and Categories of Modules*, Springer, Verlag (1991).
M. Auslander, D. A.Buchsbaum. *Unique Factorization in Regular Local Rings*, Proc. Natl. Acad. Sci. USA, **45**, (1959), 733-734.
J. Cozzens, *Homological Properties of the Ring of Differential Polynomials*, Bull. Amer. Math. Soc., **76**, number 1, (1970), 75-79.
D. S. Dummit, R. M. Foote, *Abstract Algebra*, John Wiley and Sons (2004).
Fuchs, L. *Infinite Abelian Groups Volume I*, Elsevier (1970).
Fuchs, L. *Infinite Abelian Groups Volume II*, Elsevier (1973).
L. Fuchs, *On a Directly Indecomposable Abelian Group of Power Greater than Continuum*, Acta. Math. Acad. Sci. Hungar. **8** (1957), 453-454.
L.Fuchs, *The Existence of Indecomposable Abelian Groups of Arbitrary Power*, Acta Math. Acad. Sci. Hungar. **10** (1959), 453-457.
Kaplansky, I. *Infinite abelian groups* University of Mighiga Press, Ann Arbor (1954).
Lam, TY. *Lectures on modules and rings*, Graduate Texts in Mathematics No. 189, Springer-Verlag (1999).
Lam, TY. *Exercises in modules and rings*, Problem Books in Mathematics, Springer-Verlag (2007).
A. Nowicki, *Polynomial Derivations and their Rings of Constants*, N. Copernicus University, Torun, (1994).
A. Nowicki, *Derivations of Ore Extensions of the Polynomial Ring in One Variable*, Communications in Algebra **32** number 9, (2004) 3651-3672.
A. Nowicki, *Local Derivations of Ore Extensions of the Polynomial Ring in One Variable*, Communications in Algebra **32** number 12, (2004) 4559-4571.
Rowen, Louis H. *Ring theory, vol. I, II*, Pure and Applied Mathematics, **127**, **128**, Boston, MA (1988).
Stenström, Bo *Rings and modules of quotients*, Lecture Notes in Mathematics, Vol. 237, Berlin: Springer-Verlag (1971).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Non intrusive load monitoring (NILM), or energy disaggregation, is the process of separating the total electricity consumption of a building as measured at single point into the building’s constituent loads. Previous research in the field has mostly focused on residential buildings, and although the potential benefits of applying this technology to commercial buildings have been recognised since the field’s conception, NILM in the commercial domain has been largely unexplored by the academic community. As a result of the heterogeneity of this section of the building stock (i.e., encompassing buildings as diverse as airports, malls and coffee shops), and hence the loads within them, many of the solutions developed for residential energy disaggregation do not apply directly. In this paper we highlight some insights for NILM in the commercial domain using data collected from a large smart meter deployment within an educational campus in Delhi, India, of which a subset of the data has been released for public use. We present an empirical characterisation of loads in commercial buildings, highlighting the differences in energy consumption and load characteristics between residential and commercial buildings. We assess the validity of the assumptions generally made by NILM solutions for residential buildings when applied to measurements from commercial facilities. Based on our observations, we discuss the required traits for a NILM system for commercial buildings, and run benchmark residential NILM algorithms on our data set to confirm our observations. To advance the research in commercial buildings energy disaggregation, we release a subset of our data set, called COMBED (commercial building energy data set).'
author:
- |
Nipun Batra\
\
\
Oliver Parson\
\
\
Mario Berges\
\
\
- |
Amarjeet Singh\
\
\
Alex Rogers\
\
\
bibliography:
- 'reference.bib'
title: 'A comparison of non-intrusive load monitoring methods for commercial and residential buildings'
---
Introduction
============
Energy disaggregation, also known as non-intrusive load monitoring (NILM), or non-intrusive appliance load monitoring (NIALM), can be broadly defined as a set of techniques used to obtain estimates of the electrical power consumption of individual appliances from measurements of voltage and/or current taken at a limited number of locations in the power distribution system of a building. The original ideas trace their roots back to the late 1980’s, when George Hart at the Massachusetts Institute of Technology (MIT) in the US proposed using changes in real and reactive power consumption, as measured at the utility meter, to track the operation of individual appliances in a home automatically. Since then, a great number of commercial, academic and government efforts have developed and tested methods to solve this problem [@armel_2013; @zeifman_2011; @zoha].
The majority of the solutions developed to date have focused on residential buildings, despite the fact that the potential benefits of applying this technology to commercial and/or industrial buildings have been recognised from the field’s foundation [@norford1996]. Early assessments of the technology, detailed in reports by the Electric Power Research Institute in the US [@epri_1997_1; @epri_1997_2], and later in a report prepared for the the California Energy Commission’s Public Interest Energy Research (PIER) program [@smith2003final], concluded that there were issues with the technology’s ability to deal with multi-state and variable-power devices, as well as multiple units of devices of the same make/model. These issues, although present to some extent in residential buildings, are much more prevalent in commercial facilities, which may explain why the technology development has focused primarily on the former.
Among the issues found by early assessments of the technology, the impact of complex load types on the performance of the algorithms is of particular importance. Two-state appliances (i.e., those that only operate in one of two states, typically [*on*]{} and [*off*]{}), are easier to track than those that have more complicated operations due to the number of states that they can be in and/or the possible transitions between those states. Variable loads, such as variable-speed drives, are one extreme form of a multi-state device.
Other types of loads that are more prevalent in commercial buildings can also complicate the disaggregation problem. For example, power factor correction devices, which introduce additional harmonics and noise into the building’s power distribution system, fall into this category. This is especially true when high-frequency features are used to solve the disaggregation problem. Similarly, depending on the electrical point at which the measurements are taken, the integration of additional distributed energy resources such as micro-turbines, battery storage technologies (e.g., electric vehicles), and photovoltaic panels, are likely to impact the performance of load disaggregation algorithms.
There are, however, some positive aspects for deploying NILM solutions in this sector:
1. Due to their complexity and magnitude, commercial buildings are typically managed by an experienced facility management services crew. These are likely to be the primary customers and users of NILM technologies, and given that this solution directly impacts their daily work activities, they are more likely to be motivated to embrace the technology.
2. Commercial buildings are more likely to have a detailed asset management system, which can be used to assist in creating the database of appliance signatures for the NILM system.
3. Commercial facilities are more likely to have existing building automation systems, that are used to monitor and control specific building systems. These measurement and control points can provide very useful information that can assist the disaggregation engine.
4. There are significantly more opportunities for performing automated fault detection and diagnosis and system-specific efficiency analyses.
5. A higher return on investment on each measurement point is more likely, given the value of the assets present in commercial facilities, and the amount of energy consumed by the devices.
6. Some commercial facilities tend to have more regular, periodic and predictable operation given that they are based on fixed work schedules.
7. Due to the nature of the services being provided by these facilities, automated control of end-use loads is more likely to be accepted and sought after by the building constituents (owners, managers, occupants). This, in turn, increases the value proposition for the hardware installations.
8. Regular energy audits for several commercial units (e.g. cement industry in India) are often mandated for compliance purposes. Detailed analysis from NILM can further help in simpler execution of cumbersome energy audits that currently require multiple measurements across the facility.
In this paper, we explore these challenges and opportunities in more detail, by making use of a year-long data set collected at an academic campus with dense sub-metering. While there exists a vast diversity of commercial buildings, such as airports, IT offices and malls, we shall focus on IT offices, due to the lack of publicly available data for other building types. Many of the buildings from our data set can be considered to be similar to IT offices where HVAC and IT loads are the primary energy consumers. We use our data set to highlight the key differences in energy consumption between residential and commercial buildings, specifically from a NILM perspective. Specifically, the primary contributions of our work are:
1. We describe a data set collected from an extensive deployment of smart meters across an educational campus in India. A subset of our data set, for one building, constituting data at different levels (buildings, floors and major loads), is released with this paper.
2. We assess the validity of the assumptions generally made by NILM solutions for residential buildings, when applied to commercial facilities.
3. We perform an empirical characterisation of the major loads in our data set, providing insights into the traits for a NILM system for commercial buildings.
4. We apply benchmark residential NILM algorithms to our data set to further validate the developed insights. This is the first time that these well known NILM algorithms have been applied to a commercial setting.
The remainder of this paper is structured as follows. In [Section \[related\_work\]]{} we discuss relevant energy disaggregation work from both the commercial and the residential domains. In [Section \[assumptions\]]{} we introduce the set of assumptions that are typically made in NILM solutions for residential settings. We then describe our deployment in a university campus in Delhi, India in [Section \[deployment\]]{}. Data from our deployment, together with two publicly released data sets for residential NILM are used to validate the applicability of residential NILM methods for commercial setting in [Section \[comparison\]]{}. As part of this validation, we further develop insights into useful traits for commercial NILM methods. Thereafter, in [Section \[evaluation\]]{}, we apply benchmark residential NILM algorithms to our data and validate the developed insights. Finally, conclusions and future work are discussed in [Section \[conclusion\]]{}.
Related work {#related_work}
============
The most significant work which has applied energy disaggregation to commercial buildings was completed at MIT during 2001-2003. Norford, Xing and Luo presented an approach based on the generalised likelihood ratio and applied it to disaggregate a set of chillers, fans and pumps in a single story academic test building [@norford_2001]. The authors collected aggregate and sub-metered data at 24 Hz, and identified 1 second as the minimum sample rate required to identify most appliance switch events. However, the approach required sub-metered appliance data to first be collected for training purposes, and furthermore the authors describe how the proposed approach was unable to detect appliances containing variable speed drives, such as fans, due the start up transients which lasted multiple minutes.
Norford and Lee later applied an extension of the previous approach to one floor of an office building [@norford_2001a]. The authors collected aggregate current and voltage data at 120 Hz, from which the proposed approach was able to distinguish between the different HVAC loads that were in use. However, the approach required a manual training phase in which each load was operated multiple times in isolation of the operation of other appliances. In addition, the authors also noted the difficulty of identifying slowly varying loads or multi-step change loads.
Lee, Norford and Leeb then went on to further extend this work to use both step changes in the aggregate load and longer transients to disaggregate the various loads within a building [@lee_2003]. The authors collected aggregate current and voltage data at 16 kHz from two commercial buildings: one floor of an office building and the laundry room of a dormitory. The proposed approach made extensive use of high frequency data, from which higher order harmonics could be extracted from the current signal, allowing loads to be disaggregated to a higher level of accuracy. However, the authors again highlight the need for a manual training phase required to first learn the behaviour of each load before their approach was able to disaggregate the loads. Lee et al. have also developed Gaussian random process models for loads with variable speed drives by exploiting the correlations between mean harmonic powers [@lee2005estimation]. Recently, Shao et al. [@shao] used temporal motif mining techniques for load disaggregation in both residential and commercial settings.Their approach exploits the fact that some loads in commercial buildings are run on regular schedules. While their approach is unsupervised, it requires one to specify the number of appliances which may be present in a home or the commercial building.
In summary, load disaggregation research for commercial buildings has increasingly focused on high-frequency (kHz) methods which require a manual training phase for each building due to the lack of generality of individual load features. In contrast, recent relevant work in the residential domain has started to apply low-frequency (< 1 Hz) methods to individual households which do not require a manual training phase. Such developments have been fuelled by the release of a range of residential data sets designed specifically for energy disaggregation, including REDD [@redd], BLUED [@blued], Smart\* [@smart], AMPds [@ampds], iAWE [@iawe], ECO [@eco], GREEND [@greend], Sust-data [@sustdata] and UK-DALE [@UK-DALE]. To the best of our knowledge, only two commercial data sets have been released to date. While Enernoc[^1] collected data across different industries, the data set contains only aggregate level readings and is sampled at 5 minute resolution. More recently, the BERDS [@maasoumyberds] data set was released, which provides the power consumption of a selected number of loads at 15 minute resolution. Neither of these data sets contain 1 Hz data from the majority of appliances in the monitored buildings, and therefore have limited the progress of NILM research in the commercial domain.
Assumptions in residential NILM methods {#assumptions}
=======================================
Having discussed relevant literature for NILM in commercial settings, we now introduce four key assumptions that are often made by residential disaggregation methods.
**One-at-a-time assumption:** Many energy disaggregation methods applied within the residential domain aim to detect appliance switch events from aggregate data. However, it is possible that the switch events of two or more loads might overlap (i.e., multiple appliances change state between two sequential samples of the building’s aggregate load). Furthermore, the likelihood of overlapping appliance switch events increases as the sampling rate decreases. Accordingly, many residential NILM approaches make the one-at-a-time assumption, i.e. only one appliance will change state between two consecutive samples of the aggregate load [@kolter_2012; @liang_2010a]. Although such overlaps rarely occur given a sufficiently high sampling rate, it is important that disaggregation systems operating under this assumption may still generate only minor errors as a result of overlapping events rather than failing to accurately disaggregate any appliances.
**Steady-state loads:** A common feature of many residential loads is that such appliances draw a roughly constant power while remaining in the same state (often referred to as a steady-state). This feature allows an aggregate load to be conveniently divided into periods during which each appliance remains in a single state, separated by candidates where an appliance state change could occur [@johnson_2013]. Such approaches cannot be applied to disaggregate appliances with a continuously variable load, such as plasma televisions, and furthermore often fail to disaggregate any appliances in households containing such variable loads. Such loads are relatively rare in residential buildings. However, variable speed pumps and fans are often used in HVAC systems for commercial buildings instead of fixed speed motors, as a result of their reduced energy consumption. These variable speed pumps exhibit prolonged transients which would cause steady-state detectors, that assume short duration transients, to fail [@norford1996].
**Temporal dependencies:** Residential appliance usage often follows a repeating daily and weekly schedule, as a result of the lifestyle of the household’s occupants. Such repetitive behaviour has been exploited by disaggregation approaches, and has been shown to increase the accuracy of energy disaggregation algorithms [@kim_2011; @kolter_2010]. While such daily and weekly periodicity is common in residential buildings, the individual schedules for each household often need to be learned separately due to the high variance between households. In contrast, commercial building types are more likely to have similar daily or weekly schedules (e.g., a campus of buildings which share the same opening hours or manufacturing units often operating under the same shift patterns). Previous work [@shao] also suggests exploiting fixed appliance schedules in commercial buildings. As a result, separately learning periodicity for each commercial building is likely not required.
**Feed correlations:** In addition to temporal dependencies, strong correlations between appliance usages have also been exploited in residential disaggregation. For example, the usage of appliances such as televisions and set top boxes are often highly correlated, and similarly so are the usage of computers and monitors. It has been shown that such information can be exploited by a disaggregation algorithm to disaggregate appliances even when the step changes have not been detected [@kim_2011]. However, perfect correlations between appliances can result in simultaneous start up or shut down events being produced, as might occur in commercial buildings, therefore contributing systematic violations of the one-at-a-time assumption.
We later evaluate these assumptions for commercial buildings in [Section \[comparison\]]{} using two publicly available residential NILM data sets and the data set from our deployment which we describe next.
Deployment
==========
IIIT Delhi is a modern educational campus established in 2008 and spreads over 25 acres. The campus consists of 8 buildings: an academic block, lecture block, library, facilities building, mess building, faculty housing and male and female dormitories. We leveraged the construction phase of our campus to extensively deploy smart meters across the campus. More than 200 Modbus enabled Schneider Electric EM6400 and EM6436 smart electricity meters were deployed across different buildings and sub-systems spread throughout the campus.
Based on our previous successful deployments of electricity and ambient monitoring equipment [@issnip; @iawe], we chose Raspberry Pi as our controller to collect data from Modbus enabled smart meters. We chose sMAP [@smap] as our data archiving and metadata description backend, as it is a mature platform previously proposed in the research community and has been used by other research groups[^2] for similar deployments.
Our deployment is also inspired by previous campus testbeds established by the community [@dashboard_1; @dashboard_2]. With the aim of creating a testbed for coming years, we decided to extensively instrument our campus at various metering levels. [Figure \[fig:metering\]]{} shows a simplified view of our smart meter deployment. We instrumented the 3 transformers (Commercial transformers 1 and 2, and Residential transformer in [Figure \[fig:metering\]]{}) which are the entry points from the utility. The residential transformer supplies the faculty housing and the dormitores, while the other two transformers cater to the remaining five buildings and all central loads (such as HVAC chillers, HVAC pumps and UPS). For each building, we separately monitor total power consumption, high energy consumption loads (e.g. AHU, elevators and UPS) and total consumption for each floor. In addition, certain central loads that cater to the whole campus, such as the HVAC chiller or HVAC pumps, which have a high energy consumption were also separately monitored. Many of the non-residential buildings (e.g., academic block and library) can be considered to be similar to IT offices where HVAC and IT loads are the main energy consumers. The campus deployment is summarised in [Table \[tab:deployment\]]{}. We have released one month of data from the academic block constituting data at different hierarchies (buildings, floors and major loads) for public use. More details about our data set, called COMBED (commercial building energy data set) can be found on its project page.[^3]
**Data set attribute** **Value**
----------------------------------- ------------------
\# Smart meters 200
\# Electrical parameters measured 8
\# Buildings instrumented 6
Sampling interval 30 s
Deployment period Aug 2013-present
: Summary of our campus deployment[]{data-label="tab:deployment"}
Comparison of residential and commercial buildings {#comparison}
==================================================
We now evaluate the common residential NILM assumptions introduced in [Section \[assumptions\]]{} in the context of our commercial data set described in [Section \[deployment\]]{}. Wherever possible, we make use of previously collected residential data sets while drawing comparisons. We believe that the release of such data sets has been the catalyst behind the recent expansion of the field, and therefore data sets in commercial settings might have a similar effect. As discussed previously in [Section \[introduction\]]{}, \[related\_work\] and \[deployment\], many of the buildings from our campus closely resemble common office buildings and therefore many of our findings can be generalised to other IT office buildings. We begin our analyses by comparing the total energy consumption of residential and commercial buildings.
Total energy consumption and temporal dependencies
--------------------------------------------------
[Figure \[fig:total\_power\]]{} compares the total power demand observed on a particular day across the academic and library blocks (from our deployment) with homes from the AMPds and REDD residential data sets. As expected, commercial buildings consume considerably more power than their residential counterparts throughout the day due to the presence of a greater number of loads and high power consumption systems (such as HVAC chillers and pumps). In fact, the total energy consumption on these specific days is 612 kWh and 217 kWh for the academic block and library block, and 19 kWh and 5 kWh for the AMPds and REDD homes respectively. This confirms our initial intuition, which may indicate a greater potential for energy savings in the commercial settings, through the identification of inefficient appliances or usage schedules.
Previously, in [Section \[related\_work\]]{}, we discussed how NILM approaches designed for residential buildings have used temporal dependencies in the data to enhance disaggregation performance. We now contrast the temporal dependencies in the residential and commercial settings. [Figure \[fig:heatmap\]]{} shows the hour-wise normalised energy consumption of two buildings from our deployment (academic block and library block) compared to the home from the AMPds data set. We observe that the two buildings from our deployment consume significantly greater amounts of energy during office hours (9 AM to 5 PM) on weekdays in comparison to other time periods. In contrast, the energy consumption of the home from the AMPds data set shows some consistent variation between hours of day but does not show any significant variation between weekdays and weekends. While the residential data shows some temporal dependencies (such as lower energy consumption when people leave for office), the effect (as observed in [Figure \[fig:heatmap\]]{}) is much more profound in the commercial settings.
**Insights:** Previous studies have proposed using separate NILM models for weekdays and weekends. While residential buildings may show some difference in the energy consumption between different times of day, the difference is much more prevalent in the case of commercial buildings. Based on this observation, we believe that the day of the week is an important feature to consider when performing disaggregation. In our particular data set, we believe NILM performance accuracy is likely to improve if separate models for weekdays and weekends are considered, due to the fact that our campus runs classes only on weekdays.
Number of events {#events}
----------------
Commercial buildings usually contain many more loads in comparison to their residential counterparts. Thus, it is expected that there would be more frequent appliance activity and correspondingly more appliance switch events in commercial settings. This higher activity frequency is expected to generate a higher number of simultaneous appliance switch events as a result of random collisions. Based on this intuition we now evaluate the validity of the one-at-a-time assumption discussed in [Section \[related\_work\]]{}. We define that an event has occurred when the power draw between successive samples varies by a value greater than some threshold. [Table \[tab:events\]]{} compares the number of events in a day for several buildings from our deployment to all 6 homes from the REDD data set, where a threshold of 100 W was chosen to define an event. We observe that the number of events in a commercial building can be two orders of magnitude greater than that of the homes in the REDD data set. Furthermore, as we move down the electrical metering hierarchy (from the transformer to the floor level), we see that groups of events become divided, signifying that the disaggregation problem is simplified by deploying additional meters. However, even at the floor level, the number of events in the commercial building are several times more than those observed in the residential settings.
If we increase the threshold which defines the events, the number of events reduces as shown in [Table \[tab:events\_threshold\]]{}. We believe that while extending existing event-based NILM approaches to commercial buildings, the threshold parameter may need to be carefully fine-tuned. By increasing the threshold, we reduce the number of events caused by continuously varying loads such as AHUs (discussed in [Section \[variability\]]{}) and also avoid small power events (like light switching) that may not be of much significance in commercial settings.
**Insights:** The one-at-a-time assumption often made in the context of residential NILM approaches is not expected to be valid in the commercial context due to the increased frequency of events. However, by adding more metering (transformer to building to floor), the number of simultaneous switch events is likely to decrease as the events get partitioned across different entities (e.g., floors, buildings, etc.). Furthermore, due to the presence of different categories of loads (larger power draw, varying power draw) in commercial buildings, the threshold for defining events might be significantly greater than their residential counterparts. We revisit the utility of this additional sub-metering in [Section \[evaluation\]]{}.
**Building/Entity** **median** **max**
------------------------------ ------------ ---------
**Transformers**
Transformer 1 2444.0 2596
Transformer 2 2520.0 2636
Transformer residential 2678.0 2778
**Campus buildings**
Academic block 2050.5 2270
Library block 753.0 1017
**Floors in academic block**
First floor 648.5 1274
Ground floor 1263.0 1450
Second floor 184.5 366
**Residential from REDD**
Home 1 67.0 259
Home 2 91.0 126
Home 3 87.0 285
Home 4 29.0 102
Home 5 18.5 62
Home 6 27.0 101
: Comparison of number of events in a day in commercial and residential buildings. For a valid comparison we downsampled the REDD data set from 1 second resolution to 30 second resolution to match the sampling period from our deployment. []{data-label="tab:events"}
=0.06cm
---------------- ------ ------ ------ ------ ------ ------
100 200 500 1000 2000 5000
Academic block 2050 1808 1433 1083 754 222
Library block 753 596 377 223 134 26
---------------- ------ ------ ------ ------ ------ ------
: Median of the number of events observed in the academic and library block for varying power thresholds. []{data-label="tab:events_threshold"}
Correlated data streams
-----------------------
Commercial buildings often house several subsystems which operate together. [Figure \[fig:hvac\]]{} shows different HVAC components which usually turn on and turn off at the same time. This operation is likely triggered by a fixed schedule and is therefore independent of human activity. For instance, IIITD follows a policy of fixed HVAC system running times across multiple buildings. In contrast, in residential buildings, appliances operating together are often due to human activity. For example, the television and fan may be used together during the dinner time. As a result, the use of such feed correlations are likely to contribute greater increases to the disaggregation accuracy for commercial buildings than residential buildings, as discussed in [Section \[related\_work\]]{}.
[Figure \[fig:corr\]]{} shows the correlation between the power consumption of several streams from our deployment. The high correlation amongst the different HVAC systems (AHUs on different floors of the academic block and the chiller) clearly shows that these systems often operate together. In contrast, there is little correlation between the power draw of HVAC systems and the systems catering to the residential block (residential elevators and transformer). Recent research [@fontugne2013strip] has also looked into sophisticated variants of stream correlation to identify abnormal energy consumption in buildings. We noticed the high correlation between transformer 1 and the chiller, which is a result of the chiller consuming the majority of energy supplied by this transformer. Such correlations could be used to automatically construct the metering hierarchy, by determining loads being monitored at multiple metering levels.
![Correlation heatmap between different subsystems from our deployment. +1 indicates total positive correlation and -1 indicates a total negative correlation.[]{data-label="fig:corr"}](figures/correlation.pdf)
**Insights:** Although the strong correlations between feeds in commerical buildings can be used to increase the accuracy of disaggregation algorithms, such behaviour also consistently violates the one-at-a-time assumption. As a result, NILM methods which rely on edge matching, which would consider a rising or falling edge in the aggregate power signal to be caused by a single appliance, may need to be modified to account for this. However as previously suggested by Kim et al. [@kim_2011] in the context of residential settings, the correlation among different systems can also be used to improve disaggregation performance. For example, if multiple loads in different buildings often change state simultaneously, it is highly likely to be caused by different HVAC components (AHUs, chillers, pumps) in these buildings which are often operated synchronously by the facilities. Another insight stems from the fact that commercial buildings often pay additional charges for their peak power demand usage [@compre]. The different HVAC systems starting up simultaneously all have a cumulative effect on the peak power demand. Thus, the company might incur heavy demand charges, therefore motivating the building manager to schedule the loads sequentially to lower the peak to reduce these costs. Furthermore, HVAC systems run on motors that are inductive in nature, and hence reduce the overall power factor, as shown by [Figure \[fig:pf\]]{}. Commercial buildings often use additional capacitor banks to compensate for the power factor reduction caused by inductive loads. Quick switching of large inductive loads results in quick switching of multiple capacitors in the capacitor bank that further reduces their lifetime and increase their chances of failure. Thus, distributing the operations of motorised loads also helps to improve the lifetime of the capacitive bank.
Variability in power draw {#variability}
-------------------------
![Relationship between the power draw and power factor of our campus chiller. The power factor drops significantly (from 1 to 0.9) when the chiller is on.[]{data-label="fig:pf"}](figures/pf.pdf)
Commercial buildings often employ variable frequency drives (VFDs) for optimising the energy and operation of HVAC systems [@vfd_1; @vfd_2]. These VFDs allow changes in the fan speed in accordance with the desired load to achieve the set-point temperature for different HVAC systems, and hence optimise their energy consumption. [Figure \[fig:variable\]]{} shows a histogram of the instantaneous power draw of different HVAC systems from our campus deployment over several months of operation. It can be seen that these loads draw a highly variable and continuously varying power demand, and as such are expected to be difficult to disaggregate. As discussed previously in [Section \[related\_work\]]{}, approaches for residential buildings have often modelled appliances using a set of steady states, which would be a poor representation for such variable loads. The appliances which don’t conform to steady state behavior in residential buildings include electronic equipment such as laptops and plasma televisions, few of which have a significant energy consumption. In contrast, the VFD-based systems used in commercial buildings have a significant contribution to the overall energy consumption and thus their accurate disaggregation is important.
Commercial buildings also often contain elevators, which are much less commonplace in residential buildings. [Figure \[fig:lift\]]{} shows the power consumption of an elevator in the academic block of our data set. From [Figure \[fig:lift\]]{} (a), it can be seen that elevators can draw vastly different amounts of power. Furthermore, due to the small journey duration, the power consumption of the elevator is a short time duration transient with no steady states. The variable power draw of the elevator may depend upon a variety of factors such as distance moved and weight carried, among others [@lifts_2; @lifts_1]. Additionally, it can be seen from the histogram of elevator power consumption ([Figure \[fig:lift\]]{} (b)), that most of the time the elevator consumes close to 400 W which is the base consumption when the elevator is not actively used. [Figure \[fig:lift\]]{} (c) shows the energy consumption of the elevator by the hour of day, averaged over the duration of our deployment. We can observe that the energy consumption in non-working hours is roughly 50% of the peak energy consumption (which occurs during working hours). Based on this observation, we believe that there is significant value in optimising elevator energy consumption, especially during non working hours.
**Insights:** Commercial buildings often house systems such as HVAC and elevators which can draw variable amounts of power. The presence of such variable power loads can further complicate disaggregation, especially for NILM approaches which rely on the identification of steady states. Furthermore, we saw that loads such as elevators have a significant energy consumption even when they are not actively used due to their base power demand. Since this baseload power is always on, we believe that most low frequency methods would fail to disaggregate this portion of energy. As such, existing unsupervised NILM techniques have only been applied to appliances which exhibit all of their states during routine operation, and cannot be applied to always on loads which never change state.
Where to monitor? {#where_to_monitor}
-----------------
Previous residential data sets for NILM have mostly monitored the main supply along with either circuits or appliances as individual loads for evaluation purposes. In the context of commercial buildings, there are unique aspects for monitoring power consumption. Firstly, the cost to energy consumption ratio for sub-metering is far less for commercial buildings. Secondly, due to the multi-level hierarchical metering, there are several points where we could monitor the power consumption. Thirdly, commercial buildings may house complex metering where a particular load may be fed from multiple electrical lines. Previously in [Section \[events\]]{} we had discussed how the number of events decreases as we move down the metering hierarchy from the transformer down to the floor level. We now consider a few examples and discuss the relative ease of disaggregation.
[Figure \[fig:where\]]{}a shows it may be easy to disaggregate the chiller from the campus mains even by using naive thresholding techniques (whenever the campus main power consumption is more than 250 kW, the chiller is on). However, it must be noted that it may be non-trivial to ascertain the exact power consumption of the chiller since it may be run on a variable frequency. At lower levels in the hierarchy, it is easier to disaggregate the chiller from the transformer, especially levels on which the chiller is the only major load.
On the other hand, disaggregating the elevator’s power consumption from the academic block might be more difficult ([Figure \[fig:where\]]{}b). Firstly, elevators only show short term transient power usage, which may be easily confused with appliances such as microwave or toaster. Secondly, the large number of events in the academic block could easily occlude the elevator events. Thirdly, the elevators consume a very small amount of energy of the total academic block. However, elevators in the faculty housing ([Figure \[fig:where\]]{}c) might be easier to disaggregate as the elevators are on a dedicated circuit with fewer loads.
**Insights:** We believe that minimally sensing to disaggregate a load has several subjective dimensions. As per our observations, some rule of thumb based questions which may help to answer this question are: (i) the presence (or absence) of similar behaviour loads in the stream which we intend to disaggregate, (ii) the proportion of energy a load consumes in the stream which we want to disaggregate, and (iii) the number of events in the stream which we want to disaggregate.
![Predicted power with ground truth for ground floor AHU. The thick bars in the predicted power are due to the frequent on/off switching predicted by CO, which is due to the fact that CO does not model the relationship between different samples.[]{data-label="fig:disagg_1"}](figures/disagg_1.pdf)
Evaluation of disaggregation {#evaluation}
============================
Having discussed the traits of NILM which are specific to commercial buildings, we now provide a preliminary evaluation on our data set using benchmark residential NILM algorithms provided by NILMTK [@nilmtk]. We believe that ours is the first evaluation that applies these NILM methods, proposed for residential settings, to a commercial dataset with reasonable sampling time (30 seconds). Firstly, we try to disaggregate AHU loads in the academic block considering the academic block’s total power as the input. We use an implementation of the combinatorial optimisation (CO) disaggregation algorithm provided by NILMTK to train the AHUs as two state (on and off) loads. We selected data from the month of June for our experiments and divided it equally into two halves for training and testing. Further, we downsampled the data to 1 minute resolution to synchronise the ground truth and aggregate time series. We use F-score and normalised error in assigned power (NEP) as our accuracy metrics. We follow the standard definition for NEP as defined in NILMTK: the sum of the differences between the assigned power and actual power for each appliance in each time slice, normalised by the appliance’s total energy consumption. [Table \[tab:ahu\_from\_building\]]{} shows the disaggregation performance across these AHUs. We observe that the F-score is relatively high suggesting that CO is able to detect the operation of the AHU in most cases. However, [Figure \[fig:disagg\_1\]]{} shows an example of the predicted and ground truth power for the ground floor AHU in which CO is unable to identify the AHU operation. This is due to the AHU exhibiting states which were not present in the training set, since the power demand varies with the temperature set point, and thus the finite state CO model is insufficient to accurately disaggregate the AHU power. Having discussed disaggregation performance by disaggregating academic block total power into different AHUs, we now see if this performance can be improved with increased sub-metering, as has been shown previously for residential buildings [@batra_2013]. First, we compare the AHU disaggregation performance when disaggregation is performed at the floor level instead of the building level. [Table \[tab:disagg\_different\_metering\]]{} shows that the F-score of the fifth floor AHU improves from 0.7 to 0.99 and the NEP significantly improves from 1.0 to 0.12. Thus, by traversing an additional level in the metering hierarchy, we get near perfect disaggregation accuracy for the AHU. However, we observed that the disaggregation performance remains unchanged for the chiller irrespective of whether we perform the disaggregation at the campus total level or the transformer level. This is likely due to the fact that the chiller already dominates the campus total power consumption and therefore traversing an additional level to the transformer adds little value as shown in [Figure \[fig:where\]]{}. Our observations can also be explained from an information theoretic perspective. [Table \[tab:disagg\_different\_metering\]]{} shows the entropy of the power signal measured at different locations. The disaggregation performance for the AHU improves significantly due to the large decrease in entropy moving from the building level to the floor level. On the other hand, there is insignificant change in the entropy moving from campus total to transformer, which is reflected in the disaggregation performance remaining the same. We believe that such information theoretic approaches can provide useful insights into the minimal set of sensors required to disaggregate a set of desired loads in a given building.
![Chiller dominates the total campus power consumption[]{data-label="fig:where"}](figures/where.pdf)
It is also interesting to investigate the disaggregation of the elevator’s power consumption from the building total. In this case, we applied the event-based disaggregation approach proposed by Hart [@hart_1992] due to the relatively low elevator power consumption in comparison with the building’s HVAC load. However, since the elevator power data were not time synchronised with the building level power data, it was not possible to evaluate the accuracy of disaggregation at this data resolution. Therefore, we downsampled the data to 1 minute intervals before applying the disaggregation algorithm. Such downsampling has little effect on the disaggregation performance of CO for loads such as AHUs due to their long state durations. However, this downsampling completely obscures the elevator’s power demand. Thus, we believe that greater sampling rates are needed to apply any disaggregation approaches to loads such as elevators which exhibit a short term transient power draw.
=0.2cm
**Load** **F-score** **NEP**
---------- ------------- ---------
AHU 0 0.74 0.74
AHU 1 0.62 1.42
AHU 2 0.68 1.00
AHU 5 0.70 1.00
: Disaggregation performance of different AHU loads in the academic block[]{data-label="tab:ahu_from_building"}
=0.1cm
**Location** **Entropy** **Load** **F-score** **NEP**
---------------- ------------- ---------- ------------- ---------
Academic block 15.3 AHU 5 0.70 1.00
5th floor 12.2 AHU 5 0.99 0.12
Campus total 17.5 Chiller 1 0.93
Transformer 2 17.1 Chiller 1 0.93
: Disaggregation performance at different metering levels. We used NPEET toolbox [@ver2000non] for entropy calculations.[]{data-label="tab:disagg_different_metering"}
Conclusions and Future work {#conclusion}
===========================
The majority of previous NILM research has focused on residential buildings. However, commercial buildings often consume several orders of magnitude more energy than residential buildings, and thus the return on investment for NILM is expected to be greater. Previous research in the context of residential buildings have made several assumptions: (i) appliance events happen one-at-a-time, (ii) presence of steady states, (iii) appliance usage temporal patterns and (iv) correlations amongst several loads. In order to determine the validity of these assumptions in commercial settings, we studied these assumptions from our campus deployment and contrasted them with existing residential data sets. We found that there is a significant difference in the amount of energy consumed on weekend and weekdays in our campus. Although such temporal dependencies also exist in residential setting, they are much less profound. NILM algorithms for commercial settings can exploit these temporal dependencies to enhance disaggregation accuracy. Commercial buildings usually house much larger number of loads in comparison to residential buildings and thus exhibit an order of magnitude more events than the residential buildings. This is also due to the fact that commercial buildings often house systems such as VFD based HVAC systems which can have a continuously varying power demand. Both of these factors increase the complexity of energy disaggregation in commercial settings. Thus, the common assumption of steady state power consumption and one-at-a-time events may not be valid in commercial settings. While correlations amongst appliances can be seen in residential buildings, the effect is much more profound in commercial buildings when various HVAC subsystems turn on and off at the same time. We also discussed that due to these loads all turning on together, the transient generated may significantly increase the peak power consumption of the building and thus cause the company to incur demand charges.
We also performed a preliminary evaluation of NILM on our data set where we tried to disaggregate AHUs from the building total and floor total respectively. The naive CO approach we used fails to model the continuously varying power demand of the AHU. However, the disaggregation performance improves significantly when we perform disaggregation at the floor level. We have also released a subset of our data for public use.
In the future we would like to enhance the NILM-metadata schema [@NILM_metadata] which is used in conjunction with NILMTK [@nilmtk] to support complex metering hierarchies usually present in commercial buildings. We had briefly discussed information theoretic approaches as an indicator for diaggregation performance in our paper. We would like to study these in more detail to be able to come up with models for optimal sensing. Given that our smart meters are capable of collecting data at a higher resolution of 1 Hz, we would like to explore disaggregating loads such as elevators using event based approaches.
Further, we would also like to exploit existing work in context discovery in smart commercial spaces to increase the disaggregation accuracy [@thakur; @softgreen]. Also, it is likely that existing approaches working with low frequency data may not be able to disaggregate low power consuming systems such as lighting. We would like to explore the possibility of using previous work on EMI sensing [@emisense; @electrisense] in addition to our low frequency energy disaggregation to get more detailed energy disaggregation. Additionally, we aim to make an end to end system using existing systems such as BAS [@bas] or SensorAct [@sensoract], which can collect data, disaggregate using NILMTK and perform control action or provide detailed feedback for energy optimisation.
[^1]: <http://open.enernoc.com/data/>
[^2]: For example: <http://new.openbms.org/plot/>
[^3]: <http://combed.github.io>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We perform an “archeological" study of one of the original experiments used as evidence for the static, [*time-independent*]{} Aharonov-Bohm effect. Since the experiment in question [@marton] involved a [*time varying*]{} magnetic field we show that there are problems with the explanation of this experiment as a confirmation of the static Aharonov-Bohm effect – specifically the previous analysis ignored the electric field which arises in conjunction with a time-varying magnetic flux. We further argue that the results of this experiment do in fact conform exactly to the recent prediction [@singleton; @macdougall] of a cancellation between the magnetic and electric phase shifts for the [*time-dependent*]{} Aharonov-Bohm effect. To resolve this issue a new time-dependent Aharonov-Bohm experiment is called for.'
author:
- James Macdougall
- Douglas Singleton
- 'Elias C. Vagenas'
title: 'Revisiting the Marton, Simpson, and Suddeth experimental confirmation of the Aharonov-Bohm effect'
---
The experimental work of Chambers [@chambers] is usually cited as the first experimental confirmation of the Aharonov-Bohm (AB) effect [@AB; @siday]. However, the earliest (inadvertent) test of the AB effect was an experiment by Marton, Simpson, and Suddeth (MSS) [@marton], whose set-up was designed simply to make an electron interferometer to see electron interference fringes, and this was apparently accomplished; interference fringes were seen. However, it was realized later that the electrons in [@marton] were moving through a stray 60 Hz magnetic field. Since MSS did see a fixed interference pattern, it was originally suggested that the effect of this magnetic field could be ignored given the 60 keV kinetic energy of the electrons. But with the proposal of the AB effect [@AB], it was realized that there should have been a time-varying shifting of the interference pattern which was not seen in [@marton]. Thus, the non-observation by MSS of a time-varying, shifting interference pattern appeared to be evidence against the AB effect. However, in a clever paper Werner and Brill [@werner] proposed that if one took into account the additional phase shift due to the direct Lorentz force, that under certain conditions there would be an [*almost*]{} exact cancellation of the AB phase shift versus the phase shift coming from, $q(\frac{{\bf v}}{c} \times {\bf B})$. Thus, according to Werner and Brill, the fact that MSS saw a fixed interference pattern could be taken as evidence of the phase shift predicted by the AB effect.
However, the analysis of [@werner] ignored the electric field which arises whenever there is a time-varying magnetic field. From Faraday’s law, [*i.e.*]{}, $\nabla \times {\bf E} = - \frac{1}{c}\partial _t {\bf B}$, one can see that there was an electric field present in the MSS experiment. In terms of potentials, the time-varying magnetic field is associated with a time-varying vector potential, namely ${\bf B} ({\bf x}, t) = \nabla \times {\bf A} ( {\bf x}, t)$. It is this time-varying vector potential that produces an electric field ${\bf E} = - \frac{1}{c}\partial _ t {\bf A} ( {\bf x}, t)$. Now, in [@singleton; @macdougall] a simple argument was given that the phase shift of this electric field would always exactly cancel the AB phase shift coming from the magnetic flux. Briefly, the argument of [@singleton; @macdougall] went as follows: in terms of the electric and magnetic fields one can write the AB phase shift, $\alpha$, as $$\label{phase1}
\alpha = \frac{e}{\hbar} \int _{\it C} F = \frac{e}{\hbar} \left[ \int {\bf E} \cdot d {\bf x} (c dt) + \int {\bf B} \cdot d{\bf S} \right] ~,$$ where $F$ is the Faraday two-form which in wedge notation is $(E_x dx + E_y dy + E_z dz )
\wedge (c dt) + B_x dy \wedge dz + B_y dz \wedge dx + B_z dx \wedge dy$. In the far right-hand side of Eq. we have reverted to 3-vector notation. Now, for the time-dependent Aharonov-Bohm set-up, [*i.e.*]{}, a solenoid with a time-varying flux, one has a time-varying vector potential, ${\bf A} ({\bf x}, t)$, from which one can obtain the fields via ${\bf E} = - \frac{1}{c} \partial_t {\bf A}$ and ${\bf B} = \nabla \times {\bf A}$. As pointed out in [@singleton; @macdougall], doing the time integration for the electric part of Eq. yields $$\label{e-phase}
\int {\bf E} \cdot d {\bf x} ~ (c~dt) = - \oint {\bf A} \cdot d {\bf x} ~,$$ and using Stokes’ theorem on the magnetic part of Eq. yields $$\label{b-phase}
\int {\bf B} \cdot d{\bf S} = \oint {\bf A} \cdot d{\bf x} ~.$$ Substituting the electric and magnetic parts, Eq. and Eq. , respectively, in Eq. , the total phase shift will be zero, [*i.e.*]{}, $\alpha =0$. This was the prediction for the time-dependent AB effect from [@singleton; @macdougall] and this is in fact what was seen in the MSS experiment. In [@singleton] this cancellation between the was also shown using the path length difference (and therefore phase difference) produced in the electron due to the acceleration coming from ${\bf E}$.
There are several comments to make about the above results: First the above argument only implies a canceling of the [*time-dependent*]{} part of the phase shift [*not*]{} the time-independent part,[*i.e.*]{}, the part normally referred to as the AB phase shift. If one splits the vector potential in static and time-varying parts as ${\bf A} ({\bf x}, t) = {\bf A}_0 ({\bf x}) + {\bf A}_1 ({\bf x}, t)$, it is only the contribution coming from ${\bf A}_1 ({\bf x}, t)$ which will cancel between Eq. and Eq. . In detail ${\bf B} = \nabla \times {\bf A}_0 ({\bf x}) + \nabla \times {\bf A}_1 ({\bf x}, t)$ while ${\bf E} = - \partial_t {\bf A}_1 ({\bf x}, t)$. Thus it is only the time-dependent part of the vector potential ${\bf A}_1 ({\bf x}, t)$ which takes part in the cancellation between and – the time-independent part ${\bf A}_0 ({\bf x})$ still leads to the usual static AB phase shift. This interplay between the electric and magnetic contributions found in Eq. and Eq. was noted in a different context in Ref. [@kampen] (see also [@moulopoulos]). In this paper the cancellation between Eq. and Eq. is shown without using the potentials but rather is shown to be a direct consequence of Faraday’s Law which in form notation reads $dF=0$. In detail van Kampen noted that the loop integral of the electric field could be turned into a surface integral via Stokes’ theorem as $$\label{faraday-3d}
\oint {\bf E} \cdot d{\bf x} = \int \nabla \times {\bf E} \cdot d {\bf S} =
-\frac{1}{c} \int \partial_t {\bf B} \cdot d {\bf S} ~,$$ where in the last step we have used Faraday’s law $\nabla \times {\bf E} = -\frac{1}{c} \partial_t {\bf B}$. Now applying $\int ... (c ~ dt)$ to we obtain $$\label{faraday-3d-2}
\int \left( \oint {\bf E} \cdot d{\bf x} \right) c~dt =
- \int \left( \int \partial_t {\bf B} \cdot d {\bf S} \right) dt = - \int {\bf B} \cdot d {\bf S} ~,$$ where the time integral has been undone via the time integration, and we see that this last expression in cancels the expression in . Thus following van Kampen we see that the cancellations is a consequence of the Bianchi identities $d F =0$. Second one can ask about the nature of the space-time surfaces in the expression in versus the surfaces used in the more usually encountered magnetic expression such as in . To this end we look at infinitesimal paths and surfaces for the above case and show that the cancellation of the time-dependent electric and magnetic pieces occurs. For a solenoid oriented along the $z$-axis the vector potential inside and outside a solenoid, which has a radius $\rho = R$ and a time varying flux, is given by $$\label{3-vector-a}
{\bf A}_{{\rm in}} = \frac{\rho B(t)}{2} \hat{\bf \varphi} ~~~~{\rm for ~~ \rho <R} ~~~~;~~~~~
{\bf A}_{{\rm out}} = \frac{B(t) R^2}{2 \rho} \hat{\bf \varphi} ~~~~{\rm for ~~ \rho \ge R} ~.$$ The associated magnetic (${\bf B} = \nabla \times {\bf A}$) and electric fields (${\bf E} =\partial _t {\bf A}$) are $$\label{B-field}
{\bf B}_{{\rm in}} = B(t) {\bf {\hat z}}~~~~{\rm for ~~ \rho <R} ~~~~;~~~~
{\bf B}_{{\rm out}} = 0 ~~~~{\rm for ~~ \rho \ge R} ~,$$ and $$\label{E-field}
{\bf E}_{{\rm in}} = - \frac{\rho {\dot B(t)}}{2} \hat{\bf \varphi}
~~~~{\rm for ~~ \rho <R} ~~~~;~~~~
{\bf E}_{{\rm out}} = - \frac{{\dot B(t)} R^2}{2 \rho} \hat{\bf \varphi}
~~~~{\rm for ~~ \rho \ge R} ~.$$ Now using these expressions in for some short time interval $\Delta t$, expanding the magnetic field as $B(t) = B_0 + {\dot B} \Delta t + {\cal O} (\Delta t )^2$, and taking the path to be a short circular arc segment covered in time $\Delta t$ [*i.e.*]{} $\Delta {\bf x} = (\rho \Delta \varphi ) {\hat \varphi}$ we find that the small phase for the infinitesimal path and associated area are $$\label{phase2a}
\Delta (\alpha) = \frac{e}{\hbar} \left( {\bf E} \cdot \Delta {\bf x} \Delta t +
\Delta {\bf B} \cdot \Delta {\bf S} \right) =
\frac{e}{\hbar} \left( - \frac{R^2}{2 \rho} (\dot B \Delta t)(\rho \Delta \varphi )
+ \frac{\Delta \varphi R^2}{2} (\dot B \Delta t) \right) = 0 ~.$$ Thus one can see the cancellation between the electric and magnetic pieces in for this infinitesimal path and area. Again this cancellation just involves the time-dependent part [*i.e.*]{} the $\dot B \Delta t$ part of the magnetic field expansion. The constant first term, $B_0$, would still lead the time-independent, static AB phase shift.
We now move on to the re-analysis of the MSS experiment. Originally the non-observation of a time shifting interference pattern by MSS led to the idea that this experiment had ruled out the AB effect. However, in [@werner] Werner and Brill argued that the non-observation of a time shifting interference pattern (or the fact that MSS observed a static interference pattern) was in fact positive evidence for the AB effect due to a subtle, [*almost*]{} cancellation between the magnetic AB phase shift and a phase shift coming from the $q(\frac{{\bf v}}{c} \times {\bf B})$ force on the electrons. The Werner and Brill explanation of the non-observation of a time shifting interference pattern by MSS is conceptually similar to the explanation given in [@singleton; @macdougall] and by Eq. and Eq. above. In both cases, there is an interplay between the magnetic Aharonov-Bohm shifting of the phases and a dispersive shifting of the phases due to either an electric or magnetic field through which the electrons are moving. However, if one accepts both explanations then it seems that one will have over compensated for magnetic AB phase shift and that one should then again see a time shifting interference pattern (or the interference pattern would be washed out entirely). We now re-examine the Werner and Brill analysis and show that their cancellation is highly dependent on the strength of the unknown 60 Hz magnetic field in the MSS experiment, and in fact for certain magnetic field strengths one does not get the balancing of the AB phase shift against the shift due to the $q(\frac{{\bf v}}{c} \times {\bf B})$ force.
To start the analysis we examine the re-produced figure 1 from [@werner]. In figure 1a, the electron beam is split into a diamond pattern by three crystals $C_1$, $C_2$ and $C_3$ in the absence of a magnetic field. In figure 1a, the path length $l_1 + l_2$ is equal to the path length of $m_1 + m_2$ so – modulo the complication that the beams come together with a small angle $\varphi$ – one would get constructive interference. Next, in figure 1b, there is a uniform magnetic field of magnitude, $B_0$, coming out of the page. This bends the paths $l_1, l_2, m_1, m_2$ into circular arcs of radius $$\label{radius}
R= \frac{2 \pi \hbar c}{e \lambda B_0} \approx \frac{850}{B_0} {\rm gauss \cdot cm}~.$$ In Eq. we use gaussian units and in the last approximation we have plugged in the values of the constant $\hbar, c, e$ (in gaussian units) and we have taken the de Broglie wavelength of the electrons from [@marton] as $\lambda = 4.86 \times 10^{-10}$ cm.
Now, the analysis of [@werner] calculated the change in path length produced in the paths $l_1, l_2, m_1, m_2$ due to $B_0$ and then by dividing this path length difference by $\lambda /2 \pi$ obtained a phase difference which when combined with the AB phase shift gave an [*almost*]{} cancellation. The path length differences $\Delta l_2 = l_2 ' - l_2$ and $\Delta m_1 = m_1 ' - m_1$ were both proportional to $$\label{l2m1}
\Delta l_2 \propto \frac{D^2 \tan (\theta)}{R} \;\;\; {\rm and} \;\;\; \Delta m_1 \propto \frac{D^2 \tan (\theta)}{R} ~,$$ where from Fig. 1 the angle $\theta$ is the crystal diffraction angle which for the set-up in [@marton] was $\theta \approx 2 \times 10^{-2}$ radians; $D$ was the distance between the crystals which for the set-up in [@marton] was varied between $5.0$ cm and $3.49$ cm (in subsequent estimates we take $D=5.0$ cm). The path length differences between difference $\Delta l_1 = l_1 ' - l_1$ and $\Delta m_2 = m_2 ' - m_2$ were ignored since as pointed out [@werner] these were of order $B_0 ^2$ whereas the path length differences $\Delta l_2$ and $\Delta m_1$ were of order $B_0$. In terms of $D$ and $R$ the path length differences $\Delta l_1$ and $\Delta m_2$ can be shown to have a proportionality $$\label{l1m2}
\Delta l_1 \propto \frac{D^3}{R^2} = \frac{D^2}{R} \frac{D}{R}\;\;\; {\rm and} \;\;\;
\Delta m_2 \propto \frac{D^3}{R^2} = \frac{D^2}{R} \frac{D}{R}~.$$ The reason to split off one factor of $D/R$ in Eq. is to make a comparison between these path length differences with the path length differences $\Delta m_1$ and $\Delta l_2$ from Eq. . The ratio $D/R$ is essentially the angle by which the electrons are bent by the magnetic field, $B_0$. In [@werner] this angle is called $\tau _1$ – more precisely the angle by which the path $m_1$ is bent is approximated as $\tau _1 /2$ [@werner]. Thus, both the path length differences $\Delta m_1$ and $\Delta l_2$ from Eq. and the path length differences $\Delta m_2$ and $\Delta l_1$ are of the form $\frac{D^2}{R} \times angle$, where $angle$ is either the diffraction angle $\tan (\theta) \approx \theta$ from Eq. or the deflection angle $\frac{D}{R}$ due to the magnetic field. The assumption made in [@werner], when ignoring the path length differences $\Delta l_1$ and $\Delta m_2$ relative to $\Delta l_2$ and $\Delta m_1$, is that $\frac{D}{R} \ll \theta$. Now, from [@marton] the diffraction angle was $\theta \approx 2 \times 10^{-2}$ radians and we take $D \approx 5.0$ cm. Now, even for a weak magnetic field of $B_0 = 1$ gauss (which is roughly only twice the Earth’s magnetic field strength) Eq. gives $R \approx 850$ cm and thus $\frac{D}{R} \approx 6 \times 10^{-3}$ radians. Even for such a weak field the approximation $\frac{D}{R} \ll \theta$ does not really hold (it is true $\frac{D}{R} < \theta$ but the qualification “much less than" is not correct). For a magnetic field only 10 times stronger than the Earth’s magnetic field, [*i.e.*]{}, $B_0 = 5$ gauss , gives $R \approx 170$ cm and thus $\frac{D}{R} \approx 3 \times 10^{-2}$ radians for which the neither the relationship $\frac{D}{R} \ll \theta$ nor even $\frac{D}{R} < \theta$ is valid. Thus, depending on the strength of $B_0$ (which was small, but unspecified in the MSS experiment) the approximations used by Werner and Brill to ignore $\Delta l_1$ and $\Delta m_2$ might or might not be valid. For magnetic field greater than 1 gauss the approximation becomes quickly suspect. The result is that if one has a magnetic field even as weak as $B_0 \approx 2$ gauss one should include $\Delta m_2$ and $\Delta l_1$ in calculating the total path length difference which is then $(\Delta l_2 + \Delta l_1) - (\Delta m_2 + \Delta m_1)$. Moreover, if the diffraction angle, $\theta$, is similar to the angle of deviation of the magnetic field, $\frac{D}{R}$, one finds that $\Delta l_1 \sim \Delta m_1 \sim
\Delta l_2 \sim \Delta m_2$ so that there is only a small total path length difference coming from the $q(\frac{{\bf v}}{c} \times {\bf B})$ force. In this case the cancellation of the electric phase shift Eq. with the usual magnetic phase shift Eq. is crucial to explain the results of the MSS experiment. In any case ignoring the effect of the electric field for the time varying field is not valid while ignoring the effect of $q(\frac{{\bf v}}{c} \times {\bf B})$ is justifiable depending on the strength of the unknown $B_0$.
In this paper, we have highlighted problems with previous explanations of and given an alternative explanation for the results of the MSS experiment [@marton], which was intended to study electron interference, and in fact an interference pattern was observed. However, later it was determined that the path of the electrons was contaminated with a 60 Hz “weak" magnetic field, but how weak was not specified. In light of the predictions for a quantum mechanical, AB phase shift [@AB; @siday] which would have naively predicted a shifting/time-varying interference pattern in time to the temporal variation of the flux, the observations of a static interference pattern in [@marton] seemed, at first, to provide evidence against the AB phase shift. However, later experiments [@chambers] with static magnetic fields from iron whiskers did confirm the AB phase shift. The results of the MSS experiment were explained in [@werner] as an [*almost*]{} cancellation between the AB phase shift and a phase shift coming from a path length difference produced by the $q(\frac{{\bf v}}{c} \times {\bf B})$ force on the electrons since, unlike the ideal AB set-up, the electrons in the MSS experiment did move in a region with non-zero fields – both magnetic and electric. This was our first criticism of the previous analysis of the MSS experiment and all other studies of the [*time-dependent*]{} AB effect – the effect of the electric field, which arises from the time varying magnetic flux through Faraday’s law, is completely ignored. Further, from our previous works [@singleton; @macdougall] and the calculations around Eqs. and above, we find that the phase shift coming from the electric field should [*exactly*]{} cancel the usual magnetic AB phase shift, without any need for the complicated geometrical constructions of [@werner]. Our second criticism is, depending on the strength of the “weak" magnetic field (which was not specified in either [@marton] or [@werner]) the approximation was made that the total path length difference was given by $\Delta l_2 - \Delta m_1$ – thus ignoring the contributions of $\Delta l_1$ and $\Delta m_2$ to the total path length difference. This amounts to saying that the diffraction angle, $\theta$, was much larger than the angle of derivation due to the magnetic field, $\frac{D}{R}$, [*i.e.*]{}, $\frac{D}{R} \ll \theta$. Using $D \approx 5.0$ cm, $\theta =2 \times 10^{-2}$ radians [@marton] and using Eq. for $R$, we showed the even for $B_0$ of a few gauss that the approximation $\frac{D}{R} \ll \theta$ was not valid so that one should include $\Delta l_1$ and $\Delta m_2$ in the total path length difference. Including $\Delta l_1$ and $\Delta m_2$ would decrease the total path length difference since for $\frac{D}{R} \approx \theta$ one has $\Delta l_1 \sim \Delta m_1 \sim \Delta l_2 \sim \Delta m_2$, which then gives only a small total path length difference. This would leave the proposed cancellation of the electric and magnetic phase shifts discussed in [@singleton; @macdougall] as the only explanation of the MSS experiment. Finally, the MSS experiment should be considered as an inadvertent test of the [*time-dependent*]{} AB effect rather than as a test of the [*time-independent*]{} AB effect. Here we are proposing that one should re-do a purposeful version of the [*time-dependent*]{} AB effect experiment to determine whether there is really a cancellation between the electric and magnetic phase as suggested in [@singleton; @macdougall], or if something like the resolution given in [@werner] is correct.
As a final comment we know of only one experiment which purposefully sought to test the [*time-dependent*]{} AB effect [@chentsov; @ageev], and this experiment did find no time varying shifting of the interference pattern in agreement with the prediction of [@singleton; @macdougall]. However the authors of [@chentsov; @ageev] later reported problems with the experimental set-up. This again calls for further experiments designed to specifically test the [*time-dependent*]{} AB effect to resolve this issue.
[99]{}
L. Marton, J. A. Simpson, and J. A. Suddeth, Rev. Sci. Instr. [**25**]{}, 1099 (1954).
D. Singleton and E. Vagenas, Phys. Lett. B [**723**]{}, 241 (2013).
J. MacDougall and D. Singleton, J. Math. Phys. [**55**]{}, 042101 (2014).
R.G. Chambers, Phys. Rev. Lett. [**5**]{}, 3 (1960).
Y. Aharonov and D. Bohm, Phys. Rev. [**115**]{}, 484 (1959).
W. Ehrenberg and R. E. Siday, Proc. Phys. Society B [**62**]{}, 8 (1949).
F.G. Werner and D. Brill, Phys. Rev. Lett. [**4**]{}, 344 (1960).
N.G. van Kampen, Phys. Lett. A [**106**]{}, 5 (1984).
K. Moulopoulos, J. Phys. A [**43**]{}, 354019 (2010).
Yu. V. Chentsov, Yu. M. Voronin, I. P. Demenchonok, and A. N. Ageev, Opt. Zh. [**8**]{}, 55 (1996).
A. N. Ageev, S. Yu. Davydov, and A. G. Chirkov, Technical Phys. Lett. [**26**]{}, 392 (2000).
| {
"pile_set_name": "ArXiv"
} |
addtoreset[equation]{}[section]{}
-48pt
[**Indeterministic Quantum Gravity and Cosmology**]{}\
[**IX. Nonreality of Many-Place Gravitational Autolocalization: Why a Ball Is Not Located in Different Places at Once**]{}\
[**Vladimir S. MASHKEVICH**]{}[^1]\
[*Institute of Physics, National academy of sciences of Ukraine\
252028 Kiev, Ukraine*]{}\
[**Abstract**]{}
This paper is a sequel of papers \[1-8\], being an immediate continuation and supplement to the last of them, where gravitational autolocalization of a body has been considered. A resulting solution, which describes a one-place location, has been called gravilon. Here it is shown that a gravilon is the only solution, i.e., that many-place gravitational autolocalization is unreal. This is closely related to nonreality of tunneling in the conditions under consideration.
\[b\] [12 cm]{} Why, then, do we not experience macroscopic bodies, say cricket balls, or even people, having two completely different locations at once? This is a profound question, and present-day quantum theory does not really provide us with a satisfactory answer.
Roger Penrose
In the previous paper \[8\] of this series, the possibility of a gravitational autolocalization has been established. A resulting solution describes a one-place location of a body, which is called gravilon. But in \[8\] the question on nonreality of many-place gravitational autolocalization has not been even raised. This paper makes up for this essential deficiency.
The answer to the question is positive: Many-place gravitational autolocalization is unreal. The essence of the matter is as follows.
For the sake of simplicity, we consider the case of a two-place location. A wave function of the center of mass of a ball in such a situation depends on some parameters: the distance between location places and coefficients in the linear combination of one-place functions. The requirement that the wave function be an eigenfunction of the Hamiltonian results in a relation for the parameters, which singles out a set of measure zero for the admissible values of the parameters. This result is obtained neglecting the tunneling between the locations, so that nonreality of many-place location is closely related to nonreality of tunneling in the conditions under consideration.
Thus the only solution to the problem of gravitational autolocalization is a one-place location, i.e., a gravilon.
Basic equation
==============
Let $\psi(\vec r)$ be a wave function of the center of mass of a ball with a mass $M$ \[8\]. $\psi$ must satisfy the Schr$\ddot {\rm o}$dinger equation, $$H\psi=\epsilon\psi
\label{1.1}$$ where the Hamiltonian $$H\equiv H[\psi]=-\frac{\hbar^{2}}{2M}\triangle+V(\vec r;\psi)
\label{1.2}$$ and $V$ is the potential energy of the ball in a gravitational field caused by the ball itself with the wave function $\psi$.
Two-place function
==================
For the sake of simplicity, let us consider the case of a two-place location: $$\psi\equiv
\psi(\vec r)=c_{\rm I}\psi_{\rm I}(\vec r)+
c_{\rm II}\psi_{\rm II}(\vec r)
\label{2.1}$$ where $$\psi_{\rm I}=\psi_{1}(\vec r),\quad \psi_{\rm II}=
\psi_{2}(\vec r-\vec R)
\label{2.2}$$ and $$R\gg a_{0},\:r_{0i},\qquad i=1,2,
\label{2.3}$$ here $a_{0}$ is the radius of the ball and $r_{0i}$ is an effective radius of $\psi_{i}$. We may put $$c_{\rm I}=e^{{\rm i}\alpha}\sqrt{1-b^{2}},\qquad c_{\rm II}=
e^{{\rm i}\alpha}e^{{\rm i}\beta}b,\qquad 0\le b\le 1,
\label{2.4}$$ so that, dropping $e^{{\rm i}\alpha}$, we have $$\psi\equiv\psi(\vec r;b,\beta,\vec R)=
\sqrt{1-b^{2}}\psi_{1}(\vec r)+e^{{\rm i}\beta}b
\psi_{2}(\vec r-\vec R).
\label{2.5}$$
Now, in view of eq.(\[2.3\]), we neglect the tunneling between the two locations and choose $\psi_{1},\:\psi_{2}$ to be eigenfunctions of the Hamiltonian, $$H\psi_{i}=\epsilon_{i}\psi_{i}.
\label{2.6}$$
We have $$H=H[b,\beta,\vec R],\;\psi_{1}=\psi_{1}(\vec r;b,\beta,\vec R),
\;\psi_{2}=\psi_{2}(\vec r-\vec R;b,\beta,\vec R),
\;\epsilon_{i}=\epsilon_{i}(b,\beta,\vec R).
\label{2.7}$$
The set of two-place states: measure zero
=========================================
We have from eqs.(\[2.5\]), (\[2.6\]) $$H\psi=\sqrt{1-b^{2}}\epsilon_{1}\psi_{1}+
e^{{\rm i}\beta}b\epsilon_{2}\psi_{2},
\label{3.1}$$ so that for (\[1.1\]) to be satisfied, the relation $$\epsilon_{1}(b,\beta,\vec R)=\epsilon_{2}(b,\beta,\vec R)
\equiv \epsilon(b,\beta,\vec R)
\label{3.2}$$ must be fulfilled.
Eq.(\[3.2\]) determines a set of measure zero for admissible values of the parameters $b,\beta$, and $\vec R$. This implies that many-place gravitational autolocalization is unreal.
Tunneling
=========
Let us consider the tunneling, which was neglected. The condition of this neglect is $$DV\ll \Delta\epsilon
\label{4.1}$$ where $$D\sim e^{-2\sqrt{2M(V-\epsilon)/\hbar^{2}}L},
\label{4.2}$$ is the transmission coefficient and $\Delta\epsilon$ is the distance between energy levels.
We have \[8\] $$V-\epsilon\approx V\sim \frac{\kappa M^{2}}{a},\qquad a=
a_{0}+r_{0},
\label{4.3}$$ where $\kappa$ is the gravitational constant, $$\Delta\epsilon\sim \frac{\hbar^{2}}{Ma^{2}},
\label{4.4}$$ and $$L=R.
\label{4.5}$$ Thus eq.(\[4.1\]) reduces to $$\lambda e^{-\sqrt{\lambda}R/a}\ll 1
\label{4.6}$$ with $$\lambda\sim \frac{\kappa M^{3}}{\hbar^{2}}a,
\label{4.7}$$ or $$\lambda\sim\left( \frac{M}{m_{P}} \right)^{3}
\frac{a}{l_{P}},
\label{4.8}$$ where $m_{P}$ is the Planck mass and $l_{P}$ is the Planck length.
We have by eq.(\[2.3\]) $$\frac{R}{a}\gg 1,
\label{4.9}$$ so that the condition (\[4.6\]) is fulfilled. In fact, since by \[8\] $$\lambda\gg 1
\label{4.10}$$ holds, the inequality (\[4.6\]) is very strong.
Thus nonreality of many-place gravitational autolocalization is closely related to nonreality of tunneling in the conditions under consideration.
Acknowledgment {#acknowledgment .unnumbered}
==============
I would like to thank Stefan V. Mashkevich for helpful discussions.
[99]{}
Vladimir S. Mashkevich, [*Indeterministic Quantum Gravity*]{} (gr-qc/9409010, 1994). Vladimir S. Mashkevich, [*Indeterministic Quantum Gravity II. Refinements and Developments*]{} (gr-qc/9505034, 1995). Vladimir S. Mashkevich, [*Indeterministic Quantum Gravity III. Gravidynamics versus Geometrodynamics: Revision of the Einstein Equation*]{} (gr-qc/9603022, 1996). Vladimir S. Mashkevich, [*Indeterministic Quantum Gravity IV. The Cosmic-length Universe and the Problem of the Missing Dark Matter*]{} (gr-qc/9609035, 1996). Vladimir S. Mashkevich, [*Indeterministic Quantum Gravity V. Dynamics and Arrow of Time*]{} (gr-qc/9609046, 1996). Vladimir S. Mashkevich, [*Indeterministic Quantum Gravity and Cosmology VI. Predynamical Geometry of Spacetime Manifold, Supplementary Conditions for Metric, and CPT*]{} (gr-qc/9704033, 1997). Vladimir S. Mashkevich, [*Indeterministic Quantum Gravity and Cosmology VII. Dynamical Passage through Singularities: Black Hole and Naked Singularity, Big Crunch and Big Bang*]{} (gr-qc/9704038, 1997). Vladimir S. Mashkevich, [*Indeterministic Quantum Gravity and Cosmology VIII. Gravilon: Gravitational Autolocalization*]{} (gr-qc/9708014, 1997).
[^1]: E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Remi Monceau-Baroux'
- Rony Keppens
- Zakaria Meliani
bibliography:
- 'article.bib'
title: Effect of angular opening on the dynamics of relativistic hydro jets
---
Introduction
============
Relativistic plasma jets have been exhaustively studied and observed in detail in radio galaxies such as Abell 2052, Hercules A, or Cygnus A [@Krause2005; @Krawczynski2012]. From multi-wavelength observations in specific sources we have understood important aspects of the ways in which jets are emitted by active galactic nuclei (AGNs) and form pronounced radio lobes [@blandford1974; @scheuer1973; @cohen2007; @Lister2009].
While observations give access to dynamic parameters such as the Lorentz factor $\gamma$ or an estimate of the energy flux from the overall luminosity, uncertainties remain in the properties of the jet matter itself, for instance, its composition, density or pressure and the precise topology and strength of the magnetic field. Simulation is a powerful tool to make estimates these parameters, and one can study different scenarios with them, e.g. the composition of the jet consisting of mainly electrons and protons, or electrons and positrons [@kundt1980].
Many astrophysical relativistic jets have a high Lorentz factor, with $\gamma$ varying from 5 to 30 for AGN jets [@kellermann2004]. Therefore, relativistic dynamics models are necessary to study them.
Jets associated with AGNs have been separated into different classes. Among the observational classification of radioloud AGN, we find weaksource Fanaroff-Riley class I objects (FR-I, [@fanaroff1974]), with a brighter luminosity near the origin of the jet. An example is the jet in the 3C 296 object [@laing2006]. Powerful sources are known as Faranoff-Riley class II radio galaxies (FR-II) and lobe-dominated radio-loud quasars. An example is the jet in the 3C216 object [@fejes1992]. In all cases, the jet has been shown to have a relativistic velocity with a Lorentz factor of up to 5 for FR-I and a Lorentz factor of several dozens for FR-II [@M.Marti1997; @cohen2007; @Lister2009]. We concentrate here on FR-II jet parameters.
Previous works have shown how a jet can be launched on opening magnetic field configurations [@blandford1982]. For relativistic magnetized winds from a central object, additional collimation by disk winds may be needed for their ultimate collimation, see @bogovalov2005. Those works illustrate how hard it is to obtain a fully collimated jet with an overall cylindrical geometry. Observations clearly show finite opening angles of jets, e.g. @puskarev2011 [@zezas2005; @frazer2000]. Flaring models that fit observations well were made for FR-I jets, as in the work of @laing2006. In this model, the original angle obtained from collimation in the source increases due to internal characteristics. For FR-II jets, although the beam still shows a finite width and opening angle, this conical geometry is often discarded by taking a wider width of the beam, or assuming a precession of the source [@S.M.O'Neill2005]. Few models invoke an actual conical geometry up to a few degrees [@Z.Meliani2008]. The cylindrical injection models dominate because FR-II jets are believed to be quickly re-collimated by internal shocks. Because we studied jet propagation beyond 0.5 parsec away from the central object, we ignored general relativistic as well as magnetic effects and performed pure hydro studies for varying opening angles. We assumed axisymmetry of the jet and therefore used 2D simulations. This allowed us to explore parametrically hydrodynamic relativistic simulations. Nevertheless, a more realistic study would imply 3D simulations, which we will perform in forthcoming studies.
The interest in studing relativistic jets associated with AGNs has increased recently, since jets feed back energy into the interstellar/intergalactic (ISM/IGM) medium, which makes them candidate feedback processes at play in the larger scale formation of galaxy clusters. For that reason we need to quantify the energy transfer efficiencies of AGN jets for cylindrical and conical jets.
Multiple mechanisms are thought to be responsible for the energy transfer between the jet and its surroundings. Direct energy conversion takes place at the end of the jet beam, immediately beyond the Mach disk surface of the jet (in a hydro case), where kinetic energy becomes thermal energy through efficient relativistic shocks. Hydrodynamic instabilities taking place at the interface between the shocked jet and shocked ISM matter mix the two fluids, allowing for energy transfer from the now hot outer part of the beam to the surroundings of the jet. Additionally, a bow shock forms around the entire jet cocoon, propagating in all directions and transferring energy to further regions of space. We quantify the relevance of each and how a finite opening angle at the jet inlet affects these energy transfers.
Many other dynamical features are known from observations, for instance the existence of X-ray blobs along the jet path. Different scenarios are studied to explain this emission, such as particle acceleration at bow shocks when stars cross the beam of the jet. We propose that our jet dynamics model to explain those blobs without invoking external agents such as star-related bow shocks. This explanation revisits the role of reconfinement shocks to explain these knots, as in the work by [@komissarov1997].
In the second part of this article we describe the relativistic hydrodynamic model and the way we use it with our code, MPI-AMRVAC. We discuss our settings and initial conditions and their motivation. Part 3 will compare our different simulations by giving a parametric study of the effect of an opening angle on the dynamics of our jets. This comparison studies the geometry and energy content of the jets. The fourth part discusses two aspects found in our simulations that relate to the internal structure of the jet: a scenario for X-ray blobs and the dynamically formed, layered structure of the jet. We finish with general conclusions obtained from this work.
Simulation
==========
Governing equations {#equation}
-------------------
Like in the previous work of @Z.Meliani2008, we assumed axisymmetry. We use a relativistic hydrodynamic model one with the relativistic variant of the Euler equations with the Mathews approximation [@mathews1976] to the Synge gas equation as a closure. They express the conservation of the different relevant quantities. The first equation is the conservation of particle number
$$\frac{\partial \rho \gamma}{\partial t} + \vec{\nabla} \cdot (\rho \gamma \vec{v}) = 0\,,
\label{eulerdensity}$$
where $\gamma$ is the Lorentz factor, $\rho$ is the proper density and $\vec{v}$ is the three-velocity of the fluid in the chosen lab frame. The second equation is the conservation of momentum, written as
$$\frac{\partial \vec{S}}{\partial t} + \vec{\nabla} \cdot (\vec{S} \vec{v}) + \vec{\nabla} p = 0 \,,
\label{eulermomentum}$$
where $p$ is the pressure in the fluid frame and $\vec{S}$ is the momentum density defined as
$$\vec{S} = h \gamma^2 \rho \vec{v} \,.$$
The relativistic specific enthalpy $h$ is defined as
$$h = \frac{1}{2}\left((\Gamma+1)\frac{e}{m_p}-(\Gamma-1)\frac{m_p}{e}\right) ,$$
where $\Gamma$ is a parameter, taken as the non-relativistic value of the polytropic index of the gas $\frac{5}{3}$, $e = m_p +e_{th}$ is the specific internal energy including the rest (proton) mass $m_p$ and the specific thermal energy $e_{th}$. This is equivalent to using a spatially varying, effective $\Gamma_{\mathrm{eff}} =\Gamma - \frac{\Gamma-1}{2}(1 - \frac{m_p^2}{e^2})$ that varies between its classical value of ${5}/{3}$ and its relativistic value of ${4}/{3}$. The third equation expresses the conservation of energy density
$$\frac{\partial \tau}{\partial t} + \vec{\nabla} \cdot (\vec{S} - \rho \gamma \vec{v}) = 0 \,,
\label{eulerenergy}$$
where the energy density $\tau$ in the lab frame with the rest mass contribution already subtracted, is written as
$$\tau = \rho h \gamma^2 - p - \rho \gamma\,.$$
The Mathews approximation to the Synge gas equation is our closure relation, and it writes as
$$p = \left(\frac{\Gamma - 1}{2}\right) \rho \left(\frac{e}{m_p} - \frac{m_p}{e}\right)\,.
\label{syngeequation}$$
The speed of light is taken equal to unity in our model and therefore in all equations above.
Code and implementation
-----------------------
The code we used is MPI-AMRVAC[^1] [@vac2012], designed to solve a set of equations in conservative form, as is the case for our relativistic HD model. The solver is a hybrid version of the HLLC scheme, for which a validation can be found in the work of @Z.Meliani2008. The novelty of this scheme is that it switches to a TVDLF solver where the HLLC solver could induce spurious oscillations. More validation tests for relativistic HD and MHD can be found on the code website and in the work of @vac2012.
The code incorporates block-tree adaptive mesh refinement (AMR), and we used a weighted combination of the normalized second derivative of the density variable $\rho\gamma$, the pressure $p$ and the Lorentz factor to follow the dynamics of the jet and automatically refine the grid where dynamics takes place. Additionally, a forced refinement of the grid was activated around the inlet region of the jet, up to the maximum level specified, to accurately handle the bottom boundary.
Initial conditions {#IC}
------------------
The rectangular domain was defined to be 20 by 5 pc jet beam radius in length and width, respectively $(z,r)$, with a length normalization of 1 parsec which corresponds to 400 by 100 jet beam radius as we show below. The domain did not include the central engine that launches the jet, but started at $z_0=0.5$ distance from the source. therefore we can avoid treating the close surroundings of an AGN core, which requires general relativistic theory. From a numerical point of view, all AMR sets start with a base resolution of 300 by 190 grid points, in $(z,r)$ respectively, allowing four levels of refinement (hence an effective resolution of 2400 by 1520). We made sure that the resolution allowed us to resolve the jet properly: the jet width on inlet, fixed to be $r_b=0.05$ parsec, is 15 grid points wide at the maximal level of refinement taht is enforced in the jet region. We set the density on the inlet of the jet $\rho_{\mathrm{jet}}$ such that the jet total kinetic luminosity is $L_{\mathrm{jet}}=10^{46}$ erg.s$^{-1}$ with a Lorentz factor $\gamma$ equal to 10, in a jet with radius $r_b$ given above, all values representative for an FR-II type AGN jet.
For the ISM we chose a model similar to the beta model of the King atmosphere [@king1962] profile. This assumes that the ISM has been structured into a density variation by the winds coming from the central object [@S.M.O'Neill2005; @Krause2005], resulting in a slowly decreasing density inside a core radius. Outside of this radius the density drops rapidly up to an asymptotic value. This ISM density then follows $$\rho_{\mathrm{ISM}} = \rho_0\left(1+\frac{z^2}{z_\mathrm{c}^2}\right)^{-\beta},
\label{kingatmo}$$ where $z_c$ is setting the size of the core of the galactic source, $\rho_0$ quantifies the density at the source of the jet ($\rho_0 = 5000$ g.$\mathrm{cm}^{-3}$ for cases A to E and $\rho_0=500$ g.$\mathrm{cm}^{-3}$ for cases F to H), and $\beta$ is a parameter set here to 0.75. The distance $z$ is here the axial direction instead of the actual radial distance from the central engine. This was done for numerical convenience, because it more easily handles the bottom boundary and avoids an artificial ’pixelization’ of the density background at low refinement levels. We moreover assumed that the jet will not propagate over a large radial distance, resulting in a distance to the source $d_{source}\simeq z$ for $r$ small. Normally, $z_c$ is set to a few kpc, corresponding to the average size of a galaxy. This would result in a flat density profile in our parsec scale simulation. This is the case for simulations A and B, which were injected in uniform media, denoted uniform’ in Table \[parameter\]. For simulations C, D and E we set $z_c$ to be 0.5 to investigate a possible change in the propagation of the jet as it becomes more dense than the ISM. With this setting in the domain the density profile behaves as a simple decreasing power law. This is denoted as King’ in the summary Table \[parameter\]. Finally, simulations F, G and H are in between with a $z_c$ at 5 pc, called King II’ in the summary table.
**Beam luminosity** **Domain size (pc)** **Domain resolution**
---------------------- ---------------------- -----------------------
$10^{46} erg.s^{-1}$ \[0.5 20\]x\[0 5\] 2400x1520
**Beam radius (pc)** **Beam $\gamma$**
0.05 10
: General parameters for all simulations.
**Case** **Density ratio** **Opening angle** **Density profile**
---------- ------------------- ------------------- ---------------------
A $0.01$ 5 uniform
B $0.0001$ 5 uniform
C $0.018$ 0 King
D $0.018$ 5 King
E $0.018$ 10 King
F $0.1$ 0 King II
G $0.1$ 5 King II
H $0.1$ 10 King II
: Parameters distinguish between the simulations.[]{data-label="parameter"}
The non-dimensional numeric value for the pressure was set on the whole domain to be uniformly small ($p_0 = 10^{-3} \rho_0$ for A to E and $p_0 = 10^{-2} \rho_0$ for cases F to H) so that it does not dominate the early stage of the dynamics. Having a uniform pressure also removes the need to have a gravity potential to keep the ISM in force balance. We call $\eta$ the density ratio between the jet and the ISM, i.e. $\rho_{\mathrm{jet}}/\rho_{\mathrm{ISM}}=\eta$ at the position $z=z_0$.
We chose for cases A and B a ratio $\eta$ equal to 0.01 and 0.0001, respectively, investigating the dynamics of an underdense jet. For cases C, D and E, this ratio was taken to be $\eta=0.018$. For those three cases the density of the ISM decreases rapidly in the domain and we expect the density of the jet and the ISM to match around the end of the domain. Cases F, G and H have a ratio $\eta=0.1$ and a core’ inside the domain, resulting in our simulation box in a less rapidly decreasing density profile of the ISM, the density of both jet and ISM match around the middle of the domain.
We simulated on domains $z\in[z_0,z_{\mathrm{out}}]$ and $r\in[0,r_{\mathrm{out}}]$. The boundaries at $z_{\mathrm{out}}$ and $r_{\mathrm{out}}$, as seen in figure \[jetgeo\], are open boundaries. The bottom boundary at $z_0 = 0.5$ pc is an open boundary for $r \geq r_b$ and was set to the jet injection for $r \leq r_b$, where $r_b = 0.05$ pc. The $r = 0$ boundary is the symmetry axis. We typically stopped the simulation before the jet reached the boundaries $z_{out}$ and $r_{out}$ to prevent energy from leaving the domain because we need to quantify energy deposition.
![Geometry of the jet as used in the simulation: the jet enters the domain at $z_0=0.5$ pc from the source with a width setting $r_b$ as calculated assuming a first angle of 5. A flaring-like case assumes a fixed angle $\theta$, which is added to a reference cylindrical jet, thereby introducing a ‘virtual source location’. The actual domain of the simulation is comprised between $z_0$ and $z_{out}$ for the axial direction and $r=0$ and $r_{out}$ for the radial direction. The velocity field is initiated in a finite region up to $z_1$, according to the formulae given in the text.[]{data-label="jetgeo"}](article_figure/jetgeo){height=".3\textheight"}
Fig. \[jetgeo\] gives the geometry of the jet generated with a flaring-like scenario. In the initial conditions, the jet is considered to already have propagated for sometime and to have reached $z_1=1$ pc. The initial radius of the jet at inlet is fixed at $r_b=0.05$ pc and considers the jet to be coming out of a central source located at 0.5 parsec below the grid starting at $z_0$. Then according to the flaring-like scenario, introducing a fixed angle $\theta$, we can obtain the position of a virtual source at $z_v$, which for the angles of 5 and 10 degrees is somewhere in between $z=0$ and $z=z_0$, as seen in Fig. \[jetgeo\]. This was used to initiate the entire jet region within the computational domain, by considering the flux coming from this source and adopting a constant mass flux along the section of the cone. This means that we initialized density and velocity in this part of the domain with the following equations:
$$\rho(r,z,t=0)=\rho_{\mathrm{jet}}\frac{r_b^2}{(r_b+(z-z_0)\tan(\theta))^2},
\label{}$$
$$v_\mathrm{z}(r,z,t=0)=v~\cos(\arctan(\frac{r}{z-z_\mathrm{v}})),
\label{}$$
$$v_\mathrm{r}(r,z,t=0)=v~\sin(\arctan(\frac{r}{z-z_\mathrm{v}})).
\label{}$$
Results -parametric study
=========================
Dynamics - general description {#general}
------------------------------
As previously pointed out, we aim to study the transfer of energy between the jet and the ISM. To do this, we need to trace and quantify the significance and role of the different regions of the domain that characterize the jet beam, the instability-dominated region, the shocked ISM (SISM), and the unperturbed ISM. The main feature of the region where the beam of the jet is located is its high value for the Lorentz factor. We empirically defined the threshold value to be $\gamma \geq 2$. With this criterion, the beam is well-captured without pollution from the cocoon region, which still has relativistic flows where $2 > \gamma > 1 $. The instability dominated region was located with the curl of the velocity field: because we did not include jet rotation and adopted jet axisymmetry, only the toroidal ($\phi$) component of the curl is non vanishing. We can instantaneously capture this mixing region where instabilities develop by locating where the magnitude of the curl component is higher than a threshold value of 1. The shocked region is identified by quantifying where the instantaneous energy density $\tau$ is not equal to its initial value. An example of these criteria applied to the simulation of case E at time 20 of the simulation can be seen in Fig. \[mask\]. Because the regions are not suppose to overlap, we substracted the jet from the mixing region, and both jet and mixing region from the shocked ISM.
![Example of our method to instantaneously identify jet regions for a jet opening angle of 10 at time t=20 (case E). We show the masks locating the jet (left) mixing region (center) and the shocked ISM (right). We identify the jet beam by $\gamma \geq 2$, finding the mixing region where $|(\nabla \times \vec{v})_{\phi}| > 1$ without the jet beam, and the cocoon where $\tau \neq \tau_{t=0}$ with beam and mixing region excluded.[]{data-label="mask"}](article_figure/mask-jet5){width=".4\textwidth"}
Figs. \[comparison0-10\] and \[comparison5-10\] show instantaneous comparisons between cases C, D and E, with a 0, 5, and 10opening angle with a King-type density profile, respectively. Common features can be described: the beam of the jet first shows an unperturbed region in which the matter propagated without interaction. It is then re-collimated and a shock forms that finally reaches the axis. Figure \[rhoprofile\] shows the density variation along the symmetry axis at the same time $t=20$ for the three cases C, D and E. For the finite opening angle cases D and E, this plot shows that the position of this re-collimation shock (at about $z\approx 3$) slightly changes for varying opening angle. We discuss in section \[intern\_structure\] how these shocks, which are repeated along the symmetry direction, can provide a scenario for X-ray blobs.
![Comparison of the density distribution at time t=20 for an opening angle of 0 (left - case C) and 10 (right - Case E). A wider opening angle increases the width of the jet beam that stays collimated. The maximal reach of the jet beam is inversely proportional to the opening angle. The radial reach of the shock surrounding the jet remains of the same order.[]{data-label="comparison0-10"}](article_figure/density_comparaison_0-10teta){width=".4\textwidth"}
![Comparison of the proper density distribution at time t=20 for an opening angle of 5 (left - case D) and 10 (right - Case E). Conclusions from visual investigation agree qualitatively with fig. \[comparison0-10\].[]{data-label="comparison5-10"}](article_figure/density_comparaison_5-10teta){width=".4\textwidth"}
![Density profile along the symmetry axis of the jet for cases C, D and E and for the unperturbed atmosphere. All cases show an increased density at the head of the jet beam. The maximal $z$ reach is inversely proportional to the opening angle. Case C (cylindrical case) keeps a near constant density. Cases D and E show a series of nodes with increased density. The position of the nodes depends on the opening angle: the wider the opening angle, the longer the internodal distance. The jet density in case C becomes higher than the ISM density from about $z\approx 12$, whereas for the other two cases, it remains below the ISM density profile throughout the domain.[]{data-label="rhoprofile"}](article_figure/densityprofile){width=".4\textwidth"}
The head of the jet beam sweeps up matter and the wider the opening angle of the jet, the wider this head is and the more matter is pushed away. This qualitative statement can be quantified using the masks introduced earlier: we can estimate the mass swept up by quantifying $$M_{\mathrm{swept}} = \iint_{\mathrm{beam}(t)} \rho_{\mathrm{ISM}}(r,z,t=0) \,2\pi r\,dr\,dz,$$ which is shown in Fig. \[sweptedmatter\]. A bow shock surrounds the whole structure and propagates into the ISM, transferring energy to an expanding region. By using the filters previously discussed, we performed a series of diagnostics: by integrating over the different regions, we can quantify the effect of the variation of the opening angle on, e.g. the fraction of the domain volume occupied by the mixing region, which we show in Fig. \[volumeinstab\]. We can also instantaneously locate the overall position of the frontal bow shock as seen in the right panel of Fig. \[mask\]. Therefore we can quantify the maximal radial and axial expansion of this region over time as seen in Figs. \[shockISM\] and \[hpos\]. At the dynamically evolving contact interface between matter from the jet and the swept-up ISM, instabilities develop. These are of two kinds: Rayleigh-Taylor instabilities forming at the accelerated and decelerated contact interface of the jet as it propagates into the ISM, which then are advected along the side of the jet. Additionally, Kelvin-Helmholtz instabilities form along the jet beam flanks due to the velocity gradient between the moving matter of the jet that propagates axially and the slower matter from the interaction region with a more complex velocity field. The formation of both those instabilities can be followed with the filters discussed previously. We quantify how much of the energy transfer takes place when these instabilities mix matter from the jet and from the ISM. By looking at a plot of the curl of the velocity field, we can highlight those instabilities. As expected, at the interface between the matter of the jet with a high Lorentz factor and the more complex and slower matter from the interaction region, we can see a strong shearing of the fluid with high negative value of the curl of the velocity field. We can differentiate likewise the unperturbed part of the beam, and the part of the beam where recollimating shock waves are found. The front of these shock waves, as clearly identified from a pressure plot (not shown), coincide with a locally strong shearing. The vorticity mapping makes us able to track the instabilities and their evolution in time. By integrating the instantaneous energy density over the different regions as defined by the filters, we can obtain an estimation of energy transfer to the different regions.
![Matter swept up by the jet beam for different cases. For the same density profile (cases C, D and E, or cases, F and H) wider jets sweep away more matter. We can see that the density profile clearly matters. []{data-label="sweptedmatter"}](article_figure/sweptedmatter){width=".4\textwidth"}
![Volume fraction of the mixing region up to a time t=20 for different cases. The volume depends both on the opening angle and the properties of the ISM density profile. The volume increases for a smaller opening angle although the radial reach of this region diminishes. The maximal $z$ reach of the jet is therefore dominant over the radial expansion of the mixing region.[]{data-label="volumeinstab"}](article_figure/volumeinstab){width=".4\textwidth"}
![Radial expansion of the shocked region over time for different cases. Despite initial differences, the radial width of the jet-influenced region does not vary much for a varying opening angle.[]{data-label="shockISM"}](article_figure/shockISM){width=".4\textwidth"}
![Position of the bow shock marking the head of the jet over time for different cases. The density ratio and the opening angle are both relevant for the propagation of this bow shock. The wider the opening angle, the more it decelerates. The higher the density ratio $\eta$, the slower the shock front deceleration.[]{data-label="hpos"}](article_figure/hpos){width=".4\textwidth"}
![Fraction of the domain volume occupied by the shocked ISM region up to a time t=20 for different cases. This volume fraction of the shocked region both depends on the opening angle and on the density profile of the ISM. We know from Fig. \[shockISM\] that all simulations have comparable radial reach, therefore we expect the volume of the cases to be ordered as in Fig. \[hpos\]. This is the case within the same series of simulations (cases C, D and E, and cases F and H).[]{data-label="volumeshockISM"}](article_figure/volumeshockISM){width=".4\textwidth"}
Dynamics - overall propagation distances {#propagation}
----------------------------------------
As mentioned before, Fig. \[hpos\] shows the axial offset position of the bow shock front over time for an opening angle of the jet of 0 (case C), 5 (case D) and 10 (case E) with a King-type density profile as defined in section \[IC\]. It shows the same for the pair of 0 (case F) and 10 (case H) with a King-II-type density profile. Finally, it quantifies the same position for an opening angle of 5 with a flat density profile (case A). All data sets start similarly as the head of the jet propagates at early time with the velocity obtained from the expression for pressure-matched jet propagation in 1D [@M.Marti1997]. Case A highlights how the density structure of the ISM affects the propagation. This uniform density case decelerates the head of the jet more as the influence of the ISM augments, in contrast to cases C, D and E, where the ISM density decreases. The inferred velocity of the head of the jet, quantified as the slope of the curves shown in Fig. \[hpos\], has a low Lorentz factor around 1, and the velocity within the beam of the jet itself has a Lorentz factor of 10 or more (due to a compression of the flow, which can increase the Lorentz factor above its initial value). Different effects influence the acceleration or deceleration of the head: the principal one is the continuous ram pressure force exerted by the matter from the jet. As the jet beam propagates, the ram pressure force at the Mach disk is proportional to its surface area and depends on velocity. A significant factor is the accumulated mass of matter swept up by the jet beam. This was quantified earlier in Fig. \[sweptedmatter\]. We discussed before that one of the main mechanisms of transfer of energy between the jet and the mixing region is dependent on the conversion from kinetic to thermal energy at the Mach disk, which is then mixed with the surroundings of the jet. Therefore, monitoring the velocity of the Mach disk, and the closely related propagation speed of the frontal bow shock as inferrable from Fig. \[hpos\], gives us an indication on the efficiency of this energy transfer: one would expect that the kinetic energy from the jet would be transferred into thermal energy for the cocoon.
The upper limit of propagation is a ballistic behavior of no deceleration of the jet. The closest to this is the fully cylindrical jet of case C (0, King), consistent with the least swept-up matter shown in Fig. \[sweptedmatter\]. The ram pressure felt by the head of the jet as it propagates decreases as the density of the ISM decreases in the axial direction. For cases D and E (5and 10King), we see a clear deceleration of the velocity of the head of the jet, due to a wider opening angle of the jet. Ultimately, we expect the mass of material piled up at the head of the jet to be high enough to stop its propagation, and this could happen sooner for wider opening angles and lead to shorter lengths of propagation (but this is on time-length scales beyond our simulations here). The test case scenario A with a flat density profile shows a strong deceleration as the jet encounters a constant high ISM density. As for the previous cases, the total mass of matter pushed aside by the jet increases, decelerating the head of the jet even more. This can be seen in Fig. \[hpos\] where case A (5, uniform density profile) shows the lowest curve. By comparison with case D (still 5but King density profile) it becomes evident that the density profile of the ISM is a governing effect. All these results are consistent with Fig. \[sweptedmatter\]: a wider opening angle increased the mass of matter pushed away and therefore implies a higher value of the ram pressure. For flat density profiles, the density does not decrease axially and we reach even higher values for the swept-up matter and therefore a much more significant deceleration of the head of the jet. From Fig. \[hpos\] we can expect the same relation between the opening angle for the jet and the increase of energy transfer between the jet and the ISM. We also infer that a wider opening angle for the jet translates into a reduced reach of the jet into the ISM. This calls for longer simulations to study the longer term, far distance behavior.
Cases F and H (0and 10, King II density profile) are given in all figures as the limit cases of the second series of simulations entering a varying density atmosphere. The intermediate opening angle case G (5, King II) is not quantified in the figures, but we verified that it is always intermediate between cases F and H. The conclusions obtained from the study of the second serie and their comparison between one another are similar to the ones obtained from the related series of cases C, D and E (0, 5, 10, King). Case B (5, uniform density profile, $\eta=10^{-4}$) is found to follow all trends that were identified with case A (5, uniform density profile, $\eta=10^{-2}$), only shifted in intensity due to the difference of the density ratio $\eta$.
In summary: A wider opening angle results in shorter axial propagation, while the density of the medium in which the jet propagated plays the expected role: a higher density of the external medium results in a higher deceleration.
Dynamics study - volumes and radial expansion
---------------------------------------------
As the jet propagates into the ISM, a bow shock front forms around it and propagates. This front stays always fairly close to the end of the jet beam, but radially it expands much faster than the jet which is re-collimated. As this bow shock propagates into the ISM, it transmits energy to the local matter. This is the main way to transmit energy to far radial distances. Fig. \[shockISM\] shows time evolution of the maximal radial distance reached by the SISM bounded by the bow shock, for several representative cases.
For propagation into a King atmosphere, we find that this maximal radial expansion of the SISM stays on the same order irrespective of the initial jet opening angle. While the jet beam on input is only 0.05pc for all cases, the bow shock’s sideway influence region increases to 5 parsec within the simulation time. For an opening angle of 5, but with a flat density profile instead of a King-type atmosphere (case A), the radial expansion is slower and stays within 2 pc in the same timespan. We can easily explain this behavior by considering that this shock is propagating sideways with a speed influenced by the local sound speed. As said in section \[IC\], the initial pressure in the ISM is constant, whereas for case A, density is also constant, but density decreases with axial distance for other cases. Therefore the local sound speed (which is a factor of $\frac{p}{\rho}$) in the ISM increases in the axial direction. This in combination makes the sideways expansion more dependent on the ISM circumstances than on internal beam properties. The latter, as explained previously, does affect the beam axial propagation behavior, as shown in Fig. \[hpos\], where smaller opening angles meant more deceleration. Associated with this difference in reach in the axial direction for the jet, Fig. \[volumeshockISM\] shows a higher volume fraction for the shocked region for smaller opening angles for the same density profile (compare in particular cases C, D and E, i.e. 0, 5and 10, and King-type). This figure quantifies the ratio $$\frac{V_{\mathrm{SISM}}}{V_{\mathrm{domain}}}=\frac{\iint_{{\mathrm{SISM}}(t)} 2\pi\,r\,dr\,dz}{\iint_{{\mathrm{domain}}} 2\pi\,r\,dr\,dz} \,.$$ In reality, the pressure in the ISM is not constant as we move away from the source. Therefore to link those statements to observations, we would need to study the evolution of the ratio $\frac{p}{\rho}$ of the surroundings of AGN jets, which is still a problem for future study.
From Figs. \[comparison0-10\] and \[comparison5-10\], one would expect the volume of the mixing region for cases C, D and E to be more significant for wider opening angle. But from Fig. \[volumeinstab\], which quantifies the instability dominated volume fraction, we can see that the radial expansion of the instability-influenced region is compensated for by the lesser axial propagation of the jet. By comparison, case A, which encounters a much higher external density, and therefore a much higher ram pressure, is confined to a smaller volume. For the jet beam region itself (as identified at each time with the mask as in the left panel of Fig. \[mask\]), the main factor for the variation of its volume (figure not shown) is the radial expansion set by the initial opening angle: for an opening angle of 0 (case C), the jet does not expand radially and therefore its volume increases linearly with its axial expansion. For a wider opening angle of 5 and even more for 10, the jet beam has a varying radial expansion due to recollimation by internal shocks and the varying ram pressure at its head.
In conclusion: The radial propagation of the frontal bow shock is dominated by the external medium parameters. This shock is the main means of energy transfer to large radial distances.
Temporal evolution of the energy content
----------------------------------------
We now discuss the evolution of the energy content by using various masks as described earlier in the text, to differentiate the contributions of different spatial regions (see section \[general\]). We will do so by quantifying the fraction of excess energy present in each part of the domain as $$\frac{E_{\mathrm{mask}}}{E_{\mathrm{domain}}}=\frac{\iint_{\mathrm{mask}(t)} \tau_{\mathrm{excess}}(r,z,t) \,2\pi r\,dr\,dz}{\iint_{\mathrm{domain}} \tau_{\mathrm{excess}}(r,z,t) \,2\pi r\,dr\,dz},$$ where as excess energy we define the energy in each point of the domain, minus the contribution from the ISM at the same point for $t=0$: $$\tau_{\mathrm{excess}}(r,z,t)=\tau(r,z,t)-\tau_{\mathrm{ISM}}(r,z,t=0)\, .$$ For various choices of masks, we give the temporal evolution of the energy fraction present in the different parts of the domain, namely the mixing region (Fig. \[energyinstab\]), the shocked ISM (Fig. \[energypertub\]) and their combination (Fig. \[energytransfer\]). Note that the denominator is equivalent to the time integration of the jet influx after subtracting a small loss term $E_{\mathrm{loss}}$ that leaves through the open part of the bottom boundary: $$\begin{aligned}
E_{\mathrm{domain}} &=&\int L_{\mathrm{jet}}\,dt -E_{\mathrm{loss}}\, \nonumber \\
&=& \iint_{\mathrm{domain}} \tau_{\mathrm{excess}}(r,z,t) \,2\pi r\,dr\,dz \, .\end{aligned}$$ It is in particular non-zero at $t=0$ due to the finite region of the domain where we initialize the jet properties ($z\in[z_0,z_1]$ and $r$ within the specific conical segment). For the masks mentioned (mixing region and SISM), the numerator starts at a zero value.
For Figs. \[energyinstab\] - \[thermicfraction\], the initial behavior up to time 1 in code units is mostly due to the way we initiate the simulation with a sharp-edged conical jet segment within the domain. More relevant for our discussion is the trend of these curves after this time.
In section \[propagation\] we postulated that the energy transfer in the mixing region is dominated by the energy that is transferred through the end of the jet beam (Mach disk location). Therefore, a relation should exist between the total mass of swept up matter by the beam (as quantified earlier in Fig. \[sweptedmatter\]) and the energy transferred to the mixing region. Indeed, as we can confirm from Fig. \[energyinstab\], within the same series of simulations (cases C, D and E: 0, 5and 10, King or F and H: 0, 10, King II), by keeping the density profile fixed, the wider the opening angle, the higher the mass of swept-up matter and the excess energy transfer in the mixing region. In all cases, we observe a decrease of the energy transfer rate to the mixing zone up to an asymptotic regime where the transfer becomes time-independent, e.g. cases E and H. Note that there is a clear trend for higher energy transfer for larger opening angles, minimizing the role of the ISM stratification.
Fig. \[energypertub\] shows the same quantification as before, with the mask now identifying the SISM regions. Similarly as before, wider opening angles result in higher energy transfer. In contrast to the mixing region case, the ISM stratification plays an important role in the energy transferred to the SIMS region. Indeed, Fig. \[energypertub\] shows a clear difference between cases C, D and E, and cases F and H. We point out that the energy transfer in the shocked ISM region is not dominated by the volumetric evolution of this region as seen in Fig. \[volumeshockISM\]. We can conclude that the dominant parameter for energy transfer is here the density profile of the ISM.
Combining both regions, Fig. \[energytransfer\] then quantifies how much energy is transferred to the jet beam surroundings as a whole. Contrary to our initial expectation that most of the energy transfer should occur at the Mach disk and mixing region alone, Fig. \[energytransfer\] shows that the total energy transfer is dominated by the transfer in the SISM region. We observe, on average, a ratio of 3 to 1 between the energy fraction present in the SIMS region, and its mixing region counterpart. Moreover, we identify the obtained percentage of the transferred energy as an estimate of the time-average energy transfer rate for the duration of our simulations,
$$\frac{\int L_{\mathrm{transfer}}(t) dt}{t} = \frac{E_{\mathrm{transfer}}(t)}{E_{\mathrm{domain}}(t)} \times L_{\mathrm{jet}} \,.$$
In case D (5 degree opening angle with a King density profile) we observe a transfer rate of $2.5\times 10^{45}$ erg ${\mathrm{s}}^{-1}$ (i.e. at $t=20$, 25 % of the total kinetic energy flux $E_{\mathrm{domain}}$ is injected into the surrounding medium). We can quantify the ratio over thermal and kinetic energy to the total energy even better by noting that the local energy density minus rest mass, $\tau$, can be approximately expressed by its thermal and kinetic contribution as
$$\tau_{\mathrm{thermal}} = p (\gamma^2\frac{\Gamma_{\mathrm{eff}}}{\Gamma_{\mathrm{eff}}-1}-1),$$
$$\tau_{\mathrm{kinetic}} = \gamma (\gamma-1) \rho c^2,$$
where $$\Gamma_{\mathrm{eff}} =\Gamma - \frac{\Gamma-1}{2}(1 - \frac{m_p^2}{e^2})
\label{Geffequation}$$ is the local varying effective polytropic index of the gas. In Fig. \[thermicfraction\] we use these individual contributions in the integral expressions, subtracting their initial ISM values. We then find that about 80 percent of the energy transferred to the ISM is composed of the thermal term, which again confirms the dominant contribution of the bow-shock-related heating.
![Fraction of the total domain energy excess present in the mixing region, which is dominated by instabilities, for the different cases.[]{data-label="energyinstab"}](article_figure/energyinstab){width=".4\textwidth"}
![Excess energy ratio in the shocked ISM region for different cases.[]{data-label="energypertub"}](article_figure/energypertub){width=".4\textwidth"}
![Total excess energy ratio in the mixing region and shocked ISM for different cases. By comparison with Figs. \[energyinstab\] and \[energypertub\] we conclude that the transfer is dominated by heating from the shocks in the shocked ISM region.[]{data-label="energytransfer"}](article_figure/energytransfer){width=".4\textwidth"}
![Thermal term of the total energy transfer.[]{data-label="thermicfraction"}](article_figure/thermicfraction){width=".4\textwidth"}
In conclusion: it is possible with our code to track where the energy is deposited and in which form, kinetic or thermal. For early stages of the propagation, the energy transfer is dominated by the heating by shocks which happens in the SISM region. We expect the transfer by mixing of the jet material and the surroundings to be dominant only at later times. It is also possible to quantify an average transfer rate of energy from the jet to its surroundings.
Dynamical details in conical relativistic HD jets
=================================================
Internal structure of the jet {#intern_structure}
-----------------------------
In Fig. \[rhoprofile\], we compared the density profile along the symmetry axis for several cases. By overplotting the ISM density profile, it is possible to observe the evolution of the density ratio $\eta = \rho_{\mathrm{jet}} / \rho_{\mathrm{ISM}}$. We recall that this ratio is set to 0.018 at the bottom boundary for several cases shown. For an opening angle of zero, the jet is cylindrical from the start of the domain and its density stays constant on average, whereas the density of the ISM decreases. Small fluctuations due to different internal shocks can be observed. For this case C (0, King), after t=10 the density of the jet becomes higher than that from the ISM at a distance of 11 parsec from the central engine.
When finite opening angles are present, a clear recollimation shock forms. Up to this recollimation shock, for an opening angle of 5 and 10 the jet density decreases as a factor of $\frac{1}{z^2 \tan( \theta)^2}$ with $z$ the distance to the source and $\theta$ the fixed opening angle of the jet. As $z$ increases, the recollimation shock wave propagates into the beam and compresses its density. The position at which this shock first reaches the symmetry axis evolves in time, until it finds an equilibrium recollimation position at 3 parsec for an opening angle of 5, and 4 parsec for an opening angle of 10. From time t=10 this quasi-equilibrium position is reached for both cases D and E, as in Fig. \[rhoprofile\] (similar trends emerge for the series F, G and H). A two-dimensional view of the near-stationary nodal structures is shown in Fig. \[node\] for case E at time $t=20$. This panel shows the density and effective polytropic index near the injection zone. Even if this does not dominate the dynamics, magnetic fields should be present in reality, enabling processes such as Fermi-II acceleration of particles in the nodes along the jet. This would result in an increase of the radiation processes at those positions, which is a scenario to explain the X-ray blobs found in observations of AGN jets.
We have a Synge-like closure relation, defined such that the effective polytropic index can vary between $5/3$, the classical polytropic index for a proton gas, to $4/3$, its relativistic value. We can recover the local value of this effective index and plot it following the equation \[Geffequation\]. This allows us insight into the thermodynamic condition of the gas. Figs. \[Geff0-10\] and \[Geff5-10\] show this effective polytropic index for cases C, D and E. We can see for both cases with a finite opening angle that the gas expanded and becomes cooler. This is due to adiabatic cooling in a freely expanding gas. The re-collimation shock reheats the gas and moves its polytropic index to the relativistic value of $4/3$. A zoom-in of this region was shown in Fig. \[node\] for case E. The main jet beam remains at a relativistic temperature. For an opening angle of 0 , the jet matter does not experience the adiabatic cooling phase and stays uniformly thermodynamically relativistic. We note that the core of the instability-related vortices also coincide with a mildly relativistic value of the polytropic index. In the variation of the effective polytropic index, we can also see how several secondary shocks traverse the SISM cocoon region, adding to the overall reheating.
![Details of two nodes along the axis. Internal structure of the jet beam appears with a finite opening angle: a static shock, in the lab frame, forms along the path of the jet, which can be the seat of Fermi-type-II acceleration of particles. The size of the image is $7\times 8$ pc with the two nodes at 3.5 and 7 pc along the symmetry axis.[]{data-label="node"}](article_figure/2node){width=".4\textwidth"}
![Comparison of the effective polytropic index for cases C (left) and E (right) at time 20 of the simulation. This figure reveals more clearly a series of secondary shocks that propagate in the shocked ISM region and also reheat it.[]{data-label="Geff0-10"}](article_figure/Geff_comparaison_0-10teta){width=".3\textwidth"}
![Comparison of the effective polytropic index for cases D (left) and E (right) at time 20 of the simulation.[]{data-label="Geff5-10"}](article_figure/Geff_comparaison_5-10teta){width=".3\textwidth"}
In conclusion: using of an opening angle for the injection of the jet modifies its internal structure and gives rise to recollimation nodes along its path.
Dynamic formation of a layered structure
----------------------------------------
We also point out another detail of the finite opening angle cases by showing aanother zoomed view on the near inlet expansion region for case E (10, King) in Fig. \[lowdensity\]. We repeat on the left the local polytropic index, while the right panel now gives the local vorticity value. Especially the latter panel of Fig. \[lowdensity\] hints at the presence of a sharp radial structure in the jet beam, itself surrounded by the instabilities of the mixing region. Fig. \[cutbeam\] shows a radial cut of the jet at fixed axial distances of 2, 5 and 10 parsec for cases C and E (0and 10, King). We can see at the interface between the jet and the shocked ISM that a layer forms with a width comparable with the beam width, with properties different from both media. For the cylindrical jet (opening angle of zero) this layer shows a lower density than the two surrounding regions with a clear local peak of the curl. For the 10 degrees opening angle we can see two such extrema in the local vorticity, present all along the jet from the root to the jet beam head, giving the beam a layered structure. These layered structures are sometimes invoked in studies of relativistic jets. @ostrowski2001 used this layered jet to explain increased emission at the external border of the jet beam, whereas @S.M.O'Neill2005 used it as a way to keep the jet in equilibrium with its surroundings. We recall that in our initial condition, the variation from the density of the jet to the density of the ISM is a step function. A sideways traveling rarefaction wave induces the shape we see eventually all along the entire jet beam boundary at later times. Case E is more complicated because the beam is surrounded by the growing instabilities and because of re-collimation, the edge of the beam is not radially static in space anymore.
![Presence of a dynamically formed layer surrounding the beam of the jet. This is another zoom on the inlet (as in Fig. \[node\]) for case E, now showing the (scaled) vorticity at the right.[]{data-label="lowdensity"}](article_figure/lowdensitylayer){width=".4\textwidth"}
{width=".4\textwidth"} {width=".4\textwidth"} {width=".4\textwidth"} {width=".4\textwidth"} {width=".4\textwidth"} {width=".4\textwidth"}
In conclusion: a radially layered jet structure is present also without an opening angle for the injection of the jet, but the latter does amplify this structure in a finite rarefaction zone.
Conclusion
==========
We quantified the fraction of energy present in the different jet-affected regions for realistic FR-II-type AGN jets. It was possible to show that the main mechanism of energy transfer from the jet to the ISM is through heating via shocks in the shocked ISM region, on top of mixing-related energy exchange. The energy transfer we measured for the SISM region dominates the same measurement for the mixing region with a ratio of 3 to 1 on average. It is important to note that in reality, the mechanisms studied here coexist with magnetic energy transfers and radiation transfer.
We discussed the role of finite opening angles on the dynamics of the jet, and in terms of efficiency of energy transfer. A wider opening angle increases the energy transfer but reduces the reach of the jet in its main direction of propagation and also reduces the portion of space that receives the energy. Our study followed relativistic Lorentz factor 10 jets for a lifetime of about 65 years and for a propagating distance up to 20 pc. We found that the amount of energy feedback from the jet to the ISM can vary from 10% to 70 % across all cases studied, where a total $10^{46}$ erg.s$^{-1}$ influx was assumed throughout. The jets were followed-up to asymptotic regimes where some of the energy transfer rates reached stationary values, which calls for larger scale, longer-time simulations to confirm these trends. This is because we know that astrophysical AGN jets propagate on scales of kpc to Mpc for extended lifetimes. These longer time and larger scale runs can be found in the works of [@perucho2011]. They found that for early times, the energy transfer from the jet to its surroundings is dominated by heating by shocks before mixing processes become dominant for later times. This matches our results where for early times of the jet we found a clear dominance of the SISM heating by shocks.[@perucho2011] adopted a resolution of 50 pc at 1 kpc from the source, assuming a cylindrical propagation of the jet making it complementary to our own simulations. While their study reveals the long-term, long-distance trends with merely two gridpoints in the jet inlet beam, our study quantifies the effect of finite opening angles for early-time, beam-resolved cases.
we also provided details on two dynamic phenomena due to finite opening angles. The first are the nodes corresponding to recollimation shocks forming along the jet path. These nodes can be seats of significant acceleration of particles, which in turn sets a scenario to explain the X-ray nodes in observations of AGN jets. This was originally studied in [@komissarov1997], with the results of relativistic hydro simulations translated in synthetic images. The opening angle of the jet plays an important role for the formation of these nodes, while the precise position and number of the nodes depends on the value of the opening angle. The second phenomenon we observed was the dynamically formed layered structure of the jet. Starting with a sharp profile for the density that separates the jet and its surrounding, the beam becomes surrounded by a clear density-structured layer of precisely the form needed to explain spine-sheath scenarios that are usually included for jet stability arguments or emission-related inferences. This feature does not need a finite opening angle to appear, but a higher opening angle does change its properties, such as a wider radial width and structured velocity profile.
The next step of this work is to obtain synthetic X-ray maps of these simulations to compare them directly with observations. In this paper we intentionally focused on dynamical effects and energy transfer quantifications only. Moreover, we performed a parametric survey to identify clear trends due to opening angle variations and/or ISM density profiles. In future works, we intend to target specific well-diagnosed radio sources, and extend our study to the more computationally challenging 3D large-scale scenarios.
Thanks to Oliver Porth, Alkis Vlasis and Jan Deca for their support and help. For part of the simulations, we used the infrastructure of the VSC - Flemish Supercomputer Center, funded by the Hercules Foundation and the Flemish Government - department EWI. We acknowledge financial support from FWO-Vlaanderen, grant G.0238.12 and from project GOA/2009/009 (KU Leuven).
[^1]: Code homepage at *homes.esat.kuleuven.be/\~keppens’*
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we study partial-approximate controllability of semilinear nonlocal fractional evolution equations in Hilbert spaces. By using fractional calculus, variational approach and approximating technique, we give the approximate problem of the control system and get the compactness of approximate solution set. Then new sufficient conditions for the partial-approximate controllability of the control system are obtained when the compactness conditions or Lipschitz conditions for the nonlocal function are not required. Finally, we apply our abstract results to the parial-approximate controllability of the semilinear heat equation and delay equation.'
author:
- |
N. I. Mahmudov\
[Eastern Mediterranean University\
Gazimagusa Mersin 10, Turkey\
email: [email protected]]{}
title: 'Partial-Approximate Controllability of Nonlocal Fractional Evolution Equations via Approximating Method'
---
Introduction
============
Controllability concepts are important properties of a control system that plays a substantial role in many engineering problems, such as stabilizing unstable systems using feedback control. Therefore, in recent years, controllability problems for various types of linear and semilinear (fractional) dynamical systems have been studied in many articles (see [@arti]-[@zuazua4] and the references therein). From the mathematical point of view it is necessary to distinguish between problems of exact and approximate controllability. Exact controllability allows to move the system to an arbitrary final state, while approximate controllability means that the system can be moved to an arbitrary small neighborhood of the final state. In particular, approximately controlled systems are more predominant, and very often the approximate controllability is quite adequate in applications. There are some interesting and important approximate controllability results for semilinear evolution systems in abstract spaces with a Caputo fractional derivative. Sakthivel et. al. [@mah6] initiated to study approximate controllability of fractional differential systems in Hilbert spaces. Meanwhile, Sakthivel and Ren [@sakthivel], Debbouche and Torres [@nonloc7], Mahmudov [@mah1], [@mahACM2] pay attention to the study of approximate controllability for various types of (fractional) evolution systems in abstract spaces. On the other hand, existence and approximate controllability of nonlocal fractional equations were studied in [@nonloc5], [@nonloc2], [@nonlocal8], [@mah2]
This paper is devoted to the study of partial-approximate controllability of the semilinear fractional evolution equation with nonlocal conditions,$$\begin{aligned}
^{C}D_{t}^{q}y\left( t\right) & =Ay\left( t\right) +Bu\left( t\right)
+f\left( t,y\left( t\right) \right) ,\ 0<t\leq T,\nonumber\\
y\left( 0\right) & =y_{0}-g\left( y\right) , \label{cp1}$$ where the state variable $y\left( \cdot\right) $ takes values in the Hilbert space $X$, $^{C}D_{t}^{q}$ is the Caputo fractional derivative of order $q$ with $\frac{1}{2}<q\leq1,$ $A:D\left( A\right) \subset X\rightarrow X$ is a family of closed and bounded linear operators generating a strongly continuous semigroup $S:\left[ 0,b\right] \rightarrow\mathfrak{L}\left( X\right) $, where the domain $D\left( A\right) \subset X$ is dense in $X,$ the control function $u\left( \cdot\right) $ is given in $L^{2}\left( \left[
0,b\right] ,U\right) ,$ $U$ is a Hilbert space, $B$ is a bounded linear operator from $U$ into $X,$ $f:\left[ 0,b\right] \times X\rightarrow X,$ $g:C\left( \left[ 0,b\right] ,X\right) \rightarrow X$ are given functions satisfying some assumptions to be specified later and $y_{0}$ is an element of the Hilbert space $X$.
One of the purposes of this article is to investigate the partial-approximate controllability of the system (\[cp1\]) without Lipschitz continuous or compact assumptions on the non-local term $g$. In fact, $g$ is supposed to be continuous and is completely defined by $\left[ \delta,b\right] $ for some small $\delta>0$. Meanwhile, in order to obtain the existence of solutions of the control system (\[cp1\]), we construct an approximate problem of the system (\[cp1\]) and obtain the compactness of the set of approximate solutions. It differs from the usual approach that the fixed-point theorem is applied directly to the corresponding solution operator. Our results, therefore, can be seen as the extension and development of existing results.
It has not yet been reported work on the partial-approximate controllability of fractional semilinear evolution equations with nonlocal conditions. Inspired by the aforementioned recent contributions, we propose to discuss the partial-approximate controllability of fractional evolution systems in Hilbert spaces with classical nonlocal conditions. We first impose a partial-approximate controllability of the associated linear system. Then we develop a variational approach in [@mah2], [@mahACM2] and approximating method in [@nonlocal8], and rewrite our control problem as a sequence of fixed point problems. Next using the Schauder fixed point theorem we get the existence of fixed points and show that these solutions (fixed points) steers the system to an arbitrary small neighborhood of the final state in a closed subspace.
We organize the article as follows. In Section 2, we provide preliminaries, assumptions and formulate the main result on partial-approximate controllability. In Section 3, we extend the variational method to construct approximating control for the approximating controllability problem and use the Schauder fixed point theorem to show existence of a solution for the approximating controllability problem. The partial-approximate controllability result is proved in Section 4. In this section we solve the difficulty concerning the compactness of solution operator $S\left( t\right) $ at $t=0$ by means of approximate solution set constructed in Section 3. The difficulty is due to the fact that a compact semigroup $S\left( t\right) $ is not compact at $t=0$ in an infinite dimensional space. Finally, we provide two examples to illustrate the application of the abstract results.
Statement of results
====================
Throughout this paper, let $\mathbb{N}$, $\mathbb{R}$, $\mathbb{R}_{+}$ be the set of positive integers, real numbers and positive real numbers, respectively. We denote by $X$ a Hilbert space with norm $\left\Vert
\cdot\right\Vert $, $C\left( \left[ 0,b\right] ,X\right) $ the space of all $X$-valued continuous functions on $\left[ 0,b\right] $ with the norm $\left\Vert \cdot\right\Vert _{C}$, $L^{2}(\left[ 0,b\right] ,U)$ the space of all $U$-valued square integrable functions on $\left[ 0,b\right] $, $\mathcal{L}\left( X\right) $ the space of all bounded linear operators from $X$ to $X$ with the usual norm $\left\Vert \cdot\right\Vert _{\mathcal{L}\left( X\right) }$, let $A$ be the infinitesimal generator of $C_{0}$-semigroup $\left\{ S\left( t\right) :t\geq0\right\} $ of uniformly bounded linear operators on $X$. Clearly, $M_{S}:=\sup\left\{ \left\Vert
S\left( t\right) \right\Vert _{\mathcal{L}\left( X\right) }:t\geq
0\right\} <\infty$. Let $E$ be a closed subspace of $X$ and denote by $\Pi$ the projection from $X$ onto $E$.
We recall some notations, definitions and results on fractional derivative and fractional differential equations.
\[def:1\][@pod] The Riemann-Liouville fractional order derivative of $f:\left[ 0,\infty\right) \rightarrow X$ of order $q\in\mathbb{R}_{+}$ is defined by$$^{RL}D^{q}f\left( t\right) =\frac{1}{\Gamma\left( n-q\right) }\frac{d^{n}}{dt^{n}}\int_{0}^{t}\left( t-s\right) ^{n-q-1}f\left( s\right) ds,$$ where $q\in\left( n-1,n\right) $, $n\in\mathbb{N}$.
\[def:2\][@pod] The Caputo fractional order derivative of $f:\left[
0,\infty\right) \rightarrow X$ of order $q\in\mathbb{R}_{+}$ is defined by$$^{RL}D^{q}f\left( t\right) =\frac{1}{\Gamma\left( n-q\right) }\int_{0}^{t}\left( t-s\right) ^{n-q-1}f^{\left( n\right) }\left( s\right) ds,$$ where $q\in\left( n-1,n\right) $, $n\in\mathbb{N}$.
Now, we use the probability density function to give the following definition of mild solutions to (\[cp1\]).
[@zhou1] A solution $y(\cdot;u)\in C(\left[ 0,b\right] ,X)$ is said to be a mild solution of (\[cp1\]) if for any $u\in L^{2}(\left[ 0,b\right]
,U)$ the integral equation $$y(t)=S_{q}\left( t\right) \left( y_{0}-g\left( y\right) \right)
+\int_{0}^{t}\left( t-s\right) ^{q-1}T_{q}\left( t-s\right) \left[
Bu\left( s\right) +f\left( s,y\left( s\right) \right) \right]
ds,\ \ 0\leq t\leq b, \label{cs2}$$ is satisfied. Here$$\begin{aligned}
S_{q}\left( t\right) & =\int_{0}^{\infty}\omega_{\alpha}\left(
\theta\right) S\left( t^{\alpha}\theta\right) d\theta,\ \ \ T_{q}\left(
t\right) =\alpha\int_{0}^{\infty}\theta\omega_{\alpha}\left( \theta\right)
S\left( t^{\alpha}\theta\right) d\theta,\ \ \ t\geq0,\\
\omega_{\alpha}\left( \theta\right) & =\frac{1}{\alpha}\theta
^{-1-1/\alpha}\varpi_{\alpha}\left( \theta^{-1/\alpha}\right) \geq0,\\
\varpi_{\alpha}\left( \theta\right) & =\frac{1}{\pi}\sum_{n=1}^{\infty
}\left( -1\right) ^{n-1}\theta^{-n\alpha-1}\frac{\Gamma\left(
n\alpha+1\right) }{n!}\sin\left( n\pi\alpha\right) ,\ \ \ \theta\in\left(
0,\infty\right) ,\end{aligned}$$ where is a probability density function defined on $\left( 0,\infty\right)
$, that is$$\omega_{\alpha}\left( \theta\right) \geq0,\ \ \ \theta\in\left(
0,\infty\right) ,\ \ \ \int_{0}^{\infty}\omega_{\alpha}\left( \theta\right)
d\theta=1.$$
We present some basic properties of $S_{q}$ and $T_{q}$ which will be used in the sequel [@zhou1].
- For any fixed $t\geq0$ and any $y\in X$, $\left\Vert S_{q}\left(
t\right) y\right\Vert \leq M_{S}\left\Vert y\right\Vert $ and $\left\Vert
T_{q}\left( t\right) y\right\Vert \leq\dfrac{M_{S}}{\Gamma\left( q\right)
}\left\Vert y\right\Vert .$
- $\left\{ S_{q}\left( t\right) :t\geq0\right\} $ and $\left\{
T_{q}\left( t\right) :t\geq0\right\} $ are strongly continuous.
- $\left\{ S_{q}\left( t\right) :t>0\right\} $ and $\left\{
T_{q}\left( t\right) :t>0\right\} $ are compact operators provided that $\left\{ S\left( t\right) :t>0\right\} $ is compact.
Let $y(b;u)$ be the state value of (\[cp1\]) at terminal time $b$ corresponding to the control $u$.
[@bm2] Given $b>0,$ $y_{0}\in X,$ $y_{b}\in E$ and $\varepsilon>0.$ The system (\[cp1\]) is said to be partial-approximately controllable on $\left[ 0,b\right] ,$ if there exists a control $u_{\varepsilon}\in
L^{2}\left( \left[ 0,b\right] ,U\right) $ such that the corresponding solution $y\left( t;u_{\varepsilon}\right) $ of (\[cp1\]), satisfies the conditions$$\left\Vert \Pi y\left( b;u_{\varepsilon}\right) -y_{b}\right\Vert
<\varepsilon. \label{cp3}$$
In particular, if $E=X$, the concept of partial-approximate controllability coincide with the well known concept of approximate controllability.
It is known that the system (\[ls\]) approximately controllable on $\left[
0,b\right] $ if and only if the condition $B^{\ast}T_{q}^{\ast}\left(
b-s\right) \Pi^{\ast}\varphi=0,\ 0<s<b$ implies that $\varphi=0,$ see [@curt].
Our main result is as follows:
\[thm:main\]Assume the following conditions:
1. $S\left( t\right) ,$ $t>0$ is compact operator;
2. The function $f:\left[ 0,b\right] \times X\rightarrow X$ satisfies the following
1. $f\left( \cdot,\cdot\right) :X\rightarrow X$ is jointly continuous;
2. there is a positive continuous function $n\in C\left( \left[
0,b\right] ,\mathbb{R}_{+}\right) $ such that for every $(t,y)\in\left[
0,b\right] \times X$, we have $$\left\Vert f\left( t,y\right) \right\Vert \leq n\left( t\right) .$$
3. The function $g:C\left( \left[ 0,b\right] ,X\right) \rightarrow
X$ satisfies the following
1. $g$ is continuous, there exists a positive constant $\Lambda_{g}$ such that for all $y\in X$,$$\left\Vert g\left( y\right) \right\Vert \leq\Lambda_{g}.$$
2. There is a $\delta\in\left( 0,b\right) $ such that for any $y,z\in
C\left( \left[ 0,b\right] ,X\right) $ satisfying $y\left( t\right)
=z\left( t\right) $, $t\in\left[ \delta,b\right] $, $g\left( y\right)
=g\left( z\right) $.
4. $B:U\rightarrow X$ is a linear continuous operator with $M_{B}:=\left\Vert B\right\Vert $.
5. The linear system$$y(t)=S_{q}\left( t\right) y_{0}+\int_{0}^{t}\left( t-s\right) ^{q-1}T_{q}\left( t-s\right) Bu\left( s\right) ds \label{ls}$$ is partial-approximately controllable in $\left[ 0,b\right] .$ Then the fractional control system (\[cs2\]) is partial-approximately controllable on $\left[ 0,b\right] $.
Our results are new for approximate controllability, that is for the case $E=X$.
Our results are new even for the classical evolution control system with nonlocal conditions (the case $q=1$).
One may expect the results of this paper to hold for a class of problems governed by different type of evolution systems such as Riemann-Liouville FDEs, stochastic (fractional) DEs, FDEs with infinite delay and so on.
Auxiliary lemmas
================
For $\mathbb{\varepsilon}>0$ and $n\geq1$ we introduce the following functional$$J_{\mathbb{\varepsilon},n}\left( \varphi;z\right) =\frac{1}{2}\int_{0}^{b}\left( b-s\right) ^{q-1}\left\Vert B^{\ast}T_{q}^{\ast}\left(
b-s\right) \Pi^{\ast}\varphi\right\Vert ^{2}ds+\mathbb{\varepsilon}\left\Vert
\varphi\right\Vert -\left\langle \varphi,h_{n}\left( z\right) \right\rangle
, \label{f1}$$ where $$h_{n}\left( z\right) =\Pi S_{q}\left( b\right) \left( y_{0}-S\left(
\frac{1}{n}\right) g\left( z\right) \right) +\int_{0}^{b}\Pi T_{q}\left(
b-s\right) f\left( s,z\left( s\right) \right) ds-y_{b}.$$ and the following approximating operators $Q:C\left( \left[ 0,b\right]
,X\right) \rightarrow E$ $$\begin{aligned}
Q\left( z\right) & :=\Pi S_{q}\left( b\right) \left( y_{0}-S\left(
\frac{1}{n}\right) g\left( z\right) \right) +\int_{0}^{b}\left(
b-s\right) ^{q-1}\Pi T_{q}\left( b-s\right) f\left( s,z\left( s\right)
\right) ds\\
& =:Q_{1}\left( z\right) +Q_{2}\left( z\right) .\end{aligned}$$ Set$$B\left( 0;r\right) =\left\{ x\in C\left( \left[ 0,b\right] ,X\right)
:\left\Vert x\right\Vert _{C}\leq r\right\} .$$
\[lem:rc\]The mapping $Q:B\left( 0;r\right) \rightarrow E$ is compact.
For any $\mathbb{\eta}\in\left( 0,b\right) $ and $\delta>0$, we define an operator $Q_{2}^{\eta,\delta}$ on $B\left( 0;r\right) $ by the formula$$\begin{aligned}
Q_{2}^{\eta,\delta}\left( z\right) & =q\int_{0}^{b-\mathbb{\eta}}\int_{\delta}^{\infty}\left( b-s\right) ^{q-1}\theta\omega_{q}\left(
\theta\right) \Pi S\left( \left( b-s\right) ^{q}\theta\right) f\left(
s,z\left( s\right) \right) d\theta ds\\
& =\Pi S\left( \mathbb{\eta}^{q}\delta\right) q\int_{0}^{b-\mathbb{\eta}}\int_{\delta}^{\infty}\left( b-s\right) ^{q-1}\theta\omega_{q}\left(
\theta\right) S\left( \left( b-s\right) ^{q}\theta-S\left( \mathbb{\eta
}^{q}\delta\right) \right) f\left( s,z\left( s\right) \right) d\theta
ds\\
& =:\Pi S\left( \mathbb{\eta}^{q}\delta\right) Z\left( \mathbb{\eta
}\right) .\end{aligned}$$ Bt the assumption (F)$$\left\Vert Z\left( \mathbb{\eta}\right) \right\Vert \leq bM_{S}\left\Vert
n\right\Vert _{C}.$$ Then from the compactness of $S\left( \mathbb{\eta}^{q}\delta\right) \ $and boundedness of $Z\left( \mathbb{\eta}\right) $, we obtain that the set $Q_{2}^{\eta,\delta}:B\left( 0;r\right) \rightarrow E$ is compact. Moreover, for every $z\in B\left( 0;r\right) $, we have$$\begin{aligned}
& \left\Vert Q_{2}\left( z\right) -Q_{2}^{\eta,\delta}\left( z\right)
\right\Vert \\
& \leq q\int_{0}^{b}\int_{0}^{\delta}\left( b-s\right) ^{q-1}\theta
\omega_{q}\left( \theta\right) \left\Vert \Pi S\left( \left( b-s\right)
^{q}\theta\right) \right\Vert \left\Vert f\left( s,z\left( s\right)
\right) \right\Vert d\theta ds\\
& +q\int_{b-\mathbb{\eta}}^{b}\int_{\delta}^{\infty}\left( b-s\right)
^{q-1}\theta\omega_{q}\left( \theta\right) \left\Vert \Pi S\left( \left(
b-s\right) ^{q}\theta\right) \right\Vert \left\Vert f\left( s,z\left(
s\right) \right) \right\Vert d\theta ds\\
& \leq\left\Vert \Pi\right\Vert M_{S}\left( b^{q}\int_{0}^{\delta}\theta\omega_{q}\left( \theta\right) d\theta+\frac{\eta^{q}}{\Gamma\left(
q+1\right) }\right) \left\Vert n\right\Vert _{C}\rightarrow0,\end{aligned}$$ as $\eta,\delta\rightarrow0.$Hence, there exist relatively compact operator that can be arbitrarily close to $Q_{2}$. Then, $Q_{2}$ is compact.
Since $S\left( \frac{1}{n}\right) $, $n\geq1$, is compact, $\Pi$ is linear bounded and the assumption (G) holds, it is easily seen that, the set $$\left\{ \Pi S_{q}\left( b\right) \left( y_{0}-S\left( \frac{1}{n}\right)
g\left( z\right) \right) :z\in B\left( 0;r\right) \right\}$$ is relatively compact in $X$. Thus the operator $Q=Q_{1}+Q_{2}$ is compact.
\[lem:4\]For any $n\geq1$, $h_{n}:B\left( 0;r\right) \rightarrow E$ is a continuous function.
For any $z_{l},z\in B\left( 0;r\right) $ with $\lim_{l\rightarrow\infty
}\left\Vert z_{l}-z\right\Vert _{C}=0.$ By the condition (F1), we have we can conclude that $\lim_{l\rightarrow\infty}\left\Vert f\left( \cdot,z_{l}\left(
\cdot\right) \right) -f\left( \cdot,z\left( \cdot\right) \right)
\right\Vert _{C}=0.$ On the other hand$$\begin{aligned}
\left\Vert h_{n}\left( z_{l}\right) -h_{n}\left( z\right) \right\Vert
_{E} & \leq\left\Vert \Pi S_{q}\left( b\right) S\left( \frac{1}{n}\right) \left( g\left( z_{l}\right) -g\left( z\right) \right)
\right\Vert \\
& +\left\Vert \int_{0}^{b}\left( b-s\right) ^{q-1}\Pi T_{q}\left(
b-s\right) \left[ f\left( s,z_{l}\left( s\right) \right) -f\left(
s,z\left( s\right) \right) \right] d\theta ds\right\Vert \\
& \leq M_{S}^{2}\left\Vert \Pi\right\Vert \left\Vert g\left( z_{l}\right)
-g\left( z\right) \right\Vert \\
& +\frac{M_{S}b^{q}}{\Gamma\left( q\right) }\left\Vert \Pi\right\Vert
\left\Vert f\left( \cdot,z_{l}\left( \cdot\right) \right) -f\left(
\cdot,z\left( \cdot\right) \right) \right\Vert _{C},\end{aligned}$$ which implies that$$\left\Vert h_{n}\left( z_{l}\right) -h_{n}\left( z\right) \right\Vert
_{E}\rightarrow0\ \ \ \text{as\ \ }l\rightarrow\infty.$$
From the definition of $\varphi\rightarrow J_{\mathbb{\varepsilon},n}\left(
\varphi;z\right) $, we see immediately that for any fixed $z\in B\left(
0;r\right) $ it is continuous and strictly convex.
\[lem:1\]For any $B\left( 0;r\right) $,$$\underline{\lim}_{\left\Vert \varphi\right\Vert \rightarrow\infty}\inf_{z\in
B\left( 0;r\right) }\frac{J_{\mathbb{\varepsilon},n}\left( \varphi
;z\right) }{\left\Vert \varphi\right\Vert }\geq\mathbb{\varepsilon}.
\label{f7}$$
In order to prove (\[f7\]), suppose that it is not the case. Then there exists sequences $\left\{ \varphi_{l}\right\} \subset X,$ $\left\{
z_{l}\right\} \subset B\left( 0;r\right) $, with $\left\Vert \varphi
_{l}\right\Vert \rightarrow\infty$, such that$$\underline{\lim}_{l\rightarrow\infty}\frac{J_{\mathbb{\varepsilon},n}\left(
\varphi_{l};z_{l}\right) }{\left\Vert \varphi_{l}\right\Vert }<\mathbb{\varepsilon}. \label{f2}$$ Without loss of generality, we may assume that$$h_{n}\left( z_{l}\right) \rightarrow h_{n},\ \ \text{strongly}\ \text{in\ \ }X\text{,} \label{f3}$$ for some $h_{n}\in X.$ In fact, $\left\{ h_{n}\left( z_{l}\right)
:l\geq1\right\} \subset\operatorname{Im}Q$ is relatively compact in $X$; thus, we may assume this by picking a subsequence.
Next, we normalize $\varphi_{l}:$ $\widetilde{\varphi}_{l}=\dfrac{\varphi_{l}}{\left\Vert \varphi_{l}\right\Vert }$. Since $\left\Vert \widetilde{\varphi
}_{l}\right\Vert =1,$ we can extract a subsequence (still denoted by $\widetilde{\varphi}_{l}$), which weakly converges in $X$ to an element $\widetilde{\varphi}$ in $X$. Consequently, because $S\left( t\right) ,$ $t>0$, is a compact semigroup, we see that $$B^{\ast}S^{\ast}\left( b-\cdot\right) \Pi^{\ast}\widetilde{\varphi}_{l}\rightarrow B^{\ast}S^{\ast}\left( b-\cdot\right) \Pi^{\ast}\widetilde{\varphi},\ \ \ \text{strongly\ \ in}\ C\left( \left[ 0,b\right]
,X\right) ,\ \text{as\ }l\rightarrow\infty. \label{f11}$$ From (\[f1\]), it follows that$$\frac{J_{\mathbb{\varepsilon},n}\left( \varphi_{l};z_{l}\right) }{\left\Vert
\varphi_{l}\right\Vert }=\frac{\left\Vert \varphi_{l}\right\Vert }{2}\int
_{0}^{b}\left( b-s\right) ^{q-1}\left\Vert B^{\ast}T_{q}^{\ast}\left(
b-s\right) \Pi^{\ast}\varphi\right\Vert ^{2}ds+\mathbb{\varepsilon}\left\Vert
\widetilde{\varphi}_{l}\right\Vert -\left\langle \widetilde{\varphi}_{l},h_{n}\left( z_{l}\right) \right\rangle .$$ Thus, noting that $\left\Vert \varphi_{l}\right\Vert \rightarrow\infty,$ by (\[f2\])-(\[f11\]) and the Fatou lemma$$\int_{0}^{b}\left( b-s\right) ^{q-1}\left\Vert B^{\ast}T_{q}^{\ast}\left(
b-s\right) \Pi^{\ast}\widetilde{\varphi}\right\Vert ^{2}ds\leq\underline
{\lim}_{l\rightarrow\infty}\int_{0}^{b}\left( b-s\right) ^{q-1}\left\Vert
B^{\ast}T_{q}^{\ast}\left( b-s\right) \Pi^{\ast}\widetilde{\varphi}_{l}\right\Vert ^{2}ds=0.$$ By assumption (AC) we have $\widetilde{\varphi}=0,$ and we deduce that $$\widetilde{\varphi}_{l}\rightharpoonup0\ \ \text{weakly}\ \text{in
}X\ \text{as }l\rightarrow\infty.$$ Hence$$\underline{\lim}_{l\rightarrow\infty}\frac{J_{\mathbb{\varepsilon},n}\left(
\varphi_{l};z_{l}\right) }{\left\Vert \varphi_{l}\right\Vert }\geq
\underline{\lim}_{l\rightarrow\infty}\left( \mathbb{\varepsilon}\left\Vert
\widetilde{\varphi}_{l}\right\Vert -\left\langle \widetilde{\varphi}_{l},h_{n}\left( z_{l}\right) \right\rangle \right) =\mathbb{\varepsilon},$$ which contradicts (\[f2\]) and proves the claim (\[f7\]).
For any $z\in C\left( \left[ 0,b\right] ,X\right) $, the functional $J_{\mathbb{\varepsilon},n}\left( \cdot,z\right) $ admits a unique minimum $\widehat{\varphi}_{\varepsilon,n}$ that defines a map $\Phi_{\varepsilon
,n}:C\left( \left[ 0,b\right] ,X\right) \rightarrow X$. $\Phi
_{\varepsilon,n}$ has the following properties.
\[lem:2\]There exists $R_{\varepsilon}>0$ such that $\left\Vert
\Phi_{\varepsilon,n}\left( z\right) \right\Vert <R_{\varepsilon}$ for any $z\in B\left( 0;r\right) ,$ $n\geq1.$
Let $z\in B\left( 0;r\right) .$ From Lemma \[lem:1\](b), we see that there exists a constant $R_{\varepsilon}>0$, such that$$\inf_{z\in B\left( 0;r\right) }\frac{J_{\mathbb{\varepsilon},n}\left(
\varphi;z\right) }{\left\Vert \varphi\right\Vert }\geq\frac
{\mathbb{\varepsilon}}{2},\ \ \ \left\Vert \varphi\right\Vert \geq
R_{\varepsilon}. \label{f4}$$ On the other hand, by the definition of $\Phi_{\varepsilon}$,$$J_{\mathbb{\varepsilon},n}\left( \Phi_{\varepsilon,n}\left( z\right)
;z\right) \leq J_{\mathbb{\varepsilon},n}\left( 0;z\right) =0. \label{f5}$$ Hence, combining (\[f4\]) and (\[f5\]), we have$$\left\Vert \Phi_{\varepsilon,n}\left( z\right) \right\Vert <R_{\varepsilon
},\ \ \text{for\ all}\ z\in B\left( 0;r\right) .$$
\[lem:3\]For any $z_{l},$ $z\in B\left( 0;r\right) $ satisfying$$z_{l}\longrightarrow z,\ \ \ \text{in\ \ }C\left( \left[ 0,b\right]
,X\right) \text{,}$$ it holds that$$\lim_{l\rightarrow\infty}\left\Vert \Phi_{\varepsilon,n}\left( z_{l}\right)
-\Phi_{\varepsilon,n}\left( z\right) \right\Vert =0.$$
By Lemma \[lem:2\], we have boundedness of $\widehat{\varphi}_{\varepsilon
,n,l}=\Phi_{\varepsilon,n}\left( z_{l}\right) $. Consequently, we may assume that $\widehat{\varphi}_{\varepsilon,n,l}\overset{w}{\longrightarrow
}\widetilde{\varphi}_{\varepsilon,n}$. Thus, by the definition of $J_{\mathbb{\varepsilon},n}$ and the optimality of both $\widehat{\varphi
}_{\varepsilon,n,l}=\Phi_{\varepsilon,n}\left( z_{l}\right) \ $and $\widehat{\varphi}_{\varepsilon,n}=\Phi_{\varepsilon,n}\left( z\right) $, one has$$\begin{aligned}
J_{\mathbb{\varepsilon},n}\left( \widehat{\varphi}_{\varepsilon,n};z\right)
& \leq J_{\mathbb{\varepsilon},n}\left( \widetilde{\varphi}_{\varepsilon
,n};z\right) \leq\underline{\lim}_{l\rightarrow\infty}J_{\mathbb{\varepsilon
},n}\left( \widehat{\varphi}_{\varepsilon,n,l};z_{l}\right) \\
& \leq\overline{\lim_{l\rightarrow\infty}}J_{\mathbb{\varepsilon},n}\left(
\widehat{\varphi}_{\varepsilon,n,l};z_{l}\right) \leq\lim_{l\rightarrow
\infty}J_{\mathbb{\varepsilon},n}\left( \widehat{\varphi}_{\varepsilon
,n};z_{l}\right) =J_{\mathbb{\varepsilon},n}\left( \widehat{\varphi
}_{\varepsilon,n};z\right) .\end{aligned}$$ Hence, the equalities hold in the above. That means that $\widetilde{\varphi
}_{\varepsilon,n}$ is also a minimum of $J_{\mathbb{\varepsilon},n}\left(
\cdot;z\right) $. By the uniqueness of the minimum, it is necessary that $\widetilde{\varphi}_{\varepsilon,n}=\widehat{\varphi}_{\varepsilon,n}$. Therefore$$\begin{aligned}
\lim_{l\rightarrow\infty}J_{\mathbb{\varepsilon},n}\left( \widehat{\varphi
}_{\varepsilon,n};z_{l}\right) & =J_{\mathbb{\varepsilon},n}\left(
\widehat{\varphi}_{\varepsilon,n};z\right) ,\\
\lim_{l\rightarrow\infty}\int_{0}^{b}\left( b-s\right) ^{q-1}\left\Vert
B^{\ast}T_{q}^{\ast}\left( b-s\right) \Pi^{\ast}\widehat{\varphi
}_{\varepsilon,n,l}\right\Vert ^{2}ds & =\int_{0}^{b}\left( b-s\right)
^{q-1}\left\Vert B^{\ast}T_{q}^{\ast}\left( b-s\right) \Pi^{\ast}\widehat{\varphi}_{\varepsilon,n}\right\Vert ^{2}ds,\\
\lim_{l\rightarrow\infty}\left\langle \widehat{\varphi}_{\varepsilon
,n,l},h_{n}\left( z_{l}\right) \right\rangle & =\left\langle
\widehat{\varphi}_{\varepsilon,n},h_{n}\left( z\right) \right\rangle
,\ \ \ \left\Vert \widehat{\varphi}_{\varepsilon,n}\right\Vert \leq
\lim_{l\rightarrow\infty}\left\Vert \widehat{\varphi}_{\varepsilon
,n,l}\right\Vert .\end{aligned}$$ These relations imply that$$\lim_{l\rightarrow\infty}\left\Vert \widehat{\varphi}_{\varepsilon
,n,l}\right\Vert =\left\Vert \widehat{\varphi}_{\varepsilon,n}\right\Vert .
\label{f6}$$ Because $X$ is Hilbert space, from $\widehat{\varphi}_{\varepsilon
,n,l}\overset{w}{\longrightarrow}\widehat{\varphi}_{\varepsilon,n}$ and (\[f6\]), we obtain the strong convergence of $\widehat{\varphi
}_{\varepsilon,n,l}$ to $\widehat{\varphi}_{\varepsilon,n}$.
For fixed $n\geq1$, set $\Theta_{\varepsilon,n}:C\left( \left[ 0,b\right]
,X\right) \rightarrow C\left( \left[ 0,b\right] ,X\right) $ defined by$$\left( \Theta_{\varepsilon,n}z\right) \left( t\right) =S_{q}\left(
t\right) \left( y_{0}-S\left( \frac{1}{n}\right) g\left( z\right)
\right) +\int_{0}^{t}\left( t-s\right) ^{q-1}T_{q}\left( t-s\right)
\left[ Bu_{\varepsilon,n}\left( s,z\right) +f\left( s,z\left( s\right)
\right) \right] ds, \label{fx1}$$ with $$u_{\varepsilon,n}\left( s,z\right) =B^{\ast}T_{q}^{\ast}\left( b-s\right)
\Pi^{\ast}\widehat{\varphi}_{\varepsilon,n}=B^{\ast}T_{q}^{\ast}\left(
b-s\right) \Pi^{\ast}\Phi_{\varepsilon,n}\left( z\right) . \label{c1}$$ We will prove $\Theta_{\varepsilon,n}$ has a fixed point by using Schauder’s fixed point theorem.
\[thm:app\]Assume that the hypotheses of Theorem \[thm:main\] are satisfied. Then for $n\geq1$, the approximate control operator $\Theta
_{\varepsilon,n}$ has at least one fixed point in $C\left( \left[
0,b\right] ,X\right) $.
Step 1: For any $n\geq1$, $\Theta_{\varepsilon,n}$ is continuous on $C\left(
\left[ 0,b\right] ,X\right) $.
Let $\left\{ z_{m}:m\geq1\right\} $ be a sequence in $C\left( \left[
0,b\right] ,X\right) $ with lim$_{m\rightarrow\infty}z_{m}=z$ in $C\left(
\left[ 0,b\right] ,X\right) $. By the continuity of $f$ and $u_{\varepsilon
,n}$, we deduce that $\left( f\left( s,z_{m}\left( s\right) \right)
,u_{\varepsilon,n}\left( s,z_{m}\right) \right) $ converges to $\left(
f\left( s,z\left( s\right) \right) ,u_{\varepsilon,n}\left( s,z\right)
\right) $ uniformly for $s\in\left[ 0,b\right] $, and we have$$\begin{gathered}
\left\Vert \left( \Theta_{\varepsilon,n}z_{m}\right) \left( t\right)
-\left( \Theta_{\varepsilon,n}z\right) \left( t\right) \right\Vert \\
\leq M_{S}^{2}\left\Vert g\left( z_{m}\right) -g\left( z\right)
\right\Vert +\frac{M_{S}b^{q}}{\Gamma\left( q+1\right) }\left\Vert f\left(
\cdot,z_{m}\left( \cdot\right) \right) -f\left( \cdot,z\left(
\cdot\right) \right) \right\Vert _{C}\\
+\frac{M_{S}b^{q}}{\Gamma\left( q+1\right) }\left\Vert u_{\varepsilon
,n}\left( \cdot,z_{m}\right) -u_{\varepsilon,n}\left( \cdot,z\right)
\right\Vert _{C}\rightarrow0,\ \ \ \text{as\ \ \ }n\rightarrow\infty,\end{gathered}$$ which implies that the mapping $\Theta_{\varepsilon,n}$ is continuous on $C\left( \left[ 0,b\right] ,X\right) $.
Step 2: There is a positive number $r\left( \varepsilon\right) >0$ such that $\Theta_{\varepsilon,n}$ maps $B\left( 0;r\left( \varepsilon\right)
\right) $ into itself.
We see that$$\begin{aligned}
& \left\Vert \left( \Theta_{\varepsilon,n}z_{k}\right) \left(
t_{k}\right) \right\Vert \\
& \leq M_{S}^{2}\Lambda_{g}+\frac{M_{S}b^{q}}{\Gamma\left( q+1\right)
}\left( \left\Vert n\right\Vert _{C}+M_{S}M_{B}R_{\varepsilon}\right)
:=r\left( \varepsilon\right) .\end{aligned}$$
Step 3: For any $n\geq1$, $\Theta_{\varepsilon,n}$ is compact.
At the end, applying the Schauder fixed point theorem we obtain that for each $n\geq1$, $\Theta_{\varepsilon,n}$ has at least one fixed point in $B\left(
0;r\left( \varepsilon\right) \right) $.
Assume that $z_{\varepsilon,n}\in B\left( 0;r\left( \varepsilon\right)
\right) \subset C\left( \left[ 0,b\right] ,X\right) $ is a fixed point of (\[fx1\])$$\Theta_{\varepsilon,n}z_{\varepsilon,n}=z_{\varepsilon,n}$$ and $\Phi_{\varepsilon,n}\left( z_{\varepsilon,n}\right) $ is minimizer of $J_{\mathbb{\varepsilon},n}\left( \varphi;z_{\varepsilon,n}\right) $ and $$u_{\varepsilon,n}\left( s,z_{\varepsilon,n}\right) =B^{\ast}T_{q}^{\ast
}\left( b-s\right) \Pi^{\ast}\Phi_{\varepsilon,n}\left( z_{\varepsilon
,n}\right) ,$$ is the corresponding control. Moreover, assume that $$z_{\varepsilon,n}\overset{n\rightarrow\infty}{\rightarrow}z_{\varepsilon
}\ \text{strongly}\ \text{in }C\left( \left[ 0,b\right] ,X\right)
,\ \text{as\ }n\rightarrow\infty.$$ $\Phi_{\varepsilon}\left( z_{\varepsilon}\right) $ is minimizer of $J_{\mathbb{\varepsilon}}\left( \varphi;z_{\varepsilon}\right) $ and $$u_{\varepsilon}\left( s,z_{\varepsilon}\right) =B^{\ast}T_{q}^{\ast}\left(
b-s\right) \Pi^{\ast}\Phi_{\varepsilon}\left( z_{\varepsilon}\right) ,$$ is the corresponding control.
\[lem:5\]Assume that$$\lim_{n\rightarrow\infty}\left\Vert z_{\varepsilon,n}-z_{\varepsilon
}\right\Vert _{C}=0.$$ Then $$\begin{aligned}
\Phi_{\varepsilon,n}\left( z_{\varepsilon,n}\right) & \rightharpoonup
\Phi_{\varepsilon}\left( z_{\varepsilon}\right) \ \ \text{weakly in }X,\\
\lim_{n\rightarrow\infty}\left\Vert u_{\varepsilon,n}\left( s,z_{\varepsilon
,n}\right) -u_{\varepsilon}\left( s,z_{\varepsilon}\right) \right\Vert
_{C} & =0.\end{aligned}$$
By definition of the minimizing functional $\Phi_{\varepsilon,n}\left(
z_{\varepsilon,n}\right) $ and $\Phi_{\varepsilon}\left( z_{\varepsilon
}\right) $ are minimizers of $$\begin{aligned}
J_{\mathbb{\varepsilon},n}\left( \varphi;z_{\varepsilon,n}\right) &
=\frac{1}{2}\int_{0}^{b}\left( b-s\right) ^{q-1}\left\Vert B^{\ast}T_{q}^{\ast}\left( b-s\right) \Pi^{\ast}\varphi\right\Vert ^{2}ds+\mathbb{\varepsilon}\left\Vert \varphi\right\Vert -\left\langle
\varphi,h_{n}\left( z_{\varepsilon,n}\right) \right\rangle ,\\
J_{\mathbb{\varepsilon}}\left( \varphi;z_{\varepsilon}\right) & =\frac
{1}{2}\int_{0}^{b}\left( b-s\right) ^{q-1}\left\Vert B^{\ast}T_{q}^{\ast
}\left( b-s\right) \Pi^{\ast}\varphi\right\Vert ^{2}ds+\mathbb{\varepsilon
}\left\Vert \varphi\right\Vert -\left\langle \varphi,h\left( z_{\varepsilon
}\right) \right\rangle ,\\end{aligned}$$ correspondingly. By Lemma \[lem:2\], we have boundedness of $\Phi
_{\varepsilon,n}\left( z_{\varepsilon,n}\right) $. Consequently, we may assume that $\Phi_{\varepsilon,n}\left( z_{\varepsilon,n}\right)
\rightharpoonup\widetilde{\Phi}_{\varepsilon}$ weakly in $X$. Thus, by the definition of $J_{\mathbb{\varepsilon},n}$ and the optimality of both $\Phi_{\varepsilon,n}\left( z_{\varepsilon,n}\right) $ and $\Phi
_{\varepsilon}\left( z_{\varepsilon}\right) $, one has$$J_{\mathbb{\varepsilon}}\left( \Phi_{\varepsilon}\left( z_{\varepsilon
}\right) ;z_{\varepsilon}\right) \leq J_{\mathbb{\varepsilon}}\left(
\widetilde{\Phi}_{\varepsilon};z_{\varepsilon}\right) \leq\underline{\lim
}_{n\rightarrow\infty}J_{\mathbb{\varepsilon}}\left( \Phi_{\varepsilon
,n}\left( z_{\varepsilon,n}\right) ;z_{\varepsilon}\right) , \label{qq1}$$ and$$\underline{\lim}_{n\rightarrow\infty}J_{\mathbb{\varepsilon},n}\left(
\Phi_{\varepsilon,n}\left( z_{\varepsilon,n}\right) ;z_{\varepsilon
,n}\right) \leq\overline{\lim_{n\rightarrow\infty}}J_{\mathbb{\varepsilon},n}\left( \Phi_{\varepsilon,n}\left( z_{\varepsilon,n}\right)
;z_{\varepsilon,n}\right) \leq\lim_{n\rightarrow\infty}J_{\mathbb{\varepsilon
},n}\left( \Phi_{\varepsilon}\left( z_{\varepsilon}\right) ;z_{\varepsilon
,n}\right) =J_{\mathbb{\varepsilon}}\left( \Phi_{\varepsilon}\left(
z_{\varepsilon}\right) ;z_{\varepsilon}\right) . \label{qq2}$$ Noting that $\underline{\lim}_{n\rightarrow\infty}J_{\mathbb{\varepsilon}}\left( \Phi_{\varepsilon,n}\left( z_{\varepsilon,n}\right) ;z_{\varepsilon
}\right) =\underline{\lim}_{n\rightarrow\infty}J_{\mathbb{\varepsilon},n}\left( \Phi_{\varepsilon,n}\left( z_{\varepsilon,n}\right)
;z_{\varepsilon,n}\right) $ and combining (\[qq1\]) and (\[qq2\]) we have$$\begin{aligned}
J_{\mathbb{\varepsilon}}\left( \Phi_{\varepsilon}\left( z_{\varepsilon
}\right) ;z_{\varepsilon}\right) & =J_{\mathbb{\varepsilon}}\left(
\widetilde{\Phi}_{\varepsilon};z_{\varepsilon}\right) ,\\
\underline{\lim}_{n\rightarrow\infty}J_{\mathbb{\varepsilon},n}\left(
\Phi_{\varepsilon,n}\left( z_{\varepsilon,n}\right) ;z_{\varepsilon
,n}\right) & =\overline{\lim_{n\rightarrow\infty}}J_{\mathbb{\varepsilon
},n}\left( \Phi_{\varepsilon,n}\left( z_{\varepsilon,n}\right)
;z_{\varepsilon,n}\right) =J_{\mathbb{\varepsilon}}\left( \Phi_{\varepsilon
}\left( z_{\varepsilon}\right) ;z_{\varepsilon}\right) .\end{aligned}$$ That means that $\widetilde{\Phi}_{\varepsilon}$ is also a minimum of $J_{\mathbb{\varepsilon}}\left( \cdot;z_{\varepsilon}\right) $. By the uniqueness of the minimum, it is necessary that $\widetilde{\Phi}_{\varepsilon}=\Phi_{\varepsilon}\left( z_{\varepsilon}\right) $. Therefore $$\begin{aligned}
\lim_{n\rightarrow\infty}J_{\mathbb{\varepsilon},n}\left( \Phi_{\varepsilon
,n}\left( z_{\varepsilon,n}\right) ;z_{\varepsilon,n}\right) &
=J_{\mathbb{\varepsilon}}\left( \Phi_{\varepsilon}\left( z_{\varepsilon
}\right) ;z_{\varepsilon}\right) ,\\
\lim_{n\rightarrow\infty}\int_{0}^{b}\left( b-s\right) ^{q-1}\left\Vert
B^{\ast}T_{q}^{\ast}\left( b-s\right) \Pi^{\ast}\Phi_{\varepsilon,n}\left(
z_{\varepsilon,n}\right) \right\Vert ^{2}ds & =\int_{0}^{b}\left(
b-s\right) ^{q-1}\left\Vert B^{\ast}T_{q}^{\ast}\left( b-s\right) \Pi
^{\ast}\Phi_{\varepsilon}\left( z_{\varepsilon}\right) \right\Vert ^{2}ds,\\
\lim_{n\rightarrow\infty}\left\langle \Phi_{\varepsilon,n}\left(
z_{\varepsilon,n}\right) ,h_{n}\left( z_{\varepsilon,n}\right)
\right\rangle & =\left\langle \Phi_{\varepsilon}\left( z_{\varepsilon
}\right) ,h\left( z_{\varepsilon}\right) \right\rangle ,\ \ \ \left\Vert
\Phi_{\varepsilon}\left( z_{\varepsilon}\right) \right\Vert \leq
\lim_{n\rightarrow\infty}\left\Vert \Phi_{\varepsilon,n}\left( z_{\varepsilon
,n}\right) \right\Vert .\end{aligned}$$ These relations imply that$$\lim_{n\rightarrow\infty}\left\Vert \Phi_{\varepsilon,n}\left( z_{\varepsilon
,n}\right) \right\Vert =\left\Vert \Phi_{\varepsilon}\left( z_{\varepsilon
}\right) \right\Vert . \label{ff6}$$ As $X$ is Hilbert space, from $\Phi_{\varepsilon,n}\left( z_{\varepsilon
,n}\right) \rightharpoonup\Phi_{\varepsilon}\left( z_{\varepsilon}\right) $ weakly in $X$ and (\[ff6\]), we obtain the strong convergence of $\Phi_{\varepsilon,n}\left( z_{\varepsilon,n}\right) $ to $\Phi
_{\varepsilon}\left( z_{\varepsilon}\right) $.
Proof of the main result
========================
Let $y_{0}\in X,\ y_{b}\in E$ be any given two points and $\mathbb{\varepsilon
}>0$ be any given accuracy. Then it is seen that for all $t\in\left[
0,b\right] $$$\begin{aligned}
\left( \Theta_{\varepsilon}z\right) \left( t\right) & =S_{q}\left(
t\right) \left( y_{0}-g\left( z\right) \right) +\int_{0}^{t}\left(
t-s\right) ^{q-1}T_{q}\left( t-s\right) \left[ Bu_{\varepsilon}\left(
s,z\right) +f\left( s,z\left( s\right) \right) \right] ds,\\
u\left( s,z\right) & =B^{\ast}T_{q}^{\ast}\left( b-s\right) \Pi^{\ast
}\widehat{\varphi}_{\varepsilon}=B^{\ast}T_{q}^{\ast}\left( b-s\right)
\Pi^{\ast}\Phi_{\varepsilon}\left( z\right) .\end{aligned}$$
\[thm:2\]Assume that the hypotheses of Theorem \[thm:main\] are satisfied. Then the control operator $\Theta_{\varepsilon}$ has at least one fixed point in $C\left( \left[ 0,b\right] ,X\right) $.
Let $n\geq1$ be fixed. By Theorem \[fx1\] the approximate operator $\Theta_{\varepsilon,n}$ defined by (\[fx1\]) has a fixed point, say $y_{\varepsilon,n}$. We define the approximate solution set $D$ by$$D=\left\{ y_{\varepsilon,n}\in C\left( \left[ 0,b\right] ,X\right)
:\Theta_{\varepsilon,n}y_{\varepsilon,n}=y_{\varepsilon,n},\ n\geq1\right\}
.$$
Step 1: $D\left( 0\right) $ is relatively compact in $X$.
For $y_{\varepsilon,n}\in D$, $n\geq1$, define$$\widetilde{y}_{\varepsilon,n}\left( t\right) =\left\{
\begin{tabular}
[c]{lll}$y_{\varepsilon,n}\left( t\right) ,$ & & $\delta\leq t\leq b,$\\
$y_{\varepsilon,n}\left( \delta\right) $ & & $0\leq t\leq\delta,$\end{tabular}
\ \right.$$ where $\delta$ comes from the condition (G). It is easily seen that $\left\{
\widetilde{y}_{\varepsilon,n}:n\geq1\right\} $ is relatively compact in $C\left( \left[ 0,b\right] ,X\right) $. Without loss of generality, we may suppose that $\widetilde{y}_{\varepsilon,n}\rightarrow\widetilde
{y}_{\varepsilon}\in C\left( \left[ 0,b\right] ,X\right) $ as $n\rightarrow\infty$. By assumption (G), we get that $g\left( y_{\varepsilon
,n}\right) =g\left( \widetilde{y}_{\varepsilon,n}\right) \rightarrow
g\left( \widetilde{y}_{\varepsilon}\right) $. Thus by the continuity of $S_{q}\left( t\right) $ and $g$, we have$$\begin{gathered}
\left\Vert y_{\varepsilon,n}\left( 0\right) -\left( y_{0}-g\left(
\widetilde{y}_{\varepsilon}\right) \right) \right\Vert =\left\Vert S\left(
\frac{1}{n}\right) g\left( y_{\varepsilon,n}\right) -g\left( \widetilde
{y}_{\varepsilon}\right) \right\Vert \\
\leq\left\Vert S\left( \frac{1}{n}\right) g\left( y_{\varepsilon,n}\right)
-S\left( \frac{1}{n}\right) g\left( \widetilde{y}_{\varepsilon}\right)
\right\Vert +\left\Vert S\left( \frac{1}{n}\right) g\left( \widetilde
{y}_{\varepsilon}\right) -g\left( \widetilde{y}_{\varepsilon}\right)
\right\Vert \\
\leq M_{S}\left\Vert g\left( y_{\varepsilon,n}\right) -g\left(
\widetilde{y}_{\varepsilon}\right) \right\Vert +\left\Vert S\left( \frac
{1}{n}\right) g\left( \widetilde{y}_{\varepsilon}\right) -g\left(
\widetilde{y}_{\varepsilon}\right) \right\Vert \rightarrow0,\end{gathered}$$ as $n\rightarrow\infty$. So, $D\left( 0\right) =\left\{ y_{\varepsilon
,n}\left( 0\right) =y_{0}-S\left( \frac{1}{n}\right) g\left(
y_{\varepsilon,n}\right) \right\} $ is relatively compact in $X$.
Step 2: For each $t\in\left( 0,b\right] $ the set $D\left( t\right)
:=\left\{ y_{\varepsilon,n}\left( t\right) :n\geq1\right\} $ is relatively compact in $X$.
Step 3: $D$ is equicontinuous at $t=0.$
First of all note that$$\begin{aligned}
& \sup\left\{ \left\Vert Bu_{\varepsilon,n}\left( s,y_{\varepsilon
,n}\right) +f\left( s,y_{\varepsilon,n}\left( s\right) \right)
\right\Vert :s\in\left[ 0,b\right] ,y_{\varepsilon,n}\in D\right\} \\
& \leq M_{B}^{2}M_{S}R_{\varepsilon}+\left\Vert n\right\Vert _{C}.\end{aligned}$$ For $0<t<b$, we have$$\begin{gathered}
\left\Vert y_{\varepsilon,n}\left( t\right) -y_{\varepsilon,n}\left(
0\right) \right\Vert =\left\Vert S_{q}\left( t\right) S\left( \frac{1}{n}\right) g\left( y_{\varepsilon,n}\right) -S\left( \frac{1}{n}\right)
g\left( y_{\varepsilon,n}\right) \right\Vert \\
+\frac{M_{S}}{\Gamma\left( q\right) }\int_{0}^{t}\left( t-s\right)
^{q-1}ds\sup\left\{ \left\Vert Bu_{\varepsilon,n}\left( s,y_{\varepsilon
,n}\right) +f\left( s,y_{\varepsilon,n}\left( s\right) \right)
\right\Vert :s\in\left[ 0,b\right] ,y_{\varepsilon,n}\in D\right\} \\
\leq\left\Vert \left( S_{q}\left( t\right) -I\right) S\left( \frac{1}{n}\right) g\left( y_{\varepsilon,n}\right) \right\Vert \\
+\frac{M_{S}t^{q}}{\Gamma\left( q+1\right) }\left( M_{B}^{2}M_{S}R_{\varepsilon}+\left\Vert n\right\Vert _{C}\right)\end{gathered}$$ approaches zero uniformly as $t\rightarrow0$, since $\left\{ S\left(
\frac{1}{n}\right) g\left( y_{\varepsilon,n}\right) :n\geq1\right\} $ is relatively compact in $X$. Then we obtain the set $D$ is equicontinuous at $t=0$.
Step 4: $D$ is equicontinuous on $\left( 0,b\right] $.
Let $0<t_{1}<t_{2}\leq b$ and choose $\eta>0$ such that $t_{1}-\eta>0$. Then, for any $y_{\varepsilon,n}\in D$, we have$$\begin{aligned}
& \left\Vert y_{\varepsilon,n}\left( t_{2}\right) -y_{\varepsilon,n}\left(
t_{1}\right) \right\Vert \\
& \leq\left\Vert \left( S_{q}\left( t_{2}\right) -S_{q}\left(
t_{1}\right) \right) \left( y_{0}-S\left( \frac{1}{n}\right) g\left(
y_{\varepsilon,n}\right) \right) \right\Vert \\
& +\left\Vert \int_{t_{1}}^{t_{2}}\left( t_{2}-s\right) ^{q-1}T_{q}\left(
t_{2}-s\right) \left( Bu_{\varepsilon,n}\left( s,y_{\varepsilon,n}\right)
+f\left( s,y_{\varepsilon,n}\left( s\right) \right) \right) ds\right\Vert
\\
& +\left\Vert \int_{0}^{t_{1}}\left( \left( t_{2}-s\right) ^{q-1}-\left(
t_{1}-s\right) ^{q-1}\right) T_{q}\left( t_{2}-s\right) \left(
Bu_{\varepsilon,n}\left( s,y_{\varepsilon,n}\right) +f\left(
s,y_{\varepsilon,n}\left( s\right) \right) \right) ds\right\Vert \\
& +\left\Vert \int_{0}^{t_{1}}\left( t_{1}-s\right) ^{q-1}\left(
T_{q}\left( t_{2}-s\right) -T_{q}\left( t_{1}-s\right) \right) \left(
Bu_{\varepsilon,n}\left( s,y_{\varepsilon,n}\right) +f\left(
s,y_{\varepsilon,n}\left( s\right) \right) \right) ds\right\Vert \\
& =:I_{1}+I_{2}+I_{3}+I_{4}.\end{aligned}$$
From assumption (G) and norm continuity of $S_{q}\left( t\right) $, $t>0$, it follows that$$I_{1}=\left\Vert \left( S_{q}\left( t_{2}\right) -S_{q}\left(
t_{1}\right) \right) \left( y_{0}-S\left( \frac{1}{n}\right) g\left(
y_{\varepsilon,n}\right) \right) \right\Vert \rightarrow0, \label{q1}$$ as $t_{2}-t_{1}\rightarrow0,$ uniformly for all $y_{\varepsilon,n}\in D$. By direct calculation to $I_{2}$, we obtain that$$I_{2}\leq\frac{M_{S}\left( t_{2}-t_{1}\right) ^{q}}{\Gamma\left(
q+1\right) }\left( M_{B}^{2}M_{S}R_{\varepsilon}+\left\Vert n\right\Vert
_{C}\right) \label{q2}$$
As $t_{1}>0,$we obtain$$I_{3}\leq\frac{M_{S}}{\Gamma\left( q\right) }\int_{0}^{t_{1}}\left( \left(
t_{2}-s\right) ^{q-1}-\left( t_{1}-s\right) ^{q-1}\right) ds\left(
M_{B}^{2}M_{S}R_{\varepsilon}+\left\Vert n\right\Vert _{C}\right) \label{q3}$$
Next, for $t_{1}-\eta>0$, we have$$\begin{aligned}
I_{4} & \leq\left\Vert \left( \int_{0}^{t_{1}-\eta}+\int_{t_{1}-\eta
}^{t_{1}}\right) \left( t_{1}-s\right) ^{q-1}\left( T_{q}\left(
t_{2}-s\right) -T_{q}\left( t_{1}-s\right) \right) \left( Bu_{\varepsilon
,n}\left( s,y_{\varepsilon,n}\right) +f\left( s,y_{\varepsilon,n}\left(
s\right) \right) \right) ds\right\Vert \nonumber\\
& \leq\frac{t_{1}^{q}-\eta^{q}}{q}\sup\left\{ \left\Vert Bu_{\varepsilon
,n}\left( s,y_{\varepsilon,n}\right) +f\left( s,y_{\varepsilon,n}\left(
s\right) \right) \right\Vert :s\in\left[ 0,b\right] ,y_{\varepsilon,n}\in
D\right\} \sup_{0\leq s\leq t-\eta}\left\Vert T_{q}\left( t_{2}-s\right)
-T_{q}\left( t_{1}-s\right) \right\Vert \nonumber\\
& +\frac{2M_{S}\eta^{q}}{\Gamma\left( q+1\right) }\left( M_{B}^{2}M_{S}R_{\varepsilon}+\left\Vert n\right\Vert _{C}\right) \label{q4}$$
Thus, combining the above inequalities (\[q1\])–(\[q4\]) with the continuity of $T_{q}\left( t\right) ,t>0$, in the uniform operator topology, we obtain the equicontinuity of the set $D$ on $\left( 0,b\right] $.
Therefore, the set $D$ is relatively compact in $C\left( \left[ 0,b\right]
,X\right) $ and we may assume $y_{\varepsilon,n}\rightarrow y_{\varepsilon}$ for some $y_{\varepsilon}\in C\left( \left[ 0,b\right] ,X\right) $ as $n\rightarrow\infty$. On the other hand, by Lemma \[lem:5\] $u_{\varepsilon
,n}\left( s,y_{\varepsilon,n}\right) \rightarrow u_{\varepsilon}\left(
s,y_{\varepsilon}\right) $ in $C\left( \left[ 0,b\right] ,X\right) $ as $n\rightarrow\infty.$ By taking the limit as $n\rightarrow\infty$ in $\Theta_{\varepsilon,n}y_{\varepsilon,n}=y_{\varepsilon,n}$ and using the Lebesgue dominated convergence theorem, we obtain that $$y_{\varepsilon}\left( t\right) =S_{q}\left( t\right) \left(
y_{0}-g\left( y_{\varepsilon}\right) \right) +\int_{0}^{t}\left(
t-s\right) ^{q-1}T_{q}\left( t-s\right) \left[ Bu_{\varepsilon}\left(
s,y_{\varepsilon}\right) +f\left( s,y_{\varepsilon}\left( s\right)
\right) \right] ds,$$ for $t\in\left[ 0,b\right] $, which implies that $y_{\varepsilon}$ is a mild solution of semilinear fractional control system (\[cp1\]). This completes the proof.
In view of Theorem \[thm:2\] for any $\mathbb{\varepsilon}>0$ there exists $y_{\mathbb{\varepsilon}}\in C\left( \left[ 0,b\right] ,X\right) $ such that$$y_{\mathbb{\varepsilon}}\left( t\right) =S_{q}\left( t\right) \left(
y_{0}-g\left( y_{\mathbb{\varepsilon}}\right) \right) +\int_{0}^{t}\left(
t-s\right) ^{q-1}T_{q}\left( t-s\right) \left[ Bu_{\varepsilon}\left(
s,y_{\varepsilon}\right) +f\left( s,y_{\varepsilon}\left( s\right)
\right) \right] ds,$$ where $u\left( s,y_{\mathbb{\varepsilon}}\right) =B^{\ast}S^{\ast}\left(
b-s\right) \Phi_{\varepsilon}\left( y_{\mathbb{\varepsilon}}\right) .$
Now we prove our main result.
\[**Proof of Theorem \[thm:main\]**\]By Lemma \[lem:1\](a) we know that $J_{\mathbb{\varepsilon}}$ is strictly convex. Then $J_{\mathbb{\varepsilon}}\left( \varphi;y_{\mathbb{\varepsilon}}\right) $ has a unique critical point which is its minimizer:$$\widehat{\varphi}_{\mathbb{\varepsilon}}\in X:J_{\mathbb{\varepsilon}}\left(
\widehat{\varphi}_{\mathbb{\varepsilon}};y_{\mathbb{\varepsilon}}\right)
=\min_{\varphi\in X}J_{\mathbb{\varepsilon}}\left( \varphi
;y_{\mathbb{\varepsilon}}\right) .$$ Given any $\psi\in X$ and $\lambda\in R$ we have$$J_{\mathbb{\varepsilon}}\left( \widehat{\varphi}_{\mathbb{\varepsilon}};y_{\mathbb{\varepsilon}}\right) \leq J_{\mathbb{\varepsilon}}\left(
\widehat{\varphi}_{\mathbb{\varepsilon}}+\lambda\psi;y_{\mathbb{\varepsilon}}\right)$$ or, in other words,$$\begin{aligned}
& \mathbb{\varepsilon}\left\Vert \widehat{\varphi}_{\mathbb{\varepsilon}}\right\Vert \\
& \leq\frac{\lambda^{2}}{2}\int_{0}^{b}\left( b-s\right) ^{q-1}\left\Vert
B^{\ast}T_{q}^{\ast}\left( b-s\right) \Pi^{\ast}\psi\right\Vert
^{2}ds+\lambda\int_{0}^{b}\left( b-s\right) ^{q-1}\left\langle B^{\ast}T_{q}^{\ast}\left( b-s\right) \Pi^{\ast}\widehat{\varphi}_{\mathbb{\varepsilon}},B^{\ast}T_{q}^{\ast}\left( b-s\right) \Pi^{\ast}\psi\right\rangle ds\\
& +\mathbb{\varepsilon}\left\Vert \widehat{\varphi}_{\mathbb{\varepsilon}}+\lambda\psi\right\Vert -\lambda\left\langle \psi,h\left(
y_{\mathbb{\varepsilon}}\right) \right\rangle .\end{aligned}$$ Dividing this inequality by $\lambda>0$ and letting $\lambda\longrightarrow
0^{+}$ we obtain that$$\begin{aligned}
\left\langle \psi,h\left( y_{\mathbb{\varepsilon}}\right) \right\rangle &
\leq\int_{0}^{b}\left( b-s\right) ^{q-1}\left\langle B^{\ast}T_{q}^{\ast
}\left( b-s\right) \Pi^{\ast}\widehat{\varphi}_{\mathbb{\varepsilon}},B^{\ast}T_{q}^{\ast}\left( b-s\right) \Pi^{\ast}\psi\right\rangle ds\\
& +\mathbb{\varepsilon}\lim\inf_{\lambda\longrightarrow0^{+}}\frac{\left\Vert
\widehat{\varphi}_{\mathbb{\varepsilon}}+\lambda\psi\right\Vert -\left\Vert
\widehat{\varphi}_{\mathbb{\varepsilon}}\right\Vert }{\lambda}\\
& \leq\int_{0}^{b}\left( b-s\right) ^{q-1}\left\langle B^{\ast}T_{q}^{\ast
}\left( b-s\right) \Pi^{\ast}\widehat{\varphi}_{\mathbb{\varepsilon}},B^{\ast}T_{q}^{\ast}\left( b-s\right) \Pi^{\ast}\psi\right\rangle
ds+\mathbb{\varepsilon}\left\Vert \psi\right\Vert .\end{aligned}$$ Repeating this argument with $\lambda<0$ we obtain finally that $$\left\vert \int_{0}^{b}\left( b-s\right) ^{q-1}\left\langle B^{\ast}T_{q}^{\ast}\left( b-s\right) \Pi^{\ast}\widehat{\varphi}_{\mathbb{\varepsilon}},B^{\ast}T_{q}^{\ast}\left( b-s\right) \Pi^{\ast}\psi\right\rangle ds-\left\langle \psi,h\left( y_{\mathbb{\varepsilon}}\right) \right\rangle \right\vert \leq\mathbb{\varepsilon}\left\Vert
\psi\right\Vert . \label{f9}$$ On the other hand, with $u_{\varepsilon}=B^{\ast}T_{q}^{\ast}\left(
b-s\right) \Pi^{\ast}\widehat{\varphi}_{\mathbb{\varepsilon}}$ we have$$\begin{aligned}
& \int_{0}^{b}\left( b-s\right) ^{q-1}\left\langle \Pi T_{q}\left(
b-s\right) BB^{\ast}T_{q}^{\ast}\left( b-s\right) \Pi^{\ast}\widehat
{\varphi}_{\mathbb{\varepsilon}}ds-h\left( y_{\mathbb{\varepsilon}}\right)
,\psi\right\rangle \nonumber\\
& =\left\langle \Pi y_{\mathbb{\varepsilon}}\left( b\right) -y_{b},\psi\right\rangle ,\nonumber\\
h\left( y_{\mathbb{\varepsilon}}\right) & =y_{b}-\Pi S_{q}\left(
b\right) \left( y_{0}-g\left( y_{\mathbb{\varepsilon}}\right) \right)
-\int_{0}^{b}\left( b-s\right) ^{q-1}\Pi T_{q}\left( b-s\right) f\left(
s,y_{\mathbb{\varepsilon}}\left( s\right) \right) ds. \label{f8}$$ Then, combining (\[f9\]) and (\[f8\]) we obtain that$$\left\vert \left\langle \Pi y_{\mathbb{\varepsilon}}\left( b\right)
-y_{b},\psi\right\rangle \right\vert \leq\mathbb{\varepsilon}\left\Vert
\psi\right\Vert$$ holds for any $\psi\in X$. Thus$$\left\Vert \Pi y_{\mathbb{\varepsilon}}\left( b\right) -y_{b}\right\Vert
\leq\mathbb{\varepsilon}.$$
Applications
============
**Example 1:** Consider the following initial-boundary value problem of fractional parabolic control system with Caputo fractional derivatives, based on example 2 in [@dm1]:$$\left\{
\begin{tabular}
[c]{lll}$^{C}D_{0,t}^{2/3}x\left( t,\theta\right) =\dfrac{\partial^{2}}{\partial\theta^{2}}x\left( t,\theta\right) +Bu\left( t,\theta\right)
+f\left( t,x\left( t,\theta\right) \right) ,$ & $t\in\left[ 0,b\right]
,$ & $\theta\in\left[ 0,\pi\right] ,$\\
$x\left( t,0\right) =x\left( t,\pi\right) =0,$ & $t\in\left[ 0,b\right]
,$ & \\
$x\left( 0,\theta\right) =x_{0}\left( \theta\right) -\sum_{i=0}^{m}\int_{0}^{\pi}k\left( \theta,s\right) x\left( t_{i},s\right) ds,$ &
$t\in\left[ 0,b\right] ,$ & $\theta\in\left[ 0,\pi\right] .$\end{tabular}
\ \ \ \ \ \ \ \ \ \ \ \right. \label{ex1}$$ Here, $f$ is a given function, $m$ is a positive integer, $0<t_{0}<t_{1}<...<t_{m}<b,$ $k\left( \cdot,\cdot\right) \in L^{2}\left( \left[
0,\pi\right] \times\left[ 0,\pi\right] ,R^{+}\right) .\ $
Take $X=L^{2}\left[ 0,\pi\right] $ and the operator $A:D\left( A\right)
\subset X\rightarrow X$ is defined by$$Ay=y^{\prime\prime},$$ where the domain $D\left( A\right) $ is defined by$$\left\{ x\in X:x,\ x^{\prime}\ \text{are absolutely continuous,\ \ }x^{\prime\prime}\in X,\ \ x\left( 0\right) =x\left( \pi\right) =0\right\}
.$$ Then, $A$ can be written as$$Ay=-\sum_{n=1}^{\infty}n^{2}\left\langle y,e_{n}\right\rangle e_{n},\ \ \ y\in
D\left( A\right) ,$$ where $e_{n}\left( x\right) =\sqrt{\frac{2}{\pi}}\sin nx$ is an orthonormal basis of $X$. It is well known that $A$ is the infinitesimal generator of a compact, analytic and self-adjoint semigroup $S\left(
t\right) ,$ $t>0$, in $X$ given by$$S\left( t\right) y=\sum_{n=1}^{\infty}\exp\left( -n^{2}t\right)
\left\langle y,e_{n}\right\rangle e_{n},\ \ \ y\in X.$$ Now, define an infinite-dimensional space $U$ by$$U=\left\{ u=\sum_{n=2}^{\infty}u_{n}e_{n}\left( \theta\right) :\sum
_{n=2}^{\infty}u_{n}^{2}<\infty\right\} ,$$ with the norm $\left\Vert u\right\Vert ^{2}=\sum_{n=2}^{\infty}u_{n}^{2}$. Define the operator $B:U\rightarrow X$ as follows$$Bu=2u_{2}e_{1}\left( \theta\right) +\sum_{n=2}^{\infty}u_{n}e_{n}\left(
\theta\right) .$$
The system (\[ex1\]) can be reformulated as the following nonlocal controllability problem in $X.$$$\begin{aligned}
^{C}D_{0,t}^{2/3}y\left( t\right) & =Ay\left( t\right) +Bu\left(
t\right) +f\left( t,y\left( t\right) \right) ,\nonumber\\
y\left( 0\right) & =y_{0}-g\left( y\right) ,\end{aligned}$$ where $y\left( t\right) =x\left( t,\cdot\right) ,\ f:\left[ 0,b\right]
\times X\rightarrow X$ is given by $f\left( t,y\left( t\right) \right)
=f\left( t,x\left( t,\cdot\right) \right) $ and the function $g:C\left(
\left[ 0,b\right] ,X\right) \rightarrow X$ is given by$$g\left( y\right) =\sum_{i=0}^{m}\int_{0}^{\pi}k\left( \cdot,s\right)
x\left( t_{i},s\right) ds$$
This implies that the condition (AC) is satisfied. If we assume that $f$ and $g$ satisfy the assumptions (F) and (G), then by Theorem \[thm:main\], we obtain that the system (\[ex1\]) is approximately controllable on $\left[
0,T\right] .$
**Example 2:** Consider the following affine hereditary differential system in the Hilbert space $E$ $$\begin{aligned}
y^{\prime}\left( t\right) & =Ay\left( t\right) +Ny\left( t-h\right)
+\int_{-h}^{0}M\left( \theta\right) y\left( t+\theta\right) d\theta
+Bu\left( t\right) +f\left( t,y\left( t\right) \right) ,\label{ss1}\\
y\left( 0\right) & =\xi+\int_{\delta}^{b}h\left( s,y\left( s\right)
\right) ds,\ \ y\left( \theta\right) =\eta\left( \theta\right)
,\ \ \ -h\leq\theta<0,\ h>0,\ \delta>0,\nonumber\end{aligned}$$ where $A$ is the generator of the strongly continuous compact semigroup $S\left( t\right) :E\rightarrow E,\ t>0,$ $N\in\mathcal{L}\left( E\right)
$, $M$ is a Lebesgue measurable and essentially bounded $\mathcal{L}\left(
E\right) $-valued function on $\left[ -h,0\right] $, $B\in\mathcal{L}\left( U,E\right) $, $\xi\in E$, $\eta\in L^{2}\left( \left[ -h,0\right]
,E\right) $. Introduce $M^{2}\left( \left[ -h,0\right] ,E\right)
:=E\times L^{2}\left( \left[ -h,0\right] ,E\right) $ and $\widetilde
{A}:D\left( \widetilde{A}\right) \subset M^{2}\left( \left[ -h,0\right]
,E\right) \rightarrow M^{2}\left( \left[ -h,0\right] ,E\right) $ defined by$$\widetilde{A}\left[
\begin{array}
[c]{c}\xi\\
\eta
\end{array}
\right] =\left[
\begin{array}
[c]{c}A\xi+N\eta\left( -h\right) +\int_{-h}^{0}M\left( \theta\right) \eta\left(
\theta\right) d\theta\\
\frac{d}{d\theta}\left( \eta-\xi\right)
\end{array}
\right] ,$$ where$$D\left( \widetilde{A}\right) =\left\{ \left[
\begin{array}
[c]{c}\xi\\
\eta
\end{array}
\right] :\xi\in E,\ \eta\in L^{2}\left( \left[ -h,0\right] ,E\right)
,\ \frac{d}{d\theta}\eta\in L^{2}\left( \left[ -h,0\right] ,E\right)
,\ \eta\left( 0\right) =\xi\right\} .$$$$\widetilde{y}=\left[
\begin{array}
[c]{c}y\\
\overline{y}\end{array}
\right] ,\ \widetilde{y}_{0}=\left[
\begin{array}
[c]{c}\xi\\
\eta
\end{array}
\right] ,\ \ g\left( \widetilde{y}\right) =\left[
\begin{array}
[c]{c}\int_{\delta}^{b}h\left( s,y\left( s\right) \right) ds\\
0
\end{array}
\right] ,\ \ \widetilde{B}=\left[
\begin{array}
[c]{c}B\\
0
\end{array}
\right] ,\ \ \widetilde{f}=\left[
\begin{array}
[c]{c}f\\
0
\end{array}
\right] .$$ The system (\[ss1\]) can be written in the standard form$$\begin{aligned}
\widetilde{y}^{\prime}\left( t\right) & =\widetilde{A}\widetilde{y}\left(
t\right) +\widetilde{B}u\left( t\right) +\widetilde{f}\left(
t,\widetilde{y}\left( t\right) \right) ,\label{ss2}\\
\widetilde{y}\left( 0\right) & =\widetilde{y}_{0}+g\left( \widetilde
{y}\right) .\nonumber\end{aligned}$$ Following [@delfour], we say that the system (\[ss2\]) is
- (partial-)approximately controllable on $\left[ 0,T\right] $ if $\overline{\left\{ y\left( T;u\right) :u\in L^{2}\left( \left[
0,T\right] ,U\right) \right\} }=E,$
- approximately $M^{2}$-controllable on $\left[ 0,T\right] $ if $\overline{\left\{ \widetilde{y}\left( T;u\right) :u\in L^{2}\left(
\left[ 0,T\right] ,U\right) \right\} }=M^{2}\left( \left[ -h,0\right]
,E\right) .$
Assume that $\Phi\left( t\right) $ is a solution of the equation$$\begin{aligned}
\Phi^{\prime}\left( t\right) & =A\Phi\left( t\right) +N\Phi\left(
t-h\right) +\int_{-h}^{0}M\left( \theta\right) \Phi\left( t+\theta\right)
d\theta,\ t\geq h,\\
\Phi^{\prime}\left( t\right) & =A\Phi\left( t\right) ,\ \ t<h,\ \ \Phi
\left( 0\right) =I.\\end{aligned}$$ It is shown in [@delfour], [@bm2] that the linear system corresponding to (\[ss2\]) is partial-approximately controllable if$$B^{\ast}\Phi^{\ast}\left( t-t\right) y=0,\ 0<t<T\Longrightarrow y=0.$$ Hence, the system (\[ss2\]) is partial-approximately controllable on $\left[ 0,T\right] $ provided that all conditions of Theorem \[thm:main\] are satisfied.
Conclusion
==========
In this paper, partial-approximate controllability of semilinear evolution systems in Hilbert spaces with nonlocal conditions have been investigated. Sufficient condition for the partial-approximate controllability of such systems have been established. Compared with some existing results, it can be found that the variational approach, together with the approximating technique, has been extended to consider the partial-approximate controllability of more general systems.
The key hypothesis on the nonlocal function $g$ is the assumption (G), which means that $g$ depends only on the value of $y$ in the interval $\left[
\delta,b\right] $, $\delta>0$. This hypothesis covers the situation that the nonlocal function $g$ is given by $g\left( y\right) ={\displaystyle\sum\limits_{k=1}^{p}}
c_{k}y\left( t_{k}\right) $, where $0<t_{1}<...<t_{p}<b,$ $c_{1},...c_{p}$ are given constants or by $g\left( y\right) =\int_{\delta}^{b}h\left(
s,y\left( s\right) \right) ds$, $0<\delta<b$. However, if the nonlocal function $g$ depends on the value of $y$ on the whole interval $\left[
\delta,b\right] $, such as the nonlocal function $g$ is given by $g\left(
y\right) =\int_{0}^{b}h\left( s,y\left( s\right) \right) ds,$ then even the question of existence of a mild solution and corresponding controllability issues are still open. For this reason, we are committed to studying such problems in the future.
[99]{}
G. Arthi, Ju H. Park , H.Y. Jung, Existence and controllability results for second-order impulsive stochastic evolution systems with state-dependent delay Applied Mathematics and Computation 248 (2014) 328–341.
G. Arthi. K. Balachandran, Controllability of impulsive second-order nonlinear systems with nonlocal conditions in Banach spaces, J. Control Decis. 2 (2015) 203–218.
P. Balasubramaniam, V. Vembarasan, T. Senthilkumar, Approximate controllability of impulsive fractional integro-differential systems with nonlocal conditions in Hilbert space. Numer. Funct. Anal. Optim. 35 (2014) 177–197.
A. E. Bashirov, N. I. Mahmudov, On concepts of controllability for deterministic and stochastic systems. SIAM J. Control Optim. 37 (1999) 1808–1821.
A. E. Bashirov, N. I. Mahmudov, N. Şemi, H. Etikan, Partial controllability concepts. Internat. J. Control 80 (2007) 1–7.
R. F. Curtain, H. J. Zwart, An Introduction to Infinite Dimensional Linear Systems Theory*,* Springer-Verlag/New York, 1995.
J. P. Dauer, N. I. Mahmudov, Approximate controllability of semilinear functional equations in Hilbert spaces. J. Math. Anal. Appl. 273 (2002) 310–327.
A. Debbouche, D.F.M. Torres, Approximate controllability of fractional nonlocal delay semilinear systems in Hilbert spaces. Internat. J. Control 86 (2013) 1577–1585.
A. Debbouche, D.F.M. Torres, Approximate controllability of fractional delay dynamic inclusions with nonlocal control conditions. Appl. Math. Comput. 243 (2014) 161–175.
M. C. Delfour, S. K. Mitter, Controllability, observability and optimal feedback control of affine hereditary differential systems, SIAM J. Control, 10, (1972) 298–328.
S. Ji, Approximate controllability of semilinear nonlocal fractional differential systems via an approximating method. Appl. Math. Comput. 236 (2014), 43–53.
F. D. Ge, H. C. Zhou, C. H. Kou, Approximate controllability of semilinear evolution equations of fractional order with nonlocal and impulsive conditions via an approximating technique. Appl. Math. Comput. 275 (2016) 107–120.
A. A. Kilbas, H. M. Srivastava, J. J. Trujillo, Theory and Applications of Fractional Differential Equations, Elsevier Science B.V., Amsterdam , 2006.
J, Liang, J. H. Liu, T. J. Xiao, Nonlocal Cauchy problems governed by compact operator families. Nonlinear Anal. 57 (2004) 183–189.
N. I. Mahmudov, Approximate controllability of semilinear deterministic and stochastic evolution equations in abstract spaces. SIAM J. Control Optim., 42 (2003) 1604–1622.
N. I. Mahmudov, Finite-approximate controllability of evolution equations. Appl. Comput. Math. 16 (2017), 159–167.
N. I. Mahmudov, Approximate controllability of evolution systems with nonlocal conditions, Nonlinear Anal., 68 (2008) 536-546.
N. I. Mahmudov, M. A. McKibben, On approximately controllable systems (survey). Appl. Comput. Math. 15 (2016) 247–264.
K. Naito, Approximate controllability for trajectories of semilinear control systems, J. Optim. Theory Appl., 60 (1989) 57–65.
N. Sukavanam, S. Tafesse, Approximate controllability of a delayed semilinear control system with growing nonlinear term, Nonlinear Anal. TMA .74 (2011) 6868- 6875.
A. Pazy, Semigroups of Linear Operators and Applications to Partial Differential Equations, Springer Verlag, New York, 1983.
I. Podlubny, Fractional Differential Equations, Academic Press, San Diego, 1999.
R. Sakthivel, Y. Ren, N. I. Mahmudov, On the approximate controllability of semilinear fractional differential systems, Comput. Math. Appl. .62 (2011) 451-1459.
R. Sakthivel, Y. Ren, Approximate controllability of fractional di erential equations with state-dependent delay, Results Math., 63 (2013), 949{963.
T. I. Seidmann, Invariance of the reachable set under nonlinear pertubations, SIAM J. Control and Optim., 25 (1985) 1173-1191.
X. Zhang, C. Zhu, C. Yuan, Approximate controllability of fractional impulsive evolution systems involving nonlocal initial conditions. Adv. Difference Equ. 2015, 2015:244, 14 pp.
H. X. Zhou, Approximate controllability for a class of semilinear abstract equations, SIAM Journal of Control and Optim., 21 (1983) 551-565.
Y. Zhou, F. Jiao, Nonlocal Cauchy problem for fractional evolution equations, Nonlinear Anal. RWA 11 (2010) 4465–4475.
Y. Zhou, Basic Theory of Fractional Di erential Equations, World Scientifc, Singapore, 2014.
M. Yamamoto, J. Y. Park, Controllability for parabolic equations with uniformly bounded nonlinear terms,* *J. Optim. Theory Appl., 66 (1990) 515-532.
E. Zuazua, Finite dimensional null controllability for the semilinear heat equation. J. Math. pures et appl. , 76 (1997) 570-594.
**
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Most state-of-the-art systems today produce morphological analysis based only on orthographic patterns. In contrast, we propose a model for unsupervised morphological analysis that integrates orthographic and semantic views of words. We model word formation in terms of morphological chains, from base words to the observed words, breaking the chains into parent-child relations. We use log-linear models with morpheme and word-level features to predict possible parents, including their modifications, for each word. The limited set of candidate parents for each word render contrastive estimation feasible. Our model consistently [matches or]{} outperforms five state-of-the-art systems on Arabic, English and Turkish.[^1]'
author:
- |
Karthik Narasimhan, Regina Barzilay and Tommi Jaakkola\
CSAIL, Massachusetts Institute of Technology\
[{karthikn, regina, tommi}@csail.mit.edu]{}\
bibliography:
- 'references.bib'
title: An Unsupervised Method for Uncovering Morphological Chains
---
[^1]: [Code is available at <https://github.com/karthikncode/MorphoChain>.]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this lecture I present the way stars can be modeled in two dimensions and especially the fluid flows that are driven by rotation. I discuss some of the various ways of taking into account turbulence and conclude this contribution by a short presentation of some of the first results obtained with the ESTER code on the modeling of interferometrically observed fast rotating early-type stars.'
address: 'Université de Toulouse; UPS-OMP; IRAP; Toulouse, France and CNRS; IRAP; 14, avenue Edouard Belin, F-31400 Toulouse, France'
author:
- Michel Rieutord
bibliography:
- '/home/rieutord/tex/biblio/bibnew.bib'
title: Modeling rotating stars in two dimensions
---
Introduction
============
Rotation remains, with magnetic fields, an ill-known quantity in the interiors of stars. It is however associated with fascinating objects like massive stars or the first stars of the Universe. These latter objects are indeed often called the factories of metals and because of their compactness (due to low opacity) are usually thought to have been fast rotators. However, understanding the mechanisms that lead to the enrichments of the interstellar medium with the variety of elements, requires the understanding of the mechanisms which mix the stars or which simply bring the elements from the regions of their nucleosynthesis to the surface where they can be expelled by winds.
The determination of the mean flows that pervade rotating stars is therefore an unavoidable step towards this understanding. This is why much work has been devoted to insert the effects of flows into one dimensional models [e.g. @MM00; @maeder09]. Even if this modeling succeeded in explaining a variety of effects of rotation (e.g. the ratio of blue to red supergiants in galaxies or the abundance of lithium), 1D models cannot include many specificities of rotating fluid flows. Indeed, a fluid flow is intrinsically a multidimensional phenomenon. Two dimensions of space are therefore the minimum number of dimensions to compute a rather general flow. For instance, the geostrophic flow of rotating fluids, that comes from by the domination of the Coriolis force, verifies the Taylor-Proudman theorem stating that $\dz{\rho\vv}=\vzero$. Such a condition is not compatible with the existence of a stellar core. In fact the very problem is that rotation imposes a cylindrical symmetry while gravity imposes spherical symmetry. The combination of both yields a two dimensional problem.
In the following I therefore propose to focus our attention on the mean flows that pervade a rotating star. We will thus have the opportunity to go through the processes (baroclinicity and Reynolds stresses) that drive a secular mixing of the stars. Then, we will shortly present the state of the ESTER project and some selected first results on the modeling of early-type fast rotating stars.
The problem’s formulation
=========================
In order to make a first step into the 2D-modeling of rotating stars, we shall consider an isolated, non-magnetic, not-mass-losing, early-type star. Moreover, we shall forget about any time evolution: the star is powered by nuclear reactions but these do not influence the chemical composition. However, there are some regions with turbulent flows. There we assume that the turbulence is in a statistically steady state. Such an ideal star does not exist but some stars like Vega may be close to it.
With all the foregoing precautions, we may formulate the problem of a 2D-model of a rotating star in a consistent way.
The partial differential equations that govern the steady-state of our ideal rotating star read:
= 4G\
= -+ \_\*\
= - -+\_v\
= 0. These are Poisson equation for the gravitational potential $\phi$, the energy equation involving the temperature $T$, the density $\rho$, the entropy $S$, the diffusive heat flux $\vF$ and the nuclear heat sources $\eps_*$, the momentum equation with the viscous force $\vF_v$ and the equation of mass conservation. The brackets $\moym{...}$ indicate time averages. They are necessary to remove turbulence fluctuations. Actually, all quantities are time averaged.
Equations should be completed by the prescriptions of the microphysics, namely the equation of state, the opacities, the nuclear network etc. They should also be supplemented with boundary conditions. All these are discussed in [@ELR13]. Here, we shall only focus on the problems raised by the velocity and temperature fields boundary conditions. Before that we need to concentrate on the mean flows.
The mean flows in rotating stars
================================
The mean flows that pervade a rotating star and enforce some mixing have various origin. The most common sources are the following:
- baroclinicity
- Reynolds stresses
- gravitational contraction, mass loss or mass accretion
Let us review these phenomena in more details.
Baroclinicity {#barocsect}
-------------
In fluid mechanics baroclinicity refers to the inclination of isobars and isotherms (or isopycnic, equipotential surfaces...). This is indeed a natural feature of rotating stars that was discovered long ago by [@vonzeipel]. Since much confusion has emerged after this work on the origin of the meridional circulation, I’d like to say a few words about stellar baroclinic flows in order to clarify the phenomenon and the way it should be approached.
Baroclinicity emerges in fluids because pressure and temperature obey two different and independent equations: there are some coupling through buoyancy and heat advection but these are weak. Hence, if we think to these two fields (pressure and temperature), they live independently and therefore establish there own system of isosurfaces. Usually, they do not coincide and baroclinicity arises. It arises because density depends on temperature or on both temperature and pressure. Hence, isodensity surfaces are distinct from isobars or, in other words, the two vectors $\na\rho$ and $\na P$ are not parallel. When taking the curl of the momentum equation (after division by $\rho$), one gets
$$M(\vv) = \frac{\na P\times\na\rho}{\rho^2}$$ where the right hand side is usually called the baroclinic torque and $M$ is a differential operator. The consequence of this torque is that a flow arises. In stars this flow is basically a differential rotation, namely a purely azimuthal velocity field. However, as viscosity is always present, a differential rotation, like any shear flow, transport momentum in the direction of the velocity gradients. In a steady state the momentum flux induced by viscosity must be compensated by some meridional flow. Such a balance is illustrated by the $\varphi$ component of the momentum equation, which may be written
$$v_s\partial_s(sv_\varphi) + v_z\partial_z(sv_\varphi) = \nu
s\Delta'v_\varphi$$ or, in a more condensed form,
(s\^2) = (s\^2) where we used the cylindrical coordinates $(s,\varphi,z)$ and where $\Delta'$ represents some Laplacian-like operator.
As we may note on , advection of angular momentum $\vv\cdot\na(s^2\Omega)$ just compensates the viscous force due to angular velocity gradients $\nu \na\cdot(s^2\na\Omega)$. In axisymmetric situations like that of our rotating stars, advection is that of meridional circulation. We therefore see that in a steady state meridional circulation is controlled only by viscosity.
Confusion arose in the past because it was assumed that rotation generates a thermal disequilibrium. Indeed, in a barotropic fluid
$$\Div(\khi\na T) \neq 0$$ known as von Zeipel’s paradox. The paradox was solved by saying that the meridional circulation velocity $\vv$ was such that
$$\rho T\vv\cdot\na S = \Div(\khi\na T) \andet \Div(\rho\vv) = 0$$ namely, heat advection by meridional circulation compensates the thermal imbalance and verifies mass conservation. [*Such a reasoning is incorrect*]{} because the driving of a flow is due to forces or torques, not to mass and energy conservation which cannot ensure angular momentum conservation. As shown by [@Buss81], a thermal imbalance leads to a time-evolution of the temperature field itself, which is more rapid than a mechanical rearrangement of the fluid. The foregoing discussion is detailed in [@R06b].
Reynolds stresses
-----------------
### The scale of turbulence
Baroclinic flows therefore slowly advect heat, chemicals and angular momentum. However, the differential rotation of baroclinic origin is a mere shear flow that may develop instabilities if the Reynolds number is large enough or if the Richardson number is low enough. These instabilities usually lead to turbulence. Unlike thermal convection where the linear instability drives motion at scales on the fraction of the radius, shear instabilities in a stably stratified fluid inject energy at small-scale. To see that, we recall that the Richardson criterion taking into account the high thermal diffusivity of the fluid imposes [@zahn92]
$$\frac{N^2}{S_h^2}\frac{v\ell}{\kappa} \infapp 1/4$$ where $N$ is the , $S_h$ the local shear, $v$ and $\ell$ the velocities and length scale of the eddies and $\kappa$ the heat diffusivity. This criterion originally proposed by [@townsend58], says that shear instability grows up to a scale (of the eddies) where heat diffusion is unable to smooth out the stable buoyancy of the background. Since the largest eddies have a turn-over time scale $1/S_h$, we may observe that all the eddies forced by the background shear are such that
$$\frac{N^2}{S_h^2}\frac{\ell^2}{\tau}\infapp \kappa/4, \quad {\rm with}\quad
\tau\infapp S_h^{-1}$$ or
$$\frac{\ell}{R} \infapp \lp\frac{\kappa S_h}{4R^2N^2}\rp^{\!1/2}$$ where we introduced $R$ the radius of the star. Let us put numbers in this expression. From [@ELR13], we evaluate the associated turbulent kinematic viscosity
$$\nu_t = \frac{S_h^2\kappa}{12N^2} \sim 10^6 \; {\rm cm}^2/s$$ The global shear of differential rotation is of a fraction of the rotation rate; let us write $S_h=f\Omega$. Hence we find
$$\frac{\ell}{R} \infapp
\lp\frac{\nu_t}{4fR^2\Omega}\rp^{\!1/2}=\frac{\sqrt{E}}{\sqrt{2f}}$$ where $E$ is the Ekman number based on the shear induced turbulence. Since $E< 10^{-8}$ and $f\sim 0.05$ [see @ELR13], we find that
$$\ell/R \infapp 3\times 10^{-4}$$ Hence, turbulence in vertically stratified region remains on relatively small-scales in the vertical direction. In the horizontal directions of course the scales can be large (but limited by the stability of the eddies).
As far as convection zones are concerned, nothing prevents the instability from driving all the scales where the Schwarzschild criterion predicts convection. Hence these are fully mixed.
### Reynolds stresses
The convection zones are fully mixed regions, but because of the strong turbulence they harbour, Reynolds stresses drive mean flows like the differential rotation of the Sun. These stresses and the associated flows also force some flows in the neighbouring radiative zones. The picture that emerges from these remarks is that turbulence is everywhere in stars, with variable intensity, properties, and a modeling of its effects is required. Here we first concentrate on its momentum transport that is on Reynolds stresses. If we use Reynolds decomposition, namely
$$\vv = \moym{\vv} + \vv'$$ where $\vv'$ are the velocity fluctuations with respect to the average, the Reynolds stress tensor is
$${\bf R} = \moym{\rho\vv'\otimes\vv'}$$ Usually, correlations with density are not important (a fortiori in an incompressible fluid!) and the components of the Reynolds stress tensor are often written
$$R_{ij} = \moym{v'_iv'_j}$$ In a steady state, and neglecting correlations with density, mean flows therefore verify:
$$\rho\moym{v_j}\partial_j\moym{v_i} + \partial_j(\rho\moym{v'_iv'_j}) =
-\partial_i\moym{P} + \rho g_i + F_i$$ where $\vF$ is the viscous force.
The main question with the previous equation is the expression of the Reynolds stress tensor as a function of the mean-field, the so-called closure problem of mean field equations.
A basic way of solving this problem is to assume that turbulent stresses are like fluid stresses and that they can be represented by a turbulent viscosity. We thus may write
$$\rho\moym{v_j}\partial_j\moym{v_i} =
-\partial_i\moym{P} + \rho g_i + F^{\rm T}_i$$ with
\^T &=& \_t+() +2(\_t).\
&&.+\_t() -()\_t. where $\mu_t$ is the turbulent dynamical viscosity of the gas. This expression enables local variations of the viscosity although it does not say anything on its determination. But there is a more fundamental difficulty: there is no reason why the functional form of the Reynolds stress should be
= -\_t(\_i+\_j -\_k\_[ij]{}) as expected for a newtonian fluid. The analogy of turbulence with such a fluid is very limited because this expression is derived from the assumption that fluid flows are slight perturbations of the thermodynamic equilibrium. It is unclear what is the statistical equilibrium of turbulence and how actual flows may slightly deviates from it. Moreover, turbulence particles are vortices that support long range interactions, not collisions...
Now, if we still admit , the expression of $\mu_t$ is still a problem. In the nineteentwenties, Prandtl introduced the idea of the mixing length suggesting that in turbulent shear flows
$$\mu_t = \rho\Lambda^2\left|\dz{v_x}\right|$$ where $\Lambda$ is the mixing length and $z$ the direction of the shear. This assumption leads to interesting results on wall-turbulence, but misses the point on the axis of a turbulent jet (where it predicts zero-viscosity!).
Another recipe is the Smagorinski’s prescription saying that $\mu_t=\rho\Delta^2|\moym{c_{ij}}|$ where $\moym{c_{ij}}$ is the mean shear. Such a prescription is more general and still often used in Large-Eddy Simulations (LES) as a basic subgrid model. However, as the mixing-length model, this is a local prescription that does not take care of non-local effects.
To avoid this pitfall, more sophisticated models have been developed for engineering problems. One of the most popular is the $K-\eps$ model of [@LS72]: this models adds two new equations for the evolution of $K$ the turbulent kinetic energy and $\eps$ the energy dissipation. The advection diffusion equation that controls the evolution of these quantities are derived from the evolution of second-order correlations with a simple model of the third-order correlation [@R97]. The point is that when $K$ and $\eps$ are known, the turbulent pressure and the turbulent viscosity are also known. These models may be useful to determine statistically steady inhomogeneous flows that we find in stars. However, because of the huge increase of the computing power, these models have been forsaken to the profit of LES, which are more flexible for industrial flows. In astrophysics, this raises the problem of the comparison with observations, which cannot be as detailed as laboratory experiments.
Another line of research that should be mentioned, is the one followed by Rüdiger and collaborators [see @Rudig89; @KPR94]. This is the mean-field approach of turbulence where one seeks for an expression of $\moym{v'_iv'_j}$ and other second order correlations, as functions of large-scale quantities like local rotation rate. In this approach the Reynolds tensor is not reduced to that of the functional form of a newtonian fluid. New effects like the $\Lambda$-effect or anisotropic turbulent viscosities, appear. These approaches of course raise new kinds of problems [see @snellman_etal12].
Gravitational contraction, mass accretion and mass-loss
-------------------------------------------------------
The last source of large-scale motions in a rotating star is due to expansion or contraction of the star at various phase of its evolution. In non-rotating stars these phenomena just lead to radial flows, but in a rotating star, angular momentum conservation leads to a differential rotation and an associated meridional circulation. Not much is known on these flows that are currently under investigations (Hypolite & Rieutord in preparation). Again, the associated shear will lead to some small-scale turbulence in radiative regions.
Conclusion
----------
To conclude the foregoing discussion, let us underline that we considered here only non-magnetic processes. Magnetic fields will of course complicate the picture. An open question is how all these sources compete together and of course how we should model the associated turbulence.
Heat transport
==============
convection zones
----------------
The foregoing discussion focused on turbulent viscosity and more generally on momentum transport. However, heat transport was really the first obstacle to circumvent when first stellar models were constructed. The well-known mixing-length theory, based on Prandtl’s idea, offers a plausible answer to this problem. In this approach, all the difficulties of turbulent heat transport are condensed in a dimensionless parameter of order unity $\alpha_{\rm MLT}$ that has been calibrated on the Sun. Unfortunately, such a parameter is not universal and varies from star to star, impeding any precise prediction. A clear example is given by the two stars of $\alpha$ Cen [@eggenberger_etal04].
In two dimensions, the difficulty is squared. First because the 1D-MLT has no straightforward generalisation to 2D. Turbulent transport is now also in latitude via turbulent diffusion and mean flows.
A first idea is use the “down-gradient" prescription for the heat flux (this is a prescription that is largely used in the $K-\eps$ model). Here, since thermal convection disappears when the entropy gradient vanishes or becomes positive (in the direction opposite to effective gravity), a natural way to model the convective heat flux is to assume that it is proportional to the entropy gradient, namely
$$\vF_{\rm conv} = \moym{\rho c_pT\vv} = -\khi_{\rm turb}T\na S/\calR$$ where $\calR$ is the ideal gas constant.
When reduced to 1D this prescription can be related to the classical MLT. It may therefore be viewed as a generalization of this theory. However, there are other possibilities like the introduction of an eddy conductivity $\khi_t$ such that $\vF=-\khi_t\na T$ [@Rudig89].
The surface heat transfer
-------------------------
One of the new problem raised by 2D-models concerns the implementation of the boundary conditions that should be met by the temperature field. A simple prescription that replaces a detailed atmospheric model is to impose that the star radiates locally as a black body, namely that
-\_rT = T\^4 where $\khi_r$ is the radiative conductivity, $\sigma$ Stefan-Boltzmann constant and $\vn$ the outer normal of the surface. In non-rotating models this condition is usually applied where the optical depth is $\tau_s=2/3$, but $\tau_s=1$ is also used [@Morel97]. In 2D-models, the situation is more complicated: one has to define the surface where boundary conditions (for velocity, gravitational potential and temperature) are imposed. Since the opacity is a rapidly varying quantity at the surface of a star an iso-optical-depth surface is not convenient numerically. We therefore chose to impose boundary conditions on an isobar and select the isobar that “emerges" at the pole, i.e. whose pressure is
P\_s=\_s,On this isobar $T=T_{\rm eff}$ at the pole only. As one moves towards the equator this isobar sinks into the optically thick matter because at $\tau_s=2/3$ pressure is stronger at the pole than at the equator (recall that gravity is larger at the pole). The trick here is to determine the temperature on this isobar as a function of the latitude. For this we assume that the matter lying above this isobar behaves like a polytrope of index $n$. With this hypothesis, it can be shown [cf @ELR13] that the temperature on this isobar follows the law
T\_b() = \^[1/(n+1)]{}\^[1/4]{} which fully determines the temperature field. In more sophisticated models, $T_b(\theta)$ may be determined by the modeling of the stellar atmosphere.
Computing the baroclinic flows in 2D-models
===========================================
We focus on early-type stars so as to avoid the modeling of outer convective layers. Thermal convection is only modeled in the core of the star. One-dimensional models tell us that it is very efficient and therefore the assumption of an isentropic core is quite good [@maeder09]. ESTER models therefore assume isentropic cores and no further modeling of core turbulence is included yet.
Flows in the radiative envelope
-------------------------------
Above the core, the radiative envelope is not in hydrostatic equilibrium as recalled above. Baroclinic flows pervade it and are responsible of the mixing. They represent the first difficulty for the computation of self-consistent models. When determined, we have a prediction of the differential rotation and the meridional circulation.
As we noted in sect. \[barocsect\] the determination of these flows rests on the fluid’s viscosity, which is crucial for stationary solutions. The importance of viscosity is shown by the value of the Ekman number, namely
$$E = \frac{\nu}{2\Omega R^2}$$ As shown in [@ELR13], its value is always less than 10$^{-8}$ even if shear turbulence is accounted for.
Viscous effects are therefore expected to be small. But they are crucial as they lift the degeneracy of the differential rotation as we shall show now.
Let us take the curl of the momentum equation and let us project it in the azimuthal direction on $\ephi$. If viscous or Reynolds stresses are neglected then one finds:
s = The solution of is invariant with respect to the addition of an arbitrary function of $s$; if $\Omega(s,z)$ is a solution
$$\Omega'^2(s,z) = \Omega^2(s,z)+F(s)$$ is also a solution. As shown in [@R06], viscosity lifts this kind of degeneracy. This may be understood in the following way. The differential rotation depends on $F(s)$. The balance of angular momentum advection and diffusion is given by which is completed by mass conservation. Once $\Omega$ is known, these two equations give the meridional circulation, but without matching the boundary conditions. In general, the derived meridian flow does not verify
$$\vv\cdot\vn =0$$ on the stellar surface. This condition is enforced by an Ekman boundary layer where the mass flux into the surface is identified with the mass pumping of the boundary layer. Indeed, the Ekman boundary layer completes an inviscid solution like $s\Omega(s,z)+G(s)$ in such a way that stress-free conditions are met, but the boundary layer correction $\tu_\theta,\tu_\varphi$ usually do not verify mass conservation. This latter constraint is corrected by the velocity component orthogonal to the layer, which is called the pumping. The identification of the pumping of the layer with the mass flux generated by the circulation at the stellar surface gives the differential equation for the unknown geostrophic flow that appears in inviscid solutions ($F(s)$ or $G(s)$). This differential equation is derived in [@R06], in the simplified case of an incompressible “star" not far from rigid rotation.
In modeling rotating stars in 2D we are not interested in the boundary layer because they are very thin and would require a high spatial resolution near the surface. This is why [@ELR13] derived a special boundary condition that completes and insure stress-free conditions without an explicit computation of the boundary layer flow. This boundary condition reads
$$\label{eq:bl}
E_s
s^2\vec{\hat\xi}\cdot\na\Omega+\psi\vec{\hat\tau}\cdot\na(s^2\Omega)=0
\qquad\mbox{at the surface}\;.$$
where $E_s$ is the surface Ekman number, $\vec{\hat\xi}$ and $\vec{\hat\tau}$ unit vectors normal and tangential to the surface, while $\psi$ is the stream function of the meridional circulation. It turns out that this circulation scales as the interior Ekman number, therefore this small parameter drops out of the momentum equation and only order one quantities are computed.
The Core-Envelope Interface
---------------------------
The foregoing simplification unfortunately breaks down at the core-envelope interface (CEI). Indeed, as shown by [@ELR13] the chemical evolution of the convective core, that leads to a density discontinuity at the CEI, also leads to an angular velocity discontinuity on this interface. In addition we may also consider that turbulent viscosity steeply increases when one goes from the envelope to the core. These discontinuities are triggering the so-called Stewartson layer that develops along the tangential cylinder circumventing the core. This layer may play a crucial role in the chemical enrichment of the envelope by products of the nuclear reactions. Its computation is however demanding as it scales like E$^{1/4}$ but fortunately less than the Ekman layer whose thickness scales like E$^{1/2}$. More studies are necessary to better determine the properties of this interfacial region, which is also suspected of harbouring some convective overshooting.
![Meridional view of the differential rotation of 30 ZAMS star rotating at 98% of its critical angular velocity with X=0.7 and Z=0.02.[]{data-label="Massive"}](fig/rotdiff.eps){width="\textwidth"}
Some results of the ESTER project
=================================
A detailed account of the present achievements of the ESTER project may be found in the two papers by [@ELR13] and [@REL12]. Here, we wish to give a brief summary of the performance and first results of the ESTER code[^1].
![Variations of the flux with colatitude for a ZAMS star of 30 rotating at 98% of its critical angular velocity. Solid line: the ratio of the flux to the polar flux as a function of the colatitude. Dashed line: the prediction of the von Zeipel law.[]{data-label="flux"}](fig/flux.eps){width="\textwidth"}
Presentation
------------
The ESTER code is solving the equation of stellar structure in two dimensions , assuming an isentropic convective core and a radiative envelope, thus restricted to the modeling of an early-type star. We use OPAL tables for the equation of state and opacities. Convection in layers is not computed yet : the temperature gradient is assumed to be close to the radiative one, which is fine for stars with mass larger than 1.8 .
The discretization of the PDE is based on a spectral decomposition: Chebyshev polynomials radially and spherical harmonics horizontally. The distorted shape of the star is managed through a change of variables mapping the star to the spherical geometry. Internal precision is monitored through spectral convergence, the virial test and the energy test [see @REL12 for details]. Iterations follow the Newton algorithm. Fig. \[Massive\] illustrates the result of a two-dimensional model for a massive star rotating at 98% of the critical angular velocity. We show here the differential rotation as a function of radius and latitude.
------------------------------------- ----------------------- ------- ------------------------ -------- --------------- -------
Star
Obs. Model Obs. Model Obs. Model
Mass (M$_\odot$) $2.4^{+0.23}_{-0.37}$ 2.22 2.15$^{+0.10}_{-0.15}$ 2.374 4.15$\pm$0.06 4.10
R$_{\rm eq}$ (R$_\odot$) 2.858$\pm$0.015 2.865 2.726$\pm$0.006 2.726 4.21$\pm$0.07 4.24
R$_{\rm pol}$ (R$_\odot$) 2.388$\pm$0.013 2.385 2.418$\pm$0.012 2.418 3.22$\pm$0.05 3.23
T$_{\rm eq}$ (K) 7570$\pm$124 7674 8910$\pm$130 8973 11010$\pm$520 11175
T$_{\rm pol}$(K) 9384$\pm$154 9236 10070$\pm$90 10070 14520$\pm$690 14567
L (L$_\odot$) 31.3$\pm$1 31.1 47.2$\pm$2 48.0 341$\pm$27 351
V$_{\rm eq}$ (km/s) 240$\pm$12 242 197$\pm$23 205 336$\pm$24 335
P$_{\rm eq}$ (days) 0.598 0.672 0.641
P$_{\rm pol}$ (days) 0.616 0.697 0.658
X$_\mathrm{env.}$ 0.70 0.7546 0.70
X$_\mathrm{core}$/X$_\mathrm{env.}$ 0.37 0.271 0.5
Z 0.02 0.0093 0.02
------------------------------------- ----------------------- ------- ------------------------ -------- --------------- -------
: Comparison between observationally derived parameters of the stars $\alpha$ Oph, $\alpha$ Lyr and $\alpha$ Leo and ESTER models. Data are respectively from [@monnier_etal10], [@monnier_etal12] and [@che_etal11]. Note the good matching of the models.
\[rasalhague\]
Gravity darkening law
---------------------
One of the first results of steady models has been the prediction of the gravity darkening law. Until now, the standard recipe was the von Zeipel law stating that the effective temperature is proportional to the $\frac{1}{4}$ power of the effective gravity. Fitting brightness distributions, interferometric data of rapidly rotating stars have shown that other laws like $T_{\rm eff}\propto g_{\rm eff}^{\beta}$ with adjusted $\beta$, were more appropriate. This has also been the conclusion of ESTER models with predictions on the $\beta$ value, although the models show that a power law is not exactly representing the gravity darkening [@ELR11; @ELR12]. In Fig. \[flux\], we show the dependence of the flux with colatitude for a 30 star rotating close to critical angular velocity. The von Zeipel law underestimate the flux in the equatorial regions by more than a factor 3.
Some models of nearby fast rotating stars
-----------------------------------------
Fundamentals parameters of some nearby (famous) rotating stars, derived by interferometry, have been compared successfully to ESTER models as shown by Tab. \[rasalhague\]. A further comparison is shown in Tab. \[deltavel\]. $\delta$ Vel A is an eclipsing binary that has been studied in detail by interferometry [@merand_etal11; @pribulla_etal11]. Here too, two dimensional models nicely fit the observationnally derived parameters of the two stars[^2]. In addition they suggest that they are less metallic than the Sun (Z$\simeq$ 0.011). We also note the stronger hydrogen depletion in the core of the most massive one as expected from its fastest evolution.
On the other hand the case of Achernar, the closest Be star to the solar system, is more difficult since none of the explored models really fit the parameters derived from interferometry. Fitting the polar and equatorial radii lead to rather extreme composition or mass suggesting that the star may just have left the main sequence (to which we are constrained at the moment). In view of previous results based on spherical models [@vinicius_etal06], this not totally surprising. However, this star remains a challenging case for two-dimensional models.
------------------------------------- --------------- ------- --------------- ------- ------------------------ -------
Star
Obs. Model Obs. Model Obs. Model
Mass (M$_\odot$) $2.43\pm0.02$ 2.43 2.27$\pm0.02$ 2.27 8.20
R$_{\rm eq}$ (R$_\odot$) 2.97$\pm$0.02 2.95 2.52$\pm$0.03 2.52 11.6$\pm$0.3 11.5
R$_{\rm pol}$ (R$_\odot$) 2.79$\pm$0.04 2.77 2.37$\pm$0.02 2.36 8.0$\pm$0.4 7.9
T$_{\rm eq}$ (K) 9450 9440 9560 9477 9955$^{+1115}_{-2339}$ 11250
T$_{\rm pol}$(K) 10100 10044 10120 10115 18013$^{+141}_{-171}$ 16800
L (L$_\odot$) 67$\pm$3 65.2 51$\pm$2 48.5 4500$\pm$300 3700
V$_{\rm eq}$ (km/s) 143 143 150 153 298$\pm$9 339
P$_{\rm eq}$ (days) 1.045 0.832 1.72
P$_{\rm pol}$ (days) 1.084 0.924 1.68
X$_\mathrm{env.}$ 0.70 0.70 0.74
X$_\mathrm{core}$/X$_\mathrm{env.}$ 0.10 0.30 0.05
Z 0.011 0.011 0.04
------------------------------------- --------------- ------- --------------- ------- ------------------------ -------
: Comparison between observationally derived parameters of the stars and tentative two-dimensional models. Data from $\delta$ Vel are from [@merand_etal11], those of Achernar are from [@domiciano_etal12]. The models compare nicely with observationally constrained data for the two components of $\delta$ Vel A (an eclipsing binary) but have difficulties with Achernar.
\[deltavel\]
Outlooks
========
Presently, ESTER two dimensional models are, strictly speaking, models of internal structure of rotating stars. No time evolution is included. They describe main sequence early-type stars, that is for masses larger than 1.8 M$_\odot$. Evolution along the main sequence can be mimicked by varying the hydrogen mass fraction in the convective core. Clearly, the next important steps are the extension to low mass stars, which means the computation of outer convective envelope, and the inclusion of time evolution. Other two-dimensional models are currently being developed by Deupree and coworkers [@deupree11; @deupree_etal12], but in these models the differential rotation should be prescribed.
Such internal structure models are useful to interpret the seismic frequencies of rotating stars when perturbative methods fail. Presently, two codes may deal with two dimensional models: TOP by Reese [@RLR06; @reese_etal13] or ACOR by Ouazzani [@ouazzani_etal12]. Beside asterosismology, 2D-models are important for interferometry. Indeed, the interpretation of interferometric visibilities requires the adjustement of models that include the centrifugal distortion of the stars as well as the associated gravity darkening [e.g. @DVJJA02]. Here, a future improvement will be the calculation of two-dimensional models of atmospheres.
Beyond the obvious improvements that have been mentioned above, we see that progresses in the understanding of rotating stars will have to go through a better modeling of the turbulent transport in all places where turbulence develops. This is certainly the most challenging issue of this modeling since we do not have a general theory of turbulence at hands.
I would like to warmly thank Francisco Espinosa Lara who has so much contributed to the success of the ESTER project and with whom I had so many enlighting discussions. The author acknowledges the support of the French Agence Nationale de la Recherche (ANR), under grant ESTER (ANR-09-BLAN-0140). This work was also supported by the Centre National de la Recherche Scientifique (CNRS, UMR 5277), through the Programme National de Physique Stellaire (PNPS). The numerical calculations were carried out on the CalMip machine of the “Centre Interuniversitaire de Calcul de Toulouse” (CICT) which is gratefully acknowledged.
[^1]: The ESTER code is a public domain code that is freely available at http://code.google.com/p/ester-project/
[^2]: The tidal distortion is weak, less than 10$^{-4}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The face ring of a homology manifold (without boundary) modulo a generic system of parameters is studied. Its socle is computed and it is verified that a particular quotient of this ring is Gorenstein. This fact is used to prove that the sphere $g$-conjecture implies all enumerative consequences of its far reaching generalization (due to Kalai) to manifolds. A special case of Kalai’s manifold $g$-conjecture is established for homology manifolds that have a codimension-two face whose link contains many vertices.'
author:
- |
Isabella Novik [^1]\
Department of Mathematics, Box 354350\
University of Washington, Seattle, WA 98195-4350, USA,\
`[email protected]`
- |
Ed Swartz [^2]\
Department of Mathematics,\
Cornell University, Ithaca NY, 14853-4201, USA,\
`[email protected] `
title: 'Gorenstein rings through face rings of manifolds.'
---
Introduction
============
In 1980 Stanley proved the necessity of McMullen’s conjectured description of $f$-vectors of boundaries of simplicial convex polytopes [@St80]. At about the same time, Billera and Lee demonstrated that McMullen’s conditions were sufficient [@BilleraLee]. Since then, one of the most central problems in the field of face numbers of simplicial complexes is the $g$-conjecture. In its most optimistic form it states that, just as in the case of polytope boundaries, the face ring of a homology sphere modulo a generic system of parameters has a Lefschetz element. In the middle 90’s Kalai suggested a far reaching generalization of this conjecture to all homology manifolds [@N98 Section 7]. It is a remarkable fact that all of the enumerative consequences of Kalai’s conjecture are implied by the apparently weaker $g$-conjecture. This follows from our main result, Theorem \[Gorenstein\], that a particular quotient of the face ring of a homology manifold is Gorenstein. We also verify a special case of Kalai’s conjecture when the complex has a codimension-two face whose link contains many vertices.
The main objects we will consider are Buchsbaum complexes and more specifically homology manifolds. Historically, Buchsbaum complexes were defined algebraically. Here we adopt the following theorem of Schenzel [@Sch] as our definition. Let ${{\bf k}}$ be an infinite field of an arbitrary characteristic, and let $\tilde{H}_i(\Delta)$ be the $i$-th reduced simplicial homology of $\Delta$ with coefficients in [[**k**]{}]{}.
A $(d-1)$-dimensional simplicial complex $\Delta$ is called [**Buchsbaum**]{} (over [[**k**]{}]{}) if it is pure and for every non empty face $\tau\in \Delta$, $\tilde{H}_i({\mbox{\upshape lk}\,}\tau)=0$ for all $i<d-|\tau|-1$, where ${\mbox{\upshape lk}\,}\tau =\{\sigma \in \Delta \, : \, \sigma\cap \tau=\emptyset, \,
\sigma\cup \tau \in \Delta\}$ is the link of $\tau$ in $\Delta$.
We say that $\Delta$ is a [**homology manifold**]{} over ${{\bf k}}$ (without boundary) if it is Buchsbaum and in addition $\tilde{H}_{d-|\tau|-1}({\mbox{\upshape lk}\,}\tau)\cong {{\bf k}}$ for all $\emptyset\neq \tau\in \Delta$. A [[**k**]{}]{}-[*homology sphere*]{} is a complex $\Delta$ such that for all $\tau\in \Delta,$ including $\tau=\emptyset$, $$\tilde{H}_i({\mbox{\upshape lk}\,}\tau)=
\left\{
\begin{array}{ll} 0 & \mbox{ if $i<d-|\tau|-1$}, \\
{{\bf k}}& \mbox{ if $i=d-|\tau|-1$}.
\end{array}
\right.$$ In particular, a [[**k**]{}]{}-homology sphere is a [[**k**]{}]{}-homology manifold, and a triangulation of a topological sphere (topological manifold, resp.) is a [[**k**]{}]{}-homology sphere ([[**k**]{}]{}-homology manifold, resp.) for any field [[**k**]{}]{}.
If $\Delta$ is a simplicial complex on $[n]$, then its [*face ring*]{} (or the [*Stanley-Reisner ring*]{}) is $${{\bf k}}[\Delta] := {{\bf k}}[x_1, \ldots, x_n]/I_\Delta,
\quad \mbox{where }
I_\Delta=(x_{i_1}x_{i_2}\cdots x_{i_k} : \{i_1<i_2<\cdots<i_k\}
\notin\Delta).$$ Various combinatorial and topological invariants of $\Delta$ are encoded in the algebraic invariants of ${{\bf k}}[\Delta]$ and vice versa. For instance, if $\Delta$ is a $(d-1)$-dimensional complex, then the Krull dimension of ${{\bf k}}[\Delta]$, $\dim{{\bf k}}[\Delta]$, is equal to $d$. In this case, a set of $d$ linear forms $\theta_1, \ldots, \theta_d \in {{\bf k}}[\Delta]$ is called a [*linear system of parameters*]{} (abbreviated, l.s.o.p.) if $${{\bf k}}(\Delta):={{\bf k}}[\Delta]/(\theta_1, \ldots, \theta_d)$$ has Krull dimension zero (equivalently, ${{\bf k}}(\Delta)$ is a finite-dimensional [[**k**]{}]{}-space). Assuming ${{\bf k}}$ is infinite, an l.s.o.p. always exists: a generic choice of $\theta_1, \ldots, \theta_d$ does the job.
One invariant that measures how far $\Delta$ is from being a homology sphere is the socle of ${{\bf k}}(\Delta)$, ${\mbox{\upshape Soc}\,}{{\bf k}}(\Delta)$, where for a ${{\bf k}}[x_1, \ldots, x_n]$- or ${{\bf k}}[\Delta]$-module $M$, $${\mbox{\upshape Soc}\,}M:= \{y\in M \, : \, x_i\cdot y =0
\mbox{ for all } i=1,\ldots, n\}.$$ When $\Delta$ is a homology sphere, ${\mbox{\upshape Soc}\,}{{\bf k}}(\Delta)$ is a 1-dimensional ${{\bf k}}$-space. Since ${{\bf k}}[\Delta]$ is a graded ${{\bf k}}$-algebra for any $\Delta$, the ring ${{\bf k}}(\Delta)$ and its ideal ${\mbox{\upshape Soc}\,}{{\bf k}}(\Delta)$ are graded as well. We denote by ${{\bf k}}(\Delta)_i$ and $({\mbox{\upshape Soc}\,}{{\bf k}}(\Delta))_i$ their $i$-th homogeneous components. It is well known (for instance, see [@St96 Lemma III.2.4(b)]) that for any $(d-1)$-dimensional $\Delta$, ${{\bf k}}(\Delta)_i$, and hence also $({\mbox{\upshape Soc}\,}{{\bf k}}(\Delta))_i$, vanish for all $i>d$. If $\Delta$ is a Buchsbaum complex, then for $i\leq d$, $({\mbox{\upshape Soc}\,}({{\bf k}}(\Delta))_i$ can be expressed in terms of the local cohomology modules of ${{\bf k}}[\Delta]$ with respect to the irrelevant ideal, $H^j({{\bf k}}[\Delta])$, as follows, see [@NovSw Theorem 2.2].
\[socle\] Let $\Delta$ be a $(d-1)$-dimensional simplicial complex. If $\Delta$ is Buchsbaum, then for all $0\leq i \leq d$, $$({\mbox{\upshape Soc}\,}{{\bf k}}(\Delta))_i \cong
\left(\bigoplus_{j=0}^{d-1} \binom{d}{j}H^j ({{\bf k}}[\Delta])_{i-j} \right)
\bigoplus \S_{i-d},$$ where $\S$ is a graded submodule of ${\mbox{\upshape Soc}\,}H^d({{\bf k}}[\Delta])$ and $rM$ denotes the direct sum of $r$ copies of $M$.
While Theorem \[socle\] identifies a big chunk of the socle of ${{\bf k}}(\Delta)$, its other part, $\S$, remains a mystery. Here we solve this mystery in the special case of a connected [*orientable*]{} ${{\bf k}}$-homology manifold without boundary, thus verifying Conjecture 7.2 of [@NovSw]. A connected ${{\bf k}}$-homology manifold $\Delta$ without boundary is called orientable if $\tilde{H}_{d-1}(\Delta)\cong {{\bf k}}$.
\[S\] Let $\Delta$ be a $(d-1)$-dimensional connected orientable [[**k**]{}]{}-homology manifold without boundary. Then $\dim_{{\bf k}}\S=\dim_{{\bf k}}\S_0=1$. In particular, $\dim{\mbox{\upshape Soc}\,}{{\bf k}}(\Delta)_i = \binom{d}{i}\beta_{i-1}$, where $\beta_{i-1}:=\dim_{{\bf k}}\tilde{H}_{i-1}(\Delta)$ is the $i$-th reduced Betti number of $\Delta$.
A graded [[**k**]{}]{}-algebra of Krull dimension zero is called [*Gorenstein*]{} if its socle is a 1-dimensional [[**k**]{}]{}-space (see [@St96 p. 50] for many other equivalent definitions). Let $$I:= \bigoplus_{i=0}^{d-1} {\mbox{\upshape Soc}\,}{{\bf k}}(\Delta)_i
\quad \mbox{and} \quad
\overline{{{\bf k}}(\Delta)}:={{\bf k}}(\Delta)/I.$$ (Note that the top-dimensional component of the socle is not a part of $I$). If $\Delta$ is a homology sphere, then $I=0$ and $\overline{{{\bf k}}(\Delta)}={{\bf k}}(\Delta)$ is a Gorenstein ring [@St96 Theorem II.5.1]. What if $\Delta$ is a homology manifold other than a sphere? How far is ${{\bf k}}(\Delta)$ from being Gorenstein in this case? The answer (that was conjectured in [@NovSw Conjecture 7.3]) turns out to be surprisingly simple:
\[Gorenstein\] Let $\Delta$ be a $(d-1)$-dimensional connected simplicial complex. If $\Delta$ is an orientable [[**k**]{}]{}-homology manifold without boundary, then $\overline{{{\bf k}}(\Delta)}$ is Gorenstein.
In [@Sch], Schenzel computed the Hilbert function of ${{\bf k}}(\Delta)$ for a Buchsbaum complex $\Delta$ in terms of its face and Betti numbers. It follows from Theorem \[S\] combined with Schenzel’s formula and Dehn-Sommerville relations [@Klee] that for a connected orientable homology manifold $\Delta$, the Hilbert function of $\overline{{{\bf k}}(\Delta)}$ is symmetric, that is, $$\label{symmetry}
\dim_{{\bf k}}\overline{{{\bf k}}(\Delta)}_i =
\dim_{{\bf k}}\overline{{{\bf k}}(\Delta)}_{d-i} \quad \mbox{ for all }
0\leq i \leq d.$$ As the Hilbert function of a Gorenstein ring of Krull dimension zero is always symmetric [@St96 p. 50], Theorem \[Gorenstein\] gives an alternative algebraic proof of (\[symmetry\]).
Theorems \[S\] and \[Gorenstein\] are ultimately related to the celebrated $g$-conjecture that provides a complete characterization of possible face numbers of homology spheres. The most optimistic version of this conjecture is a very strong manifistation of the symmetry of the Hilbert function. It asserts that if $\Delta$ is a $(d-1)$-dimensional homology sphere, and $\omega$ and $\Theta=\{\theta_1, \ldots, \theta_d\}$ are sufficiently generic linear forms, then for $i\leq d/2$, multiplication $$\omega^{d-2i}: {{\bf k}}(\Delta)_i \longrightarrow {{\bf k}}(\Delta)_{d-i}$$ is an isomorphism. At present this conjecture is known to hold only for the class of polytopal spheres [@St80] and edge decomposable spheres [@Murai], [@Nevo].
Kalai’s far-reaching generalization of the $g$-conjecture [@N98] posits that if $\Delta$ is an orientable homology manifold and $\omega, \Theta$ are sufficiently generic, then $$\omega^{d-2i}: \overline{{{\bf k}}(\Delta)}_i \longrightarrow
\overline{{{\bf k}}(\Delta)}_{d-i}$$ is still an isomorphism for all $i\leq d/2$. Let $h''_i = \dim_{{\bf k}}\overline{{{\bf k}}(\Delta)}_i.$ Given a system of parameters and $\omega$ which satisfy Kalai’s conjecture it is immediate that multiplication $\omega: \overline{{{\bf k}}(\Delta)}_i \to
\overline{{{\bf k}}(\Delta)}_{i+1}$ is an injection for $i < d/2.$ So, $h''_0 \le \dots \le h''_{\lfloor d/2 \rfloor},$ and examination of $ \overline{{{\bf k}}(\Delta)} / (\omega)$ shows that the nonnegative integer vector $(h''_0, h''_1 - h''_0, \dots,
h''_{\lfloor d/2 \rfloor} - h''_{\lfloor d/2 \rfloor - 1})$ is an M-vector, i.e. satisfies Macaulay’s nonlinear arithmetic conditions (see [@St96 p. 56]) for the Hilbert series of a homogeneous quotient of a polynomial ring. Applying the same reasoning to $I_{[i]}:=\oplus_{j=0}^{i} I_j$, which is also an ideal, and ${{\bf k}}(\Delta, i):= {{\bf k}}(\Delta)/I_{[i]}$, instead of $I$ and $\overline{{{\bf k}}(\Delta)} = {{\bf k}}(\Delta)/I$, the second conclusion can be strengthen to $(h''_0, h''_1-h''_0,
\ldots, h''_i-h''_{i-1}, h''_{i+1}-h''_{i}+\binom{d}{i+1}\beta_i)$ is an $M$-vector for every $i<\lfloor d/2\rfloor$. In fact, as we will see in Theorem \[connection\], these two conclusions follow from the $g$-conjecture.
For Kalai’s conjecture we prove this special case.
\[g-thm\] Let $\Delta$ be a $(d-1)$-dimensional orientable [[**k**]{}]{}-homology manifold with $d\geq 3$. If $\Delta$ has a $(d-3)$-dimensional face $\tau$ whose link contains all of the vertices of $\Delta$ that are not in $\tau$, then for generic choices of $\omega$ and $\Theta$, $\omega^{d-2}: \overline{{{\bf k}}(\Delta)}_1 \longrightarrow
\overline{{{\bf k}}(\Delta)}_{d-1}
$ is an isomorphism.
The condition that the link of $\tau$ contains all of the vertices of $\Delta$ that are not in $\tau$ is equivalent to saying that every vertex of $\Delta$ is in the star of $\tau$, ${\mbox{\upshape st}\,}\tau:=\{\sigma\in \Delta \, : \, \sigma\cup \tau\in \Delta\}$. This condition is not as restrictive as one might think: the results of [@Sw Section 5] imply that every connected homology manifold $M$ without boundary that has a triangulation, always has a triangulation $\Delta$ satisfying this condition.
The outline of the paper is as follows. We verify Theorem \[S\] in Section 2. The main ingredient in the proof is Gräbe’s explicit description of $H^d({{\bf k}}[\Delta])$ as a ${{\bf k}}[\Delta]$-module in terms of the simplicial (co)homology of the links of faces of $\Delta$ and maps between them [@Grabe]. Theorem \[Gorenstein\] is proved in Section 3 and is used to explore the relationship between the $g$-conjecture and Kalai’s conjecture. Section 4 is devoted to the proof of Theorem \[g-thm\]. The proofs of Theorems \[Gorenstein\] and \[g-thm\] rely heavily on a result from [@Sw] relating ${{\bf k}}({\mbox{\upshape lk}\,}v)$ (for a vertex $v$ of $\Delta$) to the principal ideal $(x_v) \subset {{\bf k}}(\Delta)$.
Socles of homology manifolds
============================
The goal of this section is to verify Theorem \[S\]. To do so we analyze ${\mbox{\upshape Soc}\,}H^d({{\bf k}}[\Delta])$ and prove the following.
\[socH\^d\] If $\Delta$ is a connected orientable $(d-1)$-dimensional [[**k**]{}]{}-homology manifold without boundary, then ${\mbox{\upshape Soc}\,}H^d({{\bf k}}[\Delta])_i=0$ for all $i\neq 0$.
The proof relies on results from [@Grabe]. We denote by $|\Delta|$ the geometric realization of $\Delta$. For a face $\tau\in \Delta$, let ${\mbox{\upshape cost}\,}\tau :=\{\sigma \in \Delta \, : \sigma\not\supset \tau\}$ be the contrastar of $\tau$, let $H^i(\Delta, {\mbox{\upshape cost}\,}\tau)$ be the simplicial $i$-th cohomology of a pair (with coefficients in [[**k**]{}]{}), and for $\tau \subset \sigma \in \Delta$, let ${\iota}^*$ be the map $H^i(\Delta, {\mbox{\upshape cost}\,}\sigma) \to H^i(\Delta, {\mbox{\upshape cost}\,}\tau)$ induced by inclusion ${\iota}: {\mbox{\upshape cost}\,}\tau \to {\mbox{\upshape cost}\,}\sigma$. Also, if $\emptyset \neq \tau\in \Delta$, let $\hat{\tau}$ be the barycenter of $\tau$. Finally, for a vector $U=(u_1, \ldots, u_n)\in \Z^n$, let $s(U):=\{l : u_l\neq 0\}\subseteq [n]$ be the support of $U$, let $\{e_l\}_{l=1}^n$ be the standard basis for $\Z^n$, and let $\N$ denote the set of nonnegative integers.
[[**\[Gräbe\]**]{}]{} \[Grabe\] The following is an isomorphism of $\Z^n$-graded ${{\bf k}}[\Delta]$-modules $$\label{Hochster}
H^{i+1}({{\bf k}}[\Delta]) \cong
\bigoplus_{\mbox{\tiny${\begin{array}{cc} -U\in \N^n\\
s(U)\in\Delta
\end{array}}$}}
H^i (\Delta, {\mbox{\upshape cost}\,}s(U)),$$ where the ${{\bf k}}[\Delta]$-structure on the $U$-th component of the right-hand side is given by $$\cdot x_l = \left\{ \begin{array}{lll}
\text{$0$-map}, & \mbox{ if } l\notin s(U)\\
\text{identity map}, & \mbox{ if } l\in s(U) \mbox{ and } l\in s(U+e_l)\\
{\iota}^*: H^{i}(\Delta, {\mbox{\upshape cost}\,}s(U)) \to H^i(\Delta, {\mbox{\upshape cost}\,}s(U+e_l)), &
\mbox{ otherwise}.
\end{array}
\right.$$
We remark that the isomorphism of (\[Hochster\]) on the level of vector spaces (rather than ${{\bf k}}[\Delta]$-modules) is due to Hochster, see [@St96 Section II.4].
[*Proof of Theorem \[socH\^d\]: *]{} In view of Theorem \[Grabe\], to prove that ${\mbox{\upshape Soc}\,}H^d({{\bf k}}[\Delta])_i=0$ for all $i\neq 0$, it is enough to show that for every $\emptyset \neq \tau\in\Delta$ and $l\in\tau$, the map ${\iota}^*: H^{d-1}(\Delta, {\mbox{\upshape cost}\,}\tau) \to
H^{d-1}(\Delta, {\mbox{\upshape cost}\,}\sigma)$, where $\sigma= \tau-\{l\}$, is an isomorphism. Assume first that $\sigma \neq \emptyset$. Consider the following diagram. $$\begin{CD}
H^{d-1}(|\Delta|) @>(j^\star)^{-1}>>
H^{d-1}(|\Delta|, |\Delta| - \hat{\sigma})
@>f^\star>> H^{d-1}(\Delta, {\mbox{\upshape cost}\,}\sigma) \\
@| @. @A{\iota}^\star AA \\
H^{d-1}(|\Delta|) @>(j^\star)^{-1}>>
H^{d-1}(|\Delta|, |\Delta| - \hat{\tau})
@>f^\star>> H^{d-1}(\Delta, {\mbox{\upshape cost}\,}\tau)
\end{CD}$$ The two $f^\star$ maps are induced by inclusion and are isomorphisms by the usual deformation retractions. $j^\star$ is also induced by inclusion. Since $\Delta$ is connected and orientable, all of the spaces are one-dimensional and $j^\star$ is an isomorphism, so that $(j^\star)^{-1}$ is well-defined and is an isomomorphism as well. Hence compositions $f^\star \circ (j^\star)^{-1}$ are isomorphisms. The naturality of $j^\star$ implies that the diagram is commutative. It follows that ${\iota}^\star$ is an isomorphism. If $\sigma=\emptyset$, replace in the above diagram $H^{d-1}(|\Delta|, |\Delta| - \hat{\sigma})$ with $H^{d-1}(|\Delta|)$. The same reasoning applies.
We are now in a position to complete the proof of Theorem \[S\].
[*Proof of Theorem \[S\]: *]{} Since $\S_i$ is a subspace of ${\mbox{\upshape Soc}\,}H^d({{\bf k}}[\Delta])_i$ (Theorem \[socle\]), and since the later space is the zero-space whenever $i\neq 0$ (see Theorem \[socH\^d\]), it follows that $\S_i=0$ for all $i\neq 0$. For $i=0$, we have $$\dim \S_0=\dim {\mbox{\upshape Soc}\,}{{\bf k}}(\Delta)_d= \dim {{\bf k}}(\Delta)_d=
\beta_{d-1}(\Delta)=1.$$ Here the first step is by Theorem \[socle\], the second step is a consequence of ${{\bf k}}(\Delta)_d$ being the last non vanishing homogeneous component of ${{\bf k}}(\Delta)$, and the third step is by Schenzel’s formula [@Sch] (see also [@St96 Theorem II.8.2]). The “In particular”-part then follows from Theorem \[socle\], isomorphism (\[Hochster\]), and the standard fact that $H^{i-1}(\Delta, {\mbox{\upshape cost}\,}\tau) \cong \tilde{H}_{i-|\tau|-1}({\mbox{\upshape lk}\,}\tau)$, see e.g. [@Grabe Lemma 1.3].
Gorenstein property
===================
In this section we verify Theorem \[Gorenstein\] and use it to discuss connections between various $g$-conjectures. To prove that $\overline{{{\bf k}}(\Delta)}={{\bf k}}(\Delta)/I$ is Gorenstein, where $\Delta$ is a $(d-1)$-dimensional connected orientable homology manifold and $I=\oplus_{i=0}^{d-1} {\mbox{\upshape Soc}\,}{{\bf k}}(\Delta)_i$, we have to check that the operation of moding out by $I$ does not introduce new socle elements. This turns out to be a simple application of [@Sw Proposition 4.24], which we review now.
Let $\Theta=\{\theta_1, \ldots, \theta_d\}$ be an l.s.o.p. for ${{\bf k}}[\Delta]$ and let $v$ be a vertex of $\Delta$. Fix a facet $\tau=\{v=v_1, v_2, \ldots, v_d\}$ that contains $v$. By doing Gaussian elimination on the $d\times n$-matrix whose $(i,j)$-th entry is the coefficient of $x_j$ in $\theta_i$ we can assume without loss of generality that $\theta_i=x_{v_i}+\sum_{j\notin\tau}\theta_{i,j}x_j$. Denote by $\theta'_i$ the linear form obtained from $\theta_i$ by removing all summands involving $x_j$ for $\{j\}\notin {\mbox{\upshape lk}\,}v$. Then $\Theta':=\{\theta'_2, \ldots, \theta'_d\}$ can be considered as a subset of ${{\bf k}}[{\mbox{\upshape lk}\,}v]_1$. Moreover, it is easy to check, say, using [@St96 Lemma III.2.4(a)], that $\Theta'$ forms an l.s.o.p. for ${{\bf k}}[{\mbox{\upshape lk}\,}v]$. The ring ${{\bf k}}({\mbox{\upshape lk}\,}v):={{\bf k}}[{\mbox{\upshape lk}\,}v]/(\Theta')$ has a natural ${{\bf k}}[x_1,\ldots,\hat{x_v}, \ldots, x_n]$-module structure (if $j\neq v$ is not in the link of $v$, then multiplication by $x_j$ is the zero map), and defining $$x_v \cdot y:=-\theta'_1 \cdot y \quad \mbox{ for } y\in {{\bf k}}({\mbox{\upshape lk}\,}v)$$ extends it to a ${{\bf k}}[x_1, \ldots, x_n]$-module structure. Proposition 4.24 of [@Sw] asserts the following.
\[link\] Let $\Delta$ be an orientable homology manifold. The map $$\phi: {{\bf k}}[{\mbox{\upshape lk}\,}v]/(\Theta')
\to (x_v)\left({{\bf k}}[\Delta]/(\Theta)\right)
\quad
\mbox{given by }z \mapsto x_v\cdot z,$$ is well-defined and is an isomorphism (of degree 1) of ${{\bf k}}[x_1, \ldots, x_n]$-modules. Its inverse, $x_v \cdot z \mapsto\overline{z}$, is given by replacing each occurrence of $x_v$ in $z$ with $-\theta'_1$ and setting all $x_j$ for $j\neq v$ not in the link of $v$ to zero.
[*Proof of Theorem \[Gorenstein\]: *]{} To prove the theorem it is enough to show that the socle of $\overline{{{\bf k}}(\Delta)}={{\bf k}}(\Delta)/I$, where $I=\bigoplus_{j=0}^{d-1} {\mbox{\upshape Soc}\,}{{\bf k}}(\Delta)_j$, vanishes in all degrees $j\neq d$. This is clear for $j=d-1$. For $j\leq d-2$, consider any element $y\in {{\bf k}}(\Delta)_j$ such that $x_v\cdot y\in {\mbox{\upshape Soc}\,}{{\bf k}}(\Delta)$ for all $v\in [n]$. We have to check that $y\in {\mbox{\upshape Soc}\,}{{\bf k}}(\Delta)$. And indeed, the isomorphism of Theorem \[link\] implies that $\overline{y}:=\phi^{-1}(x_v\cdot y)\in {{\bf k}}({\mbox{\upshape lk}\,}v)_j$ is in the socle of ${{\bf k}}({\mbox{\upshape lk}\,}v)$. Since ${\mbox{\upshape lk}\,}v$ is a $(d-2)$-dimensional homology sphere, ${{\bf k}}({\mbox{\upshape lk}\,}v)$ is Gorenstein, and hence its socle vanishes in all degrees except $(d-1)$-st one. Therefore, $\overline{y}=0$, and hence $x_v\cdot y=\phi(\overline{y})=0$ in ${{\bf k}}(\Delta)$. Since this happens for all $v\in [n]$, it follows that $y\in {\mbox{\upshape Soc}\,}{{\bf k}}(\Delta)$.
We now turn to discussing the sphere and manifold $g$-conjectures and the connection between them. As was mentioned in the introduction, the strongest $g$-conjecture for homology spheres and its generalization (due to Kalai) for homology manifolds asserts that if $\Delta$ is a $(d-1)$-dimensional connected orientable homology manifold, then for generically chosen $\omega$ and $\Theta=\{\theta_1, \ldots, \theta_d\}$ in ${{\bf k}}[\Delta]_1$, the map $$\cdot \omega^{d-2i}: \overline{{{\bf k}}(\Delta)}_i \to
\overline{{{\bf k}}(\Delta)}_{d-i}
\mbox{ is an isomorphism for all $i\leq \lfloor d/2\rfloor$}.$$ We refer to this conjecture as the [*strong (sphere or manifold) $g$-conjecture*]{}. If true, it would imply that $\cdot \omega: \overline{{{\bf k}}(\Delta)}_i \to
\overline{{{\bf k}}(\Delta)}_{i+1}
$ is injective for all $i < \lfloor d/2\rfloor$ and is surjective for all $i \geq \lceil d/2\rceil$. We refer to this weaker statement as the [*(sphere or manifold) $g$-conjecture*]{}. (Both, the stronger and the weaker conjectures yield exactly the same combinatorial restrictions on the face numbers of $\Delta$.) Clearly, the manifold $g$-conjecture implies the sphere $g$-conjecture. The following result shows that they are almost equivalent: the strong sphere $g$-conjecture in the middle degree implies the manifold $g$-conjecture in all degrees.
\[connection\] Let $\Delta$ be a $(d-1)$-dimensional connected orientable homology manifold. If for at least $(n-d)$ of the vertices $v$ of $\Delta$ and generically chosen $\omega$ and $\Theta'$ in ${{\bf k}}[{\mbox{\upshape lk}\,}v]_1$, the map $\cdot \omega: {{\bf k}}({\mbox{\upshape lk}\,}v)_{\lfloor (d-1)/2\rfloor} \to
{{\bf k}}({\mbox{\upshape lk}\,}v)_{\lfloor(d-1)/2\rfloor+1}$ is surjective, then $\Delta$ satisfies the manifold $g$-conjecture.
The condition on the links implies by [@Sw Theorem 4.26] that for generic choices of $\omega$ and $\Theta$ in ${{\bf k}}[\Delta]_1$, the map $\cdot \omega: {{\bf k}}(\Delta)_{\lceil d/2\rceil}
\to {{\bf k}}(\Delta)_{\lceil d/2\rceil+1}$ is surjective. Hence the map $\cdot \omega: \overline{{{\bf k}}(\Delta)}_{\lceil d/2\rceil}
\to \overline{{{\bf k}}(\Delta)}_{\lceil d/2\rceil+1}$ is surjective. Thus, $\overline{{{\bf k}}(\Delta)}/(\omega)_i$ vanishes for $i=\lceil d/2\rceil+1$, and hence also for all $i>\lceil d/2\rceil+1$, and we infer that $\cdot \omega: \overline{{{\bf k}}(\Delta)}_{i}
\to \overline{{{\bf k}}(\Delta)}_{i+1}$ is surjective for $i\geq \lceil d/2\rceil$. This in turn yields that the dual map $\cdot\omega: {\mbox{\upshape Hom}\,}_{{\bf k}}(\overline{{{\bf k}}(\Delta)}_{i+1}, {{\bf k}}) \to
{\mbox{\upshape Hom}\,}_{{\bf k}}(\overline{{{\bf k}}(\Delta)}_{i}, {{\bf k}})$ is injective for all $i\geq \lceil d/2\rceil$. Since $\overline{{{\bf k}}(\Delta)}$ is Gorenstein, ${\mbox{\upshape Hom}\,}_{{\bf k}}(\overline{{{\bf k}}(\Delta)}_{i}, {{\bf k}})$ is naturally isomorphic to $\overline{{{\bf k}}(\Delta)}_{d-i}$ (see Theorems I.12.5 and I.12.10 in [@St96]). Therefore, $\cdot \omega: \overline{{{\bf k}}(\Delta)}_{j}
\to \overline{{{\bf k}}(\Delta)}_{j+1}$ is injective for $j<\lfloor d/2 \rfloor$.
Theorem \[connection\] combined with Stanley’s $g$-theorem for polytopes [@St80], implies that every $(d-1)$-dimensional connected orientable $\Q$-homology manifold all of whose vertex links are polytopal spheres satisfies the manifold $g$-conjecture.
A special case of the $g$-theorem
=================================
In this section we prove Theorem \[g-thm\]. As in the proof of Theorem \[Gorenstein\] we will rely on Theorem \[link\] and notation introduced there. Since the set of all $(\omega, \Theta)$ for which $\cdot \omega^{d-2}: \overline{{{\bf k}}(\Delta)}_1 \to
\overline{{{\bf k}}(\Delta)}_{d-1}$ is an isomorphism, is a Zariski open set (see [@Sw Section 4]), it is enough to find one $\omega$ that does the job. Surprisingly, the $\omega$ we find is “very non-generic": as we will see, $w=x_v$ where $v\in \tau$ does the job.
[*Proof of Theorem \[g-thm\]: *]{} We use induction on $d$ starting with $d=3$. In this case, the face $\tau$ is simply a vertex, say $v$. Let $\sigma=\{v=v_1, v_2, v_3\}$ be a facet containing $v$. Let $\Theta$ be a generic l.s.o.p. for ${{\bf k}}[\Delta]$, and as in Section 3 assume that $\theta_i=x_{v_i}+\sum_{j\notin\sigma}\theta_{i,j}x_j$. Since $\dim_{{\bf k}}\overline{{{\bf k}}(\Delta)}_1=
\dim_{{\bf k}}\overline{{{\bf k}}(\Delta)}_2$ (see Eq. (\[symmetry\])) and since $\dim{\mbox{\upshape Soc}\,}{{\bf k}}(\Delta)_1=0$ (e.g. by Theorem \[S\]), we will be done if we check that the map $\cdot x_{v}: {{\bf k}}(\Delta)_1 \to \overline{{{\bf k}}(\Delta)}_2$ is injective. And indeed, if $x_{v}\cdot z \in {\mbox{\upshape Soc}\,}{{\bf k}}(\Delta)_2$ for some $z\in {{\bf k}}(\Delta)_1$, then by the isomorphism of Theorem \[link\], $\overline{z}=\phi^{-1}(x_{v}\cdot z)$ is in ${\mbox{\upshape Soc}\,}{{\bf k}}({\mbox{\upshape lk}\,}v)_1$. But the later space is the zero space, so $\overline{z}=0$ in ${{\bf k}}({\mbox{\upshape lk}\,}v)_1$, that is, $\overline{z}\in {\mbox{\upshape Span}\,}(\theta'_2, \theta'_3)$. Since ${\mbox{\upshape lk}\,}v$ contains all the vertices of $\Delta$ except $v$, it follows that $\theta'_2=\theta_2$, $\theta'_3=\theta_3$, and $z-\overline{z}$ is a multiple of $\theta_1$, and hence that $z=0$ in ${{\bf k}}(\Delta)$.
The proof of the induction step goes along the same lines: let $\tau=\{v_1, \ldots, v_{d-2}\}$, let $\sigma=\tau\cup \{v_{d-1}, v_d\}$ be any facet containing $\tau$, and let $\Theta$ be a generic ls.o.p. for ${{\bf k}}[\Delta]$. We show that for $v\in\tau$, say $v=v_1$, the map $\cdot x_{v}^{d-2}: {{\bf k}}(\Delta)_1 \to
\overline{{{\bf k}}(\Delta)}_{d-1}$ is an injection. If $x_{v}^{d-2}\cdot z \in {\mbox{\upshape Soc}\,}{{\bf k}}(\Delta)_{d-1}$ for some $z\in {{\bf k}}(\Delta)_1$, then by Theorem \[link\], $$x_{v}^{d-3}\cdot \overline{z}=
(-\theta'_1)^{d-3}\cdot \overline{z} \in {\mbox{\upshape Soc}\,}{{\bf k}}({\mbox{\upshape lk}\,}v)_{d-2}.$$ However, ${\mbox{\upshape lk}\,}v$ is a $(d-2)$-dimensional homology sphere satisfying the same assumptions as $\Delta$: the star of the face $\tau-\{v\}$ contains all the vertices of ${\mbox{\upshape lk}\,}v$. Hence by inductive hypothesis applied to ${\mbox{\upshape lk}\,}v$ with $\omega=-\theta'_1$ and $\Theta'=(\theta'_2,\ldots, \theta'_d)$, the map $\cdot (-\theta'_1)^{d-3}:
{{\bf k}}({\mbox{\upshape lk}\,}v)_1 \to \overline{{{\bf k}}({\mbox{\upshape lk}\,}v)}_{d-2}$ is an injection. Thus $\overline{z}=0$ in ${{\bf k}}({\mbox{\upshape lk}\,}v)_1$, and so $z=0$ in ${{\bf k}}(\Delta)_1$.
[**Remark**]{} Barnette’s lower bound theorem [@Ka87] asserts that for every $d$ and $n$, the $f$-vector of any $(d-1)$-dimensional connected triangulated manifold with $n$ vertices is minimized componentwise by the $f$-vector of (the boundary of) a stacked $d$-polytope with $n$ vertices, $S(d,n)$. There is a conjectural algebraic strengthening of this result based on Kalai’s notion of [*algebraic shifting*]{}, an operation that was introduced in the mid-eighties as a tool for studying $f$-numbers of simplicial complexes (see e.g. [@BjKa] and a survey paper [@Ka02]). This conjecture posits that the algebraic shifting of every connected orientable $(d-1)$-dimensional manifold on $n$ vertices contains as a subset $\Delta(S(d,n))$ — the algebraic shifting of $S(d,n)$. The complex $\Delta(S(d,n))$ was recently computed by Nevo [@Nevo Example 2.18] and independently by Murai [@Mur07]. Using their result, together with standard facts on rev-lex generic initial ideals [@Eis Prop. 15.12] and an algebraic definition of Buchsbaum complexes [@Sch Def. and Thm. 3.1(ii)], it is easy to show that in the case of symmetric algebraic shifting, a $(d-1)$-dimensional connected orientable $\Q$-homology manifold $\Delta$ satisfies the above conjecture if and only if for generically chosen $\theta_1,\ldots, \theta_d$ and $\omega$, the map $\cdot \omega^{d-2}: \overline{{{\bf k}}(\Delta)}_1 \to
\overline{{{\bf k}}(\Delta)}_{d-1}$ is a $\Q$-isomorphism. In particular, it follows that this conjecture holds for every $(d-1)$-dimensional orientable $\Q$-homology manifold satisfying assumptions of Theorem \[g-thm\].
[999]{} L. Billera and C. Lee, A proof of the sufficiency of McMullen’s conditions for $f$-vectors of simplical convex polytopes, J. Comb. Theory Series A 31 (1981), 237–255.
A. Björner and G. Kalai, An extended Euler-Poincaré theorem, Acta Math. 161 (1988), 279–303.
D. Eisenbud, [*Commutative algebra with a view toward algebraic geometry*]{}, Graduate Texts in Mathematics, 150, Springer-Verlag, New York, 1995.
H.-G. Gräbe, The canonical module of a Stanley-Reisner ring, J. Algebra 86 (1984), 272–281.
G. Kalai, Rigidity and the lower bound theorem I, Invent. Math. 88 (1987), 125–151.
G. Kalai, Algebraic shifting, in [*Computational commutative algebra and combinatorics*]{} (Osaka, 1999), 121–163, Math. Soc. Japan, Tokyo, 2002.
V. Klee, A combinatorial analogue of Poincaré’s duality theorem, Canadian J. Math. 16 (1964), 517–531.
S. Murai, Algebraic shifting of cyclic polytopes and stacked polytopes, Discrete Math. 307 (2007), 1707–1721.
S. Murai, Algebraic shifting of strongly edge decomposable spheres, arXiv:0709.4518.
E. Nevo, Algebraic shifting and $f$-vector theory, Ph.D. thesis, Hebrew University, 2007.
I. Novik, Upper bound theorems for homology manifolds, Israel J. Math. 108 (1998), 45–82.
I. Novik and E. Swartz, Socles of Buchsbaum modules, complexes and posets, arXiv:0711.0783.
P. Schenzel, On the number of faces of simplicial complexes and the purity of Frobenius, Math. Z. 178 (1981), 125–142.
R. Stanley, The number of faces of a simplicial convex polytope, Adv. in Math. 35 (1980), 236–238.
R. Stanley, [*Combinatorics and Commutative Algebra*]{}, Boston Basel Berlin: Birkh[ä]{}user, 1996.
E. Swartz, Face enumeration - from spheres to manifolds, J. Eur. Math. Soc., to appear, math.CO/0709.3998.
[^1]: Research partially supported by Alfred P. Sloan Research Fellowship and NSF grant DMS-0500748
[^2]: Research partially supported by NSF grant DMS-0600502
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate charge accretion in vicinity of a slowly rotating compact star with a non-stationary electromagnetic field. Exact solutions to the general relativistic Maxwell equations are obtained for a star formed of a highly degenerate plasma with a gravitational field given by the linearized Kerr metric. These solutions are used to formulate and then to study numerically the equations of motion for a charged particle in star’s vicinity using the gravitoelectromagnetic force law. The analysis shows that close to the star charge accretion does not always remain ordered. It is found that the magnetic field plays the dominant role in the onset of chaos near the star’s surface.'
author:
- |
BABUR M. MIRZA\
Department of Mathematics\
Quaid-i-Azam University, Islamabad. 45320. Pakistan.
title: 'CHAOTIC ACCRETION IN A NON-STATIONARY ELECTROMAGNETIC FIELD OF A SLOWLY ROTATING COMPACT STAR '
---
*Keywords:* Plasma accretion; compact stars; chaos; chaotic dynamics; magnetic field; relativistic MHD, gravitoelectromagnetic effects
Introduction
============
Accretion is a fundamental process by which stars accumulate matter from their surroundings, it therefore plays an important role in stellar formation and evolution. For normal stars accretion is principally due to the gravitational field of the star. However for compact objects (classed as neutron stars, pulsars, and white dwarfs) accretion is not only due to gravitational but also electromagnetic effects. In vicinity of these stars matter is mostly in the form of a highly conducting electron plasma, whereas the star itself is formed of a degenerate highly conducting quasi-neutral plasma. The compact state of the plasma inside the star is brought about by the intense gravitational field of the star. Exterior to the star gravitational and electromagnetic field produce combined effects on the dynamics of plasma surrounding the star. Here gravitational field has a two-fold effect on charge accretion. Firstly it effects the charge dynamics through the interaction of particle mass with the gravitational field of the star. Secondly the background spacetime also influences the electromagnetic field of the star which in turn effects the particle orbit via charge interaction. In general, under such conditions, the electromagnetic field of a compact star is not stationary but gains in intensity as the star evolves. Observationally the magnetic field intensity for a newly formed compact star to the final stages of a compact star can vary from $10^{4}G$ to $10^{12}G$ and hence varies with time$^{[1-3]}$.
In this paper we investigate the combined effects of gravitational and a non-stationary electromagnetic field on a charge particle dynamics in vicinity of a compact star. The star is considered to be formed of a degenerate highly conducting plasma. The magnetohydrodynamics (MHD) approximation is applied for the plasma forming the star. Using a generalized definition$^{[4,5]}$ of the electromagnetic field tensor for an MHD plasma in a curved spacetime the Maxwell field equations are explicitly formulated for a non-stationary axially symmetric electromagnetic field. Here the linearized Kerr metric is used to describe the gravitational field outside the star with ZAMO (zero angular momentum observers) as the stationary observer$^{[6]}$. The generalized Maxwell equation are then solved for the electric and magnetic field components using the separation of variable technique. These solutions are then used to obtain the equations of motion (geodesic equations) in linearized Kerr spacetime for a charge particle in the non-stationary electromagnetic field of the star. The numerical solutions to these equations show that at large distances from the star particle trajectories are well defined and ordered, however as the particle falls towards the star its orbit cannot be precisely predicted hence exhibits chaos. The solution show that star’s magnetic field plays the dominant role in bringing the charge accretion to a chaotic state.
The layout of the paper is as follows. In the next section generalized Maxwell equations are formulated using the generalized electromagnetic field tensor. These equations are solved in section 3 for a non-stationary axially symmetric electromagnetic field using a separation of variable ansatz for the field components. Then in section 4 equations of motion are formulated and numerically solved for a charged particle lying in the field of the star. In the conclusion (section 5) numerical results of section 4 are interpreted and the main results of the study are summarized. Throughout we have used the gravitational units in which $G=1=c$ unless mentioned otherwise.
Formulation of the Electromagnetic Field Equations
==================================================
General relativistically the gravitational field of a rotating star can be described by the Kerr metric. In the region exterior to a slowly rotating compact star of mass $M$ a linearization of the Kerr metric is given by $$ds^{2}=-e^{2\Psi(r)}dt^{2}-2\omega(r)r^{2}\sin^{2}\theta dtd\varphi+r^{2}\sin^{2}\theta d\varphi^{2}+e^{-2\Psi(r)}dr^{2}+r^{2}d\theta^{2},$$ where we have used the usual Boyer-Lindquist coordinates. Also in expression (1) $e^{2\Psi(r)}$ and $\omega(r)$ are given by the relations $e^{2\Psi
(r)}=(1-2M/r),\omega(r)\equiv d\varphi/dt=-g_{t\varphi}/g_{\varphi\varphi}$, where $\omega(r)$ is the angular velocity of a free falling frame brought into rotation by the frame dragging of the spacetime$^{[7]}$. The four velocity components $u_{r}$ and $u_{\theta\text{ }}$ of a stationary observer ZAMO circling the star at a fixed radial distance vanish, therefore using $u^{\alpha}u_{\alpha}=-1$, the components of the four velocity vector for a ZAMO are: $$u^{\alpha}=e^{-\Psi(r)}(1,0,0,\omega(r)),\quad u_{\alpha}=e^{\Psi
(r)}(-1,0,0,0).$$ For a star formed of highly degenerate plasma the electromagnetic field $(E^{\alpha},B^{\alpha})$ of the star can be determined by the Maxwell equations. In a curved spacetime, such as the Kerr spacetime the general relativistic form of the Maxwell equations is given by $$F_{\alpha\beta,\gamma}+F_{\beta\gamma,\alpha}+F_{\gamma\alpha,\beta}=0,$$$$\left( \sqrt{-g}F^{\alpha\beta}\right) _{,\beta}=4\pi\sqrt{-g}J^{\alpha},$$ where $g$ represents the determinant of the metric tensor $g_{\alpha\beta}$. Here $F_{\alpha\beta}$ is the generalized electromagnetic field tensor for an ideal magnetohydrodynamic fluid given by a unique tensor expression: $$F_{\alpha\beta}=u_{\alpha}E_{\beta}-u_{\beta}E_{\alpha}+\eta_{\alpha
\beta\gamma\delta}u^{\gamma}B^{\delta},$$ and $J^{\alpha}$ is current four vector. In general the current four vector is the sum of two terms corresponding to a convection and to a conduction current: $J^{\alpha}=\epsilon u^{\alpha}+\sigma u_{\beta}F^{\beta\alpha} $, where $\epsilon$ is the proper charge density and $\sigma$ is the conductivity of the plasma. The volume element 4-form $\eta_{\alpha\beta\gamma\delta}$ is equal to $\sqrt{-g}\epsilon_{\alpha\beta\gamma\delta}$ where $\epsilon
_{\alpha\beta\gamma\delta}$ is the Levi -Civita symbol. Since in a compact star the quasi-neutral degenerate plasma is essentially a perfect conductor, we assume that for the region external to the star $J^{\alpha}$ can be approximated to zero (see references \[8\] and \[9\] for a discussion of this often made assumption for a quasi-neutral degenerate plasma).
To determine the non-stationary electromagnetic field of the star we assume that the electric and magnetic field components are dependent not only on the polar coordinates $r$ and $\theta$ but also the time coordinate $t$ in the frame of ZAMO. We then obtain from the Maxwell equations (3) and (4) the following determining equations for the electromagnetic field in the slow rotation approximation$^{[10]}$:$$\partial_{r}(r^{2}\sin\theta u^{t}B_{{}}^{r})+\partial_{\theta}(r^{2}u^{t}\sin\theta B^{\theta})=0,$$$$r^{2}\sin\theta u^{t}\partial_{t}B^{\theta}-\partial_{r}(u_{t}E_{\varphi})=0,$$$$r^{2}\sin\theta u^{t}\partial_{t}B^{r}+u_{t}\partial_{\theta}E_{\varphi}=0,$$$$r^{2}\sin\theta u^{t}\partial_{t}B^{\varphi}+\partial_{r}(u_{t}E_{\theta
}-r^{2}\sin\theta u^{\varphi}B^{r})-\partial_{\theta}(u_{t}E_{r}-r^{2}\sin\theta u^{\varphi}B^{\theta})=0,$$$$\partial_{r}(r^{2}\sin\theta u^{t}E^{r})+\partial_{\theta}(r^{2}u^{t}\sin\theta E^{\theta})=0,$$$$r^{2}\sin\theta u^{t}\partial_{t}E^{\theta}{}-\partial_{r}(u_{t}B_{\varphi
})=0,$$$$r^{2}\sin\theta u^{t}\partial_{t}E^{r}{}+u_{t}\partial_{\theta}B_{\varphi}=0,$$$$r^{2}\sin\theta u_{{}}^{t}\partial_{t}E^{\varphi}+\partial_{r}(u_{t}B_{\theta
}-r^{2}\sin\theta u^{\varphi}E^{r})-\partial_{\theta}(u_{t}B_{r}-r^{2}\sin\theta u^{\varphi}E^{\theta})=0.$$
Solution for the Electromagnetic Field Exterior to the Star
===========================================================
The non-stationary electromagnetic field of the star is now determined by the system of equations (6) to (13). To obtain the time dependent electric and magnetic field components $E^{\alpha}$ and $B^{\alpha}$ we assume the following separation ansatz for the field components: $$E^{r}=\mathcal{A}(r)\mathfrak{a}(t)\frac{\alpha}{\sin^{2}\theta},\qquad
E^{\theta}=\mathcal{B}(r)\mathfrak{b}(t)\frac{\beta}{\sin\theta},\qquad E
_{\substack{^{{}} \\}}^{\varphi}=\mathcal{C}(r)\mathfrak{c}(t)\frac{\gamma
}{\sin^{2}\theta}$$ $$B^{r}=\mathcal{D}(r)\mathfrak{d}(t)\frac{\delta}{\sin^{2}\theta},\qquad
B^{\theta}=\mathcal{E}(r)\mathfrak{e}(t)\frac{\varepsilon}{\sin\theta},\qquad
B_{\substack{^{{}} \\}}^{\varphi}=\mathcal{F}(r)\mathfrak{f}(t)\frac{\lambda
}{\sin^{2}\theta}$$ Substituting into (6) we obtain $$\frac{d}{dr}(r^{2}u^{t}\ \mathcal{D})=0$$ This implies that $\mathcal{D}(r)=const./r^{2}u^{t}$. Also from equation (10) we obtain $\mathcal{A}(r)=const./r^{2}u^{t}$ hence the radial components of the electric and magnetic field reduce as: $$E^{r}(r,\theta,t)=\delta\mathfrak{d}(t)/r^{2}u^{t}\sin^{2}\theta,\quad
B^{r}(r,\theta,t)=\alpha\mathfrak{a}(t)/r^{2}u^{t}\sin^{2}\theta$$ From (13) it therefore follows that $$r^{2}u^{t}\mathcal{C\gamma}\frac{d\mathfrak{c}}{dt}+\frac{d}{dr}(r^{2}u_{t}\mathcal{E}\mathfrak{e}\varepsilon-r^{2}\omega u^{t}\mathcal{A}\mathfrak{a}\alpha)-u_{t}\sin\theta\frac{dB_{r}}{d\theta}=0$$ Comparing coefficients of $\sin\theta$ on both sides the last equation implies that $B_{r}$ must be independent of $\theta$. This is compatible with equation (17) iff $\delta\equiv0$. Therefore it follows that $B^{r}=0$. Similarly from (17) it follows that $E^{r}=0$. We therefore have after the separation of variables, from equation (13), (11), (9), and (7) respectively, the following determining equations $$\frac{1}{\mathfrak{e}}\frac{d\mathfrak{c}}{dt}=\frac{-\varepsilon}{\gamma
r^{2}u^{t}\mathcal{C}}\frac{d}{dr}(u_{t}r^{2}\mathcal{E})=k_{1}$$ $$\frac{1}{\mathfrak{f}}\frac{d\mathfrak{b}}{dt}=\frac{\lambda}{\beta r^{2}u^{t}\mathcal{B}}\frac{d}{dr}(u_{t}r^{2}\mathcal{F})=k_{2}$$ $$\frac{1}{\mathfrak{b}}\frac{d\mathfrak{f}}{dt}=\frac{-\beta}{\lambda
r^{2}u^{t}\mathcal{F}}\frac{d}{dr}(u_{t}r^{2}\mathcal{B})=-k_{3}$$ $$\frac{1}{\mathfrak{c}}\frac{d\mathfrak{e}}{dt}=\frac{\gamma}{\varepsilon
r^{2}u^{t}\mathcal{E}}\frac{d}{dr}(u_{t}r^{2}\mathcal{C})=-k_{4}$$ where $k_{1}$, $k_{2}$, $k_{3}$, and $k_{4}$ are the separation constants.
TIME DEPENDENT SOLUTIONS
------------------------
For the separated equations (19) to (22) we have for the time dependent parts $\mathfrak{b(}t\mathfrak{)}$, $\mathfrak{c(}t\mathfrak{)}$, $\mathfrak{e(}t\mathfrak{)}$, and $\mathfrak{f(}t\mathfrak{)}$ the following set of equations:$$\frac{d\mathfrak{c}}{dt}=k_{1}\mathfrak{e},\quad\frac{d\mathfrak{b}}{dt}=k_{2}\mathfrak{f}$$$$\frac{d\mathfrak{f}}{dt}=-k_{3}\mathfrak{b},\quad\frac{d\mathfrak{e}}{dt}=-k_{4}\mathfrak{c}$$ Differentiating the first of equations (23), and substituting in the second equation of the pair (24) we obtain $$\mathfrak{c}(t)=A_{1}\exp(-i\sqrt{k_{1}k_{4}}t)$$ Substituting back in equation (23) we obtain for $\mathfrak{e}$: $$\mathfrak{e}(t)=\frac{-i\sqrt{k_{1}k_{4}}}{k_{1}}A_{1}\exp(-i\sqrt{k_{1}k_{4}}t)$$ where $A_{1}$ is a constant. In a similar manner we obtain for $\mathfrak{f}(t)$ and $\mathfrak{b}(t)$: $$\mathfrak{b}(t)=A_{2}\exp(-i\sqrt{k_{2}k_{3}}t)$$ $$\mathfrak{f}(t)=\frac{-i\sqrt{k_{2}k_{3}}}{k_{2}}A_{2}\exp(-i\sqrt{k_{2}k_{3}}t)$$
RADIAL SOLUTIONS
----------------
The radial components $\mathcal{C}(r)$, $\mathcal{E}(r)$, $\mathcal{F}(r)$ and $\mathcal{B}(r)$ are obtainable explicitly from equations (19) to (22) as follows. First from the radial part of the separated equation (19) we have for the component $\mathcal{C}(r)$: $$\mathcal{C}(r)=\frac{\varepsilon\sqrt{1-\frac{2M}{r}}}{k_{1}\gamma r^{2}}\frac{d}{dr}(r^{2}\sqrt{1-\frac{2M}{r}}\mathcal{E})$$ Substituting into equation (22) gives an expression for $\mathcal{E}(r)$ $$\frac{d}{dr}[(1-\frac{2M}{r})\frac{d}{dr}(\sqrt{1-\frac{2M}{r}}r^{2}\mathcal{E})]=\frac{k_{1}k_{4}r^{2}\mathcal{E}}{\sqrt{1-\frac{2M}{r}}}$$ This differential equation can be solved explicitly for the component $\mathcal{E}(r)$ to give $$\begin{aligned}
\mathcal{E}(r) & =r^{-3/2}(r-2M)^{-\frac{1}{2}+2M\sqrt{k_{1}k_{4}}}C_{1}\exp(\sqrt{k_{1}k_{4}}r)-\nonumber\\
& r^{-3/2}(r-2M)^{-\frac{1}{2}-2M\sqrt{k_{1}k_{4}}}\frac{C_{2}}{2\sqrt
{k_{1}k_{4}}}\exp(-\sqrt{k_{1}k_{4}}r)\end{aligned}$$ The first part of the above solution is asymptotically unbounded, hence we take only the second part of the solution for $\mathcal{E}(r)$. Putting then for $\mathcal{E}(r)$ in equation (29) we obtain after some simplifications the component $\mathcal{C}(r)$: $$\mathcal{C}(r)=\frac{-\varepsilon C_{2}\sqrt{1-\frac{2M}{r}}}{2\gamma
r^{9/2}\sqrt{k_{1}}k_{4}^{3/2}}(r-2M)^{-\frac{1}{2}-2M\sqrt{k_{1}k_{4}}}\{r(2+\sqrt{k_{1}k_{4}}r)-3M\}\exp(-\sqrt{k_{1}k_{4}}r)$$ In a similar manner we obtain for $\mathcal{F}(r)$ and $\mathcal{B}(r)$: $$\begin{aligned}
\mathcal{F}(r) & =r^{-3/2}(r-2M)^{-\frac{1}{2}+2M\sqrt{k_{2}k_{3}}}C_{3}\exp(\sqrt{k_{2}k_{3}}r)-\nonumber\\
& r^{-3/2}(r-2M)^{-\frac{1}{2}-2M\sqrt{k_{2}k_{3}}}\frac{C_{4}}{2\sqrt
{k_{2}k_{3}}}\exp(-\sqrt{k_{2}k_{3}}r)\end{aligned}$$$$\mathcal{B}(r)=\frac{-\lambda C_{4}\sqrt{1-\frac{2M}{r}}}{2\beta r^{9/2}\sqrt{k_{2}}k_{3}^{3/2}}(r-2M)^{-\frac{1}{2}-2M\sqrt{k_{2}k_{3}}}\{r(2+\sqrt{k_{2}k_{3}}r)-3M\}\exp(-\sqrt{k_{2}k_{3}}r)$$ These solutions are single valued and bounded for $r\rightarrow\infty$.
Hence the radial solutions (31) to (34) together with the time dependent solutions (25) to (28) determine the electromagnetic field of a compact star in the region exterior to the star.
Charge Accretion in the Slow Rotation Approximation
===================================================
A linearization of the geodesic equation for a slowly rotating gravitational source is given by the gravitoelectromagnetic force of magnitude $\mathbf{G}+\mathbf{v}\times\mathbf{H}$ per unit mass where $\mathbf{G}$ and $\mathbf{H} $ are the gravitoelectric (GE) and gravitomagnetic (GM) potentials$^{[11,12]} $. In the slow rotation limit the star can be considered as a homogeneous sphere of mass $M$ and radius $R$. In this case the GE and GM potentials are given by$$\mathbf{G=}-\frac{M}{r^{2}}\hat{\mathbf{r}},\quad\mathbf{H}=-\frac{12}{5}MR^{2}(\mathbf{\Omega.r}\frac{\mathbf{r}}{r^{5}}-\frac{1}{3}\frac{\mathbf{\Omega}}{r^{3}}),$$ where $\mathbf{\Omega}$ is the angular velocity vector of the star. In the presence of an electromagnetic field the force law for a particle with mass $m$ and charge $q$ is given by $$\frac{d^{2}\mathbf{r}}{dt^{2}}=(\mathbf{G}+\mathbf{v}\times\mathbf{H)+}\frac{q}{m}\mathbf{(E+v\times B).}$$ The equations of motion can be studied numerically in Cartesian coordinates by transforming the polar coordinates into the Cartesian coordinates as $x=r\sin\theta\cos\varphi,y=r\sin\theta\sin\varphi,z=r\cos\theta$. The transformation gives for the electric field components $E^{x}=-r\sin\varphi
E^{\varphi},$ $E^{y}=r\cos\varphi E^{\varphi},$ $E^{z}=-r\sin\theta E^{\theta
},$ and similarly for the magnetic field $B^{x}=-r\sin\varphi B^{\varphi},$ $B^{y}=r\cos\varphi B^{\varphi},$ $B^{z}=-r\sin\theta B^{\theta}$. In a two dimensional plane $\theta=\pi/2$ (the equatorial plane of the star), the gravitomagnetic effects are maximum, here we get for the equations of motion for the $(x,y)$ plane: $$\frac{d^{2}x}{dt^{2}}=-g-E^{x}+(H-B^{z})\frac{dy}{dt}=-g+yE^{\varphi
}+(H+yB^{\theta})\frac{dy}{dt}$$$$\frac{d^{2}y}{dt^{2}}=-g+E^{y}-(H-B^{z})\frac{dx}{dt}=-g+xE^{\varphi
}-(H+yB^{\theta})\frac{dx}{dt}\mathbf{,}$$ where from (14) and (15) we have for the real part of $E^{\varphi}$ and $B^{\theta}$ $$\operatorname{Re}E^{\varphi}=E_{0}\frac{\sqrt{1-\frac{2M}{r}}}{r^{9/2}}(r-2M)^{-\frac{1}{2}-2Mk}\{r(2+kr)-3M\}\exp(-kr)\cos kt$$ $$\operatorname{Re}B^{\theta}=B_{0}r^{-3/2}(r-2M)^{-\frac{1}{2}-2Mk}\exp(-kr)\sin kt$$ where $r=\sqrt{x^{2}+y^{2}}$ , $k=\sqrt{k_{1}k_{4}}$, and $E_{0}\equiv-\varepsilon A_{1}C_{2}/2k_{1}k_{4}\sin^{2}\theta$ and $B_{0}\equiv\varepsilon A_{1}C_{2}/2k_{1}\sin\theta$ are constants for fixed value of $\theta$, such as in the equatorial plane of the star. Also $g\equiv
\mid\mathbf{G}\mid=M/(x^{2}+y^{2}),$ and $H=\mid\mathbf{H}\mid=\mu
/(x^{2}+y^{2})^{3/2},$ whereas $\mu=(4/5R)\Omega$, $\Omega$ being the magnitude of the angular velocity vector and $R$ is the radius of the star. Transforming the electromagnetic field components in coordinates $(x,y)$ and using expressions (37), (38) (39), and (40) we plot the trajectories of the particle for various values of the parameters $M$, $\mu$, $E_{0}$, and $B_{0}$ in figures (1) to (4); where we have taken $x(0)=2$, $y(0)=2$, $dx/dt\mid_{t=0}=0$, $dy/dt\mid_{t=0}=0$ and $k=1.$
Conclusions and Summary
=======================
In this paper we have investigated charge accretion around a slowly rotating compact star with a non-stationary electromagnetic field. In the region exterior to the star the electromagnetic field is determined by solving the generalized Maxwell equations (6) to (13). Then to study charge accretion in vicinity of a compact star, the equations of motion for a charged particle in the presence of the non-stationary electromagnetic field are solved numerically. For the external region, where the solutions are stable and convergent, our analysis shows that gravitational as well as electromagnetic field determine the charge particle trajectories. However it is the magnetic force that has the dominant effect on the transition from an ordered state of motion to a chaotic one. Close to the star particle motion is particularly sensitive to a change in the magnetic field of the star. Here an increase in the gravitational field strength (gravitoelectric as well as gravitomagnetic) directly causes a suppression in magnetically induced chaos in particle orbit (Figure 3 and 4). The process however requires a substantial (about $10$ to $100$ times) increase in the gravitoelectric field. In comparison a change in the gravitomagnetic parameter $\mu$ from $2 $ to $10$ units introduces order in particle’s orbit. On the other hand the effects of electric field (Figure 2) on the charged particle motion are more complicated. As the electric field becomes comparable to the magnetic field, the accretion trajectories become precisely determined. Crossing this value, an increase in the electric field causes the charged particle to reverse its direction of motion after which it escapes from falling into the star. Magnetic field effects on charged particle trajectory displayed in figure 1 also indicate that although an increase in the magnetic field strength makes charge accretion chaotic, the general trend is that of particle in-fall due to the dominant gravitational effects. On the other hand in the regions close to the poles of the star, where magnetic field is particularly high, it is expected that chaotic trend in accretion is more dominant provided that the electric field effects on particle’s motion are comparatively weak.
Summarizing the above conclusions, accretion at relatively large distances from the surface of a compact star is more or less ordered due to the dominant effects of gravitational as well as electric fields. However in the star’s vicinity accretion becomes chaotic depending on the magnetic field strength close to surface of the star as compared to the other forces involved, in particular the gravitational and the electric forces.
**Acknowledgements**
I thank Dr. Hamid Saleem for useful comments. I also acknowledge Quaid-i-Azam University Research Fund (URF-2006) for the financial support.
[9]{} N. K. Glendenning, *Compact Stars* (Springer-Verlag, New York, 1997).
A. Treves, L. Maraschi, and M. Abramowicz, *PASP*, **100** 427 (1988). (Reprinted in Treves *et.al*. 1989).
A. Treves, L. Maraschi, and M. Abramowicz (eds.), *Accretion: A Collection of Influential Papers* (World Scientific, New Jersey, 1989).
A. Lichnerowicz, *Relativistic Hydrodynamics and Magnetohydrodynamics* (Benjamin Press, New York, 1967).
G. F. R. Ellis, *Cargese Lectures in Physics* **6,** ed. Schatzman E. (Gordon & Breach, 1973).
B. Punsly, *Astrophys. J.* **498,** 640 (1998).
W. Rindler, *Relativity: Special, General, and Cosmological*, (Oxford University Press, New York, 2001).
J. D. Bekenstein and E. Oron, *Phys. Rev*.* ***D18**, 1809 (1978).
K. Elsässer, *Phys. Rev.* **D62,** 044007 (2000).
B. M. Mirza, *Inter. J. Mod. Phys.* **D14,** 609 (2005).
M. L. Ruggiero and A. Tartaglia, * Nuovo Cim.* **117B,** 743 (2002).
R. T. Jantzen and G.Mac Keiser (eds.), *The Seventh Marcel Grossmann Meeting on General Relativity*, Part A, (World Scientific, New Jersey, 1996).
**Figure Captions;**
Figure 1: Effects of the magnetic field on charge particle trajectories. The magnetic field strength varies from $B_{0}=10,100,1000,$ to $10000$ units, while $M=1$, $\mu=1$, and $E_{0}=1$ in gravitational units.
Figure 2: Effects of the electric field on charge particle trajectories. The electric field strength varies from $E_{0}=1,10,100,$ to $1000$ units, while $M=1$, $\mu=1$, and $B_{0}=10$ in gravitational units.
Figure 3: Effects of the gravitoelectric field on charge particle trajectories. The gravitoelectric field strength varies from $M=1,10,100,$ to $1000$ units, while $\mu=1$, $E_{0}=1$ and $B_{0}=10$ in gravitational units.
Figure 4: Effects of the gravitomagnetic field on charge particle trajectories. The gravitomagnetic field strength varies from $\mu=0.05,1,3,$ to $10$ units, while $M=1$, $E_{0}=1$ and $B_{0}=10$ in gravitational units.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Carlson and Toledo conjectured that if an infinite group $\G$ is the fundamental group of a compact Kähler manifold, then virtually $H^2(\G, \R)\not =0$. We assume that $\G$ admits an unbounded reductive rigid linear representation. This representation necessarily comes from a complex variation of Hodge structure ($\C$-VHS) on the Kähler manifold. We prove the conjecture under some assumption on the $\C$-VHS. We also study some related geometric/topological properties of period domains associated to such $\C$-VHS.'
address:
- 'IMJ, Université Paris-Diderot - Paris VII, CNRS, 175 rue du chevaleret, F-75013 Paris, France'
- 'IECN, Nancy-Université, CNRS, INRIA, Boulevard des Aiguillettes B. P. 239, F-54506 Vandœuvre-lès-Nancy, France'
author:
- Bruno Klingler
- Vincent Koziarz
- Julien Maubon
title: On the second cohomology of Kähler groups
---
Introduction {#intro}
============
The general setting for this paper is the study of topological properties of compact Kähler manifolds, in particular complex smooth projective varieties. The possible homotopy types for these spaces are essentially unknown (cf. [@si4]). Here we restrict ourselves to an a priori simpler question: the study of fundamental groups of compact Kähler manifolds (the so-called Kähler groups).
It is well-known that any finitely presented group ${\Gamma}$ can be realized as the fundamental group of a $4$-dimensional compact real manifold, or even of a symplectic complex compact $3$-fold. A classical result of Serre shows that any finite group can be realized as a projective group (i.e. the fundamental group of a complex smooth projective variety). On the other hand there are many known obstructions for an infinite finitely presented group being Kähler (we refer to [@abckt] for a panorama). Most of them come from Hodge theory (Abelian or not) [*in cohomological degree one*]{}. As a prototype: let $M$ be a compact Kähler manifold with fundamental group ${\Gamma}$. For all $i\in\N$, let $b^i(\G)=\rk H^i(\G,\Z)={\rm dim}_\R\, H^i(\G,\R)={\rm dim}_\C\, H^i(\G,\C)$. Classical Hodge theory shows that $b^1({\Gamma})=b^1(M)$ has to be even. Considering finite étale covers one obtains that $b^1({\Gamma}')$ has to be even for any finite index subgroup ${\Gamma}'$ of ${\Gamma}$. For example, the free group on $n$-generator $F_n$ is never Kähler, $n\geq 1$.
In [@Voisin], Voisin proved that there exist compact Kähler manifolds whose homotopy type cannot be realized by a smooth projective manifold. However, the question whether or not any Kähler group is a projective group is still open.
A very interesting conjecture concerning infinite Kähler groups, due to Carlson and Toledo and publicized by Kollár [@ko], deals with [*cohomology in degree $2$*]{}:
\[conj1\] Let ${\Gamma}$ be an infinite K[ä]{}hler group. Then virtually $H^2(\G,\R)\neq 0$.
We refer to Section \[explanation\] for different interpretations of this conjecture. Notice that it holds trivially for fundamental groups of complex curves. Although Kähler groups are not necessarily residually finite (and thus not necessarily linear) [@to], all known infinite such groups admit a linear representation with [*unbounded image*]{}. We will from now on restrict to this case.
The strongest evidence for Carlson-Toledo’s conjecture is the following theorem. The finite dimensional case is folkloric (we give a proof in Section \[sec2\] for the sake of completeness) and the Hilbert space case is one of the main results of Reznikov in [@Re]:
\[prop2\] Let ${\Gamma}$ be a Kähler group. If ${\Gamma}$ admits a linear representation $\rho:
{\Gamma}{\longrightarrow}G$, with $G$ the linear group ${\rm GL}(V)$ of a finite dimensional complex vector space $V$ or the isometry group of a Hilbert space, satisfying $H^1({\Gamma}, \rho) \not = 0$, then $b^2({\Gamma}) >0$.
Recall that a finitely generated group is said [*schematically rigid*]{} if all linear representations of $\G$ are [*rigid*]{}, namely if for any $n \in {{\mathbb N}}$ and any representation $\rho: {\Gamma}{\longrightarrow}\GL(n, {{\mathbb C}})$ one has $H^1({\Gamma},
\Ad \rho)=0$. As a corollary to Theorem \[prop2\] one has:
Let ${\Gamma}$ be an infinite Kähler group. If ${\Gamma}$ does not satisfy Carlson-Toledo’s conjecture then necessarily:
- ${\Gamma}$ has Kazhdan’s property $(T)$.
- ${\Gamma}$ is schematically rigid.
Let $M$ be a compact Kähler manifold and assume that ${\Gamma}=\pi_1(M)$ is schematically rigid and admits a reductive linear representation $\rho: {\Gamma}{\longrightarrow}\GL(n, \C)$. By Simpson’s theory [@Sim], the representation $\rho$ is then the monodromy of a (polarized) complex variation of Hodge structure ($\C$-VHS for short) over $M$, see Section \[var\]. The first result of this paper states that Conjecture \[conj1\] holds under some assumption on the variation of Hodge structure:
\[cartoun\] Let $M$ be a compact Kähler manifold. Assume that there exists an irreducible rigid unbounded reductive representation $\rho:\pi_1(M)\fd
\GL(n,\C)$ for some $n>0$ and that the variation of Hodge structure $(E,\t)=(\oplus_{i=0}^k E^i,\oplus_{i=0}^{k-1}\t_i)$ over $M$ which induces $\rho$ satisfies $\rk E^i=1$ for some $0<i<k$. Then $H^2(\pi_1(M),\R)\not=0$.
The theorem holds under the more general assumption that the rank of $\t_i:T_M\fd\Hom(E^i,E^{i+1})$ is $1$ for some $0\leq i\leq k-1$ (see the proof in Section \[proofofcartoun\]).
Note in particular that this result (and the discussion in section \[conj\]) implies that the Carlson-Toledo conjecture holds whenever the image of the rigid representation sits in some ${\rm U}(p,1)$.\
The fact that the representation $\rho$ comes from a $\C$-VHS implies in particular that $\rho(\G)$ sits in some $\U(p,q)$, $p+q=n$. Moreover there is a period domain $D$, which is an open orbit of the group $\PU(p,q)$ in a manifold of flags of $\C^n$, and a period map, which is a $\rho$-equivariant holomorphic superhorizontal map from the universal cover of $M$ to $D$. We call such period domain and map the $\PU(p,q)$ period domain and map associated to the $\C$-VHS, see Section \[var\] for details. It turns out that the Carlson-Toledo conjecture can be related to the question of whether the period map kills the second homotopy group $\pi_2(M)$ of $M$ or not. Indeed if the answer is yes, then the conjecture is true. Since the period map is superhorizontal, a first thing to study is the existence of non homotopically trivial superhorizontal 2-spheres in the period domain $D$. In Section \[superhori\], using Gromov’s $h$-principle we prove the following result, which is of independent geometric interest:
\[horiz\] Let $(E,\t)$ be a $\C$-VHS over a Kähler manifold and let $D$ be the $\PU(p,q)$ period domain associated to $E$. Suppose $(E,\t)$ does not satisfy the hypotheses of Theorem \[cartoun\]. Then $\pi_2(D)$ can be generated by superhorizontal 2-spheres.
It is also customary to consider “smaller” period domains that one can associate to a $\C$-VHS: the real Zariski-closure $G_0$ of $\rho({\Gamma})$ in $\GL(n,\C)$ is a group of Hodge type by [@Sim] and the period map can be chosen to land in a period domain which is an open $G_0$-orbit in some flag manifold $G/Q$, where $G$ is the complex Zariski-closure of $\rho(\G)$ and $Q$ a parabolic subgroup of $G$, see Sections \[perdo\] and \[var\]. We will comment on the generalization of Theorem \[horiz\] to such period domains at the end of the paper.
[*Acknowledgements.*]{} We wish to thank the referee for his helpful comments and for having suggested many improvements in the exposition.
What does Conjecture \[conj1\] mean? {#explanation}
====================================
In terms of group extensions
----------------------------
Recall that for a group $\G$ and an Abelian group $A$ with trivial $\Gamma$-module structure, the group $H^2({\Gamma}, A)$ classifies the central $A$-extensions $$0 {\longrightarrow}A {\longrightarrow}\tilde{\Gamma} {\longrightarrow}\Gamma {\longrightarrow}1$$ of ${\Gamma}$. As $b^2({\Gamma}) = \textnormal{rk}\, H^2({\Gamma}, {{\mathbb Z}})$, Conjecture \[conj1\] means that any infinite Kähler group admits (after maybe passing to a finite index subgroup) a non-trivial central ${{\mathbb Z}}$-extension (which does not trivialize when restricted to any finite index subgroup).
In terms of classical topology
------------------------------
For any group ${\Gamma}$ the universal coefficients exact sequence yields the isomorphism $$H^2({\Gamma}, {{\mathbb R}}) = \Hom_{{\mathbb R}}(H_2({\Gamma}, {{\mathbb R}}), {{\mathbb R}}).$$ In particular it is equivalent to show $b_2({\Gamma}) >0$ or $b^2({\Gamma})>0$.
For any reasonable topological space $M$ with fundamental group ${\Gamma}$ the universal cover $p:\tilde{M}\fd M$ is a principal ${\Gamma}$-bundle over $M$. Thus it defines (uniquely in the homotopy category) a morphism $c: M {\longrightarrow}B{\Gamma}$ from $M$ to the classifying Eilenberg-MacLane space $B{\Gamma}=
K({\Gamma},1)$. The induced morphism $c_{\star}:H_\star(M, \R) {\longrightarrow}H_\star({\Gamma}, \R)$ is easily seen to be an isomorphism in degree $1$ and an epimorphism in degree $2$: $$H_2(M, \R) {\relbar\joinrel\twoheadrightarrow}H_2({\Gamma}, \R).$$ Dually: $$H^2({\Gamma}, \R) {\lhook\joinrel\longrightarrow}H^2(M, \R).$$
How can we characterize the quotient $H_2({\Gamma}, \R)$ of $H_2(M, \R)$? In fact this quotient first appeared in Hopf’s work on the Hurewicz morphism comparing homotopy and homology:
\[Hopf\] \[Hopf\] Let $N$ be a paracompact topological space. Let $c: N {\longrightarrow}B\pi_1(N)$ be the classifying morphism and $h: \pi_\star(N) {\longrightarrow}H_\star(N, {{\mathbb Z}})$ the classical Hurewicz morphism. Then the sequence of Abelian groups $$\pi_2(N) \stackrel{h}{{\longrightarrow}} H_2(N, {{\mathbb Z}}) \stackrel{c_\star}{{\longrightarrow}} H_2(\pi_1(N),
{{\mathbb Z}}) {\longrightarrow}0$$ is exact.
Cohomologically:
\[corol3\] Let $N$ be a paracompact topological space and $\pi_2(N)
\otimes_{{\mathbb Z}}\R \stackrel{h}{{\longrightarrow}} H_2(N, \R)$ the Hurewicz morphism. Then: $$H^2(\pi_1(N), \R) = \{[\om] \in H^2(N, \R)\; |\; \forall \phi:S^2
{\longrightarrow}N, \; <[\om], \phi_\star [S^2]>=0\} \subset H^2(N, \R)\;\;,$$ where $<\cdot, \cdot>: H^2(N, \R) \times H_2(N, \R) {\longrightarrow}\R$ is the natural non-degenerate pairing between homology and cohomology.
Nowadays Theorem \[Hopf\] is a direct application of the Leray-Cartan spectral sequence.
Since $\pi_2(M)$ is nothing else than $H_2(\tilde{M}, {{\mathbb Z}})$, Carlson-Toledo’s conjecture can be restated:
\[conj2\] Let $M$ be a compact Kähler manifold with infinite fundamental group. There exists a finite étale cover $M'$ of $M$ such that the natural morphism $H_2(\tilde{M}, {{\mathbb R}}) \fd H_2(M', {{\mathbb R}})$ is not surjective, or, equivalently, such that the natural morphism $H^2(M', {{\mathbb R}}) \fd H^2(\tilde{M}, {{\mathbb R}})$ is not injective.
Note that for any compact manifold $M$ and any finite étale cover $M'$ of $M$ the arrow $H^2(M, {{\mathbb R}}) {\longrightarrow}H^2(M', {{\mathbb R}})$ is injective by the projection formula.
In terms of ${{\mathbb C}}^\star$-bundles
-----------------------------------------
Recall that for a reasonable topological space $M$ the group $H^2(M,
{{\mathbb Z}})$ canonically identifies with the group $\mathcal{L}(M)$ of principal ${{\mathbb C}}^\star$-bundles: on the one hand $H^2(M, {{\mathbb Z}}) = [M, K({{\mathbb Z}}, 2)]$, on the other hand $\mathcal{L}(M)= [M, B{{\mathbb C}}^\star]$. But both $K({{\mathbb Z}}, 2)$ and $B{{\mathbb C}}^\star$ have as canonical model the infinite projective space ${{\mathbb C}}{\mathbf P}^{\infty}$. Thus Conjecture \[conj1\] states that any infinite Kähler group ${\Gamma}$ admits a finite index subgroup ${\Gamma}'$ whose classifying space $B{\Gamma}'$ supports a non-trivial principal ${{\mathbb C}}^\star$-bundle. Let $M$ be a compact Kähler manifold with infinite fundamental group ${\Gamma}$. As $\tilde{M}$ is the homotopy fiber of $M {\longrightarrow}B\Gamma$, Conjecture \[conj2\] says there exists a finite étale cover $M'$ of $M$ and a non-trivial principal ${{\mathbb C}}^\star$-bundle on $M'$ whose pull-back to $\tilde{M}$ becomes trivial. In these statements “non-trivial” means “with non-trivial [*rational*]{} first Chern class”.
Equivalence
-----------
The equivalence of these three points of view is clear. Given $M$ and a line bundle $L$ on $M$ with associated principal ${{\mathbb C}}^\star$-bundle $L^0$, one can consider the long homotopy exact sequence for the fibration $L^0 {\longrightarrow}M$: $$\label{long}
\cdots {\longrightarrow}\pi_2(L^0) {\longrightarrow}\pi_2(M) \stackrel{\partial_L}{{\longrightarrow}}\pi_1({{\mathbb C}}^\star) = {{\mathbb Z}}{\longrightarrow}\pi_1(L^0) {\longrightarrow}\pi_1(M) {\longrightarrow}1 \;\;.$$ The boundary map $\partial_L: \pi_2(M) {\longrightarrow}{{\mathbb Z}}$ is just the first Chern class map of $L$ restricted to $\pi_2(M)$: if $[\alpha] \in \pi_2(M)$ is represented by $\alpha: S^2 {\longrightarrow}M$ then $\partial_L([\alpha])=
<\alpha^\star (c_1(L)), [S^2]> \in {{\mathbb Z}}$. As $H_2(\tilde M, {{\mathbb Z}}) = \pi_2(M)$ and $H^2(\tilde{M}, {{\mathbb R}})$ is dual to $H_2(\tilde M, {{\mathbb R}})$ the principal ${{\mathbb C}}^\star$-bundle $p^\star L^0$ is trivial if and only if $\partial_L:
\pi_2(M) {\longrightarrow}{{\mathbb R}}$ is zero. Then the long exact sequence (\[long\]) gives the short exact sequence: $$0 {\longrightarrow}{{\mathbb Z}}{\longrightarrow}\pi_1(L^0) {\longrightarrow}\pi_1(M) {\longrightarrow}1 \;\;,$$ the element of $H^2(\pi_1(M),\Z)$ classifying this extension being the first Chern class of $L$.
Carlson-Toledo’s conjecture: the non-rigid case {#sec2}
===============================================
The reductive case
------------------
The simplest instance of Theorem \[prop2\] is the following:
\[lef\] Let ${\Gamma}$ be a Kähler group. If $b^1({\Gamma})>0$ then $b^2({\Gamma})>0$.
Essentially the same proof generalizes to:
\[l1\] [([@Re])]{}. Let ${\Gamma}$ be a Kähler group. If ${\Gamma}$ admits a linear representation $\rho:
{\Gamma}{\longrightarrow}G$ with Zariski-dense image, where $G$ denotes a linear [*reductive*]{} complex algebraic group, and satisfying $H^1({\Gamma}, \rho) \not = 0$, then $b^2({\Gamma}) >0$.
Let $M$ be a compact Kähler manifold whose fundamental group is isomorphic to ${\Gamma}$. Let $E_{\rho}
{\longrightarrow}M$ be the complex local system on $M$ associated with $\rho$. Consider the following commutative diagram: $$\xymatrix{
H^{1}({\Gamma}, \rho) \times H^{1}({\Gamma}, \rho^{\star}) \ar[r] \ar@{=}[d] &H^{2}({\Gamma},
{{\mathbb C}}) \ar[r] \ar@{^{(}->}[d] &{{\mathbb C}}\ar@{=}[d]\\
H^{1}_{DR}(M, E_{\rho}) \times H^{1}_{DR}(M, E_{\rho}^{\star}) \ar[r] &H^{2}_{DR}(M, {{\mathbb C}})
\ar[r] &{{\mathbb C}}\\
}$$ where $\rho^\star$ denotes the contragredient representation of $\rho$ and the second row of the diagram is given by $$(\alpha, \beta) {\longrightarrow}\alpha \wedge \beta {\longrightarrow}\int_{M} \alpha \wedge \beta\wedge \omega_{M}^{{\rm dim}\,M-1}\;\;.$$ By the hard Lefschetz theorem generalized by Simpson to the case of reductive representations [@Sim Lemmas 2.5, 2.6] this second row defines a non-degenerate bilinear form. This implies the result.
The general linear case {#l2}
-----------------------
\[reduc\] Let ${\Gamma}$ be a Kähler group. If ${\Gamma}$ admits a linear representation $\rho: {\Gamma}{\longrightarrow}{\rm GL}(n, {{\mathbb C}})$ satisfying $H^1({\Gamma}, \rho) \not = 0$ then $b^2({\Gamma}) >0$.
Let $G= {{R}}\ltimes {{U}}$ be the Levi decomposition of the complex Zariski-closure $G$ of $\rho({\Gamma})$ in ${\rm{GL}}(n,{{\mathbb C}})$, where ${{U}}$ denotes the unipotent radical of $G$ and ${{R}}$ a Levi factor lifting the reductive quotient $G
/{{U}}$. Let $l(\rho) : {\Gamma}{\longrightarrow}{{R}}$ be the quotient representation of $\rho$. As $\rho$ has Zariski-dense image in $G$, the representation $l(\rho)$ has Zariski-dense image in the reductive group ${{R}}$.
If ${{U}}$ is non-trivial let $${{U}}= {{U}}_1
\supset \cdots \supset {{U}}_i \supset \cdots \supset {{U}}_r$$ be the central descending series for ${{U}}$ (i.e. ${{U}}_1={{U}}$, ${{U}}_{i+1}= \overline{[{{U}}, {{U}}_{i}]}$ is the Zariski-closure of the subgroup of ${{U}}$ generated by the commutators of ${{U}}$ and ${{U}}_{i}$). The ${{U}}_i$’s are unipotent, the quotients ${{A}}_i:={{U}}_i/{{U}}_{i+1}$ are non-trivial of additive type. The action of ${{R}}$ on ${{U}}$ by automorphisms preserves the central descending series. Let us define $G_i={{R}}\ltimes {{A}}_i$ and $\tau_i \in
H^{1}(G_i, {{A}}_i)$ be the canonical class of the semi-direct product $G_i$.
Let $\rho_1 : {\Gamma}{\longrightarrow}{{G}}_1$ be the quotient representation of $\rho$. If the class $\tau_{1, {\Gamma}}= \rho_{1}^{\star}(\tau_{1}) \in H^{1}({\Gamma},
({{A}}_1)_{l(\rho)})$ does not vanish, the conclusion $H^{2}({\Gamma},
{{\mathbb C}})\not = 0$ follows from Lemma \[l1\] applied to the reductive representation $({{A}}_1)_{l(\rho)}$. If $\tau_{1, {\Gamma}}=0$ then up to conjugacy by an element in ${{G}}$ one can assume that the image $\rho_{1}({\Gamma})$ is contained in ${{R}}$. Thus the image $\rho({\Gamma})$ is contained in ${{R}}\ltimes {{U}}_2$. The quotient representation $\rho_2: {\Gamma}{\longrightarrow}{{G}}_2$ of $\rho$ is then well-defined, thus also the class $\tau_{2,{\Gamma}}= \rho_{2}^{\star}(\tau_{2}) \in H^{1}({\Gamma},
({{A}}_2)_{l(\rho)})$.
By induction, for the first step $i$ where the class $\tau_{i,{\Gamma}}$ does not vanish one deduces $H^{2}({\Gamma},
{{\mathbb C}})\not = 0$ from Lemma \[l1\] applied to the reductive representation $({{A}}_i)_{l(\rho)}$. Thus one can assume that all the classes $\tau_{i,{\Gamma}}$, $1 \leq i \leq r$, vanish. Hence $\rho({\Gamma}) \subset {{R}}$, ${{U}}=\{1\}$ and ${{G}}={{R}}$. But then $\rho$ is reductive and satisfies by hypothesis $H^1({\Gamma}, \rho) \not = 0$, thus $H^{2}({\Gamma},{{\mathbb C}})\not = 0$ by Lemma \[l1\] applied to $\rho$.
The Hilbert case
----------------
The proof of Lemma \[lef\] generalizes also to the case where $\rho$ is a unitary representation in some Hilbert space. This is the main result of Reznikov in [@Re], to which we refer for a proof.
Carlson-Toledo conjecture and variations of Hodge structure {#back}
===========================================================
Period domains {#perdo}
--------------
We recall the necessary background about period domains. We refer to [@GS] for a detailed account of the subject.
Let $G$ be a connected complex semi-simple Lie group and $G_0$ a noncompact real form of $G$. We assume that $G_0$ is a group of Hodge type, namely that it has a compact maximal Abelian subgroup $T_0$ (see [@Sim 4.4.4]). Let $Q$ be a parabolic subgroup of $G$ which contains $T_0$ and such that $V_0:=G_0\cap Q$ is compact.
We choose a maximal compact subgroup $K_0$ of $G_0$ which contains $V_0$, and a maximal compact subgroup $G_u$ of $G$ which contains $K_0$. The Lie algebras of $G$, $G_0$ and $Q$ will be denoted by $\gg$, $\gg_0$ and $\qg$. Let $B$ be the Killing form on $\gg$. The Lie algebra $\gg_u$ of $G_u$ is a real form of $\gg$ and we denote complex conjugation with respect to $\gg_u$ by $\tau$. Let $\ng$ be the maximal nilpotent ideal of $\qg$ and $\vg=\qg\cap\tau(\qg)$. Then $\qg=\vg\oplus\ng$ and $\gg=\vg\oplus\ng\oplus\tau(\ng)$. The Lie algebra of $V_0$ is $\vg_0=\gg_0\cap\qg=\gg_0\cap\qg\cap\tau(\qg)=\gg_0\cap\vg$. We shall denote the Lie algebra of $K_0$ by $\kg_0$ and its complexification by $\kg$. We have $\vg\subset \kg$.
The complex analytic quotient space $D_u=G/Q$ is called a [*flag manifold*]{}. The group $G$ is a holomorphic principal $Q$-bundle over $D_u$ and the holomorphic tangent bundle of $D_u$ is the bundle associated to this principal bundle by the adjoint action of $Q$ on $\gg/\qg\simeq\tau(\ng)$. The $G_u$-orbit of $eQ\in D_u=G/Q$ is open because the real span of $\gg_u$ and $\qg$ is all of $\gg$ and it is closed, since $G_u$ is compact. Hence $G_u$ acts transitively on $D_u$ with isotropy group $V_u=G_u\cap Q=V_0$ and we identify the quotient space $G_u/V_u$ with $D_u$.
Open orbits of noncompact real forms of $G$ in the flag manifold $G/Q$ are called [*flag domains*]{} ([@FHW]). With our assumptions, this is the case of the $G_0$-orbit $D$ of $eQ$, which can be identified with the quotient space $G_0/V_0$. Because $G_0$ is of Hodge type and $V_0=G_0\cap Q$ is compact, we will call the homogeneous complex manifold $D$ a [*period domain*]{}. We say that $D$ is “dual” to $D_u$.
Note that any parabolic subgroup $Q'$ of $G$ such that $V_0=G_0\cap Q'$ induces a complex structure on $D$.
As a $C^\infty$ vector bundle, the complex tangent bundle $T_{D_u}$ (resp. $T_{D}$) to $D_u$ (resp. $D$) is associated to the principal $V_0$-bundle $G_u$ (resp. $G_0$) by the adjoint representation of $V_0$ on $\gg/\qg$. We endow the tangent space of $D_u$ at $eV_0$, which is naturally isomorphic to $\tau(\ng)$, with the ${\rm Ad} V_0$-invariant inner product $(\xi,\zeta)\mapsto-B(\xi,\tau(\zeta))$. By $G_u$-translation, this inner product gives rise to a $G_u$-invariant Hermitian metric $g_{D_u}$ on $D_u$ and by $G_0$-translation, to a $G_0$-invariant Hermitian metric $g_{D}$ on $D$.
Let $X=G_0/K_0$ be the symmetric space associated to $G_0$ and $X_u=G_u/K_0$ its compact dual. Then, we have natural fibrations $\pi_u:D_u=G_u/V_u\fd G_u/K_0=X_u$ and $\pi:D=G_0/V_0\fd G_0/K_0=X$. The fibers of $\pi_u$, resp. $\pi$, are complex submanifolds of $D_u$, resp. $D$, all isomorphic to the central fiber $K/K\cap Q=K_0/V_0$ (in both cases, although $\pi\not={\pi_u}_{|D}$) and the fibration is $G_u$-invariant, resp. $G_0$-invariant. The vectors in $T_{D}$ which are tangent to the fibers of $\pi$ form a $C^\infty$ subbundle $T^v_{D}$. A vector $\xi$ in the fiber of $T_{D}$ over $eV$, which is identified with $\gg/\qg$, is tangent to the fiber of $\pi$ if and only if $\xi\in\kg/\kg\cap\qg$. It follows that $T_{D}^v$ is associated to the $V_0$-bundle $G_0$ by the adjoint representation of $V_0$ on $\kg/\kg\cap\qg$.
The maximal compact subgroup $K_0$ of $G_0$ determines a Cartan decomposition $\gg=\kg\oplus\pg$. The adjoint action of $V_0$ on $\pg/\pg\cap\qg$ associates a $C^\infty$ vector bundle $T^h_{D}$ to the $V_0$-principal bundle $G_0$ and since $\gg/\qg=\kg/\kg\cap\qg\oplus\pg/\pg\cap\qg$, $T_{D}=T^v_{D}\oplus T^h_{D}$ is a $G_0$-invariant orthogonal splitting of the tangent bundle into two $C^\infty$ subbundles. Moreover, if we endow $X$ with its natural $G_0$-invariant metric then $\pi$ is a Riemannian submersion, meaning that $d\pi_{|T^h_{D}}:T^h_{D}\fd T_{X}$ is an isometry.
Of course, we can make the same kind of constructions on $D_u$ (with obvious changes) and we get in particular a $G_u$-invariant decomposition $T_{D_u}=T^v_{D_u}\oplus T^h_{D_u}$.
The Hermitian metric $g_{D_u}$ (resp. $g_{D}$) induces a $G_u$-invariant metric $h_{D_u}$ (resp. a $G_0$-invariant metric $h_{D}$) on the canonical bundle $K_{D_u}=\Lambda^{\dim_\C D_u} T^\star_{D_u}$ of $D_u$ (resp. the canonical bundle $K_{D}={K_{D_u}}_{|D}$ of $D$). The computations of [@GS] show that
\[cano\] The $G_u$-invariant curvature form $\Th_{D_u}(K_{D_u})$ of $(K_{D_u},h_{D_u})$ is negative, and the $G_0$-invariant curvature form $\Th_{D}(K_{D})$ of $(K_{D},h_{D})$ is positive on the horizontal space $T^h_{D}$ and negative on the vertical space $T^v_{D}$. In particular, $\log(h_{D_u}/h_{D})$ is a well-defined function on $D$ whose Levi form is positive definite on $T^h_{D}$.
It follows from [@BR Chap. 4.A] that there exists an element $\xi$ in the center of $\vg$ such that all the eigenvalues of $\ad\xi$ acting on $\gg$ belong to $\sqrt{-1}\,\Z$ and moreover, if $\gg_\ell$ denotes the eigenspace associated to the eigenvalue $\sqrt{-1}\,\ell$, then if $\ng^{(r)}$ denotes the $r$-th step of the central descending series of $\ng$, $$\qg=\bigoplus_{\ell\geq 0}\gg_\ell\ \mbox{ and }\ \ng^{(r)}=\bigoplus_{\ell\geq r}\gg_\ell.$$ For all $m>0$, let $W_m=\oplus_{\ell\geq -m}\,\gg_\ell$. Then $W_m/\qg\subset \gg/\qg$ is an ${\rm Ad}\,Q$-invariant subspace, hence defines a holomorphic subbundle of $T_{D_u}$ (or $T_{D}$). In particular, for $m=1$, we get a holomorphic subbundle $H_{D_u}$, resp. $H_{D}={H_{D_u}}_{|D}$, of $T_{D_u}$, resp. $T_{D}$, which is included in $T^h_{D_u}$, resp. $T^h_{D}$, and called the [*superhorizontal distribution*]{} of $D_u$, resp. $D$. Moreover, $\gg_{1}$ Lie-generates $\ng$ and this implies that the superhorizontal distribution satisfies the so-called bracket-generating condition (see for example [@Pansu]).
Measurable open orbits in flag manifolds {#mes}
----------------------------------------
We keep the assumptions and notation of the previous section, and we consider a new parabolic subgroup $Q'$ of $G$ which also contains the maximal Abelian subgroup $T_0$ of $G_0$, but this time we do not assume that $G_0\cap Q'$ is compact. To emphasize this difference, we set $L_0=G_0\cap Q'$.
It follows from [@FHW Sect. 4.2 and 4.5] that the $G_0$-orbit $D'$ of $eQ'$ in $D'_u$ is open and measurable, meaning that it admits a $G_0$-invariant, possibly indefinite, Kähler metric. We will identify $D'$ with $G_0/L_0$ and $D'_u$ with $G_u/L_u$ where $L_u$ is the compact real Lie subgroup $G_u\cap Q'\subset G$. $D'$ contains a special compact complex submanifold $C\simeq K_0/K_0\cap L_0$ ([@FHW Sect. 4.3]).
There is a decomposition $\qg'=\lg\oplus\ng'$, where $\lg$ is the complexification of the Lie algebra of $L_0$ and $\ng'$ is the maximal nilpotent ideal of $\qg'$. As in the case of period domains, $G$ can be seen as a holomorphic principal $Q'$-bundle on $D'_u$ and the holomorphic tangent bundle of $D'_u$ is the bundle associated to this principal bundle by the adjoint action of $Q'$ on $\gg/\qg' \simeq \tau(\ng')$. Let us mention that the underlying $C^\infty$ vector bundle on $D'$, which is associated to the principal $L_0$-bundle $G_0$, admits in general no decomposition similar to the decomposition in horizontal and vertical components. Nevertheless, we can endow the canonical bundle $K_{D'_u}$ of $D'_u$ (resp. $K_{D'}$ of $D'$) with a $G_u$-invariant (resp. $G_0$-invariant) metric $h_{D'_u}$ (resp. $h_{D'}$) and an analogous statement to Proposition \[cano\] (see [@FHW Sect. 4.7]) is available:
\[canon\] The $G_u$-invariant curvature form $\Th_{D'_u}(K_{D'_u})$ of $(K_{D'_u},h_{D'_u})$ is negative, and the $G_0$-invariant curvature form $\Th_{D'}(K_{D'})$ of $(K_{D'},h_{D'})$ has signature $(\dim_\C D'-\dim_\C C,\dim_\C C)$. In particular, $\varphi:=\log(h_{D'_u}/h_{D'})$ is a well-defined function on $D'$ whose Levi form ${\cal L}(\varphi)$ has everywhere at least $\dim_\C D'-\dim_\C C$ positive eigenvalues.
We will need later a more precise description of the tangent vectors on which ${\cal L}(\varphi)$ is positive definite. Recall that the tangent bundle of $D'$ is the bundle associated to the principal $L_0$-bundle $G_0$ by the adjoint representation of $L_0$ on $\gg/\qg'$. From the computations of [@FHW], $\Th_{D'}(K_{D'})$ is positive on any tangent vector $(e,\xi)\in G_0\times_{L_0} \gg/\qg'\simeq T_{D'}$ such that $\xi\in\pg/\pg\cap\qg'$. By $G_0$-invariance of $\Th_{D'}(K_{D'})$, it follows that it is positive on any tangent vector to $D'$ which has a representative $(g,\xi)$ with $\xi\in\pg/\pg\cap\qg'$ and then, on such a tangent vector, ${\cal L}(\varphi)$ is also positive.
Complex variations of Hodge structure and $\PU(p,q)$ period domains {#var}
-------------------------------------------------------------------
We refer to [@Gri] and [@Sim] for details on this section.
Let $M$ be a complex analytic manifold. A (polarized) complex variation of Hodge structure ($\C$-VHS for short) on $M$ is a $C^\infty$ complex vector bundle $E$ (of finite rank) with a decomposition $E=\oplus_{i=0}^k E^i$ in a direct sum of $C^\infty$ subbundles, a flat connection $\nabla$ satisfying Griffiths’s transversality condition $$\nabla:E^i\fd A^{0,1}(E^{i-1})\oplus A^{1,0}(E^i)\oplus A^{0,1}(E^i)\oplus A^{1,0}(E^{i+1})$$ and a parallel Hermitian form $h$ which makes the Hodge decomposition orthogonal and which on $E^i$ is positive if $i$ is even and negative if $i$ is odd. If we set $r_i=\rk E^i$, the monodromy representation $\rho$ of the flat connection $\nabla$ has image contained in a group isomorphic to $\U(p,q)\subset\GL(p+q,\C)$ where $p=\sum_{i\ {\rm even}}r_i$ and $q=\sum_{i\ {\rm odd}}r_i$.
Assume we have a complex variation of Hodge structure on a complex manifold $M$. We have the so-called Hodge filtration of $E$ by the subbundles $F^s=\oplus_{i\leq s} E^i$ which are holomorphic for the connection $\nabla$.
Let us fix $x_0\in M$, and let $\E=\oplus_{i=0}^k\E^i$ be the fiber of $E$ over $x_0$ endowed with the restriction of the Hermitian metric $h$ on $E$, which will be denoted by the same letter. Then, using notation which are consistent with those in Section \[perdo\], we may set $G=\PGL(\E)$, $G_0=\PU(\E,h)$, $G_u=\PU(\E)$ (w.r.t. the standard Hermitian form on $\E$), and let $Q$ be the parabolic subgroup of $G$ which stabilizes the flag ${\mathbb F}^0\subset {\mathbb F}^{1}\subset\dots\subset {\mathbb F}^{k-1}\subset{\mathbb F}^k=\E$ defined by the Hodge filtration and $V_0=V_u=G_0\cap Q={\rm P}(\U(\E^0,{h}_{|\E^0})\times\dots\times\U(\E^k,{h}_{|\E^k}))$. We obtain a period domain $D=G_0/V_0$ in the flag manifold $D_u=G/Q=G_u/V_u$. We call $D$ the $\PU(p,q)$ period domain associated to the $\C$-VHS $E$.
Choose bases of each of the $\E^i$ and put them together in order to get a basis of $\E$. Identifying now $\gg=\sg\lg(\E)$ with $\sg\lg(p+q,\C)$, the usual Cartan subalgebra $\hg\subset\gg$ is the set of (tracefree) diagonal matrices. We denote by $\Delta$ the root system associated to $\hg$ and $\Delta^+$ the set of positive roots with respect to the usual ordering of $\Delta$. Then, $\Pi=\{\a_1,\dots,\a_{p+q-1}\}\subset\Delta$ (where $\a_j({\rm diag}(h_1,\dots,h_{p+q}))=h_{j+1}-h_j)$ is a simple system. For each root $\a\in\hg^\star$, we denote by $\gg^\a$ the corresponding root-space. The parabolic subalgebra $\qg$ is associated to the subset $\Pi_\qg=\{\a_{r_0},\a_{r_0+r_1},\dots,\a_{r_0+\dots+r_{k-1}}\}\subset\Pi$ in the sense that if $\Phi=\Delta^+\cup\{\a\in\Delta\,|\,\a\in{\rm span}(\Pi\backslash\Pi_\qg)\}$ then $$\qg=\hg\oplus\bigoplus_{\a\in\Phi}\gg^\a.$$ We also have $$\vg=\hg\oplus\bigoplus_{\a\in\Phi\cap-\Phi}\gg^\a\ \ {\rm and}\ \ \ng=\bigoplus_{\substack{\a\in\Phi\\ \a\not\in-\Phi}} \gg^\a.$$ Moreover, in this case, $\gg_\ell=\oplus_{i-j=\ell}\Hom(\E^i,\E^j)\cap \sg\lg(\E)$ for all $\ell$, so that $\qg=\oplus_{i=0}^k\Hom(\E^i,{\mathbb F}^{i})\cap \sg\lg(\E)$ and the superhorizontal distribution is the holomorphic bundle associated to $(\oplus_{i=0}^{k-1}\Hom(\E^i,{\mathbb F}^{i+1})\cap \sg\lg(\E))/\qg$ by the adjoint action of $Q$.
Let $K_0={\rm P}(\U(\oplus_{i} \E^{2i},{h}_{|\oplus_{i} \E^{2i}})\times \U(\oplus_{i} \E^{2i+1},{h}_{|\oplus_{i} \E^{2i+1}}))$. This is a maximal compact subgroup of $G_0$ containing $V_0$. The symmetric space $X=G_0/K_0$ is Hermitian but we stress that $\pi:D\fd X$ is never holomorphic if $D$ is endowed with the complex structure defined by $Q$ (except if $V_0=K_0$).
Geometrically, ${\rm PU}(p,q)$-period domains and flag manifolds can be described as follows. We endow ${\mathbb E}=\C^{p+q}$ with the Hermitian form $h=\sum_{i=1}^p|z_i|^2-\sum_{i=1}^q|z_{p+i}|^2$. The flag manifold $D_u$ is simply the space of flags $\{0={F}^{-1}\subset{F}^0\subset {F}^{1}\subset\dots\subset {F}^{k-1}\subset {F}^k=\C^{p+q}\}$ of $\C^{p+q}$ with $\dim F^{i}/F^{i-1}=r_i$, $p=\sum_{i\ {\rm even}}r_i$ and $q=\sum_{i\ {\rm odd}}r_i$. The period domain $D$ is the open set of $D_u$ consisting of those flags such that for all $-1\leq i\leq k-1$, $(-1)^{i} h$ is negative definite on ${F^i}^\perp$, the orthogonal of $F^i$ in $F^{i+1}$ with respect to the indefinite form $h$. The symmetric space $X$ is the space of $p$-planes in $\C^{p+q}$ on which $h$ is positive definite. It is an open set in $X_u$, the Grassmannian of $p$-planes in $\C^{p+q}$. The fibration $\pi:D\fd X$ is given by $$\{0={F}^{-1}\subset{F}^0\subset {F}^{1}\subset\dots\subset {F}^{k-1}\subset {F}^k=\C^{p+q}\}\longmapsto\bigoplus_{i\ {\rm odd}}{F^i}^\perp,$$ whereas $\pi_u:D_u\fd X_u$ is given by $$\{0={F}^{-1}\subset{F}^0\subset {F}^{1}\subset\dots\subset {F}^{k-1}\subset {F}^k=\C^{p+q}\}\longmapsto\bigoplus_{i\ {\rm odd}}{F^i}^{\perp_u},$$ where ${F^i}^{\perp_u}$ is the orthogonal of $F^i$ in $F^{i+1}$, with respect to the standard definite Hermitian form $\sum_{i=1}^{p+q}|z_i|^2$ on $\C^{p+q}$. Remark that the central fibers of $\pi$ and $\pi_u$, namely the fibers above the $p$-plane $\{z_{p+1}=\dots =z_{p+q}=0\}$, are the same.
Now, let $\tilde M$ be the universal cover of $M$. The complex variation of Hodge structure $E$ induces a $\rho$-equivariant map $f:\tilde M\fd D$ called the [*period map*]{}, and the $\C$-VHS $E$ is said to be nontrivial if $f$ is a nonconstant map. The transversality condition implies that $f$ is holomorphic with respect to the complex structure induced by $Q$ and moreover that $df(T_{\tilde M})\subset H_{D}$: we say that $f$ is a holomorphic [*superhorizontal map*]{}.
The link between the $\C$-VHS and the period map $f$ can be seen as follows. The pull-back $f^\star H_{D}$ is a holomorphic bundle on $\tilde M$ which goes down on $M$ by $\rho$-equivariance. As such, it can be identified with $\oplus_{i=0}^{k-1}{\rm Hom}(E^i,E^{i+1})$ where the vector bundles $E^i\simeq F^i/F^{i-1}$ are endowed with the holomorphic quotient structure. The $(1,0)$ part of the differential of $f$ can hence be interpreted as a morphism $T_M\fd\oplus_{i=0}^{k-1}{\rm Hom}(E^i,E^{i+1})$. On the other hand, the flat connection $\nabla$ on $E$ gives rise to ${\cal O}_M$-linear maps $\theta_i:E^i\fd E^{i+1}\otimes\Omega^1_M$ which are holomorphic sections of ${\rm Hom}(E^i,E^{i+1})\otimes\Omega^1_M$. The so-called Higgs field $\theta:=\oplus_{i=0}^{k-1} \theta_i$ is therefore another incarnation of $df$. Note that $\t$ satisfies the relation $[\t,\t]=0$.
Changing notation, we could also have set $G=\overline{\rho(\pi_1(M))}$, the Zariski-closure of the image of the monodromy representation $\rho$ in $\PGL(\E)$, and $G_0=G\cap\PU(\E,h)$. Then by [@Sim], $G_0$ is the real Zariski closure of $\rho(\pi_1(M))$ and is a group of Hodge type. Moreover the subgroup of $G$ stabilizing the flag ${\mathbb F}^0\subset {\mathbb F}^{1}\subset\dots\subset {\mathbb F}^{k-1}\subset {\mathbb F}^k=\E$ is a parabolic subgroup $Q$ of $G$ and setting $V_0=G_0\cap Q$, we obtain another period domain $D=G_0/V_0$ in the flag manifold $G/Q$. Again, there is a period map $f$ from the universal cover of $M$ to $D$, which is $\rho$-equivariant, holomorphic and superhorizontal.
A possible strategy for proving Carlson-Toledo’s conjecture {#conj}
-----------------------------------------------------------
In [@Re], Reznikov treated a few special cases when $\G:=\pi_1(M)$ admits a rigid linear representation. More precisely, let $\rho:\G\fd\GL(m,\C)$ be a rigid, unbounded, reductive and irreducible representation. It follows from the general theory of Simpson [@Sim] that such a representation necessarily is the monodromy of a complex variation of Hodge structure over $M$. As we just explained, if $G$, resp. $G_0$, is the Zariski closure, resp. real Zariski closure, of $\rho(\G)$ in $\GL(m,\C)$, then we obtain a period map $f$ from the universal cover $\tilde M$ of $M$ to a period domain $D$ of the form $G_0/V_0$. Reznikov assumed that $G_0$ is a Hermitian group and that in fact $D$ is the Hermitian symmetric space $X$ associated to $G_0$. Then the pull-back of the Kähler form $\o_X$ by $f$, which goes down to $M$ by equivariance, gives a non trivial cohomology class in $H^2(M,\R)$ (because $f^\star\o_X$ is semi-positive and $f$ is non constant). The fact that this class comes from a class in $H^2(\G,\Z)$ follows from the fact that $\pi_2(X)=0$ (see below).
When the period domain $D$ is not a bounded symmetric domain, for example if $G_0$ is not of Hermitian type, it might still be possible to produce a class in $H^2(\G,\R)$ from the geometry of the $\C$-VHS and the period map $f:\tilde M\fd D$. Indeed, there are $G_0$-homogeneous line bundles over $D$ (for example the canonical bundle $K_D$). Let ${\bf L}\fd D$ be such a line bundle and assume that it admits a $G_0$-invariant metric whose curvature form $\sqrt{-1}\,\Theta({\bf L})$ is positive on the superhorizontal distribution (this is the case for $K_D$ by Proposition \[cano\]). The line bundle $f^\star {\bf L}\fd \tilde M$ descends to a line bundle $L\fd M$ by equivariance and, as before, $\frac{\sqrt{-1}}{2\pi}f^\star \Theta({\bf L})$ induces a non trivial cohomology class $\eta\in H^2(M,\Z)$, which is just the first Chern class of $L$. Moreover, the image of $\eta$ in $H^2(M,\R)$ does not vanish because the semi-positivity of $\frac{\sqrt{-1}}{2\pi}f^\star \Theta({\bf L})$ implies that $\eta$ cannot be a torsion class. Again, it remains to prove that $\eta$ belongs to the image of the natural inclusion $\iota:H^2(\G,\Z)\hookrightarrow H^2(M,\Z)$.
In view of the exact sequence \[long\]: $$\cdots {\longrightarrow}\pi_2(L^0) {\longrightarrow}\pi_2(M) \stackrel{\partial_L}{{\longrightarrow}}\pi_1({{\mathbb C}}^\star) = {{\mathbb Z}}{\longrightarrow}\pi_1(L^0) {\longrightarrow}\pi_1(M) {\longrightarrow}1\, ,$$ where $L^0$ is the $\C^\star$-bundle associated to $L$, it would be sufficient to show that the connecting morphism $\partial_L:\pi_2(M)\fd\Z$ is trivial for some line bundle $L\fd M$ induced by a homogeneous ${\bf L}\fd D$ as above. This is of course the case if $f_\star:\pi_2(M)\fd\pi_2(D)$ vanishes, in particular if $\pi_2(D)=0$, that is if $D=X$.
Variations of Hodge structure with a rank 1 subbundle: proof of Theorem \[cartoun\] {#proofofcartoun}
-----------------------------------------------------------------------------------
In this section, we show that a slight modification of the strategy we just described gives a proof of Theorem \[cartoun\].
We work here with the $\PU(p,q)$ period domain $D$ and map $f$ associated to the $\C$-VHS $E$ the representation $\rho$ comes from, and we keep the notation of Section \[var\]. Recall in particular that each $\theta_j:E^j\fd E^{j+1}\otimes\Omega^1_M$ deduced from the flat connection $\nabla$ on $E$ can be interpreted as a component of the differential of the period map.
By assumption, there exists $i$, $0<i<k$, such that $r_i={\rm rk}\,E^i=1$. First, we show that $\theta_{i-1}:E^{i-1}\fd E^i\otimes\Omega^1_M$ has (complex) rank 1 generically on $M$ (this kind of property was already observed by Siu, Carlson-Toledo and others). If it were not the case, $\theta_{i-1}$ would generically have rank greater than or equal to 2 ($\theta_{i-1}$ cannot vanish identically otherwise the variation of Hodge structure $E$ splits in two subvariations of Hodge structure and this is impossible since $\rho$ is assumed to be irreducible). This would imply that for a generic $x\in M$, writing $E^i_x=\C w$, there exist $v,v'\in E^{i-1}$ and $\g,\g'\in \Omega^1_{M,x}$ two linearly independent forms such that $\theta_{i-1}(v)=w\otimes\g$ and $\theta_{i-1}(v')=w\otimes\g'$. Since the forms $\g$ and $\g'$ are independent, $T_{M,x}=\Ker\g + \Ker\g'$. Now let $\zeta\in T_{M,x}$ and assume $\zeta\in\Ker\g$. Then, considering $\t_i$ as a morphism from $E^i\otimes T_M$ to $E^{i+1}$, $\t_i(w\otimes\zeta)=\t_i((\t_{i-1}(v)\xi)\otimes\zeta)$ for some $\xi\in T_{M,x}$ such that $\g(\xi)=1$. By the commutation relation $[\t,\t]=0$, $\t_i(w\otimes\zeta)= \t_i((\t_{i-1}(v)\zeta)\otimes\xi)=\t_i((\g(\zeta)w)\otimes\xi)=0$. The same argument shows that $\t_i(w\otimes\zeta)=0$ for $\zeta\in\Ker\g'$, hence for all $\zeta$: $\t_i$ vanishes. In other words, if $\theta_{i-1}$ has (complex) rank greater than or equal to 2 then $\theta_{i}$ vanishes generically, hence everywhere. As above, this contradicts the irreducibility of $\rho$.
Therefore, using again that $r_i=1$, the $\Hom(E^{i-1},E^i)$ component of $df$ also has complex rank 1 generically on $M$.
Consider the maximal parabolic subalgebra $\qg'\subset\gg$ associated to $\Pi_{\qg'}=\{\a_{r_0+\dots+r_{i-1}}\}$ (see Section \[var\]) and the corresponding parabolic subgroup $Q'\subset G$: we have $Q\subset Q'$. The projection map $\nu:D_u=G/Q\fd D'_u:=G/Q'$, which maps the flag $\{{F}^0\subset {F}^{1}\subset\dots\subset {F}^{k-1}\}$ to the $(r_0+\dots+r_{i-1})$-plane $F^{i-1}$, is holomorphic with respect to the natural complex structure on $D_u$ and $D'_u$. Its restriction to $D$, which we still denote by $\nu$, induces a holomorphic map $\nu:D\fd D'$ where $D'$ is the $G_0$-orbit of $eQ'$ in $D'_u$ (it is measurable open, see Section \[mes\]).
The map $g:=\nu\circ f:\tilde M\fd D'$ is holomorphic and since, by the very choice of $\nu$, $d\nu$ kills the $\Hom(E^{j-1},E^j)$ component of $df$ for all $j\neq i$, $g$ has rank 1. As a consequence, the fibers of $g$ are pure dimensional, of codimension 1. We can thus apply a theorem of Kaup [@Gra Theorem 6] which states that there exists a smooth complex curve $N$ and holomorphic maps $\hat g:\tilde M\fd N$, $\psi:N\fd D'$, with $\hat g$ surjective, such that $g=\psi\circ \hat g$. We claim that $N$ cannot be isomorphic to $\P^1$. Indeed, $f$ is superhorizontal, hence the Levi form ${\cal L}(\varphi)$ is positive on each nonvanishing tangent vector in the image of the differential of $g$ (see Proposition \[canon\] and the discussion following it). As a consequence, $\psi^\star\varphi$ is a subharmonic function on $N$ and its Levi form $\psi^\star{\cal L}(\varphi)$ is generically positive since $\psi$ is nonconstant. This is a contradiction if $N$ is compact.
Finally, we can apply the method described above, using $g^\star K_{D'}$ instead of $f^\star K_{D}$. The restriction of $g^\star K_{D'}$ to any 2-sphere lying in $\tilde M$ is clearly topologically trivial since $g$ factors through $N$ which is aspherical.
Superhorizontal 2-spheres in period domains {#superhori}
===========================================
Let $E$ be a $\C$-VHS over a compact Kähler manifold $M$, and $f:\tilde
M\fd D$ the corresponding period map. Since $f$ is superhorizontal, if $\tilde\sigma:S^2\fd \tilde M$ is any (smooth) lift of some element $[\sigma]\in\pi_2(M)$, $f\circ\tilde\sigma: S^2\fd D$ is a superhorizontal 2-sphere in the period domain $D$. Of course, if such superhorizontal spheres were homotopically trivial, then for every homogeneous line bundle ${\bf L}\fd D$, the connecting homomorphism $\partial_L: \pi_2(M)\fd\Z$ in the exact sequence \[long\] would be trivial and the non-vanishing of $H^2(\pi_1(M),\R)$ would easily follow from the arguments of Section \[conj\]. However, this is in general not true, and we will show in this section that in most cases, there are many non-trivial superhorizontal 2-spheres in the period domain $D$. In fact, when the hypotheses of Theorem \[cartoun\] are not met, and if $D$ is the $\PU(p,q)$ period domain associated to $E$, $\pi_2(D)$ is generated by superhorizontal spheres, see Theorem \[hori\].
Construction of horizontal maps via Gromov $h$-principle
--------------------------------------------------------
Following [@GromovPDR; @GromovCC] and [@Pansu], we explain in this section how the $h$-principle may be used to produce horizontal maps. These considerations apply to a much wider setting than superhorizontal maps in period domains, and the notation of this section might differ from those used in the rest of the paper.
Let $M$ be a real closed $k$-dimensional manifold, $N$ a real $n$-dimensional manifold, and let $H$ be a subbundle of the tangent bundle $T_N$ of $N$, called the horizontal distribution, ${\rm rk}(H)=h\geq k$. Let $\t$ be the projection $T_N\fd T_N/H$ seen as a 1-form on $N$ with values in $T_N/H$ and let $d\t:H\times H\fd T_N/H$ be defined on horizontal vector fields by $d\t(\xi,\eta)=[\xi,\eta]\mod H$ ($d\t$ is in fact a 2-form on $H$ with values in $T_N/H$).
Assume that $f$ is an $H$-horizontal immersion of $M$ into $N$, namely an immersion such that for all $x\in M$, $d_xf$ maps $T_{M,x}$ in $H_{f(x)}\subset T_{N,f(x)}$. The differential of $f$ gives rise to a map $F:M\fd {\rm Gr}_{k}(H)$, where ${\rm Gr}_{k}(H)\fd N$ is the bundle of Grassmannians of $k$-planes in $H$ over $N$. The horizontality of $f$ is equivalent to the vanishing $f^\star\t=0$. Moreover, if $f$ is $C^2$, then $f^\star d\t=0$ also holds. This means that $df$ sends the tangent spaces $T_{M,x}$ to [*isotropic*]{} subspaces of $H$, namely subspaces $S$ of $H_{f(x)}$ on which $d\t$ vanishes. Therefore, a necessary condition for a $C^2$ horizontal immersion to exist is the existence of a map $F$ from $M$ to the bundle of Grassmannian of horizontal isotropic $k$-planes over $N$.
We shall indicate the different steps of the construction of a [*folded*]{} horizontal immersion $f:M\fd N$ from the data of $F:M\fd {\rm Gr}_{k}^{reg,iso}(H)$, where ${\rm Gr}_{k}^{reg,iso}(H)\fd N$ is the bundle of [*regular*]{} horizontal isotropic $k$-planes over $N$. The regularity assumption will be explained in the next paragraphs. The meaning of a folded immersion will be given later.
First, one wants to construct germs of horizontal immersions tangent to given horizontal subspaces of $T_N$ at given points. One therefore wants to locally solve the horizontality equation ${\mathcal H}(f):=f^\star\t=0$. Since this is a local problem, we assume that $M$ is an open set in $\R^k$, $0\in M$, and that the quotient bundle $T_N/H$ is trivial on $N$, so that $\t$ and $d\t$ can be seen as globally defined $\R^{n-h}$-valued forms on $N$. Here, ${\mathcal H}$ is understood as a differential operator from the space of sections of the trivial bundle $M\times N\fd M$ (so that sections of this bundle are just maps $M\fd N$) to the space of $\R^{n-h}$-valued 1-forms on $M$. Spaces of sections are endowed with the [*fine*]{} topology: a basis of the fine $C^0$-topology on the space of continuous sections $C^0(M,E)$ of a bundle $E$ over $M$ is given by the sets $C^0(M,{\mathcal V})$ of continuous ${\mathcal V}$-valued sections of $E$, for all open sets ${\mathcal V}\subset E$. The fine $C^r$-topology on $C^r(M,E)$ is induced from the fine $C^0$-topology on the space of sections of the $r$-jets bundle $E^{(r)}\fd M$ via the natural inclusion $C^r(M,E)\subset C^0(M,E^{(r)})$.
The idea is to right invert the linearization $D_f{\mathcal H}$ of the operator ${\mathcal H}$ at $f$. If $\xi$ is a section of $f^\star T_N$, then $$D_f{\mathcal H}(\xi)=d(\iota_\xi\t)+f^\star(\iota_\xi d\t).$$ Therefore, if $\xi$ is horizontal (i.e. $\xi$ is a section of $f^\star H$), $D_f{\mathcal H}(\xi)=f^\star(\iota_\xi d\t)$. The linearization $D_f{\mathcal H}$ is seen as a bundle map from $f^\star H$ to the space $M\times\Hom(\R^k,\R^{n-h})$ of $\R^{n-h}$-valued 1-forms on $M$. That is where the regularity assumption comes from: a $k$-dimensional subspace $S$ of $T_{N,y}$ is said to be [*regular*]{} if the linear map $$\begin{array}{rcl}
H_{y} & \fd & \Hom(S,\R^{n-h})\\
v & \longmapsto & \iota_v d\t
\end{array}$$ is surjective (if $S\subset T_{N,y}$ is not horizontal, i.e. not contained in $H_y$, this definition depends on the choice of $\t$ as an $\R^{n-h}$-valued 1-form on $N$).
The set of regular $k$-dimensional subspaces is open in the bundle of Grassmannians of $k$-planes in $T_N$, and so is the set ${\rm RegIm}(M,N)$ of regular immersions $M\fd N$ in the space of maps $M\fd N$.
It turns out that the regularity assumption is enough for being able to solve the horizontality equation up to infinite order, and then to right invert the linearization of ${\mathcal H}$. Namely:
[([@Pansu Prop. 29])]{}.\[allorders\] Let $S\subset H_y$ be a $k$-dimensional regular isotropic subspace. Then there exists a germ of immersion $f:(\R^k,0)\fd
(N,y)$ which is tangent to $S$ at the origin: $d_0f(\R^k)=S$, and which satisfies the horizontality equation to all orders: ${\mathcal H}(f)(x)=o(|x|^m)$ for all $m\in\N$.
[(See [@Pansu Prop. 31])]{}.\[rightinvertible\] Assume $f:M\subset\R^k\fd N$ is a regular immersion (i.e. $d_xf(\R^k)$ is a regular $k$-dimensional subspace of $T_{N,f(x)}$ for all $x\in M$), then the map $$\begin{array}{rcl}
\Gamma(f^\star H) & \fd & \Gamma(M\times\Hom(\R^k,\R^{n-h}))\\
\xi & \longmapsto & f^\star(\iota_\xi d\t)
\end{array}$$ is right invertible.
The set of regular $k$-planes in the bundle of Grassmannians ${\rm Gr}_k(T_N)$ is open as we mentionned. Hence the set ${\rm RegInj}(T_M,T_N)$ of triples $(x,y,L)$ with $x\in M$, $y\in N$ and $L$ an injective linear map from $\R^k$ to $T_{N,y}$ with regular image, is open in the bundle $\Hom(T_M,T_N)$ over $M\times N$.
Given $(x,y,L)\in {\rm RegInj}(T_M,T_N)$, the set of right inverses of the linear map $$\begin{array}{rcl}
H_y & \fd & \Hom(\R^k,\R^{n-h}) \\
v & \longmapsto & (d\t)_y(v,L(.)) \\
\end{array}$$ is an affine space of dimension $k(n-h)(h-k(n-h))$. Therefore the collection of these spaces over ${\rm RegInj}(T_M,T_N)$ is a bundle, with contractible fibers, and hence admits a section $(x,y,L)\mapsto {\rm rightinv}(x,y,L)$.
Now, given an $\R^{n-h}$-valued 1-form $\beta$ on $M$, i.e. $\beta(x)\in\Hom(\R^k,\R^{n-h})$ for all $x\in M$, define an horizontal vector field along $f$ by $\xi_f(\beta)=(x,{\rm rightinv}(x,f(x),d_xf)(\beta(x)))$. $\xi_f$ is a right inverse of $D_f{\mathcal H}$.
Now, a theorem of Gromov in the spirit of the inverse function theorem of Nash shows that if the linearization $D_f{\mathcal H}$ of the operator ${\mathcal H}$ can be right inverted for all $f\in{\rm RegIm}(M,N)$, then locally, the operator ${\mathcal H}$ itself can be right inverted: for any regular immersion $f$, there exists a right inverse ${\mathcal H}_f^{-1}$ of ${\mathcal H}$, defined on a $C^s$-neighborhood ${\mathcal V}_f$ of ${\mathcal H}(f)$, such that ${\mathcal H}_f^{-1}({\mathcal H}(f))=f$. Moreover, ${\mathcal H}_f^{-1}$ depends smoothly on parameters. See [@GromovPDR p. 114-120] for a precise statement of the theorem and its corollaries in the form that is useful here.
This theorem and the two propositions prove that the Cauchy problem for horizontal immersions, with initial data a point $y\in N$ and a regular horizontal isotropic $k$-plane $S$ in $T_{N,y}$, has a local solution: there exists a germ of horizontal immersion $f:(\R^k,0)\fd (N,y)$ which is tangent to $S$ at the originAbelian. Indeed, Proposition \[allorders\] gives a germ of regular immersion $f_0:(\R^k,0)\fd(N,y)$ such that ${\mathcal H}(f_0)=o(|x|^m)$. Proposition \[rightinvertible\] and Gromov’s inverse function theorem [@GromovPDR Main Theorem (3), p. 117] shows that there exists a right inverse ${\mathcal H}_{f_0}^{-1}$ of ${\mathcal H}$ defined in a neighborhood of ${\mathcal H}(f_0)$. Taking a smaller neighborhood if necessary, we may assume that the identically zero 1-form belongs to this neighborhood. Applying ${\mathcal H}_{f_0}^{-1}$ gives the desired germ $f$.
In fact, we can do better: the germ $f_0$ can be deformed to a true horizontal germ [@GromovPDR Corollary (D), p. 119]. Indeed, again working on a smaller neighborhood of $0\in\R^k$ if necessary, we may assume that $(1-t){\mathcal H}({f_0})$ belongs to the given neighborhood ${\mathcal V}_{f_0}$ of ${\mathcal H}(f_0)$ for all $t\in[0,1]$ and we have the deformation $f_t:={\mathcal H}_{f_0}^{-1}((1-t){\mathcal H}({f_0}))$ from $f_0$ to $f_1$ which is horizontal.
We will say that regular horizontal immersions satisfy the [*local $h$-principle*]{}.
The right inverse ${\mathcal H}_f^{-1}$ has moreover the property of being a local operator [@GromovPDR Main Theorem (5), p. 118]: for a given Riemannian metric on $M$ and $r>0$, we may assume that for all regular immersions $f:M\fd N$, all 1-forms $\beta\in {\mathcal V}_f$ and all $x\in M$, the value of ${\mathcal H}_f^{-1}(\beta)$ at $x$ depends only on the values of $f$ and $\beta$ in a ball of radius $r$ around $x$ (of course ${\mathcal V}_f$ depends on the metric and on $r$). This in turns implies an important fact about regular horizontal immersions $M\fd N$: they are [*microflexible*]{} [@GromovPDR p. 120]. By definition, this means that given two compact sets $K'\subset K\subset M$, a continuous family of regular horizontal immersions $\{f_a\}_{a\in A}$ defined near $K$ and indexed by a polyhedron $A$, and a family $\{f_{a,t}\}_{a\in A}$ of regular horizontal immersions, defined near $K'$, continuous in $a\in A$ and $t\in[0,1]$, such that $f_{a,0}=f_{a}$ near $K'$ for all $a\in A$, there exist $\varepsilon>0$ and a continuous family $\{\tilde f_{a,t}\}_{a\in A}$ of regular horizontal maps, defined near $K$ and for $t\in[0,\varepsilon]$, such that for all $a\in A$, $\tilde f_{a,0}=f_a$ near $K$ and $\tilde f_{a,t}=f_{a,t}$ near $K'$.
Indeed, for $t$ small, we can find regular immersions $F_{a,t}$ from a neighborhood of $K$ in $M$ to $N$ such that $F_{a,t}=f_{a,t}$ near $K'$, and then, for $t$ small, we may define $\tilde f_{a,t}={\mathcal H}_{F_{a,t}}^{-1}(0)$. This gives a family of regular horizontal immersions defined near $K$ which agree with $f_{a,t}$ near $K'$ by locality.
By [@GromovCC p. 262], the local $h$-principle and the microflexibility imply that given our $k$-dimensional closed manifold $M$ and a continuous map $F$ from $M$ to the bundle ${\rm Gr}_{k}^{reg,iso}(H)\stackrel{\pi}{\fd} N$ of Grassmannians of regular $H$-horizontal isotropic $k$-planes over $N$, there is a polyhedron $W$, homotopy equivalent to $M$, and a folded horizontal immersion $f^\vee:W\fd N$ such that $f^\vee\circ e$ approximates $\pi\circ F$ (for the topology of uniform convergence and some homotopy equivalence $e:M\fd W$). Here by a folded horizontal immersion $f^\vee:W\fd N$ we mean that $f^\vee$ is continuous and is a smooth horizontal immersion on each simplex of $W$ (up to the boundary).
The construction of such a map is explained in [@Pansu] when $M=S^1$. In the proof of the following proposition, we will give a detailed construction for $M=S^2$ since this is the case we are ultimately interested in. For the general case, see Remark \[remarque1\].
\[folded\] Let $f_0:S^2\fd N$ be a continuous map and assume it admits a continuous lift $F$ to the bundle ${\rm Gr}_{2}^{reg,iso}(H)\stackrel{\pi}{\fd} N$ of Grassmannians of regular $H$-horizontal isotropic $2$-planes over $N$, such that $\pi\circ F$ is homotopic to $f_0$. Then there exists a smooth horizontal map $f:S^2\fd N$ homotopic to $f_0$.
We begin with the construction of a polyhedron $W$, homeomorphic to $S^2$, and a folded horizontal immersion $f^\vee:W\fd N$, which “approximates” the map $f_0$. From our discussion so far, we know that there exists a continuous choice of germs of horizontal immersions $f_x$, $x\in S^2$, defined on a neighborhood of $0\in\R^2$, such that $f_x(0)=\pi\circ F(0)\in N$ and $d_0f_x(\R^2)=F(x)$. We may assume that all these germs are defined in the Euclidean ball ${\bf B}$ of radius 2 around 0. Consider a fixed equilateral triangle ${\bf T}$ inscribed in the circle of radius 1 in ${\bf B}$, let ${\bf v}_r$, ${\bf v}_g$ and ${\bf v}_b$ be its vertices, and ${\bf e}_{rg}$, ${\bf e}_{gb}$, ${\bf e}_{br}$ its edges. Using microflexibility, we are going to construct families of horizontal immersions defined near ${\bf T}$ in ${\bf B}$, with prescribed boundary values on $\partial{\bf T}$, from which we shall select maps which we shall then glue together along a pattern given by a suitably chosen triangulation of $S^2$.
All the forthcoming metric notions on the sphere $S^2$ are to be understood with respect to the round metric.
Let ${\bf U}_r$, ${\bf U}_g$, ${\bf U}_b$ be small disjoint discs centered at ${\bf v}_r$, ${\bf v}_g$, ${\bf v}_b$, and ${\bf U}_{rg}$, ${\bf U}_{gb}$, ${\bf U}_{br}$ be small cylindrical neighborhoods of ${\bf e}_{rg}$, ${\bf e}_{gb}$, ${\bf e}_{br}$ in ${{\bf B}}$ such that ${\overline{\bf U}_{rg}\cap \overline{\bf U}_{gb}}\subset {\bf U}_g$, ${\overline{\bf U}_{gb}\cap \overline{\bf U}_{br}}\subset {\bf U}_b$ and ${\overline{\bf U}_{br}\cap \overline{\bf U}_{rg}}\subset {\bf U}_r$.
Let $x\in S^2$ and $\xi$ be a unit tangent vector to $S^2$ at $x$. For $t\in[0,1]$, define an horizontal immersion $f_{x,\xi,t}$ on a neighborhood of $\overline{\bf U}_r\cup \overline{\bf U}_g$ by $$f^{rg}_{x,\xi,t}=\left\{
\begin{array}{l}
f_{c_{x,\xi}(-t)} \mbox{ near } \overline{\bf U}_r\\
f_{c_{x,\xi}(t)} \mbox{ near } \overline{\bf U}_g\\
\end{array}
\right.$$ where $c_{x,\xi}$ is the unit speed geodesic going through $x$ in the direction of $\xi$.
The maps $f^{rb}_{x,\xi,t}$, $t\in[0,1]$, give a deformation of the horizontal map $f_x:{{\bf B}}\fd N$ near $\overline{\bf U}_r\cup \overline{\bf U}_g$. By microflexibility, there exists an $\varepsilon_0 >0$, uniform in $x$ and $\xi$, and a deformation of $f_x$ by horizontal immersions $\tilde f^{rg}_{x,\xi,t}:{{\bf B}}\fd N$, $t\in[0,\varepsilon_0]$, such that $\tilde f^{rg}_{x,\xi,t}=f^{rg}_{x,\xi,t}$ near $\overline{\bf U}_r\cup \overline{\bf U}_g$.
In the same way we define horizontal maps $\tilde f^{gb}_{x,\xi,t}$ and $\tilde f^{br}_{x,\xi,t}$, for $t\in[0,\varepsilon_0]$.
Let $\xi_r$, $\xi_g$ and $\xi_b$ be unit tangent vectors at $x\in S^2$. Let $m_{rg}(t)$, $m_{gb}(t)$ and $m_{br}(t)$ be the midpoints of the geodesic segments $[c_{x,\xi_r}(t),c_{x,\xi_g}(t)]$, $[c_{x,\xi_g}(t),c_{x,\xi_b}(t)]$ and $[c_{x,\xi_b}(t),c_{x,\xi_r}(t)]$. Let also $\xi_{rg}(t)$, $\xi_{gb}(t)$, $\xi_{br}(t)$ and $s_{rg}(t)$, $s_{gb}(t)$, $s_{br}(t)$ be such that $$\left\{
\begin{array}{l}
c_{m_{rg}(t),\xi_{rg}(t)}(s_{rg}(t))=c_{x,\xi_g}(t)\\
c_{m_{gb}(t),\xi_{gb}(t)}(s_{gb}(t))=c_{x,\xi_b}(t)\\
c_{m_{br}(t),\xi_{br}(t)}(s_{br}(t))=c_{x,\xi_r}(t)
\end{array}
\right.$$ Everything is made so that the maps $f^{rgb}_{x,\xi_r,\xi_g,\xi_b,t}$ given for $t\in[0,\varepsilon_1]$ for some $0<\varepsilon_1\leq \varepsilon_0$ by $$f^{rgb}_{x,\xi_r,\xi_g,\xi_b,t}=\left\{
\begin{array}{l}
\tilde f^{rg}_{m_{rg}(t),\xi_{rg}(t),s_{rg}(t)} \mbox{ near } \overline{\bf U}_{rg}\\
\tilde f^{gb}_{m_{gb}(t),\xi_{gb}(t),s_{gb}(t)} \mbox{ near } \overline{\bf U}_{gb}\\
\tilde f^{br}_{m_{br}(t),\xi_{br}(t),s_{br}(t)} \mbox{ near } \overline{\bf U}_{br}\\
\end{array}
\right.$$ are well defined on $\overline{\bf U}_{rg}\cup \overline{\bf U}_{gb}\cup \overline{\bf U}_{br}$. Indeed, for example on $\overline{\bf U}_{rg}\cap \overline{\bf U}_{gb}\subset {\bf U}_g$, we have $$\begin{array}{rcl}
\tilde f^{rg}_{m_{rg}(t),\xi_{rg}(t),s_{rg}(t)} & = & f_{c_{m_{rg}(t),\xi_{rg}(t)}(s_{rg}(t))}\\
& = & f_{c_{x,\xi_g}(t)}\\
& = & f_{c_{m_{gb}(t),\xi_{gb}(t)}(-s_{gb}(t))}\\
& = & \tilde f^{gb}_{m_{gb}(t),\xi_{gb}(t),s_{gb}(t)}
\end{array}$$
Hence these maps give a deformation of $f_x:{{\bf B}}\fd N$ near $\overline{\bf U}_{rg}\cup \overline{\bf U}_{gb}\cup \overline{\bf U}_{br}$ and by microflexibility, there exist $\varepsilon>0$, uniform in $x$, $\xi_r$, $\xi_g$, $\xi_b$, and a deformation of $f_x$ by horizontal immersions $\tilde f^{rgb}_{x,\xi_r,\xi_g,\xi_b,t}:{{\bf B}}\fd N$ defined for $t\in[0,\varepsilon]$ such that $\tilde f^{rgb}_{x,\xi_r,\xi_g,\xi_b,t}=f^{rgb}_{x,\xi_r,\xi_g,\xi_b,t}$ near $\overline{\bf U}_{rg}\cup \overline{\bf U}_{gb}\cup \overline{\bf U}_{br}$.
Now consider a triangulation ${\mathcal T}$ of $S^2$, and assume that it is geodesic, that each triangle $T$ of $\mathcal T$ is contained in a convex ball of radius $<\varepsilon/p$ for some large number $p$, and that its angles are not too obtuse (the center of the circumcircle of $T$ should also be contained in the ball of radius $<\varepsilon/p$ containing $T$). Assume moreover that $\mathcal T$ is even, namely that there is an even number of triangles around each vertex of $\mathcal T$. Triangulations satisfying our assumptions exist for any value of $p$, as is easily verified. The fact that the triangulation is even implies that $\mathcal T$ is 3-colorable, i.e. that we can color its vertices with 3 colors, say red, green and blue, so that adjacent vertices have different colors. Abelian
Choose a 3-coloring of the vertices of the triangulation ${\mathcal T}$ and, for each triangle $T$ of $\mathcal T$, let $x^T$ be the center of the circumcircle of $T$ and let $\xi^T_r$ (resp. $\xi^T_g$, $\xi^T_b$) be the unit tangent vector at $x^T$ in the direction of the red (resp. green, blue) vertex of $T$. Construct the horizontal maps $f_{T,t}=\tilde f^{rgb}_{x^T,\xi^T_r,\xi^T_g,\xi^T_b,t}$ as before. Let $t^T$ be such that $m_{rg}(t^T)$ (resp. $m_{gb}(t^T)$, $m_{br}(t^T)$) is the midpoint of the red-green (resp. green-blue, blue-red) edge of $T$ and set $f_T=f_{T,t^T}$. The maps $f_T$ are well-defined because we assumed that the triangles are very small. Moreover, if two triangles $T$ and $T'$ are adjacent, say along their red-green edges, then the maps $f_T$ and $f_{T'}$ agree on ${\bf U}_{rg}\subset{ {\bf B}}$ because they depend only on the red and green vertices they share, and on the midpoint of the red-green edge connecting them.
Consider one copy ${\bf T}_T$ of ${\bf T}$ for each triangle $T$ of the triangulation $\mathcal T$, and identify the red-green (resp. green-blue, blue-red) edge of ${\bf T}_T$ with the red-green (resp. green-blue, blue-red) edge of ${\bf T}_{T'}$ if $T$ and $T'$ intersect along their red-green (resp. green-blue, blue-red) edges. The resulting polyhedron $W$ is homeomorphic to the sphere $S^2$. Moreover, the map $f^\vee:W\fd N$ given by $f^\vee=f_T$ on ${\bf T}_T$ is well-defined, and it is the folded horizontal immersion we wanted.
Neither the polyhedron $W$ nor the map $f^\vee$ are smooth, but we can easily build a smooth horizontal map $f:S^2\fd N$ out of the $f_T$’s (this $f$ will not be an immersion).
For all vertices $v$ of the triangulation ${\mathcal T}$ of $S^2$, let $B_v$ be the ball of radius $r$ around $v$, $r$ being such that these balls are pairwise disjoint. For all edges $e:[0,L_e]\fd S^2$ of ${\mathcal T}$ parameterized by arc length, let $\eta_e$ be a unit normal vector field along $e$ and let $U_e$ be the open “tubular” set $\{c_{e(t),\eta_e(t)}(s)\ |\ t\in(r/2,L_e-r/2),\ s\in(-r',r')\}$ for some $r'<r/2$, chosen so that these sets are pairwise disjoint.
Given an edge $e$ of ${\mathcal T}$, we map the set $B_{e(0)}\cup U_e\cup B_{e(L_e)}$ to the corresponding edge of the model triangle ${\bf T}$. Say $e(0)$ is a red vertex and $e(L_e)$ is a green one. Define a map $\phi_e$ from the edge $e$ to the red-green edge ${\bf e}_{rg}:[0,\sqrt{3}]\fd \R^2$ of ${\bf T}$ as follows $$\phi_e:\ e(t)\longmapsto\left\{
\begin{array}{lr}
{\bf e}_{rg}(0)={\bf v}_r & \mbox{ if }t\in[0,r]\\
{\bf e}_{rg}(\chi_e(t)) & \mbox{ if }t\in[r,L_e-r]\\
{\bf e}_{rg}(\sqrt{3})={\bf v}_g & \mbox{ if }t\in[L_e-r,L_e]
\end{array}
\right.$$ where $\chi_e : [r,L_e-r]\fd [0,\sqrt{3}]$ is a smooth, onto, strictly increasing function all of whose derivatives vanish at $r$ and $L_e-r$. Now let $\Phi_e$ by the map from $B_{e(0)}\cup U_e\cup B_{e(L_e)}$ to ${\bf e}_{rg}$ given by: $$\Phi_e(x)=\left\{
\begin{array}{lr}
{\bf v}_r & \mbox{ if }x\in B_{e(0)}\\
{\phi_e(e(t))}\in{\bf e}_{rg} & \mbox{ if }x=c_{e(t),\eta_e(t)}(s)\in U_e\\
{\bf v}_g & \mbox{ if }x\in B_{e(L_e)}
\end{array}
\right.$$ The $\Phi_e$’s are smooth and they agree on $B_v$ whenever two edges are adjacent at a vertex $v$. Hence, if $V$ is the neighborhood of the $1$-skeleton of ${\mathcal T}$ given by the union of the $B_v$’s and the $U_e$’s, we obtain a smooth map $\Phi_V$ from $V$ to the boundary $\partial{\bf T}$ of ${\bf T}\subset\R^2$. Multiplying it by a smooth map $S^2\fd [0,1]$ which is identically 0 outside $V$ and identically 1 inside a smaller neighborhood of the 1-skeleton, we get a smooth map $\Phi:S^2\fd {\bf T}$, which sends the vertices and edges of the triangulation ${\mathcal T}$ to the vertices and edges of ${\bf T}$ according to the chosen $3$-coloring of the triangulation (note that $\Phi$ sends $S^2\backslash V$ to the center $0\in\R^2$ of $\bf T$). Moreover, by construction, all normal derivatives of $\Phi$ at the edges of ${\mathcal T}$ vanish (and $\Phi$ is constant near the vertices). This implies that setting $f_{|T}=f_T\circ \Phi$ defines a smooth horizontal map $f:S^2\fd N$.
If the triangulation ${\mathcal T}$ is chosen sufficiently fine, namely if $p$ is sufficiently big, $f$ is very close to the map $\pi\circ F$ and therefore homotopic to it, hence to $f_0$.
\[remarque1\] If $M$ is any closed $k$-dimensional manifold, the arguments given here can be adapted to prove the existence of a folded horizontal immersion (or a smooth horizontal map) $M\fd N$. It is indeed enough to have a triangulation of $M$ with an identification of its simplices with some model simplex, which will be used to glue horizontal immersions defined on copies of this model simplex, as we just did. This identification is given by a proper $(k+1)$-coloring of the vertices of the triangulation. Such triangulations, as fine as needed, always exist, since the barycentric subdivision of any triangulation of $M$ is $(k+1)$-colorable.
\[remarque2\] If instead of a map $F:M\fd {\rm Gr}^{reg,iso}_{k}(H)$, we start with a map $F:M\fd {\rm Gr}^{reg,iso}_{k,l}(H)$ in the bundle of flags of regular $k$-planes in regular horizontal isotropic $l$-planes, $l>k$, then the so-called micro-extension trick can be used to get a smooth (non folded) horizontal immersion $M\fd N$, see [@GromovCC p. 258] (and [@EM Theorem 4.6.1] for a similar statement).
Application to superhorizontal spheres in period domains
--------------------------------------------------------
Using the results of the previous section, we shall now prove that in general period domains have plenty of non-trivial superhorizontal 2-spheres.
### A model case
We begin by studying the period domain $D= \PU(2,n)/{\rm P}(\U(1)\times \U(n)\times \U(1))$. It is an open set in the flag manifold $D_u=\PU(n+2)/{\rm P}(\U(1)\times
\U(n)\times \U(1))={\rm PGL}(n+2,\C)/Q$, where $Q$ is the parabolic subgroup of ${\rm PGL}(n+2,\C)$ stabilizing the flag $\{\la e_1\ra\subset\la e_1,\ldots,e_{n+1}\ra\subset
\C^{n+2}\}$ of $\C^{n+2}$. The complex dimension of $D$ is $2n+1$ and the superhorizontal distribution $H_D$ of $D$ is a holomorphic subbundle of $T_D$ of complex rank $2n$.
We want to construct [*superhorizontal*]{} maps $S^2\fd D$, homotopic to given elements of the second homotopy group $\pi_2(D)$ of $D$. Note that $\pi_2(D)\simeq \Z$, as follows from the discussion in \[bur\]. A map $f:S^2\fd D$ is said to be superhorizontal simply if it is $H_D$-horizontal, namely if for all $x\in S^2$ and $\xi\in T_{S^2,x}$, $f_\star\xi\in H_{D,f(x)}$.
\[petit\] Let $D$ be the period domain $\PU(2,n)/{\rm P}(\U(1)\times \U(n)\times \U(1))$. For $n\geq 2$, all elements of $\pi_2(D)$ can be represented by smooth superhorizontal spheres $S^2\fd D$. For $n\geq 3$, they can be represented by smooth immersed superhorizontal spheres.
By Proposition \[folded\], all we have to prove is that the elements of $\pi_2(D)$ can be lifted to $\pi_2({\rm Gr}^{reg,iso}_{2}(H_D))$, where ${\rm Gr}^{reg,iso}_{2}(H_D)$ is the bundle of Grassmannians of regular isotropic superhorizontal 2-planes over $D$. Note that $H_D\fd D$ is a holomorphic complex vector bundle, but we consider [*real*]{} $2$-planes, namely real spans of couples of $\R$-linearly independent vectors of $H_D$. Regularity and isotropy are defined as before with respect to the 2-form $d\t:H_D\times H_D\fd T_D/H_D$. At the point $eQ\in D\subset D_u$, $H_D$ can be identified with $\C^{2n}=\C^{n}\times\C^n$ and $d\t$ restricts to the complex symplectic form $$\begin{array}{rcl}
q:(\C^{n}\times\C^n)\times (\C^{n}\times\C^n) & \fd & \C \\
((v_1,v_2),(w_1,w_2)) & \longmapsto & \trans v_1 w_2-\trans v_2 w_1
\end{array}$$
Since the superhorizontal 2-planes in $T_D$ are candidates for being the images of the tangent planes of $S^2$ by the differential of a map $f:S^2\fd D$, and since we have a natural orientation on $T_{S^2}$, we will in fact consider the bundle ${\rm Gr}^{reg,iso}_{2,or}(H_D)$ of regular isotropic superhorizontal [*oriented*]{} 2-planes. If we can lift elements of $\pi_2(D)$ to elements of $\pi_2({\rm Gr}^{reg,iso}_{2,or}(H_D))$, then using the natural projection ${\rm Gr}^{reg,iso}_{2,or}(H_D)\fd {\rm Gr}^{reg,iso}_{2}(H_D)$, we are done.
The fiber of the fibration ${\rm Gr}_{2,or}^{reg, iso}(H_D)\fd D$ at $eQ$ is ${\rm Gr}_{2,or}^{reg, iso}$, the Grassmannian of oriented (real) $2$-planes $S$ in $\C^{2n}$, isotropic for the complex symplectic form $q$ and regular in the sense that $$\begin{array}{rcl}
\C^{2n} & \fd & \Hom_\R(S,\C)\\ v & \longmapsto & (\iota_v{q})_{|S}
\end{array}$$ is surjective. It is easy to check that a real $2$-plane $S$ in $\C^{2n}$ is regular if and only if it isn’t a complex line, namely if and only if it is spanned by two $\C$-linearly independent vectors.
The long exact sequence of homotopy groups associated to the fibration is $$\cdots \fd \pi_2({\rm Gr}_{2,or}^{reg, iso}(H_D)) \fd \pi_2(D) \fd \pi_1({\rm Gr}_{2,or}^{reg,iso}) \fd \cdots$$ Therefore the elements of $\pi_2(D)$ which come from elements of $\pi_2({\rm Gr}_{2,or}^{reg, iso}(H_D))$ are those which are in the kernel of the map $\pi_2(D) \fd\pi_1({\rm Gr}_{2,or}^{reg, iso})$.
For $k\leq n$, the $q$-symplectic group ${\rm Sp}(n,\C)$ acts transitively on the set of $k$-tuples of $\C$-linearly independent vectors spanning a $k$-dimensional complex isotropic subspace in $\C^{2n}$, and the stabilizer $Q_k$ of such a $k$-tuple is topologically ${\rm Sp}(n-k,\C)\times\C^{2k(n-k)}\times\C^{{k(k+1)}/{2}}$, as a simple computation shows.
Therefore for $n\geq 2$ the stabilizer in ${\rm Sp}(n,\C)$ of an oriented regular isotropic $2$-plane is isomorphic to $\GL^+(2,\R)\times Q_2$. Hence ${\rm Gr}_{2,or}^{reg,iso}\simeq{\rm Sp}(n,\C)/(\GL^+(2,\R)\times Q_2)$. Since ${\rm Sp}(n,\C)$ is path-connected and simply connected, $\pi_1({\rm Gr}_{2,or}^{reg,iso})\simeq \pi_0(\GL^+(2,\R)\times Q_2)$. Now $\GL^+(2,\R)\times Q_2$ is path-connected, hence ${\rm Gr}_{2,or}^{reg,iso}$ is simply connected and all the elements of $\pi_2(D)$ can be lifted to elements of $\pi_2({\rm Gr}_{2,or}^{reg,iso}(H_D))$.
If $n\geq 3$, we may also look at flags of oriented $2$-planes in oriented superhorizontal regular isotropic $3$-planes: ${\rm Gr}_{2,3,or}^{reg,iso}\simeq{\rm Sp}(n,\C)/((\GL^+(2,\R)\times\R_+^\star\times\R^2)\times Q_3)$ and find that all elements of $\pi_2(D)$ can be lifted to elements of $\pi_2({\rm Gr}_{2,3,or}^{reg,iso}(H_D))$. This implies that for $n\geq 3$, all the elements of $\pi_2(D)$ can be represented by smooth immersed superhorizontal spheres, see Remark \[remarque2\].
\[n=1\] If $n=1$, one sees from the proof above that there are no regular isotropic horizontal real $2$-planes (indeed in this case a real 2-plane $S$ is isotropic if and only if it is a complex line if and only if it is not regular), hence the $h$-principle doesn’t apply (at least in its present form).
### Second homotopy groups of $\PU(p,q)$ period domains and flag manifolds {#bur}
Keeping the notation of Section \[var\], let $D$ be the $\PU(p,q)$ period domain associated to some variation of Hodge structure $E=\bigoplus_{i=0}^k E^i$ over a compact Kähler manifold $M$ and let $D_u=\PU(p+q)/{\rm P}(\Pi_{i=0}^k{\rm U}(r_i))$ be its dual.
We first sum up some results of [@BR Chap. 4.D]. Let $\gg^\a\subset\ng$ be a root-space (i.e. $\a\in\Phi$ and $-\a\not\in\Phi$). Then $\gg^\a\oplus\gg^{-\a}\oplus[\gg^\a,\gg^{-\a}]$ is a Lie algebra isomorphic to $\sg\lg(2,\C)$ which is the complexification of a subalgebra $\sg_\a$ of $\gg_u$ isomorphic to $\sg\ug(2)$. Let $S_\a$ be the corresponding copy of $\PU(2)$ and $U_\a\subset S_\a$ the connected subgroup with Lie algebra $\ug_\a=[\gg^\a,\gg^{-\a}]\cap\gg_0\subset\vg_0$. Then $U_\a$ is a copy of $U(1)$ so that $S_\a/U_\a$ is a 2-sphere. The homomorphism $i_\a:S_\a\fd G_u$ corresponding to the inclusion $\sg_\a\subset\gg_u$ defines an $i_\a$-equivariant immersion $\varphi_\a:\P^1\fd D_u$ (since $i_\a(U_\a)\subset V_0$) which is moreover totally geodesic and holomorphic.
The classes $[\Sigma_\a]$ of the spheres $\Sigma_\a:=\varphi_\a(\P^1)$ (with $\a$ as before) generate $\pi_2(D_u)$ and moreover, if $\a$ and $\b$ are such that $\a+\b$ is also a root, then $[\Sigma_{\a+\b}]=[\Sigma_\a]+[\Sigma_\b]$ in $\pi_2(D_u)$. Also, if there exists $\g\in\Phi\cap-\Phi$ such that $\a=\b+\g$ then $[\Sigma_\a]=[\Sigma_\b]$. Finally, we remark that the spheres $\Sigma_\a$ are superhorizontal for each $\a\in\Pi_\qg$. Then (see [@BR]),
\[pitwo\] $\pi_2(D_u)\simeq\Z^{k}$ and a basis of this $\Z$-module is given by $\{[\Sigma_\a]\,|\,\a\in\Pi_\qg\}$.
Define $\beta_{i}:=\a_{r_0+\dots+r_i}\in\Pi_\qg$ and for any $i< j$, set $\beta_{i,j}:=\sum_{s=r_0+\dots+r_i}^{r_0+\dots+r_j}\a_s$. Consider now the projection map $\pi_u:D_u\fd X_u$. For any $\a\in\Pi_\qg$, its restriction ${\pi_u}_{|\Sigma_\a}$ has degree one. Moreover, $\pi_2(X_u)\simeq\Z$ and for the right choice of a generator $[T]$, $[\pi_u(\Sigma_{\b_{j}})]=(-1)^j[T]$ (in fact, ${\pi_u}_{|\Sigma_{\b_{j}}}$ is alternately holomorphic and antiholomorphic). As a consequence, the morphism ${\pi_u}_\star$ induced by $\pi_u$ on the second homotopy groups is given by $$\begin{array}{rcl}
{\pi_u}_\star : \pi_2(D_u)\simeq \Z^k & \fd & \pi_2(X_u)\simeq\Z\\
\sum_{i=0}^{k-1} a_{\b_{i}}[\Sigma_{\b_{i}}] & \longmapsto & \sum_{i=0}^{k-1} (-1)^i a_{\b_{i}} [T]\;\;.
\end{array}$$ The central fiber $F$ of $\pi_u:D_u\fd X_u$ is ${\rm P}({\rm U}(p)\times {\rm U}(q))/{\rm P}(\Pi_{i=0}^k{\rm U}(r_i))$ which is nothing but ${\PU}(p)/{\rm P}(\Pi_{i\ {\rm even}}{\rm U}(r_i))\,\times\, {\PU}(q)/{\rm P}(\Pi_{i\ {\rm odd}}{\rm U}(r_i))$. It is hence isomorphic to the product of two flag manifolds: applying Proposition \[pitwo\], we get $\pi_2(F)\simeq\Z^{k-1}$ and by the above combinatorial description, it is easily seen that $\pi_2(F)\simeq{\rm Ker}\,{\pi_u}_\star$, a basis of whom is given by $$\bigl\{[\Sigma_{\b_{i}}]+[\Sigma_{\b_{i+1}}]\,|\,i=0,\dots,k-2 \bigr\}=\bigl\{[\Sigma_{\b_{i,i+1}}]\,|\,i=0,\dots,k-2 \bigr\}\;\;.$$ Remark that each representative of the latter basis is contained in the central fiber of $\pi_u$.
Since $F$ is also the central fiber of the fibration $\pi$ of the period domain $D$ over the noncompact symmetric space $X$, it is a deformation retract of $D$ and we deduce the
$\pi_2(D)\simeq\pi_2(F)\simeq\Z^{k-1}$.
### Proof of Theorem \[horiz\]
Proposition \[petit\] and the results of Section \[bur\], enable us to state the following theorem, of which Theorem \[horiz\] is a particular case:
\[hori\] Let $E=\oplus_{i=0}^k E^i$ be a $\C$-VHS and set $r_i=\rk E^i$. Let $D$ be the $\PU(p,q)$ period domain associated to $E$. Then, for each $i\in\{0,\dots,k-2\}$ such that $r_{i+1}\geq 2$, there exists a superhorizontal representative $\psi:S^2\fd D$ of $[\Sigma_{\b_{{i},{i+1}}}]$. In particular, when the hypotheses of Theorem \[cartoun\] are not met, $\pi_2(D)$ can be generated by superhorizontal 2-spheres.
If $r_{i+1}\geq 2$, $\{\b_i,\b_{i,i+1},\b_{i+1}\}$ are the simple roots of a Lie subalgebra of $\sg\lg(\E)$ isomorphic to $\sg\lg(4,\C)$ which is the complexification of a subalgebra of $\gg_0$ isomorphic to $\sg\ug(2,2)$. As in the case of $\P^1$ in Section \[bur\], this defines a totally geodesic embedding of the period domain ${\mathcal D}=\PU(2,2)/{\rm P}(\U(1)\times\U(2)\times\U(1))$ in $D$. Geometrically, if $(e^i_1,\ldots,e_{r_i}^i)$ is a basis of $\E^i$, and if ${\mathbb W}=\la e^{i}_{r_{i}}\ra\oplus\la e_1^{i+1},e_{r_{i+1}}^{i+1}\ra\oplus\la e^{i+2}_1\ra$, then $\PU({\mathbb W},h_{|{\mathbb W}})\subset \PU(\E,h)$ is the copy of $\PU(2,2)$ corresponding to this embedding.
Identifying ${\mathcal D}$ with its image in $D$, the superhorizontal distribution of ${\mathcal D}$ satisfies $H_{{\mathcal D}}\subset {H_{D}}_{|{\mathcal D}}$ and we have $\Sigma_{\b_{i,i+1}}\subset {\mathcal D}$. Hence we can apply Proposition \[petit\] in ${\mathcal D}$.
For a general period domain $D=G_0/V_0$, we can use this method to try to produce superhorizontal representatives of some elements of $\pi_2(D)$. Let us rephrase the proof we just gave. The group $\pi_2(D)$ is generated by vertical spheres. More precisely, using the notation of the end of section \[perdo\], if $\gg^\a\subset\gg_2\subset\gg$ is a root-space and $\gg^\a\oplus\gg^{-\a}\oplus[\gg^\a,\gg^{-\a}]$ is the corresponding copy of $\sg\lg(2,\C)$ in $\gg$, then exponentiating its real form $\sg\ug(2)$ gives a 2-sphere $\Sigma_\a$ in $D$, and these spheres generate $\pi_2(D)$. Now, if there is a Lie algebra homomorphism $\sg\lg(4,\C)\fd \gg$ which maps the root space $\sg\lg^{e_1-e_4}$ of $\sg\lg(4,\C)$ to $\gg^\a$ and the root spaces $\sg\lg^{e_1-e_2}$, $\sg\lg^{e_1-e_3}$, $\sg\lg^{e_2-e_4}$ and $\sg\lg^{e_3-e_4}$ to $\gg_1\subset\gg$, then as before there will be a copy of the period domain $\PU(2,2)/{\rm P}(\U(1)\times\U(2)\times\U(1))$ in $D$, with the right superhorizontality property and containing $\Sigma_\a$. Therefore Proposition \[petit\] yields a superhorizontal representative of the generator corresponding to the root-space $\gg^\a\subset \gg_2$.
It is difficult to give a general criterion for the existence of these special homomorphisms $\sg\lg(4,\C)\fd \gg$. We list a few examples (there are many others) where the method can be applied at least partially, for classical matrix period domains of the type considered by Griffiths. In this case, $G_0$ is either ${\rm Sp}(n,\R)$ or $\SO(p,q)$ with $p$ or $q$ even.
1. for $D=\SO(2p,q)/(\U(p)\times\SO(q))$ with $p\geq 2$ and $q\geq 2$, $\pi_2(D)\simeq \Z$ can be generated by a superhorizontal sphere;
2. for $D=\SO(4,4)/(\U(1)\times\U(2)\times\SO(2))$, $\pi_2(D)\simeq \Z^3$ can be generated by superhorizontal spheres;
3. for $D=\SO(2p,2q_1+q_2)/(\U(q_1)\times\U(p)\times\SO(q_2))$ with $p\geq 2$, $q_1\geq 1$ and $q_2\geq 3$, $\pi_2(D)\simeq \Z^2$ can be generated by superhorizontal spheres;
4. for $D={\rm Sp}(n,\R)/(\U(k)\times\U(n-2k)\times\U(k))$ with $k\geq 1$ and $n-2k\geq 2$, $\pi_2(D)=\Z^2$ and (at least) one generator can be taken to be a superhorizontal sphere;
5. $D={\rm Sp}(n,\R)/(\U(k)\times\U(n-k))$ with $k\geq 1$ and $n-k\geq 3$, $\pi_2(D)\simeq \Z$ can be generated by a superhorizontal sphere.
Concerning the possibility of generating the second homotopy groups of $\PU(p,q)$ period domains by horizontal spheres, there is one question which remains unanswered. In the notation of Theorem \[hori\], we do not know whether $[\Sigma_{\b_{i,i+1}}]$ can be represented by a superhorizontal 2-sphere when $r_{i+1}=1$.
To take the simplest example, it would be very interesting to know what happens for the period domain $D={\rm PU}(2,1)/{\rm P}(\U(1)\times\U(1)\times \U(1))$, whose second homotopy group is isomorphic to $\Z$. As pointed out in Remark \[n=1\], our implementation of the $h$-principle in ${\rm PU}(2,n)/{\rm P}(\U(1)\times \U(n)\times \U(1))$ fails when $n=1$.
Nevertheless, the arguments of the proof of Theorem \[cartoun\] imply that for any period domain $D$ of the form ${\rm PU}(p+q,1)/{\rm P}(\U(p)\times \U(1)\times \U(q))$ (with $p\geq 1$, $q\geq 1$), and for any holomorphic superhorizontal map $f:\tilde M\fd D$, we have $f_\star(\pi_2(M))=0$.
Indeed, let us consider the holomorphic projection $\nu:D\fd D':={\rm PU}(p+q,1)/{\rm P}(\U(p)\times \U(q,1))$ and the projection onto the symmetric space $\pi:D\fd X:={\rm PU}(p+q,1)/{\rm P}(\U(p+q)\times \U(1))$. Then, for any 2-sphere $\sigma:S^2\fd \tilde M$, $\nu\circ f\circ\sigma:S^2\fd D'$ and $\pi\circ f\circ\sigma:S^2\fd X$ are homotopically trivial (the former by the arguments of Section \[proofofcartoun\], the latter because $X$ is contractible). Let $\tau_1:S^2\times [0,1]\fd D'$ be a homotopy between $\nu\circ f\circ\sigma$ and a constant map and $\tau_2:S^2\times [0,1]\fd X$ be a homotopy between $\pi\circ f\circ\sigma$ and a constant map. For any $(s,t)\in S^2\times [0,1]$, the flag $\{0\subset\tau_1(s,t)\subset \tau_1(s,t)\oplus\tau_2(s,t)^\perp\subset\C^{p+q+1}\}$ is an element of $D$ (orthogonality is understood with respect to the indefinite Hermitian form of signature $(p+q,1)$ on $\C^{p+q+1}$), although $\tau_1(s,t)$ and $\tau_2(s,t)^\perp$ are not necessarily orthogonal. In this way we get a homotopy between $f\circ\sigma$ and a constant map.
[10 mm]{} [\[abckt\]]{} J. Amor[ó]{}s, M. Burger, K. Corlette, D. Kotschick, D. Toledo, Mathematical Surveys and Monographs 44, American Mathematical Society, Providence, RI, 1996
[\[BR\]]{} F. E. Burstall and J. H. Rawnsley, Lecture Notes in Mathematics 1424, Springer-Verlag, Berlin, 1990
[\[EM\]]{} Y. Eliashberg and N. Mishachev, Graduate Studies in Mathematics 48, American Mathematical Society, Providence, RI, 2002
[\[FHW\]]{} G. Fels, A. Huckleberry and J. A. Wolf, Progr. Math. 245, Birkhäuser Boston, Inc., Boston, MA, 2006
[\[Gra\]]{} H. Grauert, Math. Ann. [**265**]{}, 1983, 137-148
[\[Gri\]]{} P. Griffiths, Inst. Hautes Études Sci. Publ. Math. [**38**]{}, 1970, 125-180
[\[GS\]]{} P. Griffiths and W. Schmid, Acta Math. [**123**]{}, 1969, 253-302
[\[GromovPDR\]]{} M. Gromov, Ergeb. Math. Grenzgeb. (3) 9, Springer-Verlag, Berlin, 1986
[\[GromovCC\]]{} M. Gromov, Sub-Riemannian geometry, Progr. Math. 144, Birkhäuser, Basel, 1996, 79-323
[\[ko\]]{} J. Kollár, M. B. Porter Lectures, Princeton University Press, Princeton, NJ, 1995
[\[Pansu\]]{} P. Pansu, Preprint, 2006
[\[Re\]]{} A. Reznikov, Motives, polylogarithms and Hodge theory, Part II (Irvine, CA, 1998), 717–730, Int. Press Lect. Ser., 3, II, Int. Press, Somerville, MA, 2002
[\[Sim\]]{} C. T. Simpson, Inst. Hautes Études Sci. Publ. Math. [**75**]{}, 1992, 5-95
[\[si4\]]{} C. T. Simpson, Different faces of geometry, [*Int. Math. Ser.*]{} (N. Y.), Kluwer/Plenum, New York, 2004, 365-402
[\[to\]]{} D. Toledo, Publ. Math. IHES [**77**]{}, 1993, 103-119.
[\[Voisin\]]{} C. Voisin, Invent. Math. [**157**]{}, 2004, 329-343
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Dark Energy is some of the weirdest and most mysterious stuff in the universe that tends to increase the rate of expansion of the universe. Two commonly known forms of dark energy are the cosmological constant, a constant energy density filling space homogeneously, and scalar fields such as quintessence or moduli whose energy density can vary with time. We explore one particular model for dynamic dark energy; quintessence driven by a scalar dilaton field. We propose an ansatz for the form of the dilaton field, $|\phi(a)|/m\Z{P} \equiv \alpha\Z1 \ln t+ \alpha\Z2
t^n=\alpha\ln a+ \beta\, a^{2\zeta}$, where $a$ is the scale factor and $\alpha$ and $\zeta$ are parameters of the model. This phenomenological ansatz for $\phi$ can be motivated by generic solutions of a scalar dilaton field in many effective string theory and string-inspired gravity models in four dimensions. Using a compilation of current data including type Ia supernovae, we impose observational constraints on the slope parameters like $\alpha$ and $\zeta$ and then discuss the relation of our results to analytical constrains on various cosmological parameters, including the dark energy equation of state. Sensible constraints are imposed on model parameters like $\alpha$ and $\zeta$ as well as on the dark energy/dark matter couplings using results from structure formation. The constraints of this model are shown to encompass the cosmological constant limit within $1\sigma$ error bars.
author:
- 'Ishwaree P. Neupane[^1] and Holly Trowland[^2]'
title: Constraining the runaway dilaton and quintessential dark energy
---
Introduction\[intro\]
=====================
The discovery of a cosmic acceleration of the universe [@supernovae; @sn2; @sn3; @sn4] has added a new challenge for fundamental theories of physics and cosmology. NASA’s observations [@NASA; @WMAP1; @WMAP2] show that the kind of matter of which stars and galaxies are made forms less than $5\%$ of the universe’s total mass. Several independent observations indicate that about $73\%$ of the total energy density of the universe is in the form of mysterious dark energy or gravitationally repulsive energy, and about $22\%$ is in the form of non-baryonic cold dark matter particles which clump gravitationally, but which have never been directly detected. These scientific enigmas suggest we should look to new physics beyond the theories of the standard model of particle physics and Einstein’s theory of General Relativity.
The cosmic acceleration of the universe is often attributed to the existence of some form of dark energy. Understanding the nature of this dark energy is one of the most challenging theoretical problems facing present-day cosmology. There is no shortage of ideas for what dark energy might be, from the quantum vacuum to a (ultra-) light scalar field, e.g. coming from modified f(R) gravity [@Carroll:2003; @Capozziello:2003gx; @Nojiri:2003], scalar-tensor theories [@Peebles:88A; @CW; @ZWS; @Elizalde:2008], string-inspired cosmologies [@Ish05a; @CN05; @CN05b; @Ish06d; @TS06; @Ish06b], or space-time fluctuations of a holographic origin [@Miao-Li; @Wei:2007ty; @ISH07a; @ISH07b] and the likes. In some other approaches, cosmic acceleration is a manifestation of new gravitational physics, which may involve modification of gravity at long distances [@DDG], cosmology with extra dimensions [@Ish03-g; @CHNW; @Ish03f; @Ohta03b], or warped extra dimensions [@Ish09a; @Ish09b]. There are also proposals which explain the supernova observations using gravitational effects of density inhomogeneities [@Wiltshire:2007; @DW:07b]. The majority of work in the subject of cosmic acceleration, as well as the reviews of dark energy [@dark-review; @Sahni:04; @Sahni04c; @CST; @Sami:09dk], has shown that any viable theory of gravity must look in the far infrared, or on astrophysical scales, like a combination of scalar fields and gauge fields (weakly) interacting with General Relativity. Moreover, since the cosmological backgrounds of interest are time-dependent, for the consistency of any dynamical dark energy model or theory of cosmic acceleration, such scalar fields must roll but only slowly as time progresses (see Ref. [@Guo-Ohta] for some related discussion).
The $\Lambda$CDM model is the simplest known cosmological model for concurrent universe where it is assumed that most matter is in the form of cold dark matter, and dark energy is simply described by the vacuum energy $<T_{\mu\nu}>\Z{\rm vac} \propto \Lambda
g_{\mu\nu}$ in Einstein’s field equations $$R^{\mu\nu}-\frac{1}{2}R g^{\mu\nu} =8{\pi}G \left( T^{\mu\nu}_{\rm
matter} + T^{\mu\nu}_{\rm vac} \right). \label{einstein field}$$ The value of the constant $\Lambda$ is [*not*]{} set by the theory [@Weinberg:1988; @Sahni02a; @Paddy02], so it may be chosen to match with observations. Vacuum energy does not vary with space or time and is not dynamical. This solution seems to be the best-fit to most observations, especially in regard the dark energy equation of state, but it does lead to a couple of extremely unlikely scenarios leading us to question the validity of using a cosmological constant to describe dark energy. The first scenario is the “Why now?" problem. At the present epoch the universe is composed of approximately $27\%$ matter (both baryons and dark matter) and $73\%$ dark energy. If we lived in a slightly earlier or later time these values would be different by several orders of magnitude, so how come they are comparable at the current epoch? If the dark energy and matter density were similar at some earlier epoch, cosmic acceleration would have began earlier and there would not have been enough time to form structures, including our galaxy.
It is intuitive to think of the cosmological constant as a vacuum energy or the energy contained in free space. As the universe expands, more space is created so there is more gravitational vacuum energy or free space. As a result the energy density due to a cosmological constant remains the same throughout time. But, since a quantum field must be quantized at each and every point in space, it is also useful to think of a vacuum as not just empty space but containing a simple harmonic oscillator at each point. Any vibrations in this field space can propagate through space as a wave. In the absence of ordinary matter, the hypothetical simple harmonic oscillators will be in the ground state which has non-zero vacuum energy. Integrating the ground state energy over all space and renormalising it gives us the vacuum energy density. The theoretical value predicted in this way is, however, larger by a huge factor of at least $10^{15}$ than the mass scale associated with the observed value of gravitational vacuum energy density, which is $\rho\Z\Lambda^{1/4}\simeq 10^{-3}~{\rm eV}$. Although the quantity $\rho\Z\Lambda$ may conceivably be linked with small numbers such as the neutrino mass ($m_\nu \sim
\rho_\Lambda^{1/4}$), or the ratio $\rho\Z{\rm EW}/\rho\Z{\rm
Planck}\sim \rho\Z\Lambda$, a firm theoretical prediction for the value of $\rho\Z\Lambda^{1/4}\sim 10^{-3}~{\rm eV}$ is currently lacking.
The theory confronting a cosmological constant as the source of dark energy is generally called “quintessence". In models of quintessence, one assumes that dark energy is dynamical in nature instead of being a constant in time and space by utilizing a standard scalar field. In all modern unified theories including Kaluza-Klein theory, supergravity and superstring, a fundamental scalar field called the [*dilaton*]{} exists along with a spin-2 graviton – the messenger of gravitational force. In extra-dimensional theories of gravity, such a scalar field can be the radion field that encodes the distance between the two branes (or brane-antibrane pairs) in the fifth dimension. There can be many such scalar fields in higher dimensional braneworld theories that may cause cosmic acceleration if they were to self interact and generate a positive potential. A canonical scalar field $\phi$ having large self interacting potential but almost negligibly small kinetic term behaves almost as Einstein’s cosmological constant, at least, from a viewpoint of its equation of state.
In addition to exhibiting an effect on the rate of expansion of the universe, a scalar field quintessence, if it interacts with dark matter, may lead to an observable effect on how the large scale structure is growing. Recent observations show that clusters of galaxies may have formed as a result of small density fluctuations in the primordial plasma when tiny quantum fluctuations were amplified by inflation to form pockets of space that were more dense than the average energy density. Once matter begins to dominate, gravitational collapse would amplify the tiny inhomogeneities left by inflation and visible structures begin to grow into the already formed dark matter halos. In a scenario where dark energy or a scalar quintessence field interacts with dark matter, the rate of growth of the dark matter halos would be affected, and thus the formation of visible structures.
Another important motivation for considering a dynamical scalar field is an essence of one or more dynamical energy components in the early universe: so called “inflaton”. This inflaton field could easily release its potential energy to matter and radiation as the field rolls down to its lowest (or metastable) energy state. It is quite plausible that the late time quintessential scalar field (together with dilatonic dark matter) is a remnant of the inflaton field (see, e.g. [@Ish:07a]) although there is no guarantee they are the same.
In this paper we demand a specific form for the scalar field itself, which is indeed an approximation to some generic solutions for a dilaton in some effective string theory models in four dimensions. Our choice is motivated from the fact that a similar ansatz would be useful when one is considering the early inflation of the universe. Starting with the standard form of scalar-tensor theory of gravity, we will derive field equations that depend on both the scalar field and the coupling of that field to dark matter. We impose constraints on the parameters by fitting the model to observational data from the the cosmic microwave background (CMB) shift[@Komatsu:08a], baryon acoustic oscillations (BAO) [@Eisenstein:05a], SNIa (Gold+HST sample) [@Riess:06a] and SNLS [@SNLS]. Observational results from CMB can be used to tightly constrain many useful cosmological parameters, including the parameters of inflation.
The growth of structure is also expected to depend on the way that dark energy is coupled to dark matter, granting yet another constraint. Our focus in this work is to explore ways to constrain the parameters of the solutions, which include $\chi^2$ curve fittings to the WMAP+BAO+SNIa+SNLS data sets. Also, we will consider how the links between quintessential dark energy and inflation may constrain the parameters of the model. Finally we discuss on how the results on structure formations constrain the model parameters and the matter fluctuation growth rate. Using those constraints we show that the solutions of the model encompass the cosmological constant limit within $1\sigma$ error bars.
Quintessence: Theoretical motivation
====================================
Einstein formulated gravity as a tensor theory, where the field equations of general relativity can be obtained by varying the action $$\label{einstein-action1}
S= \frac{1}{2} \int \sqrt{-g}\, d^4{x}\, R + \int \kappa^2 {\cal
L}_m [\Psi_m, g_{\mu\nu}] \sqrt{-g} d^4x,$$ where $\kappa{\equiv}\,m_P^{-1}=\sqrt{8{\pi}G}$ is the inverse of the Planck mass and $G$ is Newton’s gravitational constant. $\Psi_m$ denotes the (ordinary plus) matter. The first integral is the gravitational part of the action and the second integral is the matter part, and $\mathcal{L}_m$ is the matter Lagrangian. Varying this action with the metric $g_{\mu\nu}$ gives Einstein’s field equations $$\label{field2}
\frac{1}{\kappa^2}\left({R_{\mu\nu}-\frac{1}{2}R\,g_{\mu\nu}}\right)
=-\frac{2}{\sqrt{-g}}
\frac{\delta\mathcal{L}_m}{\delta{g^{\mu\nu}}}\equiv{T_{\mu\nu}},$$ where $T_{\mu\nu}$ is the energy momentum tensor describing the density and flux of energy.
Dirac was the first to introduce a scalar field into Einstein field equations, suggesting that the Newton’s constant could be viewed as a time-dependent parameter. Thereafter Jordan and, independently, Brans and Dicke introduced a monomial scalar field known as Brans-Dicke dilaton [@BD], which also assumes a time varying gravitational constant and incorporates a scalar field into Einstein’s tensor theory. It was motivated by Mach’s principle, that “mass there influences inertia here". This was interpreted by Dicke to mean [*the gravitational constant should be a function of the mass distribution of the universe*]{}, implying that $\frac{1}{G}\propto\frac{M}{R}$, or that a scalar field somewhat plays the role of the inverse of Newton’s constant, $\frac{1}{G}\propto\varphi$. Unfortunately, the Brans-Dicke dilaton, with the action $$\label{action-Jordan}
S= \frac{1}{16\pi} \int \sqrt{-g}\, d^4{x} \left( \varphi {R} +
\frac{\omega}{\varphi}\, \partial_i\varphi\,
\partial^i\varphi \right)
+\int \sqrt{-g}\,
d^4x\, \mathcal{L}_m [g_{\mu\nu}, \Psi_m],$$ which is without a field potential for $\varphi$, is unacceptable since a massless scalar field could easily create a long range fifth force – which, however, does not seem to exist in nature. One can therefore modify the Brans-Dicke type gravitational action as $$\label{action-Jordan2}
S=\frac{1}{16\pi G} \int \sqrt{-g}\, d^4{x} \left( f(\varphi) {R}+
{\mathcal{L}}_\varphi\right)
+\int \sqrt{-g}\,
d^4x\, \mathcal{L}_m [g_{\mu\nu}, \Psi_m],$$ where ${\mathcal{L}}_\varphi$ is a scalar field Lagrangian in Jordan frame. In the particular picture of Brane-Dicke model, one defines $f(\varphi)\propto \varphi $. This is a special case of a general scalar-tensor theory in the Jordan-Fierz frame, where $\varphi$ is directly coupled to the Ricci scalar, see, e.g. Refs. [@Cho:1992; @Cho07].
By using a canonical transformation, we can write the action (\[action-Jordan\]) in the Einstein frame $$S=\int\left(\frac{R}{2\kappa^2}+\mathcal{L}_{\phi}\right)
\sqrt{-g}\,d^4{x}+{\int} \mathcal{L}_m[\widehat{g}_{\mu\nu},
\Psi_m] \sqrt{-\widehat{g}}\,d^4{x}. \label{my action}$$ The scalar field Lagrangian ${\cal L}_\phi$ is assumed to have the standard form $$\label{scalar l}
\mathcal{L}_\phi=-\frac{\gamma}{2}\,g^{\mu\nu}\partial_\mu{\phi}\,
\partial_\nu{\phi}-V({\phi}).$$ The Einstein frame metric $\widehat{g}_{\mu\nu}$ is now related to the Jordan frame metric $g_{\mu\nu}$ via $\widehat{g}_{\mu\nu}\equiv A(\phi)^2 g_{\mu\nu}$, where $A(\phi)$ is a function of $\phi$. One has $\gamma = \pm 1$ depending on whether the scalar field is the real or imaginary part of some complex modulus or axio-dilaton field present in some higher dimensional theories of gravity. Models similar to the one here were studied before, see e.g. Refs. [@Farese:2000ij; @Amendola:1999A], but here we do not constrain the form of the scalar field potential rather the evolution of the quintessence scalar field $\phi$. Moreover, we would assume that $\phi$ is a canonical scalar field ($\gamma=1$) in Einstein frame and it is coupled to both the ordinary and dark matter particles but with different gravitational couplings.
To generalize the action (\[my action\]) further, we assume that there is a number of different species of matter, including cold dark matter and baryons – say $j$ components – each coupling to the metric differently with coupling function, $A_j(\phi)$. As the simplest possibility, we choose exponential couplings $$A_j(\phi) \propto e^{\kappa Q_j\,\phi},$$ where, in general, $Q_j$ can be a function of $\phi$. The matter Lagrangian is now the sum of all components, accounting for all the different sectors: $${\cal L}_m[\tilde{g}_{\mu\nu}, \Psi_m^{(j)}]\equiv \sum_j {\cal
L}_j [e^{2\kappa Q\Z{j}\, \phi} g_{\mu\nu}, \Psi_m^{(j)}].$$ We are considering here a scenario in which a purely dark sector interaction exists, resulting from a nonminimal coupling of dark matter to a fundamental scalar field or quintessence. Such couplings give rise to additional forces on dark matter particles in addition to gravity.
In the early 1990s, Damour, Gibbons and Gundlach [@Damour:1990] showed that a cosmological model with two sectors – cold dark matter with coupling $Q_{c}(\phi)$, and normal (baryonic) matter with coupling $Q_b(\phi)$ – satisfies both the weak equivalence principle (WEP) and constraints from Solar System observations, provided that $$\label{solar-constr}
m\Z{P} \left\vert\frac{d Q\Z{b}(\phi)}{d\phi} \right\vert
\Z{\phi_0} \lesssim 10^{-2},$$ where $\phi\Z0$ is the present-day cosmological background value of $\phi$. On the other hand, since the nature of both dark energy and dark matter are still unknown, there is no physical argument that excludes a possible interaction between them. Furthermore, the magnitude of the baryonic coupling is constrained from radar time-delay measurements, $|Q_b|<0.032$ which limits the cold dark matter coupling to $|Q\Z{CDM}|<1$.
In this paper, we specialize to models where the couplings are nonnegligible only in the dark sectors, in which case the theory satisfies the WEP. As long as the scalar field is coupled non-minimally only to dark matter, the model automatically satisfies the solar system constraint [@Damour:1996; @Ish07; @Neupane07], see also Refs. [@Bean:2008ac; @Gavela1] for some related discussions.
Let us assume that the spatial curvature of our universe was erased with inflation in the early universe and choose a spatially flat Friedmann-Robertson-Walker metric, $$ds^2=-dt^2+a(t)^2 d\mathbf{x}^2,$$ where $a(t)$ is the scale factor, normalized such that $a\equiv
a_0=1$ at present, i.e. at $t=t_0$. This metric assumes the universe is homogeneous and isotropic on largest scales, which is a good approximation at large scales and is consistent with CMB observations [@WMAP1]. For simplicity, we also assume that all forms of matter, including dark energy, behave as a perfect fluid. Each form of matter may be characterized by the fluid equation of state, ${\rm w}_i\equiv {p_i}/{\rho_i}$. In particular, the cosmological constant may be thought of as a fluid with the equation of state ${\rm w}_{\Lambda}\equiv
p\Z{\Lambda}/\rho\Z{\Lambda} =-1$. Other forms of matter or energy are assumed to be perfect fluids as well, for radiation or relativistic particles, ${\rm w}_\gamma=-1/3$ and for baryons or cold dark matter ${\rm w}_m\simeq0$. The equation of state for a scalar quintessence is then defined as $${\rm
w}_\phi\equiv\frac{p_\phi}{\rho_\phi}=\frac{\frac{\gamma}{2}\dot{\phi}^2
-V(\phi)}{\frac{\gamma}{2}\dot{\phi}^2+V(\phi)},$$ which varies with time.
The variation of the action (\[my action\]) with respect to $g^{\mu\nu}$ and $\phi$ gives $$\begin{aligned}
-\frac{3}{\kappa^2} \, H^2+\frac{\gamma}{2} \, \dot{\phi}^2 +
V\left(\phi\right)+A^4\left(\phi\right)\,\sum_{i}\rho_i=0,
\label{vary1}\\
\frac{1}{\kappa^2}\,\left(2\,\dot{H}+3\,H^2\right)
+\frac{\gamma}{2}\,\dot{\phi}^2-V\left(\phi\right)
+A^4\left(\phi\right)\,\sum_{i}\left({\rm w}_i\,\rho_i\right)=0,
\label{vary2}\\
\gamma\,\left(\ddot{\phi}+3\,H\,\dot{\phi}\right)
+\frac{dV\left(\phi\right)}{d\phi}-
A^3\,\frac{dA\left(\phi\right)}{d\phi}\,\sum_{i}\left(1-3\,{\rm
w}_i\right)\,\rho_i=0, \label{vary3}\end{aligned}$$ where $H(t)\equiv \dot{a}/a$ is the Hubble parameter. For the derivation of these equations, see [@Ish07; @Neupane07; @Ish0712]. In (\[vary3\]), it can be seen that $\phi$ couples to the trace of the stress-energy tensor, $$T^\mu_{\;\mu}=g^{\mu\nu}T_{\mu\nu}=\sum_{i}\left(1-3\,{\rm
w}_i\right)\,\rho_i.$$ For radiation (${\rm w}_\gamma=1/3)$, the stress-energy tensor is traceless, so there is no interaction between dark energy and radiation. Equations (\[vary1\])-(\[vary3\]) can be supplemented by a fourth equation arising from the conservation of energy-momentum tensor, which reads in the Einstein frame as, $$\label{cons}
{\hat{T}}^\mu_{\;(m)\nu ;\,\mu}+{\hat{T}}^\mu_{\;(\phi)\nu
;\,\mu}=0,$$ where the semicolon represents a covariant derivative which may be expanded in terms of the Christoffel symbol as $\hat{T}^\mu_{\;\nu
;\mu}=\hat{T}^\mu_{\;\nu,\mu}+\hat{\Gamma}^\mu_{\;\alpha\mu}
\hat{T}^\alpha_{\;\nu}-\hat{\Gamma}^\alpha_{\;\nu\mu}
\hat{T}^\mu_{\;\alpha}$. An observer unaware of dark energy looking out at the universe would come to the conclusion that energy is not conserved. That is not the case, however; it is the combination of matter and dark energy that is conserved when there are interactions in cosmology’s dark sector. Summing the terms and remembering that $\hat{T}^\mu_{\;\nu}$ is diagonal and the relevant Christoffel symbols are $ \hat{\Gamma}^0_{\; 0\,0}=0$, and $\hat{\Gamma}^1_{\;0\,1}=\hat{\Gamma}^2_{\;0\,2}
=\hat{\Gamma}^3_{\;0\,3}=\hat{H}$ gives the conservation equation for a perfect fluid $$\label{energy cons}
\dot{\hat\rho}_i+3\,\hat{H}\hat{\rho}_i\left(1+{\rm w}_i\right)=0,$$ where $\hat{\rho}_i\propto (a\,A(\phi))^{-3 (1+{\rm w}\Z{i})}$. The fractional energy densities are defined as $$\Omega_i=\frac{\rho\Z{i}}{\rho\Z{\rm crit}}=\frac{8\pi
G\rho\Z{i}}{3 H^2},$$ so that for each component $0\leq\Omega_i\leq1$. Using the expressions $\hat{\rho}_i=A^4\rho_i$ and $\hat{a} =A(\phi)a$ and carefully considering the derivatives involved, the fluid equation becomes $$\label{vary4}
\dot{\rho_i}+3 H \left(1+{\rm w}_i\right) \rho_i
=\frac{\dot{\phi}}{A(\phi)}\,\frac{dA\left(\phi\right)}{d\phi}\,\left(1-3\,{\rm
w}_i\right)\,\rho_i.$$ Out of the four equations, (\[vary1\])-(\[vary3\]) and (\[vary4\]), only three are independent [@Ish07; @Neupane07]. That is to say, Eq. (\[vary4\]) may have been derived from (\[vary1\])- (\[vary3\]), without assuming (\[cons\]).
From Eqs. (\[vary1\]) and (\[vary2\]), we find that $$\frac{\ddot{a}}{a}=-\frac{4{\pi}G}{3}
\left[\sum_{i=1}\left(\rho_i+3\,p_i\right)+(\rho\Z{\phi}+3p\Z{\phi})\right].
\label{acceleration}$$ A species that contributes to positive cosmic acceleration ($\ddot{a}>0$) must have $(\rho+3p)<0$, i.e. the equation of state ${\rm w}$ must be $<-1/3$. This gives us limits on the range of values the equation of state which the scalar field may have over time if it is to contribute a gravitationally repulsive force, ${\rm w}_\phi < -1/3$. Indeed, the WMAP data combined with BAO and SNIa observations has put much stronger limits on the DE equation of state, at present ($z\sim 0$), it lies in the range $-1.11<{\rm
w}_{\rm DE} <-0.86\ (95\%\,CL)$ [@WMAP2]. This range includes values of equation of state less than $-1$, which is outside a theoretical limit set for a canonical scalar field and belongs to “phantom cosmology”. However, it is possible to get at low redshifts ${\rm w}_\phi <-1$ by allowing the scalar field to couple with dark matter non-minimally [@Ish07; @Neupane07], especially by allowing the coupling between $\phi$ and dark matter to grow with time, or allowing $A(\phi)^2$ to increase monotonically with $\phi$.
It is practical to express the above equations in terms of the e-folding time, defined by $N{\equiv}\,\ln[a(t)]$. This is a useful time parameter when one is considering the expansion history of the universe, including late time cosmology. We also utilize the identity $\partial\phi/\partial{N}
=\frac{1}{H}\frac{\partial\phi}{\partial t}$ and denote the differentiation with respect to $N$ by a prime, $\partial/\partial
N\equiv ^{'}$, and the time derivative represented by a dot, $\partial/\partial t \equiv {}^\cdot$. Using following definitions $$\begin{aligned}
&& \epsilon\equiv\frac{\dot{H}}{H^2}, \quad
Q\equiv\frac{d\ln\left[A\left(\phi\right)\right]}{d(\kappa\phi)},
\quad \Omega_i\equiv\frac{\kappa^2\,A^4\,\rho_i}{3\,H^2}, \quad
\Omega_{\phi}\equiv \frac{\kappa^2\,\rho_{\phi}}{3\,H^2}, \quad
{\rm w}_{\phi}\equiv \frac{p_{\phi}}{\rho_{\phi}}, \label{subs1}\end{aligned}$$ one can write the equations (\[vary1\])-(\[vary3\]) and (\[vary4\]) in the following form
$$\begin{aligned}
\sum_{i}\Omega_i+\Omega_{\phi}=1,
\label{vary6}\\
2\,\epsilon+3\,\left(1+{\rm
w}_{\phi}\right)\,\Omega_{\phi}+3\,\sum_{i}\left(1+{\rm
w}_i\right)\,\Omega_i=0,
\label{vary7}\\
\Omega_{\phi}^\prime+2\,\epsilon\,\Omega_{\phi}+3\,\Omega_{\phi}\,
\left(1+{\rm w}_{\phi}\right)+\phi^\prime\,Q\,\sum_{i}(1-3\,{\rm
w}_i)\,\Omega_i=0,
\label{vary8}\\
\sum_{i}\Omega_i^\prime+2\,\epsilon\,\sum_{i}\Omega_i +3\,\sum_{i}
\Omega_i\,\left(1+{\rm
w}_i\right)-\phi^\prime\,Q\,\sum_{i}(1-3\,{\rm w}_i)\,\Omega_i=0,
\label{vary9}\end{aligned}$$
where the sum over $i$ represents the sum over all forms of matter or energy. It is important to note that in minimal coupling case, $A(\phi)=1$, $Q$ vanishes and the field equations have one less parameter making them easier to handle. Solutions of these equations can be used to reconstruct the behavior of $\phi$ over time. The behavior of the scalar field is generally characterized by its potential, which is given by $$\label{potentiala}
\kappa^2 V\left(\phi\right) = H^2(\phi)\,\left( 3\Omega_{\phi}
-\frac{\gamma\kappa^2} {2}\,{\phi^\prime}^2\right).$$ For simplicity, one can make a specific ansatz for the form of the potential in order to reduce the number of degrees of freedom and hence solve the set of field equations analytically. For many potentials discussed in the literature, the initial value of $\phi$ and $\dot{\phi}$ must be finely tuned to obtain the correct values of $\Omega\Z\phi$ and ${\rm w}_{\phi}$ today. The tuning of the initial field expectation value is required in addition to tuning the potential parameters.
Given the many alternative form of (quintessential) potentials it is useful to try and understand the properties of DE in a model-independent manner. In this paper, we will make an ansatz for the form of the scalar field so that the form of the potential, together with a number of cosmological parameters, can be reconstructed. This is a valid way to study the effects of the scalar field on the background since it is required that the potential must be fairly flat today, or, equivalently, satisfy that $(\dot{\phi}/m_{Pl}) \ll \sqrt{3} H$. This is because the scalar field must be rolling slowly (with a small kinetic term) to force the equation of state close to $-1$ at $z\gtrsim 0$, as inferred by cosmological observations (the WMAP data combined with BAO and SNIa observations).
An unambiguous expression for the quintessence equation of state (EoS) is given by $$\label{assumption4}
{\rm w}\Z{\phi}(a)=\frac{\gamma \kappa^2
\left(\phi^\prime\right)^2-3\, \Omega_{\phi}}{3\,\Omega_{\phi}}.$$ This result is obtained by substituting the expression for $V(\phi)$, Eq. (\[potentiala\]), into the definition of ${\rm
w}_\phi$ in Eq. (\[subs1\]). In this note, rather than parametrizing the dark energy equation of state ${\rm w}_\phi$ with two or more phenomenological parameters, we will choose to parametrize the evolution of the field $\phi$ (cf see Eq. (\[ansatz\])). This, in turn, will induce a dependence of ${\rm w}_\phi$ on those parameters through Eq. (\[assumption4\]). The same applies to the scalar field potential $V(\phi)$ and the Hubble parameter which both depend on these parameters through Eqs. (\[potentiala\]) and (\[hubble\]). In our analysis, we only consider a canonical quintessence with $\gamma= 1$, and use the relation (\[assumption4\]) which is an exact and definitive expression for the equation of state.
Other parameters of interest are the effective equation of state and the deceleration parameter. The effective equation of state describes the average equation of state, taking into account all the energy components of the universe: $$\label{effective1}
{\rm w}_{\rm eff} \equiv \frac{p_{\rm tot}}{\rho_{\rm tot}}, \quad
p_{\rm tot}\equiv p_{\phi}+ \sum_{i} p_i \, e^{4\kappa Q_i\phi},
\quad \rho_{tot} \equiv \rho_{\phi}+ \sum_{i} \rho_i \,e^{4 \kappa
Q_i \phi}.$$ From Eq. (\[vary7\]) and the definitions in (\[subs1\]), we find that $$\label{ptot}
p_{\rm tot}=\frac{3\,H^2(\phi)}{\kappa^2}\,
\left(-\frac{2\,\epsilon}{3}-\Omega_{\phi}-\sum_{i}\Omega_i\right),
\quad {\rm w}_{\rm eff}=-1-\frac{2\,\epsilon}{3}.$$ Another important quantity is the deceleration parameter $$\label{q}
q\equiv -(1 +\epsilon) =
-\frac{\ddot{a}}{a\,H^2}=\frac{d}{dt}\left(\frac{1}{H}\right)-1.$$ This parameter is defined such that it is negative for an accelerating expansion ($\ddot{a}>0$) and positive for a decelerating expansion ($\ddot{a}<0$).
Motivation and Analytic solutions {#solns}
=================================
In order to be able to analytically solve the generalized field equations obtained in Sec. II, there must be more assumptions made to reduce the number of degrees of freedom. In this paper, instead of making an ansatz for the scalar field potential or a paramterization of the dark energy equation of state, we wish to reconstruct relevant cosmological parameters from some phenomenologically well-motivated Ansätze for a scalar field quintessence. Our approach is new in the existing literature.
In the literature, basically, there are two simple and convenient approximations to the form of a time-evolving scalar field quintessence, see, e.g. Refs. [@ART93; @Gasperini:01a; @Barreiro:98a]. Those approximations are $$\label{ansatz-1st-form} \kappa
|\phi| = \alpha\Z1 \ln t+ {\rm const}, \quad \kappa {|\phi|}=
\alpha\Z2 t^{m}+ {\rm const},$$ These evolutions for $\phi$ generally represent the tracking limits of some more general solutions for a runway dilaton in many string-inspired scalar-tensor theories [@Ish06b; @ART93; @Gasperini:01a]. Similar approximations have been adopted also in a few phenomenologically motivated models for quintessence, see, e.g. Refs. [@Chevallier:00; @Linder02a; @Barreiro:98a; @Picon:00a; @Picon01a; @Scherrer:2007a; @Scherrer08b; @Dutta08a].
In this paper, we will assume that the time-evolution of quintessence scalar field $\phi$ is well described by a combination of the above two specific solutions. Using a general feature of the scale factor, that $a(t)\propto t^{p}$, with different values of $p$ at different epochs, such as $p=2/3$ in a matter-dominated epoch and $p>1$ in a dark-energy-dominated epoch, the time-evolution of $\phi$ may be written as $$\label{ansatz}
\kappa\left(\phi\Z{0}-\phi(a)\right) = \alpha \ln a + \beta\,
a^{2\zeta},$$ where $\alpha$, $\beta$ and $\zeta$ are some new arbitrary constants, and $a\equiv a(t)$ is the scale factor of a four-dimensional FRW universe. This parameterization of $\phi$, which might bear some generic features of a fundamental scalar dilaton field or metric moduli in some string-inspired models, can also be motivated from other two aspects. First, it really gives a useful information as regard the dark energy equation of state once the parameters like $\alpha$ and $\zeta$, or their combination, are known (even approximately) using observational data. Second, it certainly helps to explain the cosmic coincidence problem, since the value of an arbitrary coefficient $V_0$ that arises with [*a prior*]{} choice of the scalar field potentials, such as $V(\phi)=V_0\,e^{-\lambda (\phi/m_P)}$, $V(\phi)=V_0
\phi^{-n}$ and $V(\phi)=V_0\left(1- c_0 \phi^2\right) e^{-\lambda
(\phi/M_P)}$, won’t be very important in our approach. Indeed, the parameterization (\[ansatz\]) for the evolution of $\phi$ is neither more arbitrary nor more restrictive than other parametrizations and approaches to quintessential dark energy in the literature. See Refs. [@Sahni:2006pa; @Scherrer:2007a; @Scherrer08b; @Dutta08a; @Samushia:2008fk; @Ratra08a] for some other plausible ways of reconstructing dark energy or quintessential potentials.
Our ansatz for $\phi$, i.e. (\[ansatz\]), has nonetheless some similarities with respect to numerous dark energy potentials proposed previously, which may be reconstructed by using Eq. (\[potentiala\]), or alternatively, $V(\phi)=\kappa^2
\left[(3+\epsilon)(1-\Omega_m)+\frac{1}{2} \Omega_m^\prime\right]
H^2(\phi)$. Especially, with $\zeta\simeq 0$ in (\[ansatz\]), we find that the leading term of the reconstructed potential is simple exponential in $\phi$, $$V(\phi)\equiv V_0\, e^{\,\alpha (\phi/M_P)}+ \cdots,$$ where the dots represent some other terms which could arise, for instance, due to the effects of matter-scalar interactions. Similarly, with $\alpha \simeq 0$, the scalar potential takes the form $$V(\varphi) \equiv V\Z{1}\, e^{-\,\zeta\varphi^2} \left(1-
c\Z{0}\varphi^2\right)+\cdots,\label{quin-pot2}$$ where $\varphi \equiv |\Delta\phi|/m\Z{P} =
(\phi\Z0-\phi)/m\Z{P}$ and $c\Z{0}=\frac{2\zeta^2}{3}$. A polynomial potential multiplied with exponential pre factor as above may be motivated by string theory and standard Kaluza-Klein gravity. An effective dark energy potential as above was proposed in [@Albrecht99]. In our approach, the total or effective quintessence potential is roughly given by a linear combination of the above two potentials. Interestingly, in our discussions below, the actual form of the potential is not very important, but only the values of the slope parameters $\alpha$ and $\zeta$.
In fact, using the freedom to rescale $N$ ($\equiv \ln a$) or shift $\phi$, we can set $\beta=1$ [@Ish0712]. The model then contains two free parameters, $\alpha$ and $\zeta$, so there is more freedom to tune them according to the observational constraints. In our model, there is one more degree of freedom that must be constrained. This extra parameter, which is the coupling constant between the dark matter and the scalar field, is a mixed blessing – it makes the system easier to tune but it also makes the solutions more complicated.
In order to completely solve Eqs. (\[vary6\])-(\[vary9\]), one must specify some initial conditions. These conditions can be defined on the current composition of the universe, as measured by WMAP plus other observations [@WMAP2], and are $\Omega_b\simeq
0.05$, $\Omega_{CDM}\simeq 0.22$, $\Omega_{\rm r}\simeq10^{-4}$ and $\Omega_\phi\simeq 0.73$. The baryon component (subscript $b$) is quite small compared to the cold dark matter component (subscript $CDM$), so even though there are tight limits on any baryon-dark-energy coupling, it is sufficient to assume a general dark energy coupling to $\Omega_m$, ($\Omega_m\equiv\Omega_b+\Omega_{CDM}\simeq0.27$). Henceforth the only components of the universe that are not considered negligible are dark matter and the dark energy component, up to $z\sim {\cal
O} ({10}^2)$, when radiation starts playing a major role. Also, it is assumed that the dark matter equation of state is the same as that of ordinary matter (or dust), ${\rm w}_m\simeq 0$. Using these assumptions, and for simplicity setting $\kappa=1$ from now on, the system of equations (\[vary6\]-\[vary9\]) is simplified to
$$\begin{aligned}
\Omega_m+\Omega_{\phi}=1
\label{vary10}\\
2\,\epsilon+3\,\left(1+{\rm w}_{\phi}\right)\,\Omega_{\phi}+3\,\Omega_m=0
\label{vary11}\\
\Omega_{\phi}^\prime+2\,\epsilon\,\Omega_{\phi}+3\,\Omega_{\phi}\,
\left(1+{\rm w}_{\phi}\right)+\phi^\prime\,Q\,\Omega_m=0
\label{vary12}\\
\Omega_m^\prime+2\,\epsilon\,\Omega_m+3\,\Omega_m-\phi^\prime\,Q\,\Omega_m=0.
\label{vary13}\end{aligned}$$
Equations (\[vary10\])-(\[vary13\]) can now be solved to find explicit expressions for ${\rm w}_\phi$, $\epsilon$ and $\Omega_m$. Solving the equations directly gives the solution for $\Omega_m$ and $\epsilon$ in terms of $\alpha$, $\zeta$ and $Q$ ($\alpha$ and $\zeta$ are parameters from the ansatz for $\phi$, and $Q$ is the coupling term).
The explicit solutions for $\Omega_m$ and $\epsilon$ are given by $$\label{sol-Om-int}
\Omega_m=\frac{X(N)}{C\Z0 -3\int X(N)\, dN },\quad
\epsilon=-\frac{3}{2} \Omega_m -\frac{1}{2} {\phi^\prime}^2=
-\frac{3}{2}\Omega_m -\frac{1}{2}\left(\alpha+2\zeta e^{2\zeta
N}\right)^2,$$ where $$X=X(N=\ln a) \equiv \exp\left[-(3-\alpha Q-\alpha^2)N+(2\alpha+Q)
e^{2\zeta N}+ \zeta e^{4\zeta N}\right].\label{soln-X}$$ The scale factor $a(t)$ is normalized such that $a(z=0)\equiv
a\Z0=1$, so in the past $a < 1$ and hence $N\equiv \ln a < 0$. The relationships among the scale factor $a$, redshift $z$ and e-folding time $N$ are $$\frac{a}{a_0}=\frac{1}{1+z}, \quad N=\ln(a/a_0)=-\ln(1+z).$$ So, at $t=t\Z0$, $z=N=0$. Without loss of generality, we set $a_0=1$, so $N=\ln a$. To solve the integral (\[sol-Om-int\]), in the small $\zeta$ limit, we may use the approximation $e^{\zeta
\ln a}\approx 1+ \zeta\ln a$. The solution for $\Omega_m$ is now given by $$\label{main-soln-Omm}
\Omega_m=\frac{3-\tilde{\alpha}^2-Q \tilde{\alpha}} {3+C_1
\exp{\left[\left(3-\tilde{\alpha}^2 -Q \tilde{\alpha}\right)N
\right]}},$$ where we have introduced a new variable $\tilde{\alpha}$ such that $$\tilde{\alpha}\equiv\alpha+2\zeta.$$ The solution for $\Omega_m$ automatically gives the solution for $\Omega_\phi$ which is simply $\Omega_\phi=1-\Omega_m$, in using Friedmann constraint. The integration constant $C_1$ can be fixed by using the initial conditions of the model, assuming $\Omega_m(t_0)\equiv \Omega\Z{m 0}$, where $t_0$ is the time now.
Particularly, at low redshifts, the terms quadratic (and higher powers) in ${\zeta N}$ contribute only subdominantly. \[Even at high redshifts, the terms like $(2\alpha+Q)e^{2\zeta N}$ and $\zeta e^{4\zeta N}$ in (\[soln-X\]) are only sub-leading to the first term in $X(N)$, i.e. $-(3-\alpha^2-\alpha Q) N$, since $\zeta>0$ (by assumption) and $N\equiv \ln {a}< 0$ in the past.\] To quantify this, one can write $$e^{\zeta \ln a}= 1+ \zeta \ln a + \frac{\zeta^2
\ln{a}^2}{2}+\cdots = 1+\zeta\ln a + \zeta^2 |\ln a|+\cdots.$$ For instance, in between the redshifts $z=2$ and $z=0$, $N\equiv
\ln {a}$ runs from $-1.09$ to $0$, and the solution (\[main-soln-Omm\]) remains valid for $\zeta$ as large as $\zeta
\sim 1/2$, but in the discussions below we will always assume rather implicitly that $\zeta \ll 1/2$.
The Hubble parameter may be evaluated by solving the differential equation associated with the relation $H\epsilon =H'$. This again introduces an integration constant which can be fixed by the normalization $H(t_0)\equiv H_0$. From the solutions for $\Omega_m$ and $\epsilon$, the other parameters discussed in Sec. II can also be derived. The expression for ${\rm w}_\phi$ is given by rearanging Eq. (\[vary11\]), while $V(\phi)$ is given by Eq. (\[potentiala\]) using the solutions for $\Omega_m$ and $H$.
From the second equation in (\[sol-Om-int\]), we can see that (retaining the 4d gravitational constant $\kappa$) $$\kappa^2 {\phi^\prime}^2 = 1+ q -\frac{3}{2} \Omega\Z{m}.$$ Requiring that $q<0$ (as implied by the type Ia supernovae) and imposing a generous lower bound on the value of $\Omega_m$, which is $\Omega_{0 m} >0.24$, one obtains the safe upper bound $$\kappa |\phi\Z{0}^\prime | < 0.8.$$ However, to make the present model compatible with various other data sets, including WMAP observations, one may be required to satisfy $\kappa |\phi\Z{0}^\prime | \lesssim 0.4$, which encompasses the cosmological constant limit within $1\sigma$ error.
Constraints on the model {#constraints}
========================
Constraints from supernova {#sn1a}
--------------------------
In this subsection, we use several sets of data from recent cosmological observations and put constraints on our model, limiting the values of $\alpha$, $\zeta$ and $Q$. To do this, code for $\chi^2$ curve fitting given by Nesseris and Perivolaropoulos [@limits] has been utilized along with our solution for ${\rm
w}_\phi$ and the standard form for the Hubble parameter, in terms of the two parameters $$\label{definew0}
p_1\equiv\tilde{\alpha}^2 , \quad p_2\equiv-Q\tilde{\alpha},$$ where, as before, $\tilde{\alpha}\equiv\alpha+2\zeta$. We use the Hubble parameter to fit our model to data on the expansion rate of the universe. The Hubble parameter can be found from the solution for $\epsilon$ and $\Omega_m$ and the solution to the differential equation, $\epsilon=\dot{H}/H^2$, which is the definition of $\epsilon$. After some simplification, we find that the Hubble parameter is given by $$\label{hubble}
H=H\Z0\left(\Omega_{0 m}\,a^{-3-p_2}+(1-\Omega_{0
m})\,a^{-p_1}\right)^{1/2}.$$ This result is obtained by integrating out the expression $\epsilon\equiv H^\prime/H =\left(\ln H\right)^\prime$ (see, Eq. (\[sol-Om-int\])) and substituting the expression of $\Omega_m$ from Eq. (\[main-soln-Omm\]). The frame where matter density decreases as $\rho_m\propto 1/(A(\phi) a(t))^3$ (where the scale factor is modified by the coupling $A(\phi)$) is known as the Einstein frame. In our expression for the Hubble parameter (\[hubble\]) the matter part is $\Omega_{0\,m}a^{-3-p_2}$, where the density evolution is modified by the parameter $p_2$ which is associated with the coupling $Q$. Thus this standard form of the Hubble parameter is in the Einstein frame.
As in some standard approaches, let us first drop the scalar field - dark matter coupling. The model then reduces to one-parameter parameterization of the Hubble expansion rate. Remember that the model still deviates from the $\Lambda$CDM model, since $p_1\ne 0$ even though $p_2=0$, leading to a time-varying equation of state for dark energy (see cf Fig. \[best-fit1\]). Needless to say, the $\Lambda$CDM cosmology corresponds to the choice $\tilde{\alpha}=0=Q$. This can easily be understood from the following observation. With $\kappa \phi^\prime \simeq
\alpha+2\zeta \equiv \tilde{\alpha}$, we get $$\label{main-wphi}
{\rm
w}_\phi(z)=\frac{{\tilde{\alpha}}^2-3\Omega_\phi}{3\Omega_\phi}
=\frac{(\tilde{\alpha}^2-3)
C\Z{1}-3\tilde{\alpha}Q\,(1+z)^{3-\tilde{\alpha}^2-\tilde{\alpha}Q}}{3
C\Z{1}+ 3\tilde{\alpha}(\tilde{\alpha}+Q)
(1+z)^{3-\tilde{\alpha}^2-\tilde{\alpha}Q}}.$$ $C\Z{1}$ in the above equation or in (\[main-soln-Omm\]) is fixed such that $\Omega_m(t_0)=\Omega_{m0}$ at $z=0$. Note that the dark energy EoS ${\rm w}_\phi(z)$ is time-varying as long as $\tilde{\alpha}\ne 0$.
[One notes that [*a prior*]{} choice or an arbitrary parameterization of ${\rm w}(z)$ could easily lead to an erroneous reconstruction of the dark energy equation of state. To see this, one considers the parameterization $H^2(z)=H_0^2 \left[\Omega_{0m}
(1+z)^3 +\Omega_{\rm DE}\right]$, where $\Omega_{\rm DE}\equiv
(1-\Omega_{0m})\exp\left\{3\int\frac{1+{\rm
w}(z)}{1+z}\,dz\right\}$. For example, with the ansatz $w(z)\equiv
{\rm w}_0+ \frac{z}{1+z}\,{\rm w}_1$, one finds that $\Omega_{\rm
DE}= (1-\Omega_{0m})\exp\left\{\frac{3 {\rm w}\Z{1}}{1+z}+(1+{\rm
w}_0+{\rm w}\Z{1})\ln (1+z)-{\rm w}_1+{\rm const} \right\}$. Now, with ${\rm w}_1=0$, one has $H^2(z)=H_0^2 \left[\Omega_{0m}
(1+z)^3 +(1-\Omega_{0m})\times {\rm const}\times (1+z)^{3(1+{\rm
w}_0)}\right]$, which has a form similar to Eq. (\[hubble\]), especially, with $p_2=0$, i.e. $Q=0$. This could give a wrong impression that in our model the choice $Q=0$ gives a constant equation of state for dark energy. The dark energy equation of state is not constant as long as $\tilde{\alpha}\ne 0$, or precisely, when $\alpha, \zeta\ne 0$]{}.
In Table 1 we present the best-fit values using only one parameter (i.e. $p_1\ne 0$ and $p_2=0$) first using only the Gold sample of 157 type Ia supernova data, the Supernova Legacy Survey (SNLS) data alone and then for combined data sets. The combined data includes the cosmic microwave background (WMAP) shift, baryon acoustic oscillations (BAO), suvernovae type Ia Gold sample (SNIa) and legacy survey (SNLS). The errors of these fits are shown in Fig. \[best-fit1\]. For the combined data sets $\chi^2$ is minimum when $\Omega_{0m}\approx 0.27$ where $\alpha+2\zeta\simeq0.3\pm 0.3$. For $\Omega_{0\,m}>0.29$ in the fit to the Gold sample alone, the EoS drops below -1, indicating phantom quintessence and an imaginary $\tilde{\alpha}$.
Table 1: The best fit of $\tilde{\alpha}$ to expansion history data, $Q=0$:
[|l|l|l|l|l|]{} $\Omega\Z{0 m}$& $p_1$ & $|\tilde{\alpha}| $ & $w\Z{\phi}(z=0)$ (mean) & $\chi\Z{min}^2$\
\
$ 0.26$ & $0.271\pm 0.286$ & $0.521\,(\pm 0.275)$ & $-0.878$ & $177.98$\
$0.27$ & $0.194\pm 0.298$ & $0.441\,(\pm 0.339)$ & $-0.911$ & $177.76$\
$0.28$ & $0.112\pm 0.312$ & $0.335\,(\pm 0.466)$ & $-0.948 $ & $177.54$\
\
$0.26$ & $-0.006\pm 0.274$ & $-$ & $-1.00$ & $104.15$\
$0.27$ & $-0.081\pm 0.285$ & $-$ & $-1.04$ & $104.13$\
$0.28$ & $-0.159\pm 0.297$ & $-$ & $-1.07 $ & $104.12$\
\
$0.26$ & $0.123\pm 0.178$ & $0.350\pm 0.254 $ & $-0.945$ & $283.54$\
$0.27$ & $0.098\pm 0.1842$ & $0.313\pm 0.294 $ & $-0.955$ & $283.24$\
$0.28$ & $ 0.072\pm 0.190$ & $0.268\pm 0.354 $ & $-0.967 $ & $283.61$\
It is indeed the SNLS data that lowers $w\Z{\phi}(z)$ towards the value $-1$ at present, i.e. at $z=0$, which is clearly seen from the best-fit values in Table 1. The SNLS data naively suggests a small cross over range between the cosmological constant ${\rm w}\Z\Lambda=-1$ and the phantom dark energy ${\rm w}\Z{\phi}<-1$ (or, equivalently, $p_1 <0$), but the error bars are too large. For the combined data sets (from WMAP, BAO, SNIa and SNLS) the best-fit value of ${\rm w}_\phi$ falls in the range $-0.94
> {\rm w}_\phi >-1$.
![Scalar field equation of state ${\rm w}_\phi$ as a function of redshift ($z$) with $1\sigma$ and $2\sigma$ errors (light and dark gray shades) and with EoS solutions given by $\zeta=0.2,\, 0.15,\, 0.1,\, 0$ (top to bottom; green, pink, red, blue online), $\Omega_{0_m}=0.27$. **Left plot**: SN1a+WMAP+BAO+SNLS. **Right plot**: SN1a only []{data-label="best-fit1"}](all-Om027.eps "fig:"){width="3in" height="3.2in"}0.2in ![Scalar field equation of state ${\rm w}_\phi$ as a function of redshift ($z$) with $1\sigma$ and $2\sigma$ errors (light and dark gray shades) and with EoS solutions given by $\zeta=0.2,\, 0.15,\, 0.1,\, 0$ (top to bottom; green, pink, red, blue online), $\Omega_{0_m}=0.27$. **Left plot**: SN1a+WMAP+BAO+SNLS. **Right plot**: SN1a only []{data-label="best-fit1"}](Gold-SNIa-Om027.eps "fig:"){width="3in" height="3.2in"}
From the relationships above, and also noticing that for the combined data sets the scalar field equation of state is closer to $-1$ over a longer time in Fig. \[best-fit1\], we find that the solution using all the data sets is compatible with a cosmological constant. This analysis has provided a limit to the relationship between $\alpha$ and $\zeta$ but there is still a degree of freedom for choosing the value of $\alpha$ or $\zeta$. This degree of freedom proves hard to constrain since for other observational tests, it is only required that $\alpha$ and $\zeta$ be of comparable magnitude, or that $\zeta=0$. The plots of ${\rm
w}_\phi$ in Figure \[best-fit1\] show that the chi-squared is minimized when $\zeta$ is small since cosmological reconstructions for large $\zeta$ diverge from the best-fit line. Fig.\[best-fit2\] displays the same constraints on the model, this time showing the effect of positive or negative $\zeta$. There is more divergence from the best-fit line for negative values of $\zeta$, suggesting that positive $\zeta$ is a better fit.
![As in Fig. \[best-fit1\] but with $\zeta=-0.2$ (lower line, cyan) and $\zeta=+0.2$ (upper line, blue). **Left plot**: SN1a+WMAP+BAO+SNLS. **Right plot**: SN1a only[]{data-label="best-fit2"}](all-Om027-neg.eps "fig:"){width="3in" height="3.2in"}0.2in ![As in Fig. \[best-fit1\] but with $\zeta=-0.2$ (lower line, cyan) and $\zeta=+0.2$ (upper line, blue). **Left plot**: SN1a+WMAP+BAO+SNLS. **Right plot**: SN1a only[]{data-label="best-fit2"}](Gold-SNIa-Om027-neg.eps "fig:"){width="3in" height="3.2in"}
![Best-fit plot of ${\rm w}_\phi$ with $Q\ne 0$ (i.e. $p_2\ne 0$). The continuous line is the best fit and the dashed line is for $\zeta=0.2$. The inner and outer shaded regions represent $1\sigma$ and $2\sigma$ error bars. **Left plot**: $\Omega_{0m}=0.26$. **Right plot**: $\Omega_{0m}=0.28$.[]{data-label="new-best-fit1"}](new-test26.eps "fig:"){width="2.9in" height="2.5in"}0.2in ![Best-fit plot of ${\rm w}_\phi$ with $Q\ne 0$ (i.e. $p_2\ne 0$). The continuous line is the best fit and the dashed line is for $\zeta=0.2$. The inner and outer shaded regions represent $1\sigma$ and $2\sigma$ error bars. **Left plot**: $\Omega_{0m}=0.26$. **Right plot**: $\Omega_{0m}=0.28$.[]{data-label="new-best-fit1"}](new-test28.eps "fig:"){width="2.9in" height="2.5in"}
Table 2: The best fit values for both $\tilde{\alpha}$ and $Q$ (WMAP+BAO+SNIa+SNLS):
$\Omega\Z{0 m}$ $p_1$ $p_2$ $|\tilde{\alpha}| $ $Q$ $w\Z{\phi}(z=0)$ $\chi\Z{min}^2$
----------------- ------------------ ----------------- --------------------- ------------------ ------------------ -----------------
$0.26$ $0.055\pm0.203$ $0.015\pm0.021$ $0.235\pm0.432$ $-0.065\pm0.149$ $-0.98$ $ 283.02$
$0.27$ $0.001\pm0.210$ $0.021\pm0.021$ $0.027\pm3.83$ $-0.772\pm 1.08$ $-1.00$ $282.22$
$0.28$ $-0.055\pm0.217$ $0.027\pm0.021$ - - $-1.03$ $281.98$
$0.29$ $-0.111\pm0.225$ $0.032\pm0.021$ - - $-1.05$ $282.24$
$0.30$ $-0.168\pm0.233$ $0.037\pm0.021$ - - $-1.08$ $283.00$
Our proposed model has the freedom of allowing for a non-zero coupling of dark matter to dark energy. Including this in the fit to the data sets gives the results of Table 2. Note that for $\Omega_0>0.27$ we get negative values of $p_1$. From the definition of $p_1$, this means an imaginary $\tilde{\alpha}$ and $Q$, and a negative equation of state which is phantom quintessence. The best-fit to the data is for $\Omega_{m0}\simeq 0.275$ which is apparently the best fit value for $\Lambda$CDM model obtained from the combined WMAP5+BAO+SNIa datasets [@WMAP2]. The best fit value for $\tilde{\alpha}$ is close to zero, the cosmological constant limit. $Q$ is constrained to $-0.772\pm 1.08$ but this error is too large for this to be a conclusive result. It is clear from the shaded error region in Fig. \[new-best-fit1\] that adding the extra parameter, $Q$, greatly increases the uncertainty in the fit.
Constraints from structure formation {#structure}
------------------------------------
Cosmological models are probed by observing the effects of dark energy on the expansion of the universe through measuring the distances to far off galaxies [@Cattoen:08a]. Other than this there are not many ways to check models that account for the late time acceleration of the universe. However, if dark energy does interact with dark matter even weakly, it may have an observable effect on the early stages of structure formation.
Since dark matter naturally guides the way for the formation of observed structure, any interaction between dark energy and dark matter would have an effect on the manner that visible structure was formed. In particular, we investigate how the rate of structure growth is affected by a non-zero coupling of dark matter to a fundamental scalar field.
Matter fluctuations evolve according to the standard linear differential equation $$\label{fluctuation}
\ddot{\delta}+2H\dot{\delta}=4\,\pi G \rho_m \delta,$$ where $\delta\equiv\delta\rho_m/\rho_m$ is the linear matter density contrast. This linear growth equation comes from the perturbed equations of motion of Einstein’s general relativity, see, e.g. Ref. [@weinberg]. It may be numerically solved to reconstruct the growth of these matter fluctuations, but as yet there is no physical theory that relates the matter density contrast to the matter density. In [@Lahav:1991] it was proposed that the growth rate of matter perturbations, defined as $$f\equiv\frac{\delta^\prime}{\delta}=\frac{1}{\delta}
\frac{d\delta}{d\ln a}=\frac{1}{\delta} \frac{d\delta}{dN},$$ can be characterized by the following simple expression $$\label{fom}
f=\left[\Omega_m(N)\right]^\eta.$$ This ansatz works well at low redshifts $(z\lesssim 2$) and for coupled dark energy models with a small coupling parameter. For the $\Lambda$CDM model, $\eta\simeq 0.56$ [@Lahav:1991], and for modified DGP gravity, $\eta\simeq 0.68$ [@Guzzo:2008]. However, this ansatz does have the drawback that $f\leq 1$, since $0\leq\Omega_m\leq1$. This is fine for the standard $\Lambda$CDM model which assumes $f=1$ during the matter-dominated era for high redshift, but does not hold for all models of dark energy.
For the present model of quintessential dark energy, as discussed in [@Amendola:2003a] in more detail, the linear growth equation (\[fluctuation\]) is modified to be
$$\begin{aligned}
\delta_c^{\prime\prime}+\left(2+\varepsilon+Q_c\,\phi^\prime\right)
\delta_c^\prime =\frac{3}{2}\left(1+2 Q_c^2\right)\delta_c
\Omega_c + \frac{3}{2}(1+2Q_b Q_c)\delta_b
\Omega_b,\label{modi-fluc1}\\
\delta_b^{\prime\prime}+\left(2+\varepsilon+Q_b\,\phi^\prime\right)\delta_b^\prime
=\frac{3}{2}\left(1+2 Q_b^2\right)\delta_b \Omega_b +
\frac{3}{2}(1+2Q_b Q_c)\delta_c \Omega_c,\label{modi-fluc2}\end{aligned}$$
for baryonic and CDM components respectively. The couplings constant $Q_b$ and $Q_c$ are usually coupled, so in general it is not possible to express these equations as a single differential equation. Amendola [@Amendola:2003a] made a naive estimation that in the limit $|Q_b|\ll |Q_c|$ and $\Omega_b \ll \Omega_c$, one can write $$\label{eqn-amendola}
\delta^{\prime\prime}+ (2+\epsilon+Q \,\phi^\prime)\delta^\prime
=\frac{3}{2}\Omega_m (1+2 Q^2)\delta,$$ where $Q\equiv Q_c$ and $\delta \approx
\delta_c$ (to leading order). In a sense, the baryonic component $\Omega_b$ is assumed to be negligible. But one should also note that $\delta_b$ and its derivatives are non-negligible, otherwise Eq. (\[modi-fluc2\]) would be inconsistent. Equation (\[eqn-amendola\]), which reduces to the standard result when $Q=0$, is a close approximation rather than being an exact result. In the following discussion, we will assume that $\delta_c
\Omega_c \gg \delta_b \Omega_b$.
In space-time backgrounds dominated by baryonic matter, the effect of dark matter may be neglected ($\Omega_c=0$). The effective Newton’s constant $\widehat{G}$, and two dimensionless post Newtonian parameters $\hat{\gamma}$ and $\hat{\beta}$ are related to the coupling constants $$Q(\phi)\equiv m_P \frac{\partial \ln A(\phi)}{\partial \phi},
\quad X(\phi) \equiv m_P \frac{\partial Q(\phi)}{\partial\phi}$$ via [@Damour:1993] $$\widehat{G}= G \left [ A^2(\phi) (1+ Q^2)\right]_{\phi_0}, \quad
\hat{\gamma}=1- \frac{2 Q^2}{1+ Q^2}, \quad \hat{\beta}=
1+\frac{Q^2 X}{2(1+Q^2)^2}.$$ For an exponential coupling $A(\phi)\propto e^{Q_b\kappa \phi}$, with $Q_b$ behaving (almost) as a constant, the local gravity constraint $|1-\hat{\gamma}|< 10^{-4}$ [@Bertotti:2003] implies that $Q_b^2 \equiv (1-\hat{\gamma})/(1+\hat{\gamma})<
10^{-4}$ and hence $|Q_b|<0.01$. This constraint is still weaker than the one arising from weak equivalence principle violation [@Damour:2002; @Chiba:06a]. With $Q_b\simeq $ const, one has $|1-\hat{\beta}| \simeq 0$ to a large accuracy. The effective Newton’s constant is now modified but it must be within the limit for the time variation of Newton’s constant $|\frac{d{\widehat{G}}/dt}{G}| <10^{-14}{\rm yrs}^{-1}$, see e.g. Refs. [@limits; @Nesseris07], which is generally the case when $|Q_b| \lesssim 0.01$ is satisfied. In most of the discussion below we will assume that $|Q_C|\gg |Q_b|\simeq 0$ and $Q_c \equiv
Q$.
In the nonminimally coupled theory, the linear growth rate (\[fluctuation\]) is modified from the uncoupled case; the quantity $\delta\equiv\delta\rho_m/\rho_m$ now depends on the values of $Q$ and $\phi^\prime$. The result of this is a modification of the expression for the growth rate, Eq. (\[fom\]), where it is renormalized by either the parameter $\eta\rightarrow\tilde{\eta}$ and/or a coefficient $f_0$: $$f=f_0\left[\Omega_m(N)\right]^{\tilde{\eta}}.$$ Following the analysis in Ref. [@Di-Porto07], if the growth rate is modified by a coefficient, it may be dependent on the coupling by $$\label{approx-sol}
f=\left[\Omega_m(N)\right]^\eta \left(1+c\,Q^2 +\cdots \right),$$ which reduces to (\[fom\]) for minimal coupling ($Q=0$). This expression of $f$ is merely a phenomenological fit where $\eta$ and $c$ may be determined by fitting the standard approximate solution (\[approx-sol\]) to a numerical solution to Eq. (\[modi-fluc1\]) or Eq. (\[eqn-amendola\]). The best-fit parameters found from this numerical analysis in Ref. [@Di-Porto07] are $\eta\simeq 0.56$ and $c\simeq 2.1$. The physical theory pertaining to the dependence of $\Omega_m$ and $Q$ on the perturbation growth rate is as yet unknown so this kind of generalized fit is just an approximation. Determining the growth rate over time and comparing to observational results gives limits to the model that are independent of limits from fitting the model to supernova and WMAP data sets. This grants an independent check for the consistency of the model.
To preform this check of our model, we use some known observed values for the growth rate at different low redshifts (listed below in table), from which we can determine the values of $f_0$ and $\eta$ that best match the data. There are very few measurements of the growth rate at low redshifts and often $f$ can be evaluated by assuming a $\Lambda$CDM model, for which $\eta\approx 0.56$.
Table 3: Observed matter fluctuation growth rate as compiled in [@Wei:2008; @Wei08b; @Alam:2008at].
$z$ $ 0.15$ $0.35 $ $0.55 $ $0.77 $ $1.4$ $3$
---------------- ---------------- ----------------- ---------------- ----------------- ----------------- ----------------
$f_{\rm obs} $ $0.51\pm 0.11$ $0.70\pm 0.18 $ $0.75\pm 0.18$ $0.91\pm 0.36 $ $0.90 \pm 0.24$ $1.46\pm 0.29$
It is clear from Eqs. (\[main-soln-Omm\]) that $\Omega_m$ is in fact a function of $\tilde\alpha$ and $Q$ and thus our parameterization of $f$, i.e. (\[approx-sol\]), is implicitly a function of $\tilde\alpha$. The fits are illustrated by Fig. \[structurefit\], where the best-fit parameters are $$|Q|=0.199, \quad \tilde{\alpha} = -0.010, \quad \eta=0.538.$$ In the above we have used the normalization $f_0=1+2.1Q^2$ used in [@Di-Porto07]. Within the $1\sigma$ errors, the normalization constant $f_0=1.16\pm 0.34$ is compatible with a minimally coupled quintessence scenario. The least-squares curve fitting of all three of $\tilde{\alpha}$, $Q$ and $\eta$ to the data often leads to huge ranges of possible values of these parameters. For brevity, we may restrict ourselves to fitting only two parameters such as $Q$ and $c$. The best-fit value of the normalization constant $c$ is quite sensitive to the choice of $\eta$, or vice versa. For example, assuming $\Omega_{m0}=0.27$, $\tilde{\alpha}=0$ and $\eta=0.56$ and fitting for $Q$ and $c$ gives $$Q=0.42 \pm 0.10, \qquad c=1.17 \pm 0.02.$$ Similarly, holding $\eta=0.68$ and $c=2.1$ give the values $$\quad |Q|=0.41\pm0.14, \qquad \tilde{\alpha}=0.57\pm0.38.$$ In the above analysis we have used an explicit analytical expression for $\Omega_{m}$, which is nothing but the equation (\[main-soln-Omm\]). If it is instead the parameter $\eta$ that is renormalized, say $\eta\simeq
0.68$ [@Guzzo:2008] as in modified gravity proposed by Dvali, Gabadadze and Porrati, the fit to the data in Fig. \[structurefit\] seems to be closer.
![Best fit curve to the observed values of the growth rate, with $\Omega_{0m}=0.27$. The red solid line corresponds to a best fit of Q, $\tilde{\alpha}$ and $\eta$, the blue dashed line corresponds to $\eta=0.56$ with best fit for Q and $\tilde{\alpha}$, and the blue dotted line corresponds to $\eta=0.68$ with best fit for Q and $\tilde{\alpha}$.[]{data-label="structurefit"}](structurefit.eps){width="3.0in" height="1.8in"}
Constraints from CMB
--------------------
Another test of our model to the observational constraints is to demand that the equation of state presently lies in the range, $-1.11<{\rm w}_\phi<-0.86$, as implied by the WMAP data combined with BAO and type Ia SN [@WMAP2]. This is illustrated in the contour plot, Fig. \[contour\], where the outer contour corresponds to ${\rm w}_\phi=-0.91$ and the innermost contour corresponds to ${\rm w}_\phi=-0.99 $. For ${\rm w}_\phi$ to lie within the WMAP5 limit $-1.11<{\rm w}_\phi<-0.86$, the constraint is $$\label{cmbconstraint}
\tilde{\alpha}=0(\pm 0.5)$$ where the value of the coupling has little impact. This relationship holds for small $|\zeta|\lesssim 0.1$. It is reasonable to assume that the value of the scalar field equation of state has not changed much in the recent past. Using this tighter restriction, by inspecting Fig. \[contour\], we get a tighter relationship between $\alpha$ and $\zeta$, as $z\rightarrow 3$, error in relationship (\[cmbconstraint\]) goes to $\pm 0.2$.
![**Left plot** Contour plot of ${\rm w}_\phi$ in the range $(-0.91,-0.99)$ (top to bottom) at $z=0$. **Right plot** The dark energy EoS for the two solutions of $\zeta=\pm0.05$ (solid lines) and $\zeta=\pm0.25$ (dashed lines).[]{data-label="contour"}](contour1.eps "fig:"){width="3.1in" height="2.2in"}0.1in ![**Left plot** Contour plot of ${\rm w}_\phi$ in the range $(-0.91,-0.99)$ (top to bottom) at $z=0$. **Right plot** The dark energy EoS for the two solutions of $\zeta=\pm0.05$ (solid lines) and $\zeta=\pm0.25$ (dashed lines).[]{data-label="contour"}](cmb-constraint.eps "fig:"){width="3.1in" height="2.2in"}
There is still a degree of freedom in the choice of the value of $\alpha$ or $\zeta$, so we look to a possible link between inflation and late time acceleration to set it. Quintessence is designed such that the scalar field that drives the late time acceleration may be the same, or an evolved form of the inflaton field that drove inflation. If we assume that it is in fact the same field that drives both of these accelerating epochs, we may impose extra conditions on the model by requiring the quintessence field to satisfy the constraints of inflation. One such constraining parameter is the spectral index, which describes the slope of the angular power spectrum of the CMB. The WMAP data inferred a red-tilted spectrum, $n_s<1$, which is consistent with most inflationary models. The spectral index can be approximated in the small, positive $\zeta$ limit, as [@Ish:07a] $$\label{ns}
n_s\simeq1-\alpha^2.$$ For the WMAP+BAO+SN mean value of $n_s=0.960\,^{+0.014}_{-0.013}$ [@WMAP2] and $|\alpha|=0.20\,\pm0.04$. By using the relationship between $\alpha$ and $\zeta$ determined by the combined data sets in Table 2 ($|\alpha+2\zeta|\simeq0.3$), we obtain $\zeta=\pm0.05\pm0.15$ or $\zeta=\pm0.25\pm0.15$. The evolution of the EoS for these solutions of $\zeta$ are plotted in Figure \[contour\]. The EoS for $\zeta=\pm0.25$ deviates far from -1 for a positive value of $\alpha$. Both signs of $\alpha$ are possible so the lower value $\zeta=0.05$ may be more physical.
Dynamical behaviour
===================
In this section we examine how the scalar field varies over time and how this depends on the parameters $\alpha$, $\zeta$ and $Q$. As an example we choose the value $\alpha=0.2$ and using $\alpha+2\zeta\simeq 0.3$ from Table 1, $\zeta=0.05$. From structure formation we estimate $|Q|=0.2$. In our analysis perhaps the most relevant parameter is the field velocity of $\phi$, i.e. $\phi^\prime \equiv d\phi/d\ln a$, which is a measurable quantity, at least in principle, either by accurately measuring the equation of state ${\rm w} \Z{\phi}(z)$, or by placing a constraint on the dark matter-scalar coupling $Q$, or both.
![(Left) Form of the scalar potential implied by (\[potentiala\]), with $Q=0.2, 0 $ and $-0.2$ (top to bottom) (blue, green and red). (Right plot) Shape of a reconstructed scalar field potential.[]{data-label="potential"}](potential1.eps "fig:"){width="3.0in" height="2.0in"}0.4in ![(Left) Form of the scalar potential implied by (\[potentiala\]), with $Q=0.2, 0 $ and $-0.2$ (top to bottom) (blue, green and red). (Right plot) Shape of a reconstructed scalar field potential.[]{data-label="potential"}](potential2.eps "fig:"){width="2.8in" height="2.1in"}
One can think of the scalar field “rolling down” the potential, like a skier on a mountain. It can be seen from Fig. \[potential\] that the slope of the potential now is not quite flat; rather, the scalar field is rolling with a changing velocity. This is apparent when taking the time derivative of the scalar field, $\phi'=-\alpha-2 \zeta e^{2 \zeta N}$. At $N=0$, $|\phi^\prime|=\alpha+2\zeta$. Indeed, we require that the scalar field is rolling slowly so that the kinetic energy is small compared to the potential energy. The scalar field rolls at a slowly changing velocity for a small value of $\zeta$ and the acceleration is $\phi^{\prime\prime}(N=0)=-4\zeta^2$, so we expect that the value of $\zeta^2$ to be much less than $1$. Furthermore, $\phi^{\prime\prime}$ is negative so the scalar field is slowing down, it is in a ‘freezing’ phase. The model approaches the $\Lambda$CDM model as $\alpha, \zeta \to 0$.
The shape of the reconstructed potential shown in Fig. \[potential\] is essentially an exponential. In particular, in the limit $\zeta\to 0$, it is a simple exponential, $V(\phi)\propto e^{\alpha\phi}$, and for the non-zero $\zeta$ case this is modified by extra terms proportional to $e^{c\phi}\,(\phi-\phi_0)^2$. The reconstructed potential shown on the left panel of Fig. \[potential\] displays a range of possible values of $Q$; for larger $Q$ the potential is steeper, and for smaller $Q$ it is flatter and even changes shape for a negative $Q$.
![Scalar field equation of state as a function of redshift. We take the values as $\alpha=0.2$, $\zeta=0.05$ and $\Omega_{0
m}=0.27$.[]{data-label="wde"}](3dwde_Q.eps){width="3.2in" height="2.5in"}
The equation of state for our choice of the parameters is currently within the limits from WMAP, $-1.11< {\rm w}_\phi <
-0.86$, and does not change much up to redshift $z\sim 1$. The effect of a nonzero coupling $Q$ on the equation of state is shown in Fig. \[wde\]. At low redshift, a positive value of Q drives ${\rm w}_\phi$ closer to $-1$, whereas a negative Q gives a high value of ${\rm w}_\phi$ at high redshift.
Effect on the background
========================
An intrinsic property of dark energy is that is does not interact with light. The only way that it may be observed is through its effect on the evolution of the background and its possible interaction with dark matter. In this section we look into how this model for quintessential dark energy affects the background by quantitatively looking at its effects on the rate of cosmic expansion and the fractional densities of dark matter and dark energy. A useful parameter used when considering cosmic acceleration is the “deceleration parameter", $q$. In an accelerating epoch, $q <0$. In Fig. \[qplot\], we have shown a typical variation in the value of $q$ with redshift. The important feature of this reconstruction is that $q$ drops from positive to negative at $z\approx0.7$, when cosmic acceleration began [@Melchiorri:2007]. A nonzero coupling has an effect on the redshift when cosmic acceleration started, $z_{acc}$; the positive coupling ($Q>0$) implies an earlier start, and the negative coupling ($Q<0$) implies a later start.
![(Left) Deceleration parameter for varying $\alpha$ ($\zeta=0.05$, $Q=0.2$). (Right) History of cosmic acceleration for $Q=0.2, 0$ and $-0.2$ (bottom to top) ($\zeta=0.05$, $\alpha=0.2$).[]{data-label="qplot"}](qoveralpha.eps "fig:"){width="3.0in" height="2.0in"}0.1in ![(Left) Deceleration parameter for varying $\alpha$ ($\zeta=0.05$, $Q=0.2$). (Right) History of cosmic acceleration for $Q=0.2, 0$ and $-0.2$ (bottom to top) ($\zeta=0.05$, $\alpha=0.2$).[]{data-label="qplot"}](q-plot.eps "fig:"){width="2.7in" height="2.0in"}
An alternative way of looking at the start of cosmic acceleration is to see when the effective equation of state drops below $-1/3$ in Fig. \[wdeweff\]. At high redshift, the effective equation of state goes towards $0$, which is expected in the matter-dominated epoch (${\rm w}_m\simeq0$). For all values of $\alpha$, the effective equation of state goes to $\sim 0$ during the matter dominated epoch, but for large values of $\alpha$, ${\rm w}_{\rm
eff}>0 $. This is because the relative energy density of dark matter and dark energy now is fixed and the matter equation of state is constant, but the scalar equation of state is closer to $0$ for a larger $\alpha$, as shown in Fig. \[contour\]. At least, at low redshifts ($z<1$), the effective equation of state does not vary much for different values of the coupling.
![(Left plot) The scalar field equation of state and (right plot) the effective equation of state with $\zeta=0.05$, $\alpha=0.2$ and $Q=-0.2, 0$ and $+0.2$ (from top to bottom).[]{data-label="wdeweff"}](wdeweff.eps "fig:"){width="2.8in" height="2.1in"}0.1in ![(Left plot) The scalar field equation of state and (right plot) the effective equation of state with $\zeta=0.05$, $\alpha=0.2$ and $Q=-0.2, 0$ and $+0.2$ (from top to bottom).[]{data-label="wdeweff"}](weffover-z.eps "fig:"){width="2.8in" height="2.1in"}
The relative energy density of dark matter to dark energy is affected by changes in the $\alpha$ parameter, as seen in Fig. \[density\]. For smaller values of $\alpha$, the matter density tends towards a lower value at high redshift. That is, due to the Friedmann constraint (\[vary10\]), there is a higher proportion of dark energy during the matter dominated epoch for a small $\alpha$ component. Here, it is clear that there has been an assumption made for the matter density today. The right hand side plot of Fig. \[density\] illustrates the effect of a non-minimal coupling $Q$ on the evolution of matter density. The effect is not great within this small range, but there is a difference. A smaller value of the coupling parameter, $Q$, leads to an increase in the proportion of matter in the matter-dominated epoch. This study is valid for a range of low redshift, but will break down at high redshift when we enter the radiation-dominated epoch because we have made the assumption that the radiation component is negligible.
![(Left plot) Matter density for varying $\alpha$ ($\zeta=0.05$, $Q=0.2$) with $\Omega_{0m}=0.27$. (Right plot) Matter density for $Q=-0.2, 0$ and $0.2$ (top to bottom; online: red, green and blue) ($\zeta=0.05$, $\alpha=0.2$).[]{data-label="density"}](density2.eps "fig:"){width="2.8in" height="2.0in"}0.1in ![(Left plot) Matter density for varying $\alpha$ ($\zeta=0.05$, $Q=0.2$) with $\Omega_{0m}=0.27$. (Right plot) Matter density for $Q=-0.2, 0$ and $0.2$ (top to bottom; online: red, green and blue) ($\zeta=0.05$, $\alpha=0.2$).[]{data-label="density"}](density.eps "fig:"){width="2.8in" height="2.0in"}
One requirement of the potential is that it should dominate the kinetic term so that the scalar field equation of state, ${\rm
w}_{\phi}$, may be driven towards $-1$. It can be seen that this is satisfied by examining Figure \[wdeweff\]. Recently, the authors of [@Sahni:2008xx] have found two new diagnostic tools for distinguishing quintessential dark energy from Einstein’s cosmological constant. Their method may be extended to a non-minimally coupled theory for which $|Q|>0$. Here we simply note that for small values of $\alpha$ and $\zeta$, satisfying the constraints given in table 2, the constraints of the present model encompass the cosmological constant limit at the $1\sigma$ level at low redshifts.
Conclusion
==========
In this paper, a scalar-tensor theory of quintessential dark energy has been reconstructed. This reconstruction rests on the general action (\[my action\]), in which the tensor theory of general relativity is modified by including a fundamental scalar field. The scalar field affects the gravitational part of the action by introducing a gravitationally repulsive term which drives the cosmic acceleration, and is coupled to the matter part of the action, giving rise to a nontrivial dark matter - dark energy interaction. As for all quintessence models, this action reduces to the one of general relativity, when the scalar field is time-independent and the coupling of dark matter to a scalar field (or dark energy) is minimal.
We made an ansatz for the form of the scalar field itself, in order to be able to reconstruct relevant cosmological parameters, including $\Omega_m, \Omega_\phi, \epsilon$ and ${\rm w}_\phi$. The ansatz $|\phi_0 - \phi|/m\Z{P}= {\alpha} \ln a+ \exp
\left(2\zeta\ln a \right)+{\rm const}$ for the scalar field was motivated by generic solutions of the dilaton field in some effective string theory models in four dimensions. The two new parameters, $\alpha$ and $\zeta$ as well as the matter scalar coupling parameter, $Q$, were constrained using several cosmological methods. Primarily, the relationship between $\alpha$ and $\zeta$ was constrained by curve-fitting to various cosmological data sets. For example, with $\Omega_{0m}=0.27$ and minimal coupling, using just the SN1a data, minimizing $\chi^2$ gave $\alpha+2 \zeta=\pm 0.44\,(\pm 0.34)$ and using all the data sets available gave $\alpha+2 \zeta=\pm 0.3\,(\pm 0.3)$. We also considered the effect of dark energy on the growth of large scale structure. The effect of a nonminimal coupling on the standard expression for the rate of growth of matter perturbations, $f=\left[\Omega_m(N)\right]^\eta$, was considered. A modified ansatz for the dependence of matter density and dark matter-dark energy coupling on the matter fluctuation growth rate was introduced and then fitted to some observed values of the growth rate at low redshift. This analysis gave a best-fit minimum value for the coupling $Q=0.2\,\pm 0.2$ (with the input $\Omega_{0
m}\simeq 0.27$), which is compatible with the minimally coupled scalar field case. These methods of constraining the variables of the model have succeeded in constraining $\alpha+2\zeta$, but failed in constraining the separate values of $\alpha$ and $\zeta$. This is a direct result of taking the small $\zeta$ limit when analyzing the solution for $\Omega_m$. Nevertheless, by assuming the same field for the inflaton as quintessence the individual values of $\alpha$ and $\zeta$ were constrained using the spectral index of the CMB. We find $\alpha=0.20\pm0.04$ and $\zeta=0.04\pm0.15$.
Our analysis showed that the present model of coupled quintessence is compatible with a cosmological constant (where $\alpha, \zeta =
0$ and $Q=0$), within $1\sigma$ error bars. This means that dark energy may still simply be a cosmological constant, but even a small deviation away from this limit results in a very different source of cosmic acceleration, with time varying behavior. The number of parameters of this model may make it hard to rule out, as they may be tuned to satisfy many more constraints. In this work, the matter scalar coupling parameter was chosen to be a constant. This was not motivated physically but was chosen for the solvability of the system. Indeed, there is no physical reason why the dark energy - (dark) matter interaction should not vary in time, it is in all likelihood a function of $\phi$. There needs to be further study to explore this $\phi$ dependence of $Q$. Also, the links between quintessential dark energy and inflation deserve further investigation.
[99]{}
A. G. Riess [*et al.*]{} (Supernova Search Team Collaboration), Astron. J. [**116**]{}, 1009 (1998) \[arXiv:astro-ph/9805201\]. S. Perlmutter [*et al.*]{} (Supernova Cosmology Project Collaboration), Astrophys. J. [**517**]{}, 565 (1999) \[arXiv:astro-ph/9812133\]. A.G. Riess [*et al.*]{} (Supernova Search Team Collaboration), Astrophys. J. [**607**]{}, 665 (2004) \[astro-ph/0402512\]. R. A. Knop [*et al.*]{} Astroph. J. [**598**]{}, 102(K) \[arXiv:astro-ph/0309368\].
http://lambda.gsfc.nasa.gov/
D. N. Spergel [*et al.*]{} \[WMAP Collaboration\], [*First Year Wilkinson Microwave Anisotropy Probe (WMAP) Observations: Determination of Cosmological Parameters*]{}, Astrophys. J. Suppl. [**148**]{}, 175 (2003). D. N. Spergel [*et al.*]{} \[WMAP Collaboration\], [*Wilkinson Microwave Anisotropy Probe (WMAP) three year results: implications for cosmology*]{}, Astrophys. J. Suppl. [**170**]{}, 377 (2007) \[arXiv:astro-ph/0603449\]. S. M. Carroll, V. Duvvuri, M. Trodden and M. S. Turner, [*Is cosmic speed-up due to new gravitational physics?*]{}, Phys. Rev. D [**70**]{}, 043528 (2004) \[arXiv:astro-ph/0306438\]. S. Capozziello, V. F. Cardone, S. Carloni and A. Troisi, [*Curvature quintessence matched with observational data*]{}, Int. J. Mod. Phys. D [**12**]{}, 1969 (2003) \[arXiv:astro-ph/0307018\]. S. Nojiri and S. D. Odintsov, [*Modified gravity with negative and positive powers of the curvature: Unification of the inflation and of the cosmic acceleration*]{}, Phys. Rev. D [**68**]{}, 123512 (2003) \[arXiv:hep-th/0307288\]; P. J. E. Peebles and B. Ratra, [*Cosmology with a time variable cosmological [‘constant’]{}*]{}, , L17 (1988). C. Wetterich, [*Cosmology and the Fate of Dilatation Symmetry*]{}, Nucl. Phys. B [**302**]{}, 668 (1988). I. Zlatev, L. M. Wang and P. J. Steinhardt, [*Quintessence, Cosmic Coincidence, and the Cosmological Constant*]{}, Phys. Rev. Lett. [**82**]{}, 896 (1999) \[arXiv:astro-ph/9807002\]. E. Elizalde, S. Nojiri, S. D. Odintsov, D. Saez and V. Faraoni, [*Reconstructing the universe history, from inflation to acceleration, with phantom and canonical scalar fields*]{}, Phys. Rev. D [**77**]{}, 106005 (2008) \[arXiv:0803.1311 \[hep-th\]\]. S. Nojiri, S. D. Odintsov and M. Sasaki, [*Gauss-Bonnet dark energy*]{}, Phys. Rev. D [**71**]{}, 123509 (2005) \[arXiv:hep-th/0504052\]. B. M. N. Carter and I. P. Neupane, [*Dynamical relaxation of dark energy: Solution to both the inflation and the cosmological constant*]{}, Phys. Lett. B [**638**]{}, 94 (2006) \[arXiv:hep-th/0510109\]. B. M. N. Carter and I. P. Neupane, [*Towards inflation and dark energy cosmologies from modified Gauss-Bonnet theory*]{}, JCAP [**0606**]{}, 004 (2006) \[arXiv:hep-th/0512262\]. I. P. Neupane, [*Towards inflation and accelerating cosmologies in string-generated gravity models*]{}, arXiv:hep-th/0605265. S. Tsujikawa and M. Sami, [*String-inspired cosmology: Late time transition from scaling matter era to dark energy universe caused by a Gauss-Bonnet coupling*]{}, JCAP [**0701**]{}, 006 (2007) \[arXiv:hep-th/0608178\]. I. P. Neupane, [*On compatibility of string effective action with an accelerating universe*]{}, Class. Quant. Grav. [**23**]{}, 7493 (2006) \[arXiv:hep-th/0602097\]. M. Li, [*A model of holographic dark energy*]{}, Phys. Lett. B [**603**]{}, 1 (2004) \[arXiv:hep-th/0403127\]. H. Wei and R. G. Cai, [*A New Model of Agegraphic Dark Energy*]{}, Phys. Lett. B [**660**]{}, 113 (2008) \[arXiv:0708.0884\]. I. P. Neupane, [*Remarks on Dynamical Dark Energy Measured by the Conformal Age of the Universe*]{}, Phys. Rev. D [**76**]{}, 123006 (2007) \[arXiv:0709.3096\]. I. P. Neupane, [*A Note on Agegraphic Dark Energy*]{}, Phys. Lett. B [**673**]{} (2009) 111 \[arXiv:0708.2910\]. C. Deffayet, G. R. Dvali and G. Gabadadze, [*Accelerated universe from gravity leaking to extra dimensions*]{}, Phys. Rev. D [**65**]{}, 044023 (2002) \[arXiv:astro-ph/0105068\]. P. K. Townsend and M. N. R. Wohlfarth, [*Accelerating cosmologies from compactification*]{}, Phys. Rev. Lett. [**91**]{}, 061302 (2003) \[arXiv:hep-th/0303097\]. C. M. Chen, P. M. Ho, I. P. Neupane and J. E. Wang, [*A note on acceleration from product space compactification*]{}, JHEP [**0307**]{}, 017 (2003) \[arXiv:hep-th/0304177\]. I. P. Neupane, [*Accelerating cosmologies from exponential potentials*]{}, Class. Quant. Grav. [**21**]{}, 4383 (2004) \[arXiv:hep-th/0311071\]. N. Ohta, [*A study of accelerating cosmologies from superstring/M theories*]{}, Prog. Theor. Phys. [**110**]{} (2003) 269 \[arXiv:hep-th/0304172\].
I. P. Neupane, [*Simple cosmological de Sitter solutions on dS$_4 \times Y_6$ spaces*]{}, Class. Quan. Grav. [**27**]{} (2010) 045011 \[arXiv:0901.2568\]; I. P. Neupane, [*Accelerating universe from warped extra dimensions*]{}, Class. Quant. Grav. [**26**]{} (2009) 195008 \[arXiv:0905.2774\]. D. L. Wiltshire, [*Exact solution to the averaging problem in cosmology*]{}, Phys. Rev. Lett. [**99**]{} (2007) 251101 \[arXiv:0709.0732\]. D. L. Wiltshire, [*Dark energy without dark energy*]{}, arXiv:0712.3984 \[astro-ph\]. P. J. E. Peebles and B. Ratra, [*The cosmological constant and dark energy*]{}, Rev. Mod. Phys. [**75**]{} (2003) 559 \[arXiv:astro-ph/0207347\]. U. Alam, V. Sahni, T. D. Saini and A. A. Starobinsky, [*Is there Supernova Evidence for Dark Energy Metamorphosis?*]{}, Mon. Not. Roy. Astron. Soc. [**354**]{} (2004) 275 \[arXiv:astro-ph/0311364\]. V. Sahni, [*Dark matter and dark energy*]{}, Lect. Notes Phys. [**653**]{} (2004) 141 \[arXiv:astro-ph/0403324\]. E. J. Copeland, M. Sami and S. Tsujikawa, [*Dynamics of dark energy*]{}, Int. J. Mod. Phys. D [**15**]{} (2006) 1753 \[arXiv:hep-th/0603057\]. M. Sami, [*Dark energy and possible alternatives*]{}, arXiv:0901.0756 \[hep-th\]. Z. K. Guo, N. Ohta and Y. Z. Zhang, [*Parametrization of quintessence and its potential*]{}, Phys. Rev. D [**72**]{} (2005) 023504 \[arXiv:astro-ph/0505253\]. S. Weinberg, [*The cosmological constant problem*]{}, Rev. Mod. Phys. [**61**]{} (1989) 1. V. Sahni, [*The cosmological constant problem and quintessence*]{}, Class. Quant. Grav. [**19**]{} (2002) 3435 \[arXiv:astro-ph/0202076\]. T. Padmanabhan, [*Cosmological constant: The weight of the vacuum*]{}, Phys. Rept. [**380**]{} (2003) 235 \[arXiv:hep-th/0212290\]. I. P. Neupane, [*Reconstructing a model of quintessential inflation*]{}, Class. Quant. Grav. [**25**]{} (2008) 125013 \[arXiv:0706.2654\]. E. Komatsu [*et al.*]{} \[WMAP Collaboration\], [*Five-Year Wilkinson Microwave Anisotropy Probe (WMAP) Observations:Cosmological Interpretation*]{}, arXiv:0803.0547 \[astro-ph\]. D. J. Eisenstein [*et al.*]{} \[SDSS Collaboration\], [*Detection of the Baryon Acoustic Peak in the Large-Scale Correlation Function of SDSS Luminous Red Galaxies*]{}, Astrophys. J. [**633**]{} (2005) 560 \[arXiv:astro-ph/0501171\]. A. G. Riess [*et al.*]{}, [*New Hubble Space Telescope Discoveries of Type Ia Supernovae at $z > 1$: Narrowing Constraints on the Early Behavior of Dark Energy*]{}, Astrophys. J. [**659**]{} (2007) 98 \[arXiv:astro-ph/0611572\]. P. Astier [*et al.*]{} \[The SNLS Collaboration\], [*The Supernova Legacy Survey: Measurement of $\Omega_M$, $\Omega_\Lambda$ and ${\rm w}$ from the First Year Data Set*]{}, Astron. Astrophys. [**447**]{} (2006) 31 \[arXiv:astro-ph/0510447\]. C. Brans and R. H. Dicke, [*Mach’s principle and a relativistic theory of gravitation*]{}, Phys. Rev. [**124**]{} (1961) 925. Y. M. Cho, [*Reinterpretation of Jordan-Brans-Dicke theory and Kaluza-Klein cosmology*]{}, Phys. Rev. Lett. [**68**]{} (1992) 3133. Y. M. Cho and J. H. Kim, [*Dilatonic dark matter and its experimental detection*]{}, Phys. Rev. D [**79**]{} (2009) 023504 \[arXiv:0711.2858\].
G. Esposito-Farese and D. Polarski, Phys. Rev. D [**63**]{} (2001) 063504 \[arXiv:gr-qc/0009034\]. L. Amendola, [*Coupled quintessence*]{}, Phys. Rev. D [**62**]{} (2000) 043511 \[arXiv:astro-ph/9908023\].
T. Damour, G. W. Gibbons and C. Gundlach, [*Dark matter, time varying G, and a dilaton field*]{}, Phys. Rev. Lett. [**64**]{} (1990) 123. T. Damour and G. Esposito-Farese, [*Tensor-scalar gravity and binary-pulsar experiments*]{}, Phys. Rev. D [**54**]{} (1996) 1474 \[arXiv:gr-qc/9602056\]. B. M. Leith and I. P. Neupane, [*Gauss-Bonnet cosmologies: crossing the phantom divide and the transition from matter dominance to dark energy*]{}, JCAP [**0705**]{} (2007) 019 \[arXiv:hep-th/0702002\]. I. P. Neupane, [*Constraints on Gauss-Bonnet Cosmologies*]{}, arXiv:0711.3234 \[hep-th\]. R. Bean, E. E. Flanagan, I. Laszlo and M. Trodden, [*Constraining interactions in cosmology’s dark sector*]{}, Phys. Rev. D [**78**]{} (2008) 123514 \[arXiv:0808.1105\]. M. B. Gavela, D. Hernandez, L. L. Honorez, O. Mena and S. Rigolin, [*Dark coupling*]{}, arXiv:0901.1611 \[astro-ph\]. I. P. Neupane and C. Scherer, [*Inflation and Quintessence: Theoretical Approach of Cosmological Reconstruction*]{}, JCAP [**0805**]{}(2008) 009 \[arXiv:0712.2468\]. M. Chevallier and D. Polarski, [*Accelerating universes with scaling dark matter*]{}, Int. J. Mod. Phys. D [**10**]{} (2001) 213 \[arXiv:gr-qc/0009008\]. E. V. Linder, [*Exploring the expansion history of the universe*]{}, Phys. Rev. Lett. [**90**]{} (2003) 091301 (2003) \[arXiv:astro-ph/0208512\]. V. Sahni, A. Shafieloo and A. A. Starobinsky, [*Two new diagnostics of dark energy*]{}, Phys. Rev. D [**78**]{} (2008) 103502 \[arXiv:0807.3548\]. C. Cattoen and M. Visser, [*Cosmographic Hubble fits to the supernova data*]{}, Phys. Rev. D [**78**]{} (2008) 063501 \[arXiv:0809.0537\]. S. Weinberg, [*Gravitation and Cosmology*]{}, John Wiley and Sons, United State, 1972.
O. Lahav, P. B. Lilje, J. R. Primack and M. J. Rees, [*Dynamical effects of the cosmological constant*]{}, Mon. Not. Roy. Astron. Soc. [**251**]{} (1991) 128. L. Guzzo [*et al.*]{}, [*A test of the nature of cosmic acceleration using galaxy redshift distortions*]{}, Nature [**451**]{} (2008) 541 \[arXiv:0802.1944\]. L. Amendola, [*Linear and non-linear perturbations in dark energy models*]{}, Phys. Rev. D [**69**]{} (2004) 103524 \[arXiv:astro-ph/0311175\]. T. Damour and K. Nordtvedt, [*Tensor - scalar cosmological models and their relaxation toward general relativity*]{}, Phys. Rev. D [**48**]{} (1993) 3436. B. Bertotti, L. Iess and P. Tortora, [*A test of general relativity using radio links with the Cassini spacecraft*]{}, Nature [**425**]{} (2003) 374. T. Damour, F. Piazza and G. Veneziano, [*Runaway dilaton and equivalence principle violations*]{}, Phys. Rev. Lett. [**89**]{} (2002) 081601 \[arXiv:gr-qc/0204094\]. T. Chiba, T. Kobayashi, M. Yamaguchi and J. Yokoyama, [*Time variation of proton-electron mass ratio and fine structure constant with runaway dilaton*]{}, Phys. Rev. D [**75**]{} (2007) 043516 \[arXiv:hep-ph/0610027\].
S. Nesseris and L. Perivolaropoulos, [*The limits of extended quintessence*]{}, Phys. Rev. D [**75**]{} (2007) 023517 \[arXiv:astro-ph/0611238\]. S. Nesseris and L. Perivolaropoulos, [*Testing LCDM with the Growth Function: Current Constraints*]{}, Phys. Rev. D [**77**]{} (2008) 023504 \[arXiv:0710.1092\]. C. Di Porto and L. Amendola, [*Observational constraints on the linear fluctuation growth rate*]{}, Phys. Rev. D [**77**]{} (2008) 083508 \[arXiv:0707.2686\]. I. Antoniadis, J. Rizos and K. Tamvakis, [*Singularity - free cosmological solutions of the superstring effective action*]{}, Nucl. Phys. B [**415**]{} (1994) 497 \[arXiv:hep-th/9305025\]. M. Gasperini, F. Piazza and G. Veneziano, [*Quintessence as a runaway dilaton*]{}, Phys. Rev. D [**65**]{} (2002) 023508 \[arXiv:gr-qc/0108016\]. T. Barreiro, B. de Carlos and E. J. Copeland, [*Stabilizing the Dilaton in Superstring Cosmology*]{}, Phys. Rev. D [**58**]{} (1998) 083513 \[arXiv:hep-th/9805005\]. B. Boisseau, G. Esposito-Farese, D. Polarski and A. A. Starobinsky, [*Reconstruction of a scalar-tensor theory of gravity in an accelerating universe*]{}, Phys. Rev. Lett. [**85**]{} (2000) 2236 \[arXiv:gr-qc/0001066\]. C. Armendariz-Picon, V. F. Mukhanov and P. J. Steinhardt, [*Essentials of k-essence*]{}, Phys. Rev. D [**63**]{} (2001) 103510 \[arXiv:astro-ph/0006373\].
V. Sahni and A. Starobinsky, [*Reconstructing Dark Energy*]{}, Int. J. Mod. Phys. D [**15**]{} (2006) 2105 \[arXiv:astro-ph/0610026\]. R. J. Scherrer and A. A. Sen, [*Thawing quintessence with a nearly flat potential*]{}, Phys. Rev. D [**77**]{} (2008) 083515 \[arXiv:0712.3450\]. R. J. Scherrer and A. A. Sen, [*Phantom Dark Energy Models with a Nearly Flat Potential*]{}, Phys. Rev. D [**78**]{} (2008) 067303 \[arXiv:0808.1880\]. S. Dutta and R. J. Scherrer, [*Hilltop Quintessence*]{}, Phys. Rev. D [**78**]{} (2008) 123525 \[arXiv:0809.4441\]. N. J. Nunes, D. F. Mota [*Structure Formation in Inhomogeneous Dark Energy Models*]{}, Mon. Not. Roy. Astron. Soc. [**368**]{} (2006) 751 \[astro-ph/0409481\].
L. Samushia and B. Ratra, [*Constraints on Dark Energy from Galaxy Cluster Gas Mass Fraction versus Redshift data*]{}, Astrophys. J. [**680**]{} (2008) L1 \[arXiv:0803.3775\]. A. J. Albrecht and C. Skordis, [*Phenomenology of a realistic accelerating universe using only Planck-scale physics*]{}, Phys. Rev. Lett. [**84**]{} (2000) 2076 \[arXiv:astro-ph/9908085\].
H. Wei, [*Growth Index of DGP Model and Current Growth Rate Data*]{}, Phys. Lett. B [**664**]{} (2008) 1 \[arXiv:0802.4122\]. H. Wei and S. N. Zhang, [*How to Distinguish Dark Energy and Modified Gravity?*]{}, Phys. Rev. D [**78**]{} (2008) 023011 \[arXiv:0803.3292\]. U. Alam, V. Sahni and A. A. Starobinsky, [*Reconstructing Cosmological Matter Perturbations using Standard Candles and Rulers*]{}, Astrophys. J. [**704**]{} (2009) 1086 \[arXiv:0812.2846\]. A. Melchiorri, L. Pagano and S. Pandolfi, [*When Did Cosmic Acceleration Start?*]{}, Phys. Rev. D [**76**]{} (2007) 041301 \[arXiv:0706.1314\].
[^1]: [email protected]
[^2]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the effect of different feedback prescriptions on the properties of the low redshift ($z\leq1.6$) forest using a selection of hydrodynamical simulations drawn from the Sherwood simulation suite. The simulations incorporate stellar feedback, AGN feedback and a simplified scheme for efficiently modelling the low column density forest. We confirm a discrepancy remains between *Cosmic Origins Spectrograph* (COS) observations of the forest column density distribution function (CDDF) at $z \simeq 0.1$ for high column density systems ($N_{\rm HI}>10^{14}\rm\,cm^{-2}$), as well as velocity widths that are too narrow compared to the COS data. Stellar or AGN feedback – as currently implemented in our simulations – have only a small effect on the CDDF and velocity width distribution. We conclude that resolving the discrepancy between the COS data and simulations requires an increase in the temperature of overdense gas with $\Delta=4$–$40$, either through additional photo-heating at $z>2$ or fine-tuned feedback that ejects overdense gas into the IGM at just the right temperature for it to still contribute significantly to the forest. Alternatively a larger, currently unresolved turbulent component to the line width could resolve the discrepancy.'
author:
- |
Fahad Nasir$^{1}$[^1], James S. Bolton$^{1}$, Matteo Viel$^{2,3,4}$, Tae-Sun Kim$^{5}$, Martin G. Haehnelt$^{6}$, Ewald Puchwein$^{6}$ & Debora Sijacki$^{6}$\
$^{1}$School of Physics and Astronomy, University of Nottingham, University Park, Nottingham, NG7 2RD, UK\
$^{2}$SISSA - International School for Advanced Studies, Via Bonomea 265, I-34136 Trieste, Italy\
$^{3}$INAF - Osservatorio Astronomico di Trieste, Via G.B. Tiepolo 11, I-34131 Trieste, Italy\
$^{4}$INFN - National Institute for Nuclear Physics, Via Valerio 2, I-34127 Trieste, Italy\
$^{5}$Department of Astronomy, University of Wisconsin-Madison, 475 N. Charter St., Madison, WI 53706, USA\
$^{6}$Kavli Institute for Cosmology and Institute of Astronomy, Madingley Road, Cambridge, CB3 0HA, UK
bibliography:
- 'bibliography.bib'
title: 'The effect of stellar and AGN feedback on the low redshift Lyman-$\alpha$ forest in the Sherwood simulation suite'
---
\[firstpage\]
methods: numerical – intergalactic medium – quasars: absorption lines
Introduction {#sec:intro}
============
The forest is an indispensable tool for probing the evolution and distribution of gas in the intergalactic medium [see @Meiksin2009; @McQuinn2016 for reviews]. At intermediate redshifts ($2\la
z\la5$), the majority of the intergalactic medium (IGM) by volume is in the form of warm ($T\sim 10^{4}\rm\,K$) diffuse gas that is kept photo-ionised by the metagalactic ultraviolet background (UVB). The latest hydrodynamical simulations of the IGM that model the forest of absorption produced by this gas are in very good agreement with a wide range of spectroscopic data from optical, ground-based telescopes [e.g. @Bolton_2017MNRAS]. However, at lower redshifts ($z\la 2$), due to the interplay between cosmic expansion, the growth of structure and the declining intensity of the UVB, the forest thins out and becomes more transmissive [@Danforth_2016ApJ]. By $z\sim 0$ a significant fraction of the baryons in the IGM are in a collisionally ionised, shock heated phase with temperatures $T\sim 10^{5}$–$10^{7}\rm\,K$ [@Cen1999]. In the present day Universe, the forest therefore traces less than one half of the remaining diffuse gas in the IGM [@Lehner_2007ApJ; @Shull_2012ApJ].
Over the past two decades, the properties of absorption arising from the low redshift IGM have been studied with observations made by a series of space-based UV spectrographs . Most recently, the *Cosmic Origins Spectrograph* [COS, @Green_2012ApJ], mounted on the *Hubble Space Telescope* (HST) has collected many more active galactic nuclei (AGN) spectra with improved signal-to-noise, providing new opportunities for absorption line studies of the low redshift forest [@Shull_2014ApJ; @Danforth_2016ApJ; @Pachat_2016MNRAS]. Some of the key questions addressed by recent investigations are related to the nature of these absorbers, and their relationship to the evolution of the UVB and galactic feedback [see e.g. @Dave_2010MNRAS; @Tepper-Garcia_2012MNRAS; @Kollmeier_2014ApJ; @Shull_2015ApJ; @Khaire_2016MNRAS; @Viel_2016; @Gurvich_2016].
The mostly widely used measurements of absorbers at $z<2$ are the column density distribution function (CDDF), the velocity (Doppler) widths of the absorption lines and the evolution of the absorption line number density with redshift. The CDDF shape and normalisation is determined by the density distribution of absorbers and the intensity of the UVB. The velocity width distribution provides constraints on the gas temperature. Furthermore, both observables are in principle sensitive to galactic feedback. In this context, hydrodynamical simulations have significantly aided our understanding and interpretation of these observational data .
Of particular note is the recent comparison of the CDDF for absorbers at $z\simeq 0.1$ observed with COS [@Danforth_2016ApJ] to hydrodynamical simulations. @Kollmeier_2014ApJ reported that the UVB intensity needed to match the COS measurement of the CDDF is a factor of five larger than predicted by empirically calibrated UVB models [@Haardt_2012ApJ HM12]. The claimed “photon underproduction crisis” prompted further investigation into this apparent discrepancy. Large uncertainties in the production rate of ionising photons by quasars and star forming galaxies, and possibly also the role of AGN feedback, have been suggested as the likely cause [@Khaire2015; @Shull_2015ApJ; @Wakker_2015ApJ; @Gurvich_2016; @Gaikwad_2016].
@Viel_2016 (hereafter V17) also recently considered this problem using hydrodynamical simulations including Illustris [@Vogelsberger2014] and a sub-set of the Sherwood simulations [@Bolton_2017MNRAS]. They concluded the HM12 photo-ionisation rate needed to be increased by a factor of two in the hydrodynamical simulations to match the observed CDDF, smaller than the factor of five found by @Kollmeier_2014ApJ. However, in contrast to other recent studies, V17 also focussed on the distribution of line velocity widths [see also @Gaikwad17]. V17 found that the simulations struggled to reproduce the velocity width distribution at $z\simeq0.1$, producing line widths that were too narrow compared to the COS data. This was attributed to gas in the simulations that was either too cold to produce the observed line widths, or too highly (collisionally) ionised to produce absorption due to vigorous AGN feedback heating the gas to $T>10^{5}\rm \,K$ (particularly in the Illustris simulation). V17 concluded that the line width distribution at $z\simeq 0.1$ is a valuable diagnostic for models of feedback in the low redshift Universe.
The goal of the present work is to the extend the V17 analysis, and present in full the low redshift ($0<z<2$) IGM models that V17 used from the Sherwood simulation suite [@Bolton_2017MNRAS]. Following V17, we perform the same Voigt profile analysis on the simulations and COS data to minimise potential biases in the comparison due to line fitting. This furthermore enables a detailed assessment of the effect that our stellar and AGN feedback implementation has on the forest.
The paper is organised as follows. In Section 2, we briefly describe the simulations used in this work. In Section 3, we compare the simulations to current observational data at $z<2$ and assess the impact of different galactic feedback prescriptions on forest observables, including the the velocity width distribution. The physical origin of the absorbers is examined in Section 4. Finally, we summarise our results in Section 5. A series of convergence tests are given in the Appendix. Throughout this paper we refer to comoving distances with the prefix “c”. The cosmological parameters used throughout this work are $\Omega_{\rm m}=0.308$, $\Omega_{\rm
\Lambda}=0.692$, $h=0.678$, $\Omega_{\rm b}=0.0482$, $\sigma_{\rm
8}=0.829$ and $n=0.961$ consistent with the best fit $\Lambda {\rm
CDM}+Planck+WP+highL+BAO$ cosmological parameters .
Methodology
===========
Hydrodynamical simulations {#sec:sim}
--------------------------
The simulations used in this work are summarised in Table \[tab:simulation\]. These were performed using <span style="font-variant:small-caps;">P-Gadget-3</span>, a modified version of the publicly available code <span style="font-variant:small-caps;">Gadget-2</span> [@Springel_2005MNRAS]. The initial conditions were generated at $z=99$ on a Cartesian grid using the <span style="font-variant:small-caps;">N-GenIC</span> code [@Springel_2005Nat] using transfer functions generated by <span style="font-variant:small-caps;">CAMB</span> [@Lewis_2000ApJ]. A detailed description of these models may be found in the Sherwood overview paper [@Bolton_2017MNRAS].
----------------- --------------------- --------------------- ----------------------
Name $M_{\rm dm} $ $M_{\rm gas}$ $l_{\rm soft}$
$[h^{-1} M_{\sun}]$ $[h^{-1} M_{\sun}]$ $[h^{-1} \rm{ckpc}]$
80-512-ps13 $2.75\times 10^8$ $5.10\times 10^7$ $6.25$
80-512-ps13+agn $2.75\times 10^8$ $5.10\times 10^7$ $6.25$
80-512 $2.75\times 10^8$ $5.10\times 10^7$ $6.25$
80-1024 $3.44\times 10^7$ $6.38\times 10^6$ $3.13$
80-256 $2.20\times 10^9$ $4.08\times 10^8$ $12.50$
40-512 $3.44\times 10^7$ $6.38\times 10^6$ $3.13$
20-256 $3.44\times 10^7$ $6.38\times 10^6$ $3.13$
----------------- --------------------- --------------------- ----------------------
: The hydrodynamical simulations used in this work. The first column uses the naming convention *L-N-param*. Here $L$ is the box size in units of , $N$ is the cube root of the total number of gas particles and *-param* is associated with the subgrid treatment of star formation and feedback. Here $-ps13$ refers to the star formation and energy driven outflow model of @Puchwein_2013MNRAS, while *-ps13+agn* also includes AGN feedback based on @Sijacki2007. All other models instead convert all cold ($T<10^{5}\rm\,K$), dense ($\Delta>1000$) gas into collisionless particles. The remaining columns give the mass of each dark matter and gas particle in $h^{-1} M_{\sun}$ and the gravitational softening length in $h^{-1} \rm{ckpc}$. All simulations are performed to $z=0$.
\[tab:simulation\]
The photo-ionisation and photo-heating rates are calculated using a spatially uniform UVB model for emission from star forming galaxies and quasars [@Haardt_2012ApJ] applied in the optically thin limit. A small increase to the photo-heating rate, $\epsilon_{\rm HeII}=1.7\epsilon_{\rm HeII}^{\rm HM12}$ at $2.2<z<3.4$ was applied to better match observational measurements of the IGM temperature at $z>2$ [@Becker_2011MNRAS]. The ionised fractions for hydrogen and helium are obtained assuming ionisation equilibrium. Metal line cooling is not included, although this is not expected to impact significantly on the forest at the redshifts considered here [@Tepper-Garcia_2012MNRAS]. In five of the simulations, gas particles with temperatures $T<10^5$K and densities $\Delta=\rho/\langle \rho \rangle >1000$ are converted into collisionless star particles [the <span style="font-variant:small-caps;">QUICKLYA</span> method, see @Viel2004] to speed up the calculation. We will show this choice makes very little difference to forest systems at low column densities, $N_{\rm HI}<10^{14.5}\rm\,cm^{-2}$, in our analysis. However, galactic winds and AGN feedback may have an impact on high column density lines that probe the (circumgalactic) gas in the vicinity of dark matter haloes. Therefore, two models that follow star formation and feedback are included. The first model (80-512-ps13) includes energy driven galactic outflows following @Puchwein_2013MNRAS. The star formation model is based on @Springel_2003MNRAS, but assumes a Chabrier rather than Salpeter initial mass function (IMF) and a galactic wind velocity that is directly proportional to the escape velocity of the galaxy. These choices increase the available supernovae feedback energy by a factor of two, as well as raising the mass-loading of winds in low-mass galaxies. The second run (80-512-ps13+agn) also incorporates AGN feedback, again following @Puchwein_2013MNRAS. This AGN feedback model is based on @Sijacki2007. In the “quasar” mode, when accretion rates are above $0.01$ of the Eddington rate, only $0.5$ per cent of the accreted rest mass energy is thermally coupled to the surrounding gas. For lower accretion rates a more efficient coupling of AGN jets in the “radio” mode is assumed and $2$ per cent of the rest mass energy is used for recurrently injecting hot AGN bubbles.
The evolution of the low-density IGM thermal history at $z<2$ in a sub-set of the simulations is displayed in Figure \[fig:ToGamma\]. This is parameterised as a power law temperature-density ($T-\Delta$) relation for the low density ($\Delta<1$–$10$) IGM, $T=T_0
\Delta^{\gamma-1}$. Here $\Delta=\rho/\langle\rho\rangle$ is the gas density normalised by the background density, $T_0$ is the gas temperature at mean density and $\gamma-1$ is the slope of the relation [@Gnedin_1998MNRAS; @Sanderbeck_2016MNRAS]. Note that the 80-512-ps13, 80-512-ps13+agn and 80-512 models have very similar thermal histories; this suggests gas at the mean background density is not strongly impacted by feedback. Additionally, by comparing these models to the 80-1024 simulation (solid black line), it is clear that the lower resolution simulations are not converged, with volume weighted temperatures at mean density $\sim 800\rm\,K$ hotter in the 80-512 runs (further details may be found in the Appendix).
There are currently no detailed observational constraints on the temperature of the low density IGM at $z<1.6$, but there are a few measurements from optical data at $1.6\leq z \leq 2$. In the upper left panel of Figure \[fig:ToGamma\] the red triangles show the IGM temperature at mean density from @Boera_2014. This study uses the curvature of the forest transmitted flux to measure the gas temperature at the characteristic density, $T(\bar{\Delta})$, probed by the forest absorption. We derive values for $T_0$ from these data assuming the $\gamma-1$ for the 80-1024 model shown in the lower left panel, such that $T_0=T(\bar{\Delta})/\bar{\Delta}^{\gamma-1}$. Additionally, the cyan squares display the @Boera_2014 measurements obtained assuming the effective optical depth, (where $ = -\ln \langle F \rangle$ and $\langle F
\rangle$ is the mean transmission in the forest), evolution used in an earlier curvature analysis performed by @Becker_2011MNRAS ([E. Boera, private communication]{}). The effective optical depth sets the characteristic gas density probed by the forest, and this comparison gives an indication of the likely systematic uncertainty in $\tau_{\rm eff}$ arising from line of sight variance. For completeness, we also directly compare the simulations to the $T(\bar{\Delta})$ measurements in the right panel of Figure \[fig:ToGamma\]. The measurements bracket all the simulations at $z>1.6$. The models are also formally consistent with the constraints on $\gamma-1$ from @Schaye_2000MNRAS and @Ricotti_2000ApJ shown in the lower left panel of Figure \[fig:ToGamma\]. These were obtained from an analysis of the lower envelope of the column density – velocity width plane obtained from Voigt profile fits to the forest, although the error bars on these measurements are very large. Note also the simulations do not follow non-equilibrium ionisation effects, and therefore produce a temperature-density relation slope that is steeper by $\Delta(\gamma-1)\sim 0.1$ at $z=1$–$2$ compared to the recent study by @Puchwein_2015MNRAS.
In all cases the simulated $T_{0}$ monotonically decreases toward lower redshift as the IGM cools due to adiabatic expansion, reaching $\sim 4000\rm\,K$ by $z=0$. The agreement with measurements of the IGM thermal state at $1.6<z<2$ suggests the photo-heating rates used in the UVB model are broadly correct at these (and higher) redshifts. Significantly higher temperatures at mean density than those predicted by the models at $z<1.6$ would require an additional, unidentified source of heating in the low density IGM.
Mock forest spectra
-------------------
Mock forest spectra were obtained by extracting $5000$ lines of sight parallel to each of the box axes on regularly spaced grids of $50^{2}$, $40^{2}$ and $30^{2}$, each with $2048$ pixels. The optical depths, $\tau^{\alpha}_{\rm HI}$, are obtained using the method described by @Theuns_1998MNRAS combined with the Voigt profile approximation described by @Tepper-Garcia_2006MNRAS. The transmitted flux, $F$, in each pixel is then $F=e^{- \tau^{\rm \alpha}_{\rm HI}}$, and the effective optical depth evolution with redshift is averaged over all sight lines. In our analysis we also make use of the fact that the pixel optical depths are inversely proportional to the photo-ionisation rate, $\tau^{\alpha}_{\rm HI} \propto \Gamma_{\rm HI}^{-1}$. This enables the optical depths (and hence the evolution) to be rescaled to match different UVB models in post-processing [e.g. @Rauch_1997ApJ].
In Figure \[fig:obs\_cmp\] (left panel), the $\tau_{\rm eff}$ evolution for a variety of sub-grid physics and UVB models are compared with forest observations. The results shown by the black curves use the HM12 UVB model, whereas the red triple-dot dashed and blue dashed curves are obtained by scaling the simulated optical depths to match the photo-ionisation rates in the earlier @Haardt_2001 [HM01] and @Faucher_2009ApJ UVB models for emission from star-forming galaxies and quasars. The evolution of in the 80-1024 model when using the HM01 UVB is in reasonable agreement with observational measurements from @Kirkman_2007MNRAS (except the dip at $z\simeq1.4$) and @Kim_2007MNRAS at $1.7<z<2.0$, but underpredicts at $z<0.4$. In contrast, the fiducial HM12 UVB predicts values that are too high, particularly $z<1.5$; a closely related result was noted by @Kollmeier_2014ApJ when analysing the CDDF. None of the UVB models is able to reproduce a evolution that is in good agreement with the more recent COS data from @Danforth_2016ApJ at $z\leq0.4$. There is also a difference between these data and the measurements from @Kirkman_2007MNRAS at $z\sim 0.4$, possibly due to systematic differences in the measurements and cosmic variance. The @Faucher_2009ApJ UVB is closer to these data than HM12 at $z<0.5$, but overpredicts at all redshifts including $z=2$. Interestingly, the differences between models with different feedback implementations are small, indicating that it is the uncertain amplitude of the UVB rather than galactic feedback that is largely responsible for this discrepancy [see also figure 1 in @Dave_2010MNRAS].
This is further illustrated in the right panel of Figure \[fig:obs\_cmp\] where observational estimates of the photoionisation rate, $\Gamma_{\rm -13} = \Gamma_{\rm HI}/10^{\rm
-13} {\rm s}^{\rm -1}$, are compared to the three UVB models. The measurements by @Shull_2015ApJ [@Gaikwad_2016] and V17, based on an analysis of the COS CDDF, are in mutual agreement and predict photo-ionisation rates at $z<0.4$ intermediate between HM01 and HM12. The value proposed by @Kollmeier_2014ApJ (cyan triangle) lies very close to the HM01 UVB at $z=0.1$. The red inverted triangles with large error bars display earlier constraints on $\Gamma_{\rm HI}$ obtained from the proximity effect by @Scott_2002ApJ. The blue rectangle corresponds to the recent, independent constraint from @Fumagalli_2017 using observations of H$\alpha$ fluorescence from in a nearby galactic disk [see also @Adams2011].
Differing assumptions regarding the relative contributions of star forming galaxies and quasars and their spectral shape are largely responsible for differences between existing UVB models. Given these considerable differences, efforts to calibrate the photo-ionisation rate to match and CDDF measurements at $z<0.4$ [e.g. @Khaire2015; @Madau2015], as well as renewed attempts to obtain measurements of $\Gamma_{\rm HI}$ in the redshift interval $0.4< z < 2$, are desirable.
Voigt profile analysis {#sec:cddf_data}
======================
We now proceed to present the results of the Voigt profile analysis of our mock forest spectra. We fit lines at three redshifts, $z=[0.1,\,1,\,1.6]$. When comparing simulations to each other or to data compilations that have a range of resolution and signal-to-noise properties, the mock spectra are post-processed by convolving them with an instrument profile with a Full Width Half Maximum (FWHM) of $7$ and rebinned to a pixel size of 3 . A uniform Gaussian distributed noise is added with a signal-to-noise of S/N$=50$ per pixel. The mean transmission of the spectra are rescaled to match $=[0.012,\,0.043,\,0.086]$ at $z=[0.1,\,1,\,1.6]$. These values correspond to the 80-1024 model evolution for the HM01 UVB for emission from star forming galaxies and quasars, which is in good agreement with the @Kirkman_2007MNRAS and @Kim_2007MNRAS $\tau_{\rm eff}$ measurements in the left panel of Figure \[fig:obs\_cmp\].
The exception to this is when we directly compare the simulations to the V17 COS measurements of the CDDF and velocity width distribution at $z=0.1$. In this case, we follow V17 and convolve the spectra with the LF1 G130M COS line spread function including scattering at wavelength $1350\,$[^2]. The spectra are rebinned to pixels of $\sim 7.2 \rm \, km\,s^{-1}$ and a Gaussian distributed signal-to-noise of S/N$=30$ per resolution element (or equivalently $\sim 20$ per pixel) is then added. The effective optical depth is scaled to $=0.021$ ($\langle F
\rangle =0.979$). This corresponds to the value V17 required to match the observed CDDF in the range $10^{13} \leq N_{\rm HI}/{\rm cm}^{-2}
\leq 10^{14}$.
The COS observations of the CDDF and velocity width distribution at $z=0.1$ are as described by V17; further details regarding the data reduction and AGN spectra can also be found in @Wakker_2015ApJ and Kim et al. (in prep). The data comprises of 704 lines within column densities $10^{12.5} \leq N_{\rm HI}/\rm cm^{-2} \leq 10^{14.5}
$ with a total redshift path length of $\Delta z=4.991$, covering the forest at $0<z<0.2$. The Voigt profile fits are obtained using absorption only.
The fitting of Voigt profiles to the mock spectra is performed with <span style="font-variant:small-caps;">VPFIT</span> [@VPFIT], which deconvolves the (already convolved) mock spectra with the instrument profile to obtain intrinsic line widths. Importantly, this approach matches that used to fit the COS observational data at $z=0.1$. A total path length of $10^5$ is fitted for each model in segments of $20$ to obtain our simulated line lists. All lines within $50\,$ of the start and end of each segment are ignored to avoid spurious line fits due to edge effects. Finally, to estimate the sample variance in the COS observational data, we adopt the approach of @Rollinde_2013MNRAS and bootstrap resample the mock spectra with replacement over the observed path length $\Delta z=4.991$. The sampling is performed 1000 times over the total simulated path length of $10^{5}h^{-1}\rm\,cMpc$.
The CDDF
--------
\[fig:CDDF\_Phys\]
Figure \[fig:CDDF\_COS\] displays the comparison between the CDDF, defined as the number of absorbers per unit log column density per unit redshift, $f(N_{\rm HI},z)=\partial^2 N/\partial \log N_{\rm
HI}\partial z$, from the Sherwood models and the COS observations described by V17 at $z=0.1$ (see also figure 1 in V17). Note that only the 80-512-ps13+agn model was examined in detail by V17; in this work we present a comparison of the additional Sherwood models in which star formation, feedback and numerical resolution have been changed while other parameters remain fixed. We also display the independent measurements of the CDDF from @Danforth_2016ApJ (red triangles) and @Gaikwad_2016 (orange circles), although we caution that differences in the line fitting methodologies used in these studies will bias any direct comparison with our simulations.
The simulations fail to capture the correct number of absorbers at $N_{\rm HI}<10^{13.2}\rm\,cm^{-2}$, by as much as a factor $\sim2$ at $10^{13}\rm\,cm^{-2}$. Note that the simulations are well converged with mass resolution – this may be assessed by comparing the 80-512 (blue dashed curve) and 80-1024 (orange dot-dashed curve) models. However, the CDDF is very sensitive to the signal-to-noise and spectral resolution at low column densities (cf. Figure \[fig:CDDF\_Phys\] where S/N$=50$ and FWHM=$7\rm\,km\,s^{-1}$), and the observational data are incomplete at $N_{\rm HI}<10^{13}\rm\,cm^{-2}$. This discrepancy therefore reflects differences between the true noise on the data and the idealised, uniform noise added to the mock spectra and is probably not significant. In contrast, the agreement between the V17 data and simulations at $10^{13.2}<N_{\rm HI}/\rm cm^{-2}<10^{14}$ is within the $2\sigma$ uncertainty from sample variance, shown by the grey shading in the lower panel. At $N_{\rm HI}>10^{14}\rm\,cm^{-2}$, however, the number of absorbers are underpredicted by a factor $\sim
2$ and the slope of the observed CDDF is shallower than the simulated prediction. A similar discrepancy was present in the recent analysis by @Shull_2015ApJ using simulations performed with the Eulerian hydrodynamical code <span style="font-variant:small-caps;">enzo</span>, as well as in the study by @Gurvich_2016 with the moving mesh code <span style="font-variant:small-caps;">Arepo</span>, suggesting this discrepancy is not unique to <span style="font-variant:small-caps;">P-Gadget-3</span> simulations. In contrast @Gaikwad17 recently found better agreement with the CDDF at $N_{\rm HI}>10^{14}\rm\,cm^{-2}$ using <span style="font-variant:small-caps;">Gadget-2</span> adiabatic hydrodynamical simulations post-processed to incorporate radiative cooling, photo-heating and ionisation. Note, however, the effect of gas pressure on the distribution will not be correctly captured in these models, complicating interpretation of this result.
Lastly, it is clear that the various feedback scenarios explored by the simulations impact very little overall on the CDDF at low column densities, particularly at $N_{\rm HI}<10^{14}\rm\,cm^{-2}$. Outflows have a similarly small impact on the low column forest at $z>2$; these systems are associated with the diffuse IGM rather than circumgalactic gas [@Theuns2002; @Viel_2013MNRAS and see Section \[sec:gas\]]. Any differences between the feedback models are within the expected 1$\sigma$ sample variance.
This is further emphasised in Figure \[fig:CDDF\_Phys\] where we examine the absorption systems at column densities $
10^{12.2}\leq N_{\rm HI}/\rm{cm}^{-2} \leq 10^{17}$ at $z=[0.1,\,1.0,\,1.6]$. Note that we adopt S/N$=50$ per pixel and a instrument resolution $\rm FWHM=7$ here. The lower panels display the residual with respect to the 80-512-ps13 model, and grey shading shows the Poisson error. The <span style="font-variant:small-caps;">QUICKLYA</span> method (80-512, blue dashed-line) agrees well with the more sophisticated sub-grid physics models for weak absorption systems with $N_{\rm
HI}<10^{13.5}\rm\,cm^{-2}$ at $z=0.1$ and $N_{\rm
HI}<10^{14.5}\rm\,cm^{-2}$ at $z=1.6$. However, higher column density lines are either underpredicted or absent relative to the other models. Multi-phase star formation and stellar feedback (80-512-ps13, black curve) results in a factor of 2-5 more absorption systems at $N_{\rm HI}\ga 10^{14}\rm\,cm^{-2}$. A similar result was noted by @Bolton_2017MNRAS for the CDDF at $2<z<3$, although the column density threshold where <span style="font-variant:small-caps;">QUICKLYA</span> underpredicted the CDDF was higher, $N_{\rm HI}\simeq
10^{14.5-15}\rm\,cm^{-2}$. The lack of $N_{\rm HI}\ga
10^{14}\rm\,cm^{-2}$ absorbers is in part because of missing cold dense gas in the <span style="font-variant:small-caps;">QUICKLYA</span> model, but also because winds redistribute gas into the diffuse IGM and introduce peculiar velocities that can increase the equivalent widths of the lines [@Dave_2010MNRAS; @Meiksin2015].
Finally, including AGN feedback in the Sherwood models yields similar results for the CDDF relative to the stellar feedback only model, except at $z=0.1$ where the incidence of high column density systems is slightly lower (although the total number of absorbers here is small). A stronger effect was noted by @Gurvich_2016, who found that the more vigorous AGN feedback in the Illustris simulation suppresses the CDDF at $N_{\rm HI}=10^{14}$–$10^{15}\rm\,cm^{-2}$ by a factor of two due to changes in the density and temperature of the gas associated with these absorbers. The higher temperatures reduce the opacity by lowering the radiative recombination rate and (for $T\ga 10^{5}\rm\,K$) collisionally ionising the gas. In contrast, @Tepper-Garcia_2012MNRAS [see their figure C1] find stellar and AGN feedback have a negligible effect on the CDDF at $z=0.25$ using simulations drawn from the OWLS project [@Schaye2010]. These differences likely reflect variations in the sub-grid physics, as the precise impact of feedback will depend on how efficiently gas is ejected from dark matter haloes.
In contrast to the present study, @Tepper-Garcia_2012MNRAS also found reasonably good agreement with the CDDF measurement at $10^{14}\leq N_{\rm HI}/\rm cm^{-2} \leq 10^{14.5}$ from @Lehner_2007ApJ based on an earlier compilation of FUSE (*Far Ultraviolet Spectroscopic Explorer*, FWHM$\sim20
\rm\,km\,s^{-1}$) and HST/STIS (*Space Telescope Imaging Spectrograph*, FWHM$\sim 7\rm\,km\,s^{-1}$) observations at $z \la
0.4$. However, the models in @Tepper-Garcia_2012MNRAS underpredicted the incidence of saturated lines at $N_{\rm HI}\sim
10^{14.5}\rm\,cm^{-2}$; their only line fits underestimated the column density of these absorbers since the @Lehner_2007ApJ HST/STIS data are obtained from and higher order Lyman lines. The @Lehner_2007ApJ CDDF furthermore only includes absorbers with velocity widths $b_{\rm HI}<40$ and relative errors less than 40 per cent on $N_{\rm HI}$ and $b_{\rm HI}$ [i.e. figure 14 in @Lehner2008erratum]. We have confirmed we obtain similar agreement with the @Lehner_2007ApJ CDDF measurements and our simulations (for $\rm S/N=50$ and FWHM=$7\rm\,
km\,s^{-1}$) to that obtained by @Tepper-Garcia_2012MNRAS when selecting only absorbers with $b_{\rm HI}<40\rm\,km\,s^{-1}$.
The velocity width distribution
-------------------------------
\[fig:BDist\_Phys\]
\[fig:dndz\]
We next turn to consider the distribution of line velocity widths, $b_{\rm HI}$. Figure \[fig:BDist\_COS\] displays a comparison between the V17 COS measurements and simulations at $z=0.1$. We again show the independent measurements from @Danforth_2016ApJ (red triangles) and @Gaikwad_2016 (orange circles) for comparison. Overall, the models slightly underpredict the incidence of broad lines at $b_{\rm HI}=40$–$60$ and overpredict at $b_{\rm HI}<30$, although generally the data lie within the expected 95 per cent scatter arising from sample variance in the reference model. It is important to note, however, that unlike the CDDF (where the effect of feedback is comparable to or smaller than the effect of mass resolution) the line widths are not well converged with mass resolution (see also the Appendix later). The distribution from the 80-1024 simulation (orange dot-dashed curve) differs significantly from the 80-512 model (blue dashed curve), particularly at $b_{\rm HI}\la 30\,$, with a median value of $30.8$ (cf. $33.5$ for the 80-512 model and $36.2$ for the COS data). Higher resolution simulations will therefore predict line widths that are narrower than observed. As pointed out by V17, this suggests there may be additional physics that broadens the absorption lines that is not incorporated into the simulations at present.
We may nevertheless still examine the relative difference between the different feedback models. The incorporation of multi-phase star formation and supernovae-driven winds (black curve) makes very little difference compared to the <span style="font-variant:small-caps;">QUICKLYA</span> model. The addition of AGN feedback (red dotted curve) also has a small effect, increasing the number of broad lines at $b_{\rm HI}>50\rm\,km\,s^{-1}$ with a median[^3] b-parameter $35.3$.
In comparison to earlier work, @Dave_2010MNRAS also reported a small difference of $\sim1$ in the median $b_{\rm HI}$ values predicted by their constant wind and momentum-driven wind models at $z=0$. However, their velocity widths are generally higher then we find here, with a median value of $\sim43$. The mass resolution of their simulations is around factor of two higher than this work ($M_{\rm gas}=2.49\times 10^{7}h^{-1}M_{\odot}$) although performed in smaller volumes $\,(48h^{-1}\rm\,cMpc)^{3}$. This difference is therefore most likely because @Dave_2010MNRAS do not deconvolve their Voigt profile fits with the instrument profile (see the first paragraph in section 4 of that paper). @Tepper-Garcia_2012MNRAS similarly find that AGN feedback produces a small increase in the typical line widths at $z=0.25$ using the OWLS simulations, and report a median b-parameter of $29.4\,$ from mock spectra with FWHM=$7$ and a S/N$=50$ per pixel. This is similar to the value of $27.7$ we obtain from our AGN feedback (80-512-ps13+agn) model using the same signal-to-noise and instrument resolution. As @Tepper-Garcia_2012MNRAS also use SPH simulations performed at slightly lower mass resolution compared to this work ($M_{\rm gas}=8.7\times 10^{7}h^{-1}M{\odot}$), their line widths will be similarly under-resolved. More recently, @Gaikwad17 also found their simulated velocity width distribution predicts lines that are narrower than the COS data. It is possible that the lack of any coupling between radiative cooling, photo-ionisation and the hydrodynamical response of the gas in their simulations may exaggerate this difference further, however.
The impact of feedback on the velocity distribution at $z=[0.1,\,1.0,\,1.6]$ is displayed in Figure \[fig:BDist\_Phys\], for FWHM$=7\rm\,km\,s^{-1}$ and $\rm S/N=50$ per pixel. Again it is clear that the velocity width distribution is not converged with mass resolution. By comparing the 80-1024 (orange dot dashed line) and 80-512 (blue dashed line) the line widths differ significantly at $b_{\rm HI}<30\rm\,km\,s^{-1}$. This implies a comparison between simulations and observations requires a mass resolution of at least $m_{\rm gas} \simeq 6\times 10^{6}h^{-1}M_{\odot}$. However, as for the CDDF, the simulations with varying sub-grid physics performed at the same mass resolution are very similar. The only exception is at $z=0.1$ when hot gas from AGN feedback produces slightly broader lines.
The evolution of number density of absorbers {#sec:dndz}
--------------------------------------------
Lastly we examine the number of absorbers per unit redshift, $dN/dz$. Figure \[fig:dndz\] displays $dN/dz$ for low ($10^{13.1}
\le N_{\rm HI}/\rm{cm}^{-2} \leq 10^{14}$, upper panel) and high ($10^{14} < N_{\rm HI}/\rm{cm}^{-2}\leq 10^{17}$, lower panel) column density absorbers, overlaid with a compilation[^4] of observational data from . In this instance we do not attempt to match the resolution and signal-to-noise of the simulations to the data compilation, which was obtained with a variety of instruments. We instead use FWHM$=7\,$ and S/N$=50$ per pixel and assume the HM01 UVB for emission from star forming galaxies and quasars. We also show the more recent COS measurements[^5] from @Danforth_2016ApJ at $0\leq z \leq0.47$. These are formally consistent with the earlier data compilation at $z\leq 0.3$, but with an increased incidence of lines at $0.3< z<0.5$. Note also we use only fits for comparison to the observations, which may complicate the comparison to the stronger, saturated lines with column densities inferred using higher order Lyman lines. We have therefore tested this by also directly integrating the number densities along the simulated sight-lines to obtain $dN/dz$ [see also @Gurvich_2016]. This procedure yields similar results to the Voigt profile fits.
The number of low column density forest absorbers (top panel) is not sensitive to the differences in star formation and feedback prescriptions in the models. This is consistent with the CDDF, where galactic winds and AGN feedback do not significantly impact on the gas distribution at the low column densities that probe the diffuse IGM [see also @Dave_2010MNRAS]. The simulations are in good agreement with the observational data at $z>1$, but underpredict the number of weak lines at $z=0.1$. To investigate this further, we have rescaled the mean transmission in the 80-512-ps13 model to $\langle F
\rangle=0.979$ at $z=0.1$ (blue triangle), which V17 find provides a better match to CDDF at $10^{13} < N_{\rm HI}/\rm{cm}^{-2}
<10^{14}$. The improved agreement again demonstrates this discrepancy is due to the uncertain UVB amplitude; the HM01 model overpredicts the photo-ionisation rate at $z\la 0.5$.
In contrast to the low column density systems, the simulated predictions for higher column density forest absorbers exhibit some differences with the choice of star formation and feedback model at $z<1.5$. The 80-512-ps13 model produces a factor of $\sim2$ higher $dN/dz$ relative to 80-512 at $z\simeq0$, with the AGN feedback intermediate between the two. However, all models fail to reproduce the observational data at $z\leq 1.5$. Rescaling the mean transmission at $z=0.1$ slightly improves agreement but does not remove the discrepancy. In combination with the CDDF, this strongly suggests the simulations are not correctly capturing the saturated absorption systems that are still optically thin to Lyman continuum radiation.
Identifying the gas responsible for low redshift forest absorbers {#sec:gas}
=================================================================
--------- ----------------- ---------------------------------- ---------------------------------------- ----------------------------------- ---------------------------------- ------- --
Name Model Diffuse WHIM Hot Halo Condensed Stars
$\Delta<\Delta_{\rm th},T<10^5$K $\Delta<\Delta_{\rm th},T>10^5{\rm K}$ $\Delta >\Delta_{\rm th},T>10^5$K $\Delta>\Delta_{\rm th},T<10^5$K
$z=0.1$ 80-512 36.5 21.3 17.1 0.9 24.2
80-512-ps13 36.6 24.3 22.4 9.1 7.7
80-512-ps13+agn 36.5 36.9 20.0 3.3 3.3
$z=1.0$ 80-512 51.7 15.4 12.6 1.9 18.4
80-512-ps13 51.9 17.0 16.3 10.6 4.3
80-512-ps13+agn 51.6 19.2 17.7 8.5 3.0
$z=1.6$ 80-512 61.8 11.3 8.9 2.8 15.1
80-512-ps13 62.0 12.3 11.3 11.5 2.8
80-512-ps13+agn 62.0 12.7 11.7 11.1 2.4
--------- ----------------- ---------------------------------- ---------------------------------------- ----------------------------------- ---------------------------------- ------- --
\[tab:feedback\]
In order to more closely examine the physical origin of the absorption in our models and assess the effect of feedback, we consider the typical temperature and density of gas in the <span style="font-variant:small-caps;">QUICKLYA</span> model (80-512) and the @Puchwein_2013MNRAS energy driven winds model with (80-512-ps13+agn) and without (80-512-ps13) AGN feedback. We have divided the gas into four categories based on temperature, $T_{\rm
th}=10^{5}\rm\,K$, and normalised density thresholds, $\Delta_{\rm
th}$, following @Dave_2010MNRAS. These are i) diffuse gas ($\Delta<\Delta_{\rm th}, T<10^5$K) ii) the warm-hot IGM (WHIM, $\Delta<\Delta_{\rm th}, T>10^5$K) iii) hot halo gas ($\Delta>\Delta_{\rm th}, T>10^5$K, and iv) condensed gas ($\Delta>\Delta_{\rm th}, T<10^5$K). In Table \[tab:feedback\], the percentage of the total baryon mass in each category is shown at $z=0.1,\,1$ and $1.6$ where $\Delta_{\rm th}=97.2,\,65.9$ and $62$, respectively [here $\Delta_{\rm th}$ is computed using equations 1 and 2 in @Dave_2010MNRAS]. We have also included the percentage of gas that has been converted into collsionless star particles.
The redshift evolution of gas in the diffuse phase is very similar for all models, with no noticeable difference due to feedback. This diffuse gas follows a power law temperature-density relationship, $T=T_{0}\Delta^{\gamma-1}$ (see Figure \[fig:ToGamma\]). However, there is a significant increase in the mass fraction in the WHIM for the model with AGN feedback at $z=0.1$, with a corresponding decrease in the condensed phase fraction and stellar mass. Including multi-phase star formation and stellar feedback produces more gas in the hot halo and condensed phases compared to the <span style="font-variant:small-caps;">QUICKLYA</span> model at $z=0.1$. In contrast, the <span style="font-variant:small-caps;">QUICKLYA</span> model produces significantly more stars at $z\leq 1$ and – while capturing the diffuse IGM reasonably well – will produce unreliable results for gravitationally bound gas at $\Delta>\Delta_{\rm th}$.
The mass fractions for the 80-512-ps13 model are broadly similar to the “vzw” momentum driven winds model described in @Dave_2010MNRAS, with slightly more ($\sim 5$ per cent) gas in the dense phases with $\Delta>\Delta_{\rm th}$. There is also reasonable agreement between the 80-512-ps13 model and the post-processed adiabatic simulations described by @Gaikwad_2016 at $z=0$, with around 4 per cent more (less) gas in the WHIM (hot halo) phase. This suggests the bulk of the hot gas with $T>10^{5}\rm\,K$ in the models without AGN feedback is produced by gravitational infall. The largest difference is the condensed phase which contains 18.8 per cent of the mass at $z=0$ in @Gaikwad_2016; this is because dense gas is not converted into stars in their adiabatic simulations.
The connection between the density and temperature of the gas and the column density and velocity widths of the absorbers at $z=0.1$ is summarised in Figure \[fig:nhi\_den\]. Along each simulated line of sight we have computed the optical depth weighted temperature, $T_{
\tau}$, and density, $\Delta_{ \tau}$, where $X_{\tau,\rm
j}=\sum_{\rm i}\tau_{\rm i,j}X_{\rm i}/\sum_{\rm i} \tau_{\rm i,j}$ and $\tau_{\rm i,j}$ is optical depth at the $i^{\rm th}$ pixel in real space contributing to the $j^{\rm th}$ pixel in velocity space [@Schaye_1999MNRAS]. We then associate $T_{ \tau}$ and $\Delta_{
\tau}$ with and at the location of the line centres identified by <span style="font-variant:small-caps;">VPFIT</span>. This approach takes redshift space distortions due to the peculiar motion of the gas into account. As before, following V17 we only consider systems with relative errors less then 50 per cent on $b_{\rm HI}$ and $N_{\rm HI}$.
The top panel of Figure \[fig:nhi\_den\] displays the relationship between $N_{\rm HI}$ and $\Delta_{\rm \tau}$ in the simulations. The data are coloured according to their velocity widths. The red dashed line shows the power-law relationship between and $\Delta$ predicted by @Schaye_2001ApJ, assuming the typical size of an absorber is the local Jeans scale:
$$N_{\rm HI} = \frac{2.44\times 10^{13}\rm\,cm^{-2} }{\Gamma_{-13}} \left(\frac{\Delta}{10}\right)^{3/2} \left(\frac{T_{4}}{2}\right)^{-0.26} \left(\frac{1+z}{1.1}\right)^{9/2}. \label{eq:NH1Delta}$$
Here $\Gamma_{-13}=\Gamma_{\rm HI}/10^{-13}\rm\,s^{-1}$ is the photo-ionisation rate (where $\Gamma_{-13} = 0.78$ at $z=0.1$ in the 80-512-ps13 simulation for $\langle F \rangle =0.979$) and $T_{4}=T/10^{4}\rm\,K$ is the gas temperature.
This provides a good description of the scaling between gas density and column density [see also @Dave_1999ApJ; @Dave_2010MNRAS; @Tepper-Garcia_2012MNRAS]. The majority of the absorption lines at $N_{\rm
HI}=10^{13}$–$10^{14}\rm\,cm^{-2}$ correspond to gas densities $\Delta=4$–$40$ in the simulations (V17), with typical velocity widths of $b_{\rm HI}=20$–$40\rm\,km\,s^{-1}$. In all models there are a small number of broad lines, $b_{\rm HI}>40\rm\,km\,s^{-1}$, with column densities $N_{\rm HI}=10^{12.5}$–$10^{13.5}\rm\,cm^{-2}$ that are associated with hot $T\sim 10^{5}\rm\,K$ gas in the WHIM with $\Delta \simeq 1$–$100$ [see also @Tepper-Garcia_2012MNRAS]. The effect of feedback on the diffuse, low column density gas is rather subtle; there are slightly more broad lines lying above the $N_{\rm HI}-\Delta_{\rm \tau}$ correlation for the AGN feedback model, consistent with additional heating. Although there is significantly more hot gas in the WHIM by mass at $z=0.1$ when including AGN feedback, much of this is also collisionally ionised. In all instances, however, the higher column density systems, $N_{\rm
HI}>10^{14}\rm\,cm^{-2}$ correspond to gas densities at $\Delta
\ga 20$. The simulations fail to correctly capture the number density of these systems, some of which are at densities associated with circumgalactic gas.
The lower panel in Figure \[fig:nhi\_den\] shows the corresponding values of $T_{\rm \tau}$ at the line centres against the velocity widths. The red dashed line displays the expected velocity width under the assumption of pure thermal broadening:
$$b_{\rm HI} = 18.2 \rm\,km\,s^{-1} \left(\frac{T_{4}}{2}\right)^{1/2}. \label{eq:bH1T}$$
In general, although the minimum line widths trace the expectation for thermal broadening, most of the absorbers are significantly broader, particularly for low column density lines originating from low density gas [see also @Garzilli_2015MNRAS for a recent discussion at $z\sim
3$]. Many of the absorbers are therefore broadened through the additional effects of gas pressure (i.e. Jeans smoothing). The absorbers with $N_{\rm
HI}=10^{13}$–$10^{14}\rm\,cm^{-2}$ are associated with gas with a broad range of temperatures, but most are in the range $T=10^{4}$–$10^{4.5}\rm\,K$. The velocity widths of these lines remain too narrow with respect to the COS data. Again, the effect of stellar and AGN feedback on the temperature of the absorbers is small.
Conclusions
===========
We have investigated the impact of different feedback prescriptions on the properties of the low redshift forest, using a selection of hydrodynamical simulations drawn from the Sherwood simulation suite. The simulations adopt stellar feedback, AGN feedback and a simplified scheme for modelling the low column density forest that converts all gas with $\Delta>1000$ and $T<10^{5}\rm\,K$ into collisionless star particles. We have examined the CDDF, velocity width distribution, the redshift evolution in the line number density, $dN/dz$, and the relationship between the column density, velocity widths, and the underlying gas density and temperature.
Our analysis confirms the finding by V17 that the velocity widths of simulated absorption lines with gas densities $\Delta=4$–$40$ are too narrow when compared with COS data. We also stress that this difference will be exacerbated further for simulations performed at higher mass resolution than considered here; the line widths are under-resolved at our fiducial resolution of $M_{\rm gas}=5.1\times
10^{7}h^{-1}M_{\odot}$ (see also the Appendix). We also find that absorbers with column densities $N_{\rm HI}>10^{14}\rm\,cm^{-2}$, which correspond to gas with $\Delta \ga 20$ in our simulations, are underestimated, in agreement with earlier work [@Shull_2015ApJ; @Gurvich_2016].
On the other hand, using simulations that vary the feedback prescription while keeping other parameters fixed, we find that the impact of galactic winds and AGN feedback is generally very modest. AGN feedback only impacts on the absorbers at very low redshift, $z=0.1$, by producing some broader lines and slightly reducing the incidence of systems with $N_{\rm HI}>10^{14}\rm cm^{-2}$ with respect to a model with only stellar feedback. There is no noticeable impact on the CDDF due to stellar or AGN feedback at $N_{\rm HI}<10^{14}\rm\,cm^{-2}$. Furthermore, we find a simplified scheme that ignores feedback and star formation altogether adequately capture the properties of absorption lines with $N_{\rm
HI}<10^{14}\rm\,cm^{-2}$, although the incidence of stronger systems is underestimated compared to models that include feedback. Similarly, the line number density and its evolution with redshift, $dN/dz$, is in good agreement with observational data for low column density systems, $10^{13.1}\leq N_{\rm HI}/\rm cm^{-2} \leq 10^{14}$, once the amplitude of the UVB is rescaled to match the CDDF. In contrast, for high column density systems, $10^{14}\leq N_{\rm HI}/\rm cm^{-2} \leq
10^{17}$, the simulations underpredict the incidence of absorbers at $z=0.1$ and $z=1$ by a factor of $\simeq 2.2$ and $3$, respectively.
This demonstrates that stellar and AGN feedback – as currently implemented in our simulations – have a limited impact on the forest. We conclude that resolving the discrepancy between COS data and the simulations requires an increase in the temperature of gas with $\Delta=4$–$40$, either through (i) additional photo-heating at $z>2$, (ii) additional fine-tuned feedback that ejects overdense gas into the IGM at just the right temperature for it to still contribute significantly to the forest or alternatively (iii) a larger, currently unresolved turbulent component to the line widths [@Oppenheimer2009; @Iapichino2013; @Gaikwad17]. Note, however, the first possibility would require a UVB with a harder spectrum at $z>2$, potentially producing gas temperatures that are too high with respect to existing observational constraints from the forest at $1.6\leq z \leq 3$ [@Becker_2011MNRAS; @Boera_2014; @Bolton_2014MNRAS]. Our results provide further motivation for studying the impact of state-of-the-art feedback models on the properties of the low redshift forest.
Acknowledgments {#acknowledgments .unnumbered}
===============
The hydrodynamical simulations used in this work were performed with supercomputer time awarded by the Partnership for Advanced Computing in Europe (PRACE) 8th Call. We acknowledge PRACE for awarding us access to the Curie supercomputer, based in France at the Tre Grand Centre de Calcul (TGCC). This work also made use of the DiRAC High Performance Computing System (HPCS) and the COSMOS shared memory service at the University of Cambridge. These are operated on behalf of the STFC DiRAC HPC facility. This equipment is funded by BIS National E-infrastructure capital grant ST/J005673/1 and STFC grants ST/H008586/1, ST/K00333X/1. We thank Volker Springel for making <span style="font-variant:small-caps;">P-GADGET-3</span> available. FN is supported by a Vice-Chancellor’s Scholarship for Research Excellence. JSB acknowledges the support of a Royal Society University Research Fellowship. MV and TSK are supported by the FP7 ERC grant “cosmoIGM” and the INFN/PD51 grant. TSK also acknowledges the NSF-AST118913. MGH and EP acknowledge support from the FP7 ERC Grant Emergence-320596, and EP gratefully acknowledges support by the Kavli Foundation. DS acknowledges support by the STFC and the ERC starting grant 638707 “Black holes and their host galaxies: co-evolution across cosmic time”.
numerical convergence
=====================
We have performed convergence tests with mass resolution and box size for the forest CDDF and velocity width distribution at the redshifts considered in this paper. These are displayed in Figures \[fig:CDDF\_RRes\]-\[fig:BDist\_RBox\]. The CDDF is well converged with mass resolution and box size within the range $
10^{12.7}\leq \ N_{\rm HI}/\rm{cm}^{-2} \leq 10^{14}$. However, at $N_{\rm HI}>10^{14}\rm\,cm^{-2}$ ($N_{\rm HI}<10^{12.7}\rm\,cm^{-2}$) the number of lines are over (under) estimated at lower mass resolution. This suggests that numerical resolution is unlikely to be the cause of the discrepancy between the COS observations and our simulations at $N_{\rm HI}>10^{14}\rm\,cm^{-2}$.
In Figure \[fig:BDist\_RRes\], however, it is clear that mass resolution significantly impacts on the velocity width distributions, with line widths that are systematically overestimated at our fiducial resolution of $M_{\rm gas}=5.1\times 10^{7}h^{-1}M_{\odot}$ (80-512). A mass resolution of at least $M_{\rm gas}=6.4\times
10^{6}h^{-1}M_{\odot}$ is required for an acceptable level of convergence. This means the discrepancy between the COS observations of the velocity width distribution will be even larger for higher resolution simulations. The line width distribution is generally well converged with box size, as shown Figure \[fig:BDist\_RBox\], with the exception of lines at $b_{\rm
HI}<20$in the $20h^{-1}\rm\,cMpc$ box.
Finally, Figures \[fig:BDist\_RRes\]-\[fig:BDist\_RBox\] and Table \[tab:conv\] display convergence tests for the volume weighted IGM thermal state with box size and mass resolution. The results are well converged with box size, but demonstrate the temperature of the IGM at mean density is overestimated by $10$–$20$ per cent at our fiducial resolution (80-512). We have checked, however, that the *mass weighted* temperature obtained from the gas particles is well converged with mass resolution, suggesting this difference originates from the larger smoothing lengths used in the lower resolution models. Note also that $T(\bar{\Delta})$, shown in the middle panels of Figures \[fig:BDist\_RRes\]-\[fig:BDist\_RBox\], corresponds to $T(\bar{\Delta})=T_{0}\bar{\Delta}^{\gamma-1}$, and therefore reflects the convergence of both $T_{0}$ and $\gamma-1$. Additionally, we observe from Table \[tab:conv\] that as mass resolution is increased using the <span style="font-variant:small-caps;">QUICKLYA</span> model, a greater proportion by mass of the baryons are converted to collisionless star particles. Note, however, this is already a significant overestimate and the results of the <span style="font-variant:small-caps;">QUICKLYA</span> model will be unreliable for gas at $\Delta>\Delta_{\rm th}$.
--------- --------- ---------------------------------- ---------------------------------------- ----------------------------------- ---------------------------------- ------- --
Name Model Diffuse WHIM Hot Halo Condensed Stars
$\Delta<\Delta_{\rm th},T<10^5$K $\Delta<\Delta_{\rm th},T>10^5{\rm K}$ $\Delta >\Delta_{\rm th},T>10^5$K $\Delta>\Delta_{\rm th},T<10^5$K
$z=0.1$ 80-1024 35.0 20.1 15.3 0.7 28.8
80-512 36.5 21.3 17.1 0.9 24.2
80-256 39.7 22.6 20.0 1.2 16.6
40-512 34.6 20.5 15.3 0.8 28.8
20-256 35.3 19.4 15.2 0.7 29.3
$z=1.0$ 80-1024 49.6 14.5 11.1 1.6 23.2
80-512 51.7 15.4 12.6 1.9 18.4
80-256 55.4 16.4 14.9 2.1 11.2
40-512 49.2 15.0 11.1 1.5 23.2
20-256 49.7 13.5 11.5 1.6 23.7
$z=1.6$ 80-1024 59.3 10.7 7.7 2.3 20.0
80-512 61.8 11.3 8.9 2.8 15.1
80-256 65.7 12.3 10.6 2.9 8.4
40-512 58.9 11.1 7.8 2.2 19.9
20-256 59.2 10.2 7.9 2.3 20.4
--------- --------- ---------------------------------- ---------------------------------------- ----------------------------------- ---------------------------------- ------- --
\[tab:conv\]
[^1]: E-mail: [email protected]
[^2]: <http://www.stsci.edu/hst/cos/performance/spectral_resolution/>
[^3]: This is consistent with V17, who reported an increase by roughly $2$ in the median b-parameter when including AGN feedback. Note that V17 incorrectly reported this quantity as the peak of the velocity width distribution rather than the median.
[^4]: The individual measurements are from .
[^5]: Note the @Danforth_2016ApJ measurements in the upper panel of Fig. \[fig:dndz\] are for the slightly different column density range $10^{13} \le N_{\rm
HI}/\rm{cm}^{-2} \leq 10^{14}$, rather than $10^{13.1} \le
N_{\rm HI}/\rm{cm}^{-2} \leq 10^{14}$ used for the data compilation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Composite pulses provide a simple means for constructing quantum logic gates which are robust to small errors in the control fields used to implement them. Here I describe how antisymmetric composite [<span style="font-variant:small-caps;">not</span>]{} gates can be nested to produce gates with arbitrary tolerance of errors.'
author:
- 'Jonathan A. Jones'
title: Nested composite NOT gates for quantum computation
---
[^1]
Quantum information processing requires the ability to transform the state of physical systems representing quantum bits through a series of unitary transformations representing quantum logic gates [@Bennett2000]. In real physical systems such gates will be vulnerable to systematic errors arising from imperfections in the control fields, and it is useful to design gates which are robust to such imperfections. This can be achieved through composite pulses, sometimes called composite rotations [@Levitt1979; @Levitt1986; @Cummins2003; @Jones2011; @Merrill2012; @Jones2013; @Husain2013].
Although many composite pulses have been developed in the context of nuclear magnetic resonance (NMR) experiments, their potential applications are much wider, including electron paramagnetic resonance [@Morton2005a], muon spin resonance [@Clayden2012], ion traps [@Gulde2003], neutral atoms [@Rakreungdet2009], atom interferometers [@Butts2013], and optical retarders [@Ardavan2007b; @Ivanov2012]. Note that for applications in quantum information it is necessary to use Class A composite pulses [@Levitt1986], sometimes called general rotors, which correct imperfections in the underlying gate; simpler “point-to-point” pulses which only correct the effects of the gate on particular initial states are not suitable.
I will describe each rotation in a composite pulse by its propagator $$\theta_\phi={{\textrm e}}^{-{{\textrm i}}\theta\sigma_\phi/2}=\cos(\theta/2)\openone-{{\textrm i}}\sin(\theta/2)\sigma_\phi$$ $$\sigma_\phi=\cos\phi\,\sigma_x+\sin\phi\,\sigma_y$$ where $\theta$ is the rotation angle and $\phi$, the pulse phase, fixes the rotation axis in the $xy$-plane. As usual propagators must be written with time running from right to left, the reverse of the usual order for pulse sequences. In the presence of pulse strength errors, when the amplitude of the driving field deviates from its nominal value by some fraction $\epsilon$, the rotation angle $\theta$ of each pulse is also increased by the same fraction. In particular an error prone pulse can be written as $$\pi'_\phi=[\pi(1+\epsilon)]_\phi=\pi_\phi(\epsilon\pi)_\phi$$ so that the propagator for an error prone pulse can be written as a product of the desired and error components.
Here, as elsewhere [@Jones2013; @Husain2013], I will concentrate on attempts to construct [<span style="font-variant:small-caps;">not</span>]{} gates, that is $\pi_0$ rotations, using only $\pi_\phi$ rotations, as composite pulses of this kind have many desirable properties. Note that [<span style="font-variant:small-caps;">not</span>]{} gates play an important role in experimental techniques such as dynamical decoupling [@Viola1999; @Souza2011b; @Souza2012]. A number of relatively short composite pulses of this kind have been described, and it is known that some of these, most notably 1, can be nested to produce longer pulse sequences with much higher error tolerance [@Husain2013]. Here I explain why nesting is successful for 1 and for other antisymmetric composite pulses, but does not work for most other types of composite pulse. I do not consider here the use of NMR phase cycles and supercycles [@Levitt1981; @Shaka1987a], which have proved useful in some NMR implementations of dynamical decoupling [@Souza2012].
The [${\textrm{F}_{1}}$]{} composite $\pi$ pulse
================================================
The 1 composite pulse [@Wimperis1991] implements a $\pi_0$ rotation as a sequence of five $\pi$ pulses, with the propagator $$\begin{aligned}
V_{\textrm{F}_1}&=\pi'_{\phi_5}\pi'_{\phi_4}\pi'_{\phi_3}\pi'_{\phi_2}\pi'_{\phi_1}\\
&=\pi_{\phi_5}\delta_{\phi_5}\pi_{\phi_4}\delta_{\phi_4}\pi_{\phi_3}\delta_{\phi_3}\pi_{\phi_2}\delta_{\phi_2}\pi_{\phi_1}\delta_{\phi_1}\end{aligned}$$ where $\delta=\epsilon\pi$. Using the identity $$\theta_\beta\,\pi_\alpha=\pi_\alpha\,\theta_{2\alpha-\beta}
\label{eq:pimove}$$ allows the $\pi$ pulses to be propagated through the sequence, in effect separating the evolution into two parts, the desired overall evolution and an error term $$V_{\textrm{F}_1}=(\pi_{\phi_5}\pi_{\phi_4}\pi_{\phi_3}\pi_{\phi_2}\pi_{\phi_1})(\delta_{\phi'_5}\delta_{\phi'_4}\delta_{\phi'_3}\delta_{\phi'_2}\delta_{\phi'_1})$$ where the desired evolution depends on the phases of the pulses as applied, but the error term depends on the phase of these pulses in the interaction frame [@Levitt1986] or toggling frame [@Suter1987a; @Odedra2012b], given by [@Jones2013] $$\phi'_j=(-1)^{(j+1)}\phi_j+\sum_{k<j}(-1)^{(k+1)}2\phi_k. \label{eq:phitog}$$
The 1 composite pulse uses *antisymmetric* pulse phases, so that $\phi_5=-\phi_1$, $\phi_4=-\phi_2$ and $\phi_3=0$, which has two important consequences. Firstly the desired evolution term is always equal to a $\pi_0$ pulse, whatever values are chosen for $\phi_1$ and $\phi_2$, and these phases can be chosen to create error tolerance. Secondly the toggling frame phases are *symmetric*, and are given by $$\begin{aligned}
\phi'_5=\phi'_1&=\phi_1\\
\phi'_4=\phi'_2&=2\phi_1-\phi_2\\
\phi'_3&=2\phi_1-2\phi_2.\end{aligned}$$ This relationship between antisymmetric pulse phases and symmetric toggling frame phases underlies the success of antisymmetric composite [<span style="font-variant:small-caps;">not</span>]{} gates.
To produce a good composite pulse requires simply that the error term be made as close as possible to the identity operation. It is tempting to expand the error propagator as Taylor series in $\epsilon$ and then sequentially remove error terms, but it is more insightful to consider the underlying generator of the error propagator. This can be written as $$\delta_{\phi'_5}\delta_{\phi'_4}\delta_{\phi'_3}\delta_{\phi'_2}\delta_{\phi'_1}=\exp\{-{{\textrm i}}(\Delta_1+\Delta_2+\dots)\}$$ where the individual error terms can be calculated using the Baker–Campbell–Hausdorff (BCH) relation [@Campbell1897; @EBWbook] as $$\Delta_1=\frac{\delta}{2}\sum_j\sigma_{\phi'_j}
\label{eq:del1}$$ and $$\begin{aligned}
\Delta_2&=-\frac{{{\textrm i}}\delta^2}{8}\sum_j\sum_{k<j}\big[\sigma_{\phi'_j},\sigma_{\phi'_k}]\\
&=-\frac{\delta^2}{4}\sum_j\sum_{k<j}\sin(\phi'_j-\phi'_k)\,\sigma_z.\label{eq:del2}\end{aligned}$$ Note that in NMR treatments it is customary to perform these calculations using average Hamiltonian theory based on the Magnus expansion [@Haeberlen1968; @Tycko1984; @Tycko1985a; @Odedra2012b; @EBWbook], and the numbering convention differs by one, so that these errors are referred to as zero and first order errors respectively.
From the form of Eq. \[eq:del2\] it can be seen that the second order error term will vanish for sequences such as 1 which have symmetric toggling frame phases. Thus if the phases can be chosen such that the first order error, Eq. \[eq:del1\], is suppressed, then the resulting composite pulse will have third order error terms. The corresponding propagator fidelity [@Jones2011], which can in this case be calculated from the trace of the error propagator, will be correct up to sixth order infidelity. This can be achieved by choosing $\phi_1=3\psi$ and $\phi_2=\psi$ with $\psi=\pm\arccos(-1/4)$, a result which can be understood geometrically by considering the first order error as a vector sum [@Jones2013]. The 1 pulse can be summarised by listing its phases $$\mbox{\boldmath$(\phi)$}=(\phi_1,\, \phi_2,\, \phi_3,\, \phi_4,\, \phi_5)=(3\psi,\,\psi,\,0,\,-\psi,\,-3\psi)$$ or the corresponding phases in the toggling frame $$\mbox{\boldmath$(\phi')$}=(3\psi,\,5\psi,\,4\psi,\,5\psi,\,3\psi)$$ where Eq. \[eq:phitog\] can be used to interconvert these two forms. Note that the choice of plus or minus sign for $\psi$ is arbitrary, but must be made consistently.
To make further progress it is necessary to consider the form of the higher order error terms, $\Delta_3$ and so on. Although the detailed forms of these higher terms swiftly become complicated, it is possible to make some general statements. In particular, the vanishing of $\Delta_2$ is simply a specific example of a more general property, that all even order terms disappear in the expansion of a time symmetric sequence of operators [@Wang1972; @Burum1981; @Merrill2012]. The third order term will be a sum of double commutators, and so will be a sum of $\sigma_\phi$ terms, and similarly for all odd-order terms, while the missing even order terms would be proportional to $\sigma_z$. Thus all the error terms for the 1 composite pulse lie in the $xy$-plane, a fact whose significance will soon become clear.
The [${\textrm{F}_{n}}$]{} family
=================================
1 is simply the first member of a family [@Wimperis1991; @Husain2013] of composite [<span style="font-variant:small-caps;">not</span>]{} gates obtained by recursively nesting the 1 pulse. There has been considerable interest within the NMR community in designing high order robust pulses by recursive or iterative expansions [@Levitt1983; @Tycko1984], usually concentrating on the form of the error propagator [@Husain2013; @Odedra2012b]. Naive iterative expansions are sometimes useful in the design of point-to-point pulses [@Tycko1984], but less successful for Class A general rotors [@Husain2013].
The 1 composite pulse can, however, be extended by nesting with phase reversals. (The first extension to obtain 2 has been known for many years [@Wimperis1991], and this was subsequently extended and generalised [@Husain2013].) The phase patterns are given by $$\mbox{\boldmath$(\phi_{n+1})$}=(3\psi+\mbox{\boldmath$\phi_{n}$},\,\psi-\mbox{\boldmath$\phi_{n}$},\,\mbox{\boldmath$\phi_{n}$},\,-\psi-\mbox{\boldmath$\phi_{n}$},\,-3\psi+\mbox{\boldmath$\phi_{n}$})
\label{eq:Fnn}$$ with $\mbox{\boldmath$(\phi_{0})$}=(0)$, so that 1 is the sequence of 5 pulses described above, while 2 is a sequence of 25 pulses. The origin of this formula is more easily seen by considering the corresponding toggling frame phases $$\mbox{\boldmath$(\phi'_{n+1})$}=(3\psi+\mbox{\boldmath$\phi'_{n}$},\,5\psi+\mbox{\boldmath$\phi'_{n}$},\,4\psi+\mbox{\boldmath$\phi'_{n}$},\,5\psi+\mbox{\boldmath$\phi'_{n}$},\,3\psi+\mbox{\boldmath$\phi'_{n}$})$$ which makes clear that 2 is obtained by applying the 1 pattern to 1 *in the toggling frame*, and so on.
The error propagator for 2 is most simply analysed by breaking it into five equal parts. Each of these is equivalent to the error propagator for 1, except that the overall phase of each error propagator is shifted by some angle. The outer sequence of phases in the 2 pulse cycles these phases such as to cause *any* error terms in the $xy$-plane to be cancelled to first order, and the overall symmetry of the error propagator ensures that these terms will also be cancelled to second order. The first error term which therefore survives in the 2 error propagator is the ninth order term, arising from a third order combination of third order terms. The next member of the family, 3, can be analysed in precisely the same way, leaving only error terms of order 27 and higher. At each stage all the remaining errors are odd-order and found in the $xy$-plane, and so are cancelled to third order by the next iteration. Thus, as previously conjectured [@Husain2013], the infidelity of a general [${\textrm{F}_{n}}$]{} sequence is of order $2\times3^n$.
More complex pulses
===================
As noted previously [@Husain2013] a similar pattern is seen for other antisymmetric composite $\pi$ pulses, such as the [${\textrm{G}_{n}}$]{} family of targeted composite pulses, and the [${\textrm{N}_{n}}$]{} family of narrowband pulses, as well as for combinations of these pulses in which different phase patterns are applied at each iteration. The same behaviour is also seen for antisymmetric versions of the [${\textrm{W}_{n}}$]{} family of $\pi$ pulses [@Husain2013]. All these results are easily understood when the error propagator is examined in the toggling frame: each pattern of phases produces an error propagator which only has terms in the $xy$-plane, and each outer pattern modulates the size of inner error terms in the obvious way. This view of nested pulses completely explains previous observations, such as the GF nested pulse having sixth order infidelity around $\epsilon=0$ and perfect fidelity at $\epsilon\approx\pm0.720$ [@Husain2013].
A similar approach can be used to tackle off-resonance errors, which occur in experimental approaches such as NMR when the control field is not quite in resonance with the qubit transition. In this case the error term for a single rotation is more complicated than for pulse strength errors [@Jones2013], but the error term still lies in the $xy$-plane, with both the rotation angle and the phase of the error term being functions of the size of the off-resonance error, such that even-order error terms lie parallel to the main pulse, while odd-order terms lie perpendicular to it.
Correcting such errors requires a slightly more complex approach than for pulse strength errors[@Jones2013; @Odedra2012b], but can be achieved using longer composite pulses, such as the ASBO-9 family [@Odedra2012b], which comprises nine $\pi$ pulses with antisymmetric phases and which corrects both pulse strength errors and off-resonance errors up to and including second order terms, so that the composite pulse has sixth order infidelity with respect to both errors. As before, ASBO-9 pulses can be nested in the obvious way, resulting in a sequence of 81 $\pi$ pulses with infidelities of order 18 with respect to both errors. As usual ASBO-9 can be combined with other composite pulses to get more complex patterns of error suppression.
Symmetric pulses
================
All the composite pulses considered so far are antisymmetric pulses, which leads to symmetric phases in the toggling frame. Symmetric composite pulses frequently have better responses to simultaneous pulse strength and off-resonance errors [@Jones2013], and it would be desirable to be able to nest such pulses to achieve higher error suppression. This approach is not, however, successful.
Consider the symmetric five pulse sequence [@Jones2013] $$\mbox{\boldmath$(\phi)$}=(\alpha,\,\beta,\,2\beta-2\alpha,\,\beta,\,\alpha)$$ with pulse phases $\beta=2\alpha\pm\arccos[-(1+2\cos\alpha)/2]$ and $\alpha=\mp2\arcsin(\sqrt[4]{5/32})$, which suppresses both first and second order pulse strength errors. The lowest surviving error term is a third order term in the $xy$-plane, and nesting the composite pulse will cancel this term. However because the toggling frame phases are not symmetric (they are in fact antisymmetric in this case) the higher even order errors are not automatically suppressed, and so the nested sequence will contain a fourth order error term. Similar arguments apply to other symmetric and asymmetric composite pulses. This does not, of course, completely rule out the possibility of developing a recursive procedure for expanding such pulses, but does explain why the current approach will not lead there.
Conclusions
===========
The previous conjecture [@Husain2013] that [${\textrm{F}_{n}}$]{} composite pulses could be indefinitely nested using Eq. \[eq:Fnn\] has now been confirmed, as has the equivalent behaviour of other antisymmetric sequences of $\pi$ pulses. These composite pulses have symmetric error phases in the toggling frame, leading to a cascading suppression of amplitude error terms. The general [${\textrm{F}_{n}}$]{} pulse contains $5^n$ component pulses and suppresses all error terms below order $3^n$. Thus to achieve an error of the form $O(\epsilon^q)$ requires a pulse sequence of length about $q^p$ with $p=\ln5/\ln3\approx1.47$, much shorter than some previous estimates [@Brown2004]. The same approach can be used to achieve simultaneous suppression of amplitude and off-resonance errors. This result, however, is limited to composite $\pi$ pulses.
It has also been conjectured that the [${\textrm{W}_{n}}$]{} approach [@Husain2013] allows composite pulses to be constructed with length linear in the desired error suppression, although only low members of the family were explicitly located. Recently Low *et al.* have shown [@Low2013] that higher members do indeed exist, although actually locating them remains a difficult problem. Thus the nesting approach is still the only technique known for explicitly constructing arbitrary precision composite pulses.
[33]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, **** (), <http://arxiv.org/abs/1203.6392>.
, ****, ().
, , , ****, ().
, , , , , , ****, ().
, , , ****, ().
, , , , , , , , , ****, ().
, , , , , , ****, ().
, , , , , , ****, ().
, ****, ().
, , , , , ****, ().
, , , ****, ().
, , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, , , ** (, ).
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , (), <http://arxiv.org/abs/1307.2211>.
[^1]: Telephone +44 1865 272247, FAX: +44 1865 272400
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose a novel approach to continuum modelling of dynamics of crystal surfaces. Our model follows the evolution of an ensemble of step configurations, which are consistent with the macroscopic surface profile. Contrary to the usual approach where the continuum limit is achieved when typical surface features consist of many steps, our continuum limit is approached when the number of step configurations of the ensemble is very large. The model is capable of handling singular surface structures such as corners and facets and has a clear computational advantage over discrete models.'
address:
- 'Department of Physics of Complex Systems, Weizmann Institute of Science, Rehovot, 76100, Israel.'
- 'Department of Physics of Complex Systems, Weizmann Institute of Science, Rehovot, 76100, Israel.'
author:
- Navot Israeli
- Daniel Kandel
date: 'January 1, 2004'
title: Configurational Continuum modelling of crystalline surface evolution
---
Introduction
============
The behavior of classical physical systems is typically described in terms of equations of motion for discrete microscopic objects (e.g. atoms). The dynamics of the microscopic objects is usually very erratic and complex. Nevertheless, in many cases a smooth behavior emerges when the system is observed on macroscopic length and time scales (e.g. in fluid flow through a pipe). A fundamental problem in physics is to understand the emergence of the smooth macroscopic behavior of a system starting from its microscopic description. A useful way to address this problem is to construct a continuum, coarse-grained model, which treats the dynamics of the macroscopic, smoothly varying, degrees of freedom rather than the microscopic ones. The derivation of continuum models from the microscopic dynamics is far from trivial. In most cases it is done in a phenomenological manner by introducing various uncontrolled approximations.
In this work we address the above problem in the context of the dynamics of crystal surfaces. The evolution of crystal surfaces below the roughening transition proceeds by the motion of discrete atomic steps which are separated by high symmetry orientation terraces. One can model step motion by solving the diffusion problem of adatoms on the terraces with appropriate boundary conditions at step edges. This approach was introduced long ago by Burton, Cabrera and Frank [@BCF], and was further developed by other authors [@WilliamsReviews]. The resulting models specify the normal velocity of each step in the system as a function of its position and shape and as a function of position and shapes of neighboring steps. These step flow models are capable of describing surface evolution on the mesoscopic scale with significant success [@Fu_PRL77; @Tanaka_PRL78]. However, step flow models pose a serious challenge for numerical computations, and can be solved only for small systems.
Several attempts were made to construct continuum models for stepped surfaces [@Mullins_JAP28; @Ozdemir_PRB42; @Nozieres_JPI48; @Lancon_PRL64; @Uwaha_JPSJ57; @Chame_BulChemComm; @HagerSpohn_SurfSci324; @cone_prl; @cone; @sine_scaling; @1D_scaling; @Bonzel_AppPhys35; @Bonzel_SurfSci336; @Ramana_PRB62], in order to understand their large scale properties. The general idea behind these attempts is that step flow can be treated continuously in regions where every morphological surface feature is composed of many steps. If we label surface steps by the index $n$, the continuum limit in these models is obtained by taking $n$ to be continuous. In what follows we will refer to these models as the [*conventional*]{} approach.
In the literature, there are two methods to derive conventional continuum models. One method is to write down the discrete step equations of motion and then transform them into a partial differential equation by taking the step index $n$ to be continuous [@Nozieres_JPI48; @cone_prl; @cone; @sine_scaling; @1D_scaling]. The second method is to start with a continuous surface free energy density and derive a surface dynamic equation that minimizes it[@Ozdemir_PRB42; @Lancon_PRL64; @HagerSpohn_SurfSci324; @Bonzel_AppPhys35; @Bonzel_SurfSci336; @Ramana_PRB62]. These two methods are complementary provided that: 1. The free energy density of the second method is the continuum analog of the free energy of an array of discrete steps. 2. The two methods uses the same mass transport mechanism. Such continuum models are fairly successful in describing the evolution of smooth surfaces with very simple morphologies. However, they suffer from fundamental drawbacks, which do not allow generalizations to more complex and realistic situations.
The most severe drawback is that below the roughening temperature, crystal surfaces have singularities in the form of corners and macroscopic facets. The latter are a manifestation of the cusp singularity of the surface free energy at high symmetry crystal orientations. The assumption that every surface feature is composed of many steps clearly breaks down on macroscopic facets where there are no steps at all. Thus, existing continuum models fail conceptually near singular regions. Several authors have tried to overcome this problem by solving a continuum model only in the non-singular parts of the surface and then carefully match the boundary conditions at the singular points or lines[@HagerSpohn_SurfSci324; @sine_scaling; @cone; @1D_scaling]. In most cases however it is not at all clear how these matching conditions can be derived. This difficulty is a fundamental drawback of conventional continuum models and not merely a technicality. As we argue below, boundary conditions at the singular points or lines cannot be derived in the context of conventional continuum models.
To see why this is true consider the situation near a facet edge. The step at the facet edge is special and obeys a unique equation of motion. In contrast to steps in the sloping parts of the surface which all have two neighboring steps of the same sign, a facet step has only one neighbor of the same sign and potentially a second neighbor of an opposite sign. There might also be special physical conditions such as surface reconstruction that add to the uniqueness of a facet step. As we found in several cases [@cone_prl; @cone; @sine_scaling; @1D_scaling], the unique behavior of a facet step sensitively determines the amount of material emitted or absorbed at the facet and the rate at which steps cross the facet and annihilate. When going to the continuum limit these quantities serve as flux boundary conditions at the singularity. The problem is that conventional continuum models are derived from the equations of motion (or from the surface free energy density) that apply away from the facet and are therefore ignorant of the special behavior of facet steps. Thus, the boundary conditions at the facet edge must be derived from a careful analysis of the discrete dynamics of faces steps. However, in going to the continuum in the conventional way, one loses the information regarding the position of individual steps and the discrete analysis cannot be performed.
Another approach for dealing with surface singularities is to round the surface free energy cusp [@Bonzel_AppPhys35; @Bonzel_SurfSci336; @Ramana_PRB62], approximating true facets by relatively flat but analytic regions. This method avoids the need of specifying explicit boundary conditions at the singularity by assuming analyticity of the surface. The correct surface behavior is then expected to be captured in the limit of vanishing cusp rounding. This procedure completely ignores the key role of facet steps and implicitly assumes that the surface free energy derived for non singular orientations determines the dynamics on facets as well. This assumption is generally false due to the same reasons discussed above. An example for a case where cusp rounding leads to erroneous results can be found in Appendix \[cusp\_rounding\_appendix\].
In this work we propose a conceptually new definition of the continuum limit, which we term Configurational Continuum[@prl_configcont]. Configurational Continuum allows construction of continuum models, which are free of all the limitations of conventional continuum models discussed above. It provides a rigorous way of deriving the continuum model directly from the discrete step equations of motion. Like other continuum models, Configurational Continuum has a clear computational advantage over the discrete step model due to the small number of discretization points it requires for the description of smooth surface regions in a numerical scheme.
Configurational continuum {#config_cont_introduction}
=========================
In order to overcome the limitations of conventional continuum models we now propose a conceptually new definition of the continuum limit for stepped surfaces. Our key observation is that a continuous surface height profile can be represented by many similar, but not identical, step configurations. Figure \[ensemble\] is a one dimensional demonstration of this point. It shows a continuous height function, $h(x)$, of position $x$ (thick solid line), and three valid microscopic representations of this profile as step configurations. The main idea of this work is to define the height profile in the continuum limit as the upper envelope of the discrete height functions of an [*ensemble*]{} of many such step configurations.
To construct the configurations of the ensemble let $a$ be the height of a single step and $N$ the number of configurations in the ensemble. We construct the ensemble so that the height difference $\delta h$, between two adjacent configurations is $a/N$, as depicted in Fig. \[ensemble\]. The continuum limit is obtained when $N \rightarrow \infty$. The generalization to higher dimensions is straightforward.
The dynamics of the continuum model is as follows. Each step configuration of the ensemble evolves according to the microscopic dynamics. As a result, the envelope of discrete height functions changes with time, thus defining the evolution of the continuous height function $h({\vec{r}})$ where ${\vec{r}}$ is a vector in the high symmetry $xy$ plane. There is one technical complication which might arise if two steps from different configurations cross each other. Such an event would make $h({\vec{r}})$ a multi valued function of position and requires a more general mathematical description of the surface. For simplicity we ignore this and assume that $h({\vec{r}})$ remains single valued.
There is a crucial assumption hidden in this definition of the continuum model. We postulate that our construction leads to a mathematically well defined height function at all times. When does this assumptions hold? Consider two initially similar configurations of the ensemble. Our continuum limit is well defined provided these two configurations have similar microscopic dynamics and hence remain similar at later times. Note however that this assumption has to hold in the conventional continuum definition as well, and therefore does not put additional restrictions on our model. In fact, if two initially similar configurations evolve very differently, one must abandon the continuum limit and follow the specific microscopic configuration of interest, using discrete dynamics.
We now derive the evolution equation for the continuous height, $h({\vec{r}},t)$, at position ${\vec{r}}$ and time $t$. As a basis for the derivation we assume knowledge of the discrete equations of motion for the underlying step flow model. These equations of motion specify the normal velocity of a step that passes through $({\vec{r}},t)$ as a function of the local step configuration. We denote this dependency by writing $$\vec{v}\left({\vec{r}},t\right)=\vec{v}\left(C_{{\vec{r}},t}\right),
\label{general_step_flow}$$ where $C_{{\vec{r}},t}$ denotes the configuration of steps in the region that influences the step velocity at $({\vec{r}},t)$. In most models of step flow this region of influence covers a small number of neighboring steps. Note that in the context of the discrete step model, $C_{{\vec{r}},t}$ is the actual configuration of steps in the system. When going to the continuum we will be interested in the ensemble of configurations $\{C_{{\vec{r}},t}\}$ which are consistent with $h({\vec{r}},t)$.
The continuous height $h({\vec{r}},t)$ changes with time due to the flow of steps through ${\vec{r}}$, and due to nucleation and annihilation of steps. At this stage we disregard nucleation processes and include them later. First, we consider positions which are not local extrema of the height profile. It is obvious from the construction of the Configurational Continuum, that for each point ${\vec{r}}$ there is exactly one configuration $C_{{\vec{r}},t}$ in the ensemble, which has a step that passes through ${\vec{r}}$ at time $t$. That step lies along the unique equal-height contour line, which passes through ${\vec{r}}$. As is demonstrated in Fig. \[ensemble\], the exact positions of neighboring steps in the configuration $C_{{\vec{r}},t}$ can be calculated from the knowledge of $h({\vec{r}},t)$, and the fact that in this configuration there is a step at ${\vec{r}}$. Hence, we can use the discrete step model Eq. (\[general\_step\_flow\]) and calculate the normal velocity of the step $\vec{v}\left(C_{{\vec{r}},t}\right)$. Note that at different positions, $\vec{v}\left(C_{{\vec{r}},t}\right)$ is the normal velocity of steps which may belong to [*different*]{} configurations in the ensemble.
Next we define the directional gradient in the direction from which steps flow towards ${\vec{r}}$ $$\nabla h_{-\hat{v}}({\vec{r}},t)\equiv -\hat{v}({\vec{r}},t)\lim_{\epsilon\to
0_+}\frac{h\left({\vec{r}}-\epsilon
\cdot\hat{v}({\vec{r}},t)\right)-h\left({\vec{r}},t\right)}{\epsilon}\;,$$ where $\hat{v}({\vec{r}},t)=\frac{\vec{v}\left(C_{{\vec{r}},t}\right)}{\left|\vec{v}\left(C_{{\vec{r}},t}\right)\right|}$. This is useful for the calculation of the current of steps arriving at ${\vec{r}}$: $$J({\vec{r}},t)=\frac{N}{a}\left|\nabla h_{-\hat{v}}({\vec{r}},t)\cdot
\vec{v}\left(C_{{\vec{r}},t}\right)\right|\;,$$ where we have used the fact that the local step density is $\left|\nabla h_{-\hat{v}}({\vec{r}},t)\right|N/a$. Note that $J$ is the current of steps belonging to all configurations in the ensemble, and not to one particular configuration. Since each step (from any configuration), which passes through ${\vec{r}}$ changes the height of the ensemble envelope by $a/N$, the continuous height profile obeys the evolution equation $$\frac{\partial h({\vec{r}},t)}{\partial t}=-\nabla
h_{-\hat{v}}({\vec{r}},t)\cdot \vec{v}\left(C_{{\vec{r}},t}\right)\;.
\label{dynamic_equation}$$
The above derivation of the evolution in the continuum is not valid at local extrema of the surface, because generally one cannot define a unique equal-height contour line which passes through such a point. To avoid the problem, we define $\partial
h/\partial t$ at local extrema as the limit of the height time derivative as one approaches these points. This limiting procedure is justified, since there are no microscopic realizations of the surface with steps exactly at the local extrema.
At this point we emphasize that the Configurational Continuum evolution is formally identical to the evolution of the discrete step model. This statement is almost trivial, since the definition of Eq. (\[dynamic\_equation\]) follows the envelope of the ensemble of configurations, and each configuration evolves with step velocities calculated from the discrete step model. Thus Eq.(\[dynamic\_equation\]) is exact. Moreover, the use of directional derivatives in the derivation of Eq.(\[dynamic\_equation\]) makes it valid even at singular surface regions such as corners or facet edges. Similarly to other continuum models it is solved numerically by discretization of space, which is the only approximation involved in such solutions.
What is the relation between Configurational Continuum and conventional continuum models? In regions where $h({\vec{r}},t)$ is analytic, the evolution equation (\[dynamic\_equation\]) reduces to the continuity equation $$\frac{\partial h({\vec{r}},t)}{\partial t}=- \nabla h({\vec{r}},t) \cdot
\vec{v}\left(C_{{\vec{r}},t}\right)\;. \label{analytic_dyn_eq}$$ If $h({\vec{r}},t)$ is sufficiently smooth, $\vec{v}\left(C_{{\vec{r}},t}\right)$ can be approximated as a local function of $h$ and its spatial derivatives, as is commonly done in conventional continuum models. Making this approximation will therefore recover the conventional continuum approach. We can conclude that in analytic surface regions the conventional continuum approach approximates the Configurational Continuum model and that the approximation quality depends on the smoothness of the surface. However, near corners, facets or regions where the profile is not smooth, one cannot reconstruct the microscopic step configuration from the local value of $h$ and its spatial derivatives. In these regions $\vec{v}\left(C_{{\vec{r}},t}\right)$ contains non local information and as a result Eq.(\[dynamic\_equation\]) cannot even be written as a differential equation.
Is there any computational gain in using such a continuum model? After all, we replaced a discrete model, which follows the evolution of a single microscopic step configuration, by a model which follows a whole ensemble of step configurations. The key point is that we do not have to follow all the steps of all configurations. To calculate $\partial h({\vec{r}},t)/ \partial t$, it is enough to locally follow the single configuration which has a step that passes trough ${\vec{r}}$ at time $t$. In addition, the continuum evolution equation is solved on a grid, and the density of grid points can be very small in regions where the continuous height profile is smooth. The smoothness of the profile allows very accurate interpolation between these points. Only near singular points or lines we have to use a rather dense grid, and there is no computational gain in these regions. In practice, the total number of grid points used can be orders of magnitude smaller than the number of points one has to use in order to follow the evolution of a single microscopic step configuration.
So far we ignored the possibility of island or void nucleation. It is possible to include island or void nucleation in our model provided that we have a microscopic description for these events which determines the nucleation probability in a given step configuration. Within our continuum approach, the nucleation probability at a point on the continuous surface is the ensemble average of the microscopic nucleation probabilities at this point. For demonstration proposes we consider a simple model where the probability for the nucleation of an island on a terrace grows as the square of the local concentration of diffusing adatoms. Information regarding the values of terrace adatom concentrations is already contained in the underlying step flow model Eq.(\[general\_step\_flow\]), since it is used in the calculation of adatom fluxes into and out of steps.
Numerical solutions of the Configurational Continuum {#config_cont_application}
====================================================
We now apply our approach to a few simple cases. First, we consider a conic structure which consists of circular concentric steps. This crystalline cone was studied in Refs. [@cone_prl; @cone] where we wrote a one-dimensional step flow model for the step radii in the absence of growth flux. In the diffusion limited case where adatom diffusion across terraces is the rate limiting process, the equation of motion for the step radii $r_n$ read: $$\begin{aligned}
\frac{dr_1}{dt}&=&\frac{D_s C^{eq}_{\infty} \Omega}{k_B T
r_1}\frac{\mu_1-\mu_2}{\ln\left(r_1/r_2\right)}\;,
\nonumber \\
\frac{dr_n}{dt}&=&\frac{D_s C^{eq}_{\infty} \Omega}{k_B T
r_n}\left(\frac{\mu_n-\mu_{n+1}}{\ln\left(r_n/r_{n+1}\right)}-\frac{\mu_{n-1}-\mu_{n}}{\ln\left(r_{n-1}/r_n\right)}\right)\;,
\;\;\; n>1\;, \label{cone_step_velocity}\end{aligned}$$ with the step chemical potentials $\mu_n$ given by $$\mu_n = \frac{\Omega \Gamma}{r_n} + \Omega G \left[
\frac{2r_{n+1}}{r_{n+1}+r_n} \cdot
\frac{1}{\left(r_{n+1}-r_n\right)^3} -
\frac{2r_{n-1}}{r_n+r_{n-1}} \cdot
\frac{1}{\left(r_n-r_{n-1}\right)^3}
\right]\;. \label{cone_step_chemical_potential}$$ In the above expressions $D_s$ is the adatom diffusion constant, $\Omega$ is the atomic area of the crystal and $C^{eq}_{\infty}$ is the equilibrium concentration in the vicinity of a straight isolated step. $\Gamma$ is the step line tension, $G$ is the step-step interaction strength, $T$ is the temperature and $k_B$ is the Boltzmann constant. Eq.(\[cone\_step\_chemical\_potential\]) can be used for calculating the chemical potential of the top step, $\mu_1$, by omitting the second interaction term.
Numerical integration of the above step model shows that the cone decays through the collapse of individual islands. During the decay a facet develops at the top of the cone with a radius that grows with time as $t^{1/4}$. Similar equations can be written in the presence of growth flux. With flux, the cone grows except at the peak which initially decays and then saturates. A facet forms at the peak after saturation.
The continuum model we solved in the context of this example is a fully two dimensional model, which can, in principle, develop non radially symmetric morphologies. The microscopic dynamics we used were a two dimensional generalization of the microscopic equations for step radii of the discrete one dimensional model given above. In particular the step chemical potential was generalized to $$\mu \left({\vec{r}}\right)=\Omega \Gamma \kappa\left({\vec{r}}\right)+\Omega
G \left( \frac{\exp \left[ \frac{ \left| {\vec{r}}_d-{\vec{r}}\right| \cdot
\kappa \left( {\vec{r}}_d \right) }{2} \right] }{ \left| {\vec{r}}_d-{\vec{r}}\right|^3}-\frac{\exp \left[ -\frac{ \left| {\vec{r}}_u-{\vec{r}}\right|
\cdot \kappa \left( {\vec{r}}_u \right) }{2} \right] }{ \left| {\vec{r}}_u-{\vec{r}}\right|^3} \right)\;, \label{exp_chemical_potential}$$ where $\kappa\left({\vec{r}}\right)$ is the local step curvature. ${\vec{r}}_d$ and ${\vec{r}}_u$ are the coordinates at which the lower and upper neighboring steps are closest to the step at ${\vec{r}}$. It is assumed that these neighbors have the same sign as the step at ${\vec{r}}$ and that steps of opposite signs do not interact (in which case the relevant interaction terms in Eq.(\[exp\_chemical\_potential\]) are omitted). This expression was derived assuming an $l^{-2}$ repulsion between two segments of two different steps which are separated by a distance $l$. Under this assumption, Eq. (\[exp\_chemical\_potential\]) is exact to first order in the curvature of neighboring steps and qualitatively captures the interaction when they have large curvatures.
The equations describing adatom diffusion on terraces were solved assuming the steps at ${\vec{r}}$, ${\vec{r}}_d$ and ${\vec{r}}_u$ are circular and concentric and that the radius of the step at ${\vec{r}}$ is $1/
\kappa({\vec{r}})$. Any microscopic dynamics, such as a full solution of the diffusion equation on each terrace, or a more detailed calculation of the interactions between steps can easily be used in the framework of our model. For the sake of demonstrating the validity of Configurational Continuum the simple dynamics we chose are sufficient.
Figure \[cone\_decay\] shows a comparison between a numerical solution of the discrete step model and a cross section from the two dimensional solution of our continuum model in the absence of growth flux and when nucleation of new steps is not allowed. Clearly the continuum model captures the main features of the surface evolution. In particular the width and height of the top facet are in excellent agreement with the discrete model. Fig. \[growing\_cone\] shows a similar comparison in the presence of growth flux. Again the agreement is quite impressive. In this case we allowed new islands to nucleate and our simulation indicates that nucleation events occur on the top facet once it becomes large enough. There is hardly any nucleation on the finite slope regions, because the steps there efficiently absorb the excess material. Figure \[growing\_cone\](c) shows an island on the top facet, which nucleated and started growing.
We now turn to the more demanding example of bidirectional sinusoidal grating relaxation. Here the initial surface height profile of wave length $L$ is given by $$h_L(\vec{r},t=0)=h_0 \sin\left(\frac{2\pi x}{L}\right)
\sin\left(\frac{2 \pi y}{L}\right)\;.$$ The relaxation of this profile towards a flat surface was studied by Rettori and Villain [@Rettori_Vilain], who gave an approximate solution to a step flow model, in the limit where the interaction between steps can be neglected with respect to the step line tension. The decay of bidirectional sinusoidal profiles was also studied numerically [@Ramana_PRB62]. We now apply our model to this problem assuming diffusion limited kinetics without deposition flux or island nucleation.
For weak interactions between steps, the surface height evolves according to $\partial h_L/\partial t\propto \nabla^2 \kappa \sim
L^{-3}$, where $\kappa$ is the step curvature. We therefore expect the following scaling law for $h_L(\vec{r},t)$: $$h_L(\vec{r},t)=h_{L=1}(\vec{r}/L,t/L^3)\;. \label{shape_scaling}$$
Figure \[low\_g\_egg\] shows the data collapse of cross sections of profiles resulting from our continuum model. The different symbols correspond to profiles of different wave lengths at time $t=t_0 L^3$ for some fixed $t_0$. The quality of the data collapse shows that the scaling scenario (\[shape\_scaling\]) holds very accurately. Note that large facets have developed at the surface extrema, and they are connected by very steep slopes. This shape does not agree with Rettori and Villain’s heuristic argument [@Rettori_Vilain], which predicts that facets appear also near $h=0$ lines. Nevertheless, their prediction that after a short transient the amplitude of the height profile decays as $t/L^3$ is in agreement with both Eq. (\[shape\_scaling\]) and our numerical solutions.
Figure \[high\_g\_egg\] shows results for a much stronger repulsive interactions between steps, where the scaling law (\[shape\_scaling\]) clearly does not hold. Fig.\[high\_g\_egg\] (a) shows profiles of different wavelengths which have relaxed to half of the initial amplitude. Here profiles with a smaller wavelength have small facets and moderate slopes. This happens because repulsion between steps becomes increasingly important as the profile wavelength is reduced. At long wavelengths the weak step-step interaction limit of Fig.\[low\_g\_egg\] is recovered. Fig. \[high\_g\_egg\] (b) shows the different amplitudes as a function of scaled time $t/L^3$.
Conclusions
===========
We proposed a novel continuum model for the evolution of sub roughening crystal surfaces. Our model, which we term Configurational Continuum, is derived directly from the underlying dynamics of atomic steps. Unlike conventional continuum models, Configurational Continuum is fully consistent with the step dynamics and is capable of handling singular surface features such as facets and corners.
The key idea in our model is to view the continuous surface profile as the envelope of an ensemble of step configurations which are all consistent with the continuous profile. Knowing the ensemble envelope, it is always possible to reconstruct individual configurations and evolve them in time. This evolution of individual configurations determines the evolution of the ensemble envelope which within our model is interpreted as the evolution of the profile. The continuum limit in our model is naturally realized because the continuous profile induces a continuum of possible step configurations.
Like other continuum models, Configurational Continuum has a computational advantage over the underlying discrete step model. When solved on a computer, it is possible to use a sparse grid in regions where the profile is very smooth. However, near corners or facets our model requires a fine discretization grid. The fine grid is necessary in order to faithfully reconstruct the step configurations which are possible near a singular point. Our model thus has the important property of being capable of describing step flow on different scales in a consistent way. In smooth surface regions, Configurational Continuum provides a coarse grained numerical description of surface dynamics. However it is still capable of accounting for the unique behavior of steps near singular points or lines.
The problem of connecting between different scales in dynamical systems is not limited to the evolution of surfaces. This problem is widespread in physics, engineering and biology as well as in other fields. Our hope is that the ideas at the basis of Configurational Continuum can be applied in other multi-scale problems.
Failure of the cusp rounding method. {#cusp_rounding_appendix}
====================================
In this appendix we give an example which shows how rounding of the surface free energy cusp can lead to erroneous solutions for surface evolution. As an example system we again choose the crystalline cone studied in Refs.[@cone_prl; @cone]. For simplicity we consider the diffusion limited case. This system is convenient since it exhibits scaling. In the scaling limit conventional continuum modelling of the cone becomes exact [@cone_prl; @cone] and we can concentrate on effects introduces by the cusp rounding method. We start by solving a model with a rounded free energy cusp and later compare the solution of this model with the relevant discrete step model and with the solution of the Configurational Continuum.
In the conventional continuum approach, the continuous free energy density of sub roughening surfaces has a cusp singularity at the high symmetry orientation. In a coordinate system where $(x,y)$ is the high symmetry plane and $h(x,y)$ is the surface profile, the projected (on $(x,y)$) surface free energy density assumes the form [@Gruber_JPCS28]: $${\mathcal F}\left(x,y\right)={\mathcal F}_0+\Gamma\left|\nabla h(x,y)\right|+\frac{G}{3}\left|\nabla h(x,y)\right|^3\;.
\label{cusped_free_energy}$$ This form is the continuum analog of the free energy of an array of steps with line tension $\Gamma$ and an inverse square repulsive step-step interaction of strength $G$.
The singular nature of the surface free energy complicates the modelling of surface evolution. Several authors have tried to overcome this problem by rounding the cusp with a small parameter $\alpha$: $${\mathcal F}_{\alpha}\left(x,y\right)={\mathcal
F}_0+\Gamma\left[\left(\nabla
h(x,y)\right)^2+\alpha^2\right]^{1/2}+\frac{G}{3}\left[\left(\nabla
h(x,y)\right)^2+\alpha^2\right]^{3/2}\;.
\label{rounded_free_energy}$$ The hope behind this regularization scheme is that in the limit $\alpha\rightarrow 0$ the resulting model captures the correct surface dynamics.
Surface dynamics is derived from ${\mathcal F}_{\alpha}$ as follows. Taking the functional derivative of ${\mathcal
F}_{\alpha}$ we obtain the surface chemical potential: $$\mu_{\alpha}=\frac{\delta {\mathcal F_{\alpha}}}{\delta h}=-\Omega \left(\frac{\partial}{\partial x}\frac{\partial {\mathcal F}_{\alpha}}{\partial h_x}+\frac{\partial}{\partial y}\frac{\partial {\mathcal F}_{\alpha}}{\partial h_y} \right)\;,
\label{continuus_mu}$$ where $h_x=\partial h/\partial x$ and $h_y=\partial h/\partial y$. For a radially symmetric profile $h(r,t)$ we find that $$\mu_{\alpha}=-\frac{\Omega}{\sqrt{h_{r}^2+\alpha^2}}\left[\left(\frac{h_{r}}{r}+h_{rr}\right)\left(\Gamma+G\left(h_{r}^2+\alpha^2\right)\right)+h_{r}^2h_{rr}\left(G-\frac{\Gamma}{h_{r}^2+\alpha^2}\right)\right]\;,
\label{rounded_chemical_potential}$$ where $h_{r}=\partial h/\partial r$ and $h_{rr}=\partial^2
h/\partial r^2$.
In diffusion limited kinetics variations in the chemical potential give rise to currents which are proportional to the chemical potential gradient. For our radially symmetric profile we can consider only the radial component of this current $$J=-\frac{D_s \tilde{C}^{eq}}{k_BT}\frac{\partial
\mu_{\alpha}}{\partial r}\;.$$ The dynamic equation for the profile can now be written using the continuity equation $$\frac{\partial h}{\partial t}=-\Omega\cdot\nabla J=\frac{\Omega
D_s \tilde{C}^{eq}}{k_BT}\frac{1}{r}\frac{\partial}{\partial
r}\left(r\frac{\partial \mu_{\alpha}}{\partial r}\right)\;.
\label{round_cusp_dyn_eq}$$
The $\alpha \rightarrow 0$ limit of Eq. (\[round\_cusp\_dyn\_eq\]) automatically gives the correct equilibrium crystal shapes because Eq. (\[rounded\_free\_energy\]) goes to Eq.(\[cusped\_free\_energy\]) in this limit. We want to check whether this limit also gives the correct (consistent with step flow) dynamics.
We applied Eq. (\[round\_cusp\_dyn\_eq\]) to a crystalline cone $h(r,t=0)\sim -r$. The similarity sign here indicates that the tip of the initial profile was smoothed in order to have an analytic surface. Analyticity at the origin was also used as a boundary condition for the surface evolution. Numerical solutions show that, at long times the profile slope obeys a scaling law $$h_r(r,t)=-F(r t^{-1/4})\;. \label{cusp_rounding_scaling_ansatz}$$ This behavior agrees with the scaling properties exhibited by a discrete step flow model of the same surface structure [@cone_prl; @cone]. In Fig. \[round\_cusp\_scaling\] we show the resulting scaled slopes (dots) for different values of the cusp rounding parameter $\alpha$. These long time solutions are not sensitive to the initial smoothing of the cone. As $\alpha$ is reduced we observe the appearance of a very flat region around the origin. This flat region supposedly becomes a true facet in the $\alpha=0$ limit.
For small values of $\alpha$, solutions of the dynamic equation (\[round\_cusp\_dyn\_eq\]) approach scaling very slowly. For this reason it becomes increasingly difficult to probe the $\alpha=0$ scaling state. In order to reach smaller values of $\alpha$ we continued in the following way. We assumed that the scaling ansatz (\[cusp\_rounding\_scaling\_ansatz\]) holds and used it to transform Eq. (\[round\_cusp\_dyn\_eq\]) into an ordinary differential equation for the scaling function $F(\xi)$ in the scaling variable $\xi=rt^{-1/4}$. Replacing $h_r(r,t)$ in Eq.(\[round\_cusp\_dyn\_eq\]) by the scaling function $F(\xi)$ we obtain the following equation: $$\begin{aligned}
-\frac{1}{4}F'&=&\frac{\Omega D_s
\tilde{C}^{eq}}{k_BT}\frac{d}{d\xi}\left[\frac{1}{\xi}\frac{d}{d\xi}\left(\xi\frac{d\eta_{\alpha}}{d\xi}\right)\right]\;,
\nonumber \\
\eta_{\alpha}&=&-\frac{\Omega}{\sqrt{F^2+\alpha^2}}\left[\left(\frac{F}{\xi}+F'\right)\left(\Gamma
+G\left(F^2+\alpha^2\right)\right)\right. \nonumber \\
&&\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\left.+F^2F'\left(G-\frac{\Gamma}{F^2+\alpha^2}\right)\right]\;,
\label{round_cusp_scaling_eq}\end{aligned}$$ with $F'=dF/d\xi$.
Solutions of the above equation for the large values of $\alpha$ agree with the scaling states of the dynamic equation (\[round\_cusp\_dyn\_eq\]). The dashed lines in Fig.\[round\_cusp\_scaling\] show the resulting scaling functions for smaller values of $\alpha$. Finally in order to determine the $\alpha=0$ limit we estimated the scaled position of the facet edge from the $\alpha\neq 0$ solutions. This position selects[@cone_prl; @cone] a unique scaling solution for the $\alpha=0$ case of Eq. (\[round\_cusp\_scaling\_eq\]). The resulting scaling function is shown by the solid line in Fig.\[round\_cusp\_scaling\]. By our procedure this function is an approximation for the true scaling function of the system according to the cusp rounding method.
The solid line in Fig. \[round\_cusp\_scaling\] should be compared with the behavior of a system of discrete steps. For this purpose we introduce the following step model: $$\begin{aligned}
\frac{dr_1}{dt}&=&\frac{D_s C^{eq}_{\infty} \Omega}{k_B T
r_1}\frac{\mu_1-\mu_2}{\ln\left(r_1/r_2\right)}\;,
\nonumber \\
\frac{dr_n}{dt}&=&\frac{D_s C^{eq}_{\infty} \Omega}{k_B T
r_n}\left(\frac{\mu_n-\mu_{n+1}}{\ln\left(r_n/r_{n+1}\right)}-\frac{\mu_{n-1}-\mu_{n}}{\ln\left(r_{n-1}/r_n\right)}\right)\;,
\;\;\; n>1, \nonumber \\
\mu_n &=& \frac{\Omega \Gamma}{r_n} + \Omega G \left[
\frac{2r_{n+1}}{r_{n+1}+r_n} \cdot
\frac{1}{\left(r_{n+1}-r_n\right)^3} +\frac{r_{n+1}}{\left(r_{n+1}^2-r_n^2\right)^2} \right. \nonumber \\
&&-\left.\frac{2r_{n-1}}{r_n+r_{n-1}} \cdot
\frac{1}{\left(r_n-r_{n-1}\right)^3}+\frac{r_{n-1}}{\left(r_{n}^2-r_{n-1}^2\right)^2}
\right]\;.
\label{cone_step_velocity}\end{aligned}$$ This is the same step model studied in section \[config\_cont\_application\] and in Refs. [@cone_prl; @cone] with modified step chemical potentials. The modification has a small effect on the model behavior and does not introduce any qualitative changes. In particular, this step model obeys the same scaling properties that were studied in Refs. [@cone_prl; @cone], i.e., the density of steps in this model scales according to $D(r,t)=F_{discrete}(rt^{-1/4})$. In addition, applying the scaling analysis of Refs. [@cone_prl; @cone] to this modified step model results in an ordinary differential equation for the scaling function $F_{discrete}$ which is identical to the $\alpha=0$ limit of Eq.(\[round\_cusp\_scaling\_eq\]). This fact gives us a basis for comparison between the step model and the cusp rounding scheme. Identifying the step density of the discrete model with the slope of the continuum model we can finally compare the limiting solution from the cusp rounding method with the scaled density of steps. In Fig. \[cusp\_rounding\_comparison\] we show that these two functions do not coincide. The scaled position of the facet edge in the cusp rounding method is about $40\%$ larger than the one in the discrete system. This means that the difference between the two facets in real space diverges at long times. The height at the origin according to the cusp rounding method will suffer from the same errors. By assuming analyticity of the profile throughout the limiting procedure of the cusp rounding method, we have imposed erroneous boundary conditions at the facet edge.
Figure \[cusp\_rounding\_comparison\] also shows a one dimensional solution of the Configurational Continuum model for this cone system. The Configurational Continuum model predicts scaling of the slope as well. The discrepancy between the resulting scaling function and the discrete step system is much smaller and is consistent with what one would expect from discretization errors.
[10]{}
W. K. Burton, N. Cabrera and F. C. Frank, Philos. Trans. R. Soc. London, Ser. A [**243**]{}, 299 (1951).
For general reviews see H.-C. Jeong and E. D. Williams, Surf. Sci. Rep. [ **34**]{}, 171 (1999); E. D. Williams, Surf. Sci. [**299/300**]{}, 502 (1994).
E. S. Fu, M. D. Johnson, D. -J. Liu, J. D. Weeks, and E. D. Williams, Phys. Rev. Lett. [**77**]{}, 1091 (1996).
S. Tanaka, N. C. Bartelt, C. C. Umbach, R. M. Tromp, and J. M. Blakely, Phys. Rev. Lett. [**78**]{}, 3342 (1997).
W. W. Mullins, J. Appl. Phys. [**28**]{}, 333 (1957).
M. Ozdemir and A. Zangwill, Phys. Rev. B [**42**]{}, 5013 (1990).
P. Nozières, J. Phys. I France [**48**]{}, 1605 (1987).
F. Lançon and J. Villain, Phys. Rev. Lett. [**64**]{}, 293 (1990).
M. Uwaha, J. Phys. Soc. Jpn. [**57**]{}, 1681 (1988).
A. Chame, S. Rousset, H. P. Bonzel, and J. Villain, Bul. Chem. Commun. [ **29**]{}, 398 (1996/1997).
J. Hager and H. Spohn, Surf. Sci. [**324**]{}, 365 (1995).
N. Israeli and D. Kandel, Phys. Rev. Lett. [**80**]{} 3300 (1998).
N. Israeli and D. Kandel, Phys. Rev. B [**60**]{} 5946 (1999).
N. Israeli and D. Kandel, Phys. Rev. B [**62**]{} 13707 (2000).
N. Israeli, H.-C. Jeong, D. Kandel and J. D. Weeks, Phys. Rev. B [**61**]{} 5698 (2000).
H. P. Bonzel, E. Preuss and B Steffen, Appl. Phys. A [**35**]{}, 1 (1984).
H. P. Bonzel and E. Preuss, Surf. Sci. [**336**]{}, 209 (1995).
M. V. Ramana Murty, Phys. Rev. B [**62**]{}, 17004 (2000).
N. Israeli and D. Kandel, Phys. Rev. Lett. [**88**]{}, 116103 (2002).
A. Rettori and J. Villain, J. Phys. (Paris) [**49**]{}, 257 (1988).
M. M. Gruber and W. W. Mallins, J. Phys. Chem. Solids [**28**]{}, 875 (1966).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The clustering for functional data with misaligned problems has drawn much attention in the last decade. Most methods do the clustering after those functional data being registered and there has been little research using both functional and scalar variables. In this paper, we propose a simultaneous registration and clustering (SRC) model via two-level models, allowing the use of both types of variables and also allowing simultaneous registration and clustering. For the data collected from subjects in different unknown groups, a Gaussian process functional regression model with time warping is used as the first level model; an allocation model depending on scalar variables is used as the second level model providing further information over the groups. The former carries out registration and modeling for the multi-dimensional functional data (2D or 3D curves) at the same time. This methodology is implemented using an EM algorithm, and is examined on both simulated data and real data.'
author:
- Pengcheng Zeng
- 'Jian Qing Shi [^1]'
- 'Won-Seok Kim'
nocite: '[@Cou90]'
title: 'Simultaneous Registration and Clustering for Multi-dimensional Functional Data'
---
[*Keywords*]{}: Allocation model, Curve clustering, EM algorithm, Functional data analysis, Gaussian process functional regression model, Registration, Simultaneous registration and clustering, Time warping.
Introduction {#sec:intro}
============
Functional data analysis (FDA) has many applications in almost every branch of science, such as engineering, medicine, geology and biology. Basically, it aims at dealing with the analysis of data in the form of images, curves and shapes. But it always comes along with some challenges, like warping time (or misaligned), observation noise and infinite-dimensionality of function spaces. Among those challenges, data registration plays an important role in terms of preprocessing [@Ram05]. In curves, the lateral displacement termed as phase variation, as opposed to amplitude variation in curve height, has drawn much attention. It is necessary to remove the phase variation from the amplitude in a desirable fashion, since it always increase data variance, distorts principal components and make the underlying data structures unclear [@Mar15]. Time warping function is then introduced for this purpose. If we denote the system time or internal time scale as $t$, the underlying time process shared by all the observations, the functional relationship $g^{-1}(t)$, called time-warping function, represent the clock time or individual-specific time scale, varying on from another. The choice of warping function is dependent on the particular application context and the methods of estimating $g^{-1}(t)$ differ. Generally, among those strategies, are seeking various metric for $g^{-1}(t)$, such as $L^{2}$ distance [@Ger04; @Tan08], similarity index [@San09] and Fisher-Rao metric [@Sri11b], as well as modeling $g^{-1}(t)$ directly or indirectly and estimating it through MLE [@Lars:2016] or Bayesian inference [@Che16]. Most of the existing methods target the registration of one dimensional functional data and only a few of them, e.g. [@San09], [@Sri11a] and [@Che16], can be applied to multi-dimensional case.
On the other hand, the clustering for functional data has been receiving much attention. The aim is to group a set of data such that data within groups (clusters) are more similar than across groups with respect to a metric. It is often used as a preliminary step for data exploration by identifying particular patterns to provide the user with convenient interpretation. It is generally a difficult task due to the lack of distances or estimation from noise data and a definition for the probability of a functional variable. Loads of different approaches have been proposed. For instance, the nonparametric methods defining specific distance or dissimilarities, like hierarchical clustering [@Hart78; @Fer06] and k-means algorithm [@Hart78; @Iev13; @Tar03; @Tok07], and the distribution-based clustering, such as modeling principal components [@Del10; @Bou11] or basis expansion coefficients [@Jam03; @Sam11]) with mixture Gaussian distributions, or curves themselves by mixture Gaussian process [@Shi:05; @Shi:08].
[0.39]{}
[0.46]{}
In our study of the motion analysis of hyoid bone observed from X-ray video clips, the trajectories can be thought of as 2D functional data (see Figure \[fig:realdata\], the background of the data will be discussed in Section \[sec:num\_ana\]). There exists obvious misaligned problems for those curves in both vertical and horizontal variation. Usually, curve alignment is performed as a preprocessing technique and the clustering is conducted afterwards. This way is not efficient, since a subject belonging to which cluster is closely related to how it unfolds its progression pace. Another challenging problem for this study is that the heterogeneity of regression relationships among different groups. It consists in both the potential time warping for curves corresponding to the subjects and the subjects’ scalar variables such as initial level of disease, gender, age and the characteristics of those trajectories themselves, like motion time, average speed and range of motion. Therefore, simultaneous curve registration and clustering by considering all those factors seems to be a better way for modeling the functional data.
There are some research work on handling the similar problems. For instance, [@Liu:09] developed a framework that allows for simultaneously aligning and clustering k-centers functional data. But their model did not use any subject specific information (scalar variables) and assumed the heterogeneity among groups just depends on the curves themselves; similar idea is also used in $k$-means alignment for curve clustering by [@San10]. On the other hand, [@Shi:08] proposed a hierarchical mixture of Gaussian process (GP) functional regression models with an allocation model to do curve prediction and clustering. They used the functional covariates to reconstruct the response curve and the personal scalar variables, such as height and gender, to deal with heterogeneity of the regression relationships among different groups. However, their method did not consider the misaligned problem. In addition, most related models are limited to one dimensional curve.
The purpose of our study is to address the above problems by constructing one hierarchical mixture of models for the sake of simultaneous curve registration and clustering. In the first level, we assume a simultaneous registration and functional regression model. And an allocation model with scalar variables is introduced as the second level model. Non-linear functional random effects are modeled by a GP in the first level model to address the heterogeneity among subjects. This also allows to estimate subject-specific warping function. The main contributions of this paper are: (i) simultaneously carrying out registration and modeling for multi-dimensional functional data (2D or 3D curves) allowing variation among subjects, and (ii) the use of both functional and scalar covariates while conducting clustering.
The paper is organized as follows. Section \[subsec:mode\] defines simultaneous registration and clustering (SRC) models via two-level models. We discuss the estimation and the details of implementation in Section \[subsec:esti\] and Section \[subsec:impl\] respectively. The problem of model selection and the related methods are discussed in Section \[subsec:mod\_sel\]. Section \[sec:num\_ana\] presents a number of examples with simulated data and real data. A short summary and discussion are given in Section \[sec:dis\].
The simultaneous registration and clustering method {#sec:meth}
===================================================
Suppose there are $N$ subjects coming from $K$ different groups, ${\mbox{\boldmath ${x}$}}_{1}, {\mbox{\boldmath ${x}$}}_{2},\dots, {\mbox{\boldmath ${x}$}}_{N}$ being the observations of 2D continuous curves, where ${\mbox{\boldmath ${x}$}}_{i} = ({\mbox{\boldmath ${x}$}}_{1i}(t), {\mbox{\boldmath ${x}$}}_{2i}(t))^{\intercal}$, ${\mbox{\boldmath ${x}$}}_{1i}(t)$ and ${\mbox{\boldmath ${x}$}}_{2i}(t)$ are the corresponding $x$-coordinates and $y$-coordinates of ${\mbox{\boldmath ${x}$}}_{i}$. Let ${\mbox{\boldmath ${v}$}}_{1},\dots, {\mbox{\boldmath ${v}$}}_{N}$ be the observed scalar variables. Suppose there are $n_{i}$ time points on which the $i$-th curve is measured. The data set is $$D = \{({\mbox{\boldmath ${x}$}}_{i}(t_{ij}), {\mbox{\boldmath ${v}$}}_{i}); i = 1,\dots, N; j = 1,\dots, n_{i}\}.$$ We introduce a latent indicator variable ${\mbox{\boldmath ${z}$}}_{i} = (z_{i1}, \dots, z_{iK})^{\intercal}$ for the $i$-th subject where $z_{ik}$ takes value $1$ if they are in the $k$-th group and $0$ otherwise.
The model {#subsec:mode}
---------
In our study of 2D curves, we will use the preprocessing procedure Generalized Procrustes Analysis (GPA) [@Gower:1975] to address part of registration problems in advance except warping. Conventionally, most methods tried to complete all the registration problems including warping before clustering. This is not the best way since different warping functions may need to be used in different clusters, yet we have no such information before clustering, and heterogeneity among different subjects should also be considered. Thus, a hierarchical structure defined by two levels of models is proposed.
We start with the first level model for the continuous curve as follows $$\label{eq:mixt}
x_{ai}(t)|_{z_{ki}=1} = (\tau_{ak}\circ g_{ki})(t) + r_{aki}(t) + \epsilon_{ai}(t), \hspace{5mm} i = 1,\dots,N,$$ where $a = 1$ or 2 represents $x$- or $y$-coordinates of ${\mbox{\boldmath ${x}$}}_{i}(t)$. The item $(\tau_{ak}\circ g_{ki})$ denotes functional composition: $(\tau \circ g)(t) = \tau(g(t))$, where $g_{ki}(t)$ is the inverse of a warping function. $\tau_{ak}(\cdot)$ is a fixed but unknown nonlinear mean curve, which can be approximated by a set of basis functions, the details will be given in the next subsection. The variation among different subjects is modeled by a non-linear functional random-effects, $r_{aki}(t)$, by a Gaussian process with zero-mean and a parametric covariance function $S$ [@Shi:12]. The error item $\epsilon_{ai}(t)$ is assumed to be Gaussian white noise with variance $\sigma^{2}$.
Following the previous discussion, we need to use different warping function in different cluster, and we also need to consider the variation among different subjects, and thus, we allow warping function depending on $k$ and $i$. Further, we assume $$g_{ki}(t) = t + w_{k}(t) + w_{ki}(t),$$ where $w_{k}(t)$ is the fixed part and $w_{ki}(t)$ is the random part in terms of different subjects. Instead of making assumption for curves $w_{k}(t)$ and $w_{ki}(t)$, we first discretize them by a set of fixed parameters, for example, by ${\mbox{\boldmath ${w}$}}_{k} = (w_{k}(t_{1}), \dots, w_{k}(t_{n_{v}}))$ and ${\mbox{\boldmath ${w}$}}_{ki} = (w_{ki}(t_{1}), \dots, w_{ki}(t_{n_{w}}))$ respectively. We then model ${\mbox{\boldmath ${w}$}}_{ki}$ by a Gaussian distribution with zero mean and a parametric covariance function ${\mbox{\boldmath ${H}$}}$. We can also define the warping function as simple as a horizontal shift, i.e. $g_{ki}(t) = t + b_{ki}$ or a linear stretch of the curves, i.e. $g_{ki}(t) = (1 + b_{ki})t + c_{ki}$, where $b_{ki}$ and $c_{ki}$ are both one dimensional unknown parameter. Those linear warping functions have been examined by others [@Liu:09; @San10].
We define a logistic allocation model in the second level model for the latent indicator variable in the form $$\label{eq:allo}
p(z_{ki} = 1) = \pi_{ki} = \frac{\text{exp}\{{\mbox{\boldmath ${v}$}}^{\intercal}_{i}{\mbox{\boldmath ${\beta}$}}_{k}\}}{1 + \sum_{j =1}^{K-1}\text{exp}\{{\mbox{\boldmath ${v}$}}^{\intercal}_{i}{\mbox{\boldmath ${\beta}$}}_{j}\}}, \hspace{5mm} i = 1, \dots, N; \hspace{5mm} k = 1,\dots, K-1,$$ with $p(z_{Ki} = 1) = \pi_{Ki} = 1 - \sum_{l=1}^{K-1}\pi_{li}$, where $\{{\boldsymbol}{\beta}_{k}, k = 1, \dots, K-1\}$ are unknown parameters to be estimated. We can also replace model (\[eq:allo\]) by other models, e.g. Potts model [@Gre:00]. The information of scalar variables is integrated with functional variables via the two-level models (\[eq:mixt\]) and (\[eq:allo\]). The reason of using both types of variables is that the variation between subjects does not usually depend on the curve data only, summary statistics or some subject-specific variables do provide useful information, like the scenario in Figure \[fig:rec01\] and Figure \[fig:rec02\] in the simulated example and Table \[tab:cluster\_realdata\] and Figure \[fig:real1\] in the real data analysis. The introduction of the latent indicator variable is very useful in the implementation; see the details below.
We call the models defined in (\[eq:mixt\]) and (\[eq:allo\]) as simultaneous registration and clustering (*SRC*) models.
Estimation {#subsec:esti}
----------
The discrete form of model (\[eq:mixt\]) for the $i$th curve data ${\mathbf}{x}_{ai} = \big(x_{ai}(t_{i1}),\dots, x_{ai}(t_{in_{i}})\big)^{\intercal}$ can be expressed as follows $$\label{eq:mixt3}
{\mbox{\boldmath ${x}$}}_{ai}|_{z_{ki} = 1} = {\mbox{\boldmath ${\tau}$}}_{ak}(g_{ki}) + {\mbox{\boldmath ${r}$}}_{aki} + {\mbox{\boldmath ${\epsilon}$}}_{i}, \hspace{5mm} \text{for}\hspace{2mm} a = 1, 2; \hspace{2mm} k = 1,\dots,K,$$ where ${\mbox{\boldmath ${\tau}$}}_{ak}(g_{ki}) = \Big(\tau_{ak}\big(g_{ki}(t_{i1})\big), \dots, \tau_{ak}\big(g_{ki}(t_{in_{i}})\big)\Big)^{\intercal}$, and ${\mbox{\boldmath ${r}$}}_{aki}$ and ${\mbox{\boldmath ${\epsilon}$}}_{i}$ are both $n_{i}$-dimensional column vector. In this paper, we respectively set ${\mbox{\boldmath ${S}$}}$ as the Matern covariance function with parameters ${\mbox{\boldmath ${\rho}$}}_{s}$ and ${\mbox{\boldmath ${H}$}}$ as the unstructured covariance function or Brownian covariance function with parameter ${\mbox{\boldmath ${\rho}$}}_{h}$ [@Raket], and they can be estimated by the data; the details are provided in the next subsection. Other covariance functions can also be used [@ShiB11]. Let ${\mbox{\boldmath ${S}$}}_{aki}$ and ${\mbox{\boldmath ${H}$}}_{ki}$ be the covariance matrix of ${\mbox{\boldmath ${r}$}}_{aki}$ and ${\mbox{\boldmath ${w}$}}_{ki}$ respectively, which can be calculated by the corresponding covariance function. We model $\tau_{ak}(t)$ using $q$ basis functions $\{\psi_{1}(t), \dots, \psi_{q}(t)\}$ with weights ${\mbox{\boldmath ${d}$}}_{ak} = (d_{ak1}, \dots, d_{akq})^\intercal$. Thus, ${\mbox{\boldmath ${\tau}$}}_{ak}(g_{ki}) = {\mbox{\boldmath ${\Psi}$}}_{ki}{\mbox{\boldmath ${d}$}}_{ak}$ where ${\mbox{\boldmath ${\Psi}$}}_{ki} = [{\mbox{\boldmath ${\Psi}$}}_{ki1}, \dots, {\mbox{\boldmath ${\Psi}$}}_{kiq}]_{n_{i} \times q}, {\mbox{\boldmath ${\Psi}$}}_{kil} = (\psi_{l}(g_{ki}(t_{i1})), \dots, \psi_{l}(g_{ki}(t_{in_{i}})))^{\intercal}, l=1,\dots,q$. In this paper, we use a smooth non-linear deformation for the curves, which is produced by a cubic Hermite spline [@Raket].
The unknown parameters from the $k$-th component for the $a$-coordinates ($x$- or $y$-coordinates) of the $i$-th curve are denoted by ${\mbox{\boldmath ${\theta}$}}_{aki} \overset{\Delta}{=} \{{\mbox{\boldmath ${d}$}}_{ak},{\mbox{\boldmath ${w}$}}_{k}, {\mbox{\boldmath ${w}$}}_{ki}, {\mbox{\boldmath ${\rho}$}}_{s}, {\mbox{\boldmath ${\rho}$}}_{h}, \sigma\}$. Let ${\mbox{\boldmath ${\theta}$}}_{ki}$ be the vector of $\{{\mbox{\boldmath ${\theta}$}}_{aki}, a =1, 2\}$. We can similarly define ${\mbox{\boldmath ${\theta}$}}_{i} = \{{\mbox{\boldmath ${\theta}$}}_{ki}, k = 1, \dots, K\}$, ${\mbox{\boldmath ${\theta}$}} = \{{\mbox{\boldmath ${\theta}$}}_{i}, i = 1, \dots, N\}$ and ${\mbox{\boldmath ${\beta}$}} = \{{\mbox{\boldmath ${\beta}$}}_{k}, k = 1, \dots, K-1\}$. The Gaussian mixture distribution for the $i$-th curve data can be written in the form $$p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}_{i},{\mbox{\boldmath ${\beta}$}}) = \sum_{k=1}^{K}\pi_{ki}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}_{ki}),\hspace{5mm} i = 1, \dots, N,$$ where $p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}_{ki}) = p({\mbox{\boldmath ${x}$}}_{1i}|{\mbox{\boldmath ${\theta}$}}_{1ki})p({\mbox{\boldmath ${x}$}}_{2i}|{\mbox{\boldmath ${\theta}$}}_{2ki})$. We assume ${\mbox{\boldmath ${x}$}}_{1i}$ and ${\mbox{\boldmath ${x}$}}_{2i}$ are conditional independent given those parameters. The log-likelihood of $({\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}})$ is therefore $$L({\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}}) = \sum_{i=1}^{N}\text{log}\Big\{\sum_{k=1}^{K}\pi_{ki}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}_{ki})\Big\}.$$
It is quite tricky to conduct the estimation due to the large number of unknown parameters. EM algorithm will be adopted in this paper. We have defined the latent indicator variable ${\mbox{\boldmath ${z}$}}_{i}$, which is treated as missing. The joint likelihood function of ${\mbox{\boldmath ${x}$}}$ and ${\mbox{\boldmath ${z}$}}$, where ${\mbox{\boldmath ${z}$}} = \{{\mbox{\boldmath ${z}$}}_{i}; i = 1, \dots, N\}$, takes the form $$p({\mbox{\boldmath ${x}$}},{\mbox{\boldmath ${z}$}}|{\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}}) = p({\mbox{\boldmath ${z}$}}|{\mbox{\boldmath ${\beta}$}})p({\mbox{\boldmath ${x}$}}| {\mbox{\boldmath ${z}$}}, {\mbox{\boldmath ${\theta}$}}) = \prod_{i=1}^{N}\prod_{k=1}^{K}\pi_{ki}^{z_{ki}}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}_{ki})^{z_{ki}}.$$ Taking the logarithm, we have the log-likelihood for complete data $({\mbox{\boldmath ${x}$}}, {\mbox{\boldmath ${z}$}})$ $$\begin{aligned}
L_{c}({\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}}) = \text{log}p({\mbox{\boldmath ${x}$}},{\mbox{\boldmath ${z}$}}|{\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}}) = \sum_{i=1}^{N}\sum_{k=1}^{K}z_{ki}\bigg(\text{log}\pi_{ki} + \text{log}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}_{ki})\bigg).
\end{aligned}$$ The expected value of the complete log-likelihood with respect to ${\mbox{\boldmath ${z}$}}$ is given by $${\label{eq:mle}}
\begin{split}
E_{{\mbox{\boldmath ${z}$}}}\{L_{c}({\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}})\} & = \sum_{k=1}^{K}\sum_{i=1}^{N}E(z_{ki}|{\mbox{\boldmath ${x}$}}, {\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}})\bigg(\text{log}\pi_{ki} + \text{log}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}_{ki})\bigg)\\
& = \sum_{k=1}^{K}\sum_{i=1}^{N}M_{ki}\bigg(\text{log}\pi_{ki} + \text{log}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}_{ki})\bigg),
\end{split}$$ where $$M_{ki} \overset{\Delta}{=} E(z_{ki}|{\mbox{\boldmath ${x}$}}, {\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}}) = \frac{\pi_{ki}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}_{ki})}{\sum_{j=1}^{K}\pi_{ji}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}_{ji})},\hspace{5mm} i = 1, \dots, N; k = 1,\dots, K.$$ The derivation of $M_{ki}$ is given by \[ap:deri\]. The procedure of EM algorithm includes
1. Initialize ${\mbox{\boldmath ${\theta}$}}^{(l)}$ and ${\mbox{\boldmath ${\beta}$}}^{(l)}$ and evaluate the $M_{ki}$ (E-step) $$M_{ki} = \frac{\pi_{ki}^{(l)}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}^{(l)}_{ki})}{\sum_{j=1}^{K}\pi_{ji}^{(l)}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}^{(l)}_{ji})} .$$
2. Fix $M_{ki}$ and maximize $Q({\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}})$ with respect to ${\mbox{\boldmath ${\theta}$}}$ and ${\mbox{\boldmath ${\beta}$}}$ $$Q({\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}}) \overset{\Delta}{=} \sum_{k=1}^{K}\sum_{i=1}^{N}M_{ki}\bigg(\text{log}\pi_{ki} + \text{log}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}_{ki})\bigg),$$ leading to ${\mbox{\boldmath ${\theta}$}}^{(l+1)}$ and ${\mbox{\boldmath ${\beta}$}}^{(l+1)}$ (M-step).
The technical details are given in the next subsection.
Implementation {#subsec:impl}
--------------
In E-step, we first initialize the weights $M_{ki}$. In practice, we choose $M_{ki}^{(0)} \sim U(0,1)$ for the purpose of simplicity. Each $M_{ki}$ is then divided by their summation $\sum_{k=1}^{K}M_{ki}$ and we set $\pi^{(0)}_{ki} = M_{ki}^{(0)}$. In M-step, there are no analytic solutions to the maximization of $Q({\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}})$ with respect to ${\mbox{\boldmath ${\theta}$}}$, so that we use the following algorithms. Maximizing $Q({\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}})$ with respect to ${\mbox{\boldmath ${\theta}$}}$ given the current weights $M_{ik}$ is equivalent to maximizing $$\sum_{k=1}^{K}\sum_{i=1}^{N}M_{ki}\bigg(\sum_{a=1}^{2}\big(\text{log}p({\mbox{\boldmath ${x}$}}_{ai}|{\mbox{\boldmath ${\theta}$}}_{aki})\big)\bigg).$$ Borrowing the idea from [@Lars:2016], all the parameters within ${\mbox{\boldmath ${\theta}$}}$ are estimated iteratively through three conditional models. In order to simplify the likelihood computations, all the random effects are scaled by a noise standard deviation $\sigma$. The norm induced by a full-rank covariance matrix ${\mbox{\boldmath ${B}$}}$ is denoted by $||{\mbox{\boldmath ${A}$}}||^{2}_{{\mbox{\boldmath ${B}$}}} = {\mbox{\boldmath ${A}$}}^{T}{\mbox{\boldmath ${B}$}}^{-1}{\mbox{\boldmath ${A}$}}$.
**(i) Estimating the fixed effects ${\mbox{\boldmath ${\tau}$}}_{ak}$**
Given ${\mbox{\boldmath ${\theta}$}}^{(l)}$, we have ${\mbox{\boldmath ${x}$}}_{ai}|_{z_{ki} = 1} \sim N_{n_{i}}({\mbox{\boldmath ${\Psi}$}}_{ki}{\mbox{\boldmath ${d}$}}_{ak}, {\mbox{\boldmath ${I}$}}_{n_{i}}+{\mbox{\boldmath ${S}$}}_{aki}), a = 1, 2; i = 1, \dots, N$. ${\mbox{\boldmath ${I}$}}_{n_{i}}$ denotes the $n_{i} \times n_{i}$ identity matrix. So the negative log likelihood for the weights ${\mbox{\boldmath ${d}$}}_{ak}$ (its square magnitude is penalized by a weighting factor $\eta$) is proportional to $$L({\mbox{\boldmath ${d}$}}_{ak}) = \sum_{i=1}^{N}M_{ki}||{\mbox{\boldmath ${x}$}}_{ai} - {\mbox{\boldmath ${\Psi}$}}_{ki}{\mbox{\boldmath ${d}$}}_{ak}||^{2}_{{\mbox{\boldmath ${I}$}}_{n_{i}}+{\mbox{\boldmath ${S}$}}_{aki}} + \eta{\mbox{\boldmath ${d}$}}_{ak}^\intercal{\mbox{\boldmath ${d}$}}_{ak}, \hspace{0.5cm} a = 1, 2; \hspace{0.5cm} k = 1, \dots, K.$$ This gives the estimator $$\hat{{\mbox{\boldmath ${d}$}}}_{ak} = ({\mbox{\boldmath ${\Psi}$}}_{k}^\intercal (\frac{{\mbox{\boldmath ${I}$}}_{n} + {\mbox{\boldmath ${S}$}}_{ak}}{{\mbox{\boldmath ${M}$}}_{k}})^{-1}{\mbox{\boldmath ${\Psi}$}}_{k} + \eta {\mbox{\boldmath ${I}$}}_{q})^{-1}{\mbox{\boldmath ${\Psi}$}}_{k}^\intercal (\frac{{\mbox{\boldmath ${I}$}}_{n} + {\mbox{\boldmath ${S}$}}_{ak}}{{\mbox{\boldmath ${M}$}}_{k}})^{-1}{\mbox{\boldmath ${x}$}}_{a}, \hspace{0.5cm} a = 1, 2; \hspace{0.5cm} k = 1, \dots,K,$$ where ${\mbox{\boldmath ${\Psi}$}}_{k} = [{\mbox{\boldmath ${\Psi}$}}_{k1}^{\intercal},\dots,{\mbox{\boldmath ${\Psi}$}}_{kN}^{\intercal}]^\intercal \in {\mbox{\boldmath ${R}$}}^{m \times q}$, $m = \sum_{i = 1}^{N}n_{i}$, ${\mbox{\boldmath ${x}$}}_{a} = ({\mbox{\boldmath ${x}$}}_{a1}^{\intercal}, \dots, {\mbox{\boldmath ${x}$}}_{aN}^{\intercal})^{\intercal}$ and $$\frac{{\mbox{\boldmath ${I}$}}_{n}+{\mbox{\boldmath ${S}$}}_{ak}}{{\mbox{\boldmath ${M}$}}_{k}} \overset{\Delta}{=} \begin{bmatrix}
({\mbox{\boldmath ${I}$}}_{n_{1}}+{\mbox{\boldmath ${S}$}}_{ak1})/M_{k1} & & \\
& \ddots & \\
& & ({\mbox{\boldmath ${I}$}}_{n_{N}}+{\mbox{\boldmath ${S}$}}_{akN})/M_{kN}
\end{bmatrix} \in {\mathds}{R}^{m \times m}.
\label{mk}$$
**(ii) Estimating warping parameters ${\mbox{\boldmath ${w}$}}_{k}$ and ${\mbox{\boldmath ${w}$}}_{ki}$**
Given ${\mbox{\boldmath ${\theta}$}}^{(l)}$ and $\hat{{\mbox{\boldmath ${d}$}}}_{ak}$, we have the joint probability density function of $({\mbox{\boldmath ${x}$}}_{ai}, {\mbox{\boldmath ${w}$}}_{ki})$ given by $$p({\mbox{\boldmath ${x}$}}_{ai}, {\mbox{\boldmath ${w}$}}_{ki}) = p({\mbox{\boldmath ${x}$}}_{ai}|{\mbox{\boldmath ${w}$}}_{ki})*p({\mbox{\boldmath ${w}$}}_{ki}) \sim N_{n_{i}}({\mbox{\boldmath ${\Psi}$}}_{ki}\hat{{\mbox{\boldmath ${d}$}}}_{ak}, {\mbox{\boldmath ${I}$}}_{n_{i}}+{\mbox{\boldmath ${S}$}}_{aki}) * N_{n_{w}}(0, {\mbox{\boldmath ${H}$}}_{ki}).$$ So, we can simultaneously estimate the fixed warping effects ${\mbox{\boldmath ${w}$}}_{k}$ and predict the random warping effects ${\mbox{\boldmath ${w}$}}_{ki}$ from the joint conditional negative log posterior. It is proportional to $$\label{eq:post}
L({\mbox{\boldmath ${w}$}}_{k}, {\mbox{\boldmath ${w}$}}_{ki}) = \sum_{a=1}^{2}\sum_{i=1}^{N}M_{ki}||{\mbox{\boldmath ${x}$}}_{ai} - {\mbox{\boldmath ${\Psi}$}}_{ki}\hat{{\mbox{\boldmath ${d}$}}}_{ak}||^{2}_{{\mbox{\boldmath ${I}$}}_{n_{i}}+{\mbox{\boldmath ${S}$}}_{aki}} + 2\sum_{i=1}^{N}M_{ki}||{\mbox{\boldmath ${w}$}}_{ki}||^{2}_{{\mbox{\boldmath ${H}$}}_{ki}}, \hspace{0.3cm} k = 1, \dots, K,$$ where ${\mbox{\boldmath ${\Psi}$}}_{ki}$ is determined by $n_{i}$ discrete values of the inverse of warping function $g_{ki}(t)$ which is totally characterized by ${\mbox{\boldmath ${w}$}}_{k}$ and ${\mbox{\boldmath ${w}$}}_{ki}$ as aforementioned. By minimizing $L({\mbox{\boldmath ${w}$}}_{k}, {\mbox{\boldmath ${w}$}}_{ki})$ we can obtain the estimation of ${\mbox{\boldmath ${w}$}}_{k}$ and the prediction of ${\mbox{\boldmath ${w}$}}_{ki}$.
**(iii) Estimating the variance parameters $\sigma^{2}, {\mbox{\boldmath ${\rho}$}}_{s}$ and ${\mbox{\boldmath ${\rho}$}}_{h}$**
By using the first-order Taylor approximation of model (\[eq:mixt3\]) in the the random warping parameters ${\mbox{\boldmath ${w}$}}_{ki}$ around a given prediction ${\mbox{\boldmath ${w}$}}_{ki}^{0}$ (${\mbox{\boldmath ${w}$}}^{0}_{ki}$ is specified by the estimate of ${\mbox{\boldmath ${w}$}}_{ki}$ from **(ii)** in the current iteration), we can write this model as a vectorized linear mixed-effects model $$\label{eq:mod3}
{\mbox{\boldmath ${x}$}}_{a}|_{z_{ki}=1} \approx {\mbox{\boldmath ${G}$}}_{ak} + {\mbox{\boldmath ${B}$}}_{ak}({\mbox{\boldmath ${W}$}}_{k} - {\mbox{\boldmath ${W}$}}_{k}^{0}) + {\mbox{\boldmath ${r}$}}_{ak} + {\mbox{\boldmath ${\epsilon}$}},\hspace{0.5cm} a = 1, 2; \hspace{0.5cm} k = 1, \dots,K,$$ where ${\mbox{\boldmath ${x}$}}_{a} = \{{\mbox{\boldmath ${x}$}}_{ai}, i = 1, \dots, N\}$, with effects given by $$\begin{aligned}
{\mbox{\boldmath ${G}$}}_{ak} &= \Big\{{\mbox{\boldmath ${\Psi}$}}_{ki}|_{g_{ki} = g_{ki}^{0}}{\mbox{\boldmath ${d}$}}_{ak}\Big\}_{ij} \in {\mathds}{R}^{m},\\
{\mbox{\boldmath ${B}$}}_{ak} &= \text{diag}({\mbox{\boldmath ${B}$}}_{aki})_{i} \in {\mathds}{R}^{m\times Nn_{w}},\\
{\mbox{\boldmath ${B}$}}_{aki} &= \bigg\{\partial_{g_{ki}}\Big(\tau_{ak}\big(g_{ki}(t_{j})\big)\Big)\Big|_{g_{ki} = g_{ki}^{0}}\Big(\nabla_{{\mbox{\boldmath ${w}$}}_{ki}}\big(g_{ki}(t_{j})\big)\Big)^\intercal\Big|_{{\mbox{\boldmath ${w}$}}_{ki} = {\mbox{\boldmath ${w}$}}_{ki}^{0}}\bigg\}_{j} \in {\mathds}{R}^{n_{i} \times n_{w}},\\
{\mbox{\boldmath ${W}$}}_{k} &= ({\mbox{\boldmath ${w}$}}_{ki})_{i} \sim N_{Nn_{w}}(0, \sigma^{2}{\mbox{\boldmath ${I}$}}_{N}\otimes {\mbox{\boldmath ${H}$}}_{n_{w} \times n_{w}}), \hspace{3mm} {\mbox{\boldmath ${W}$}}^{0}_{k} = ({\mbox{\boldmath ${w}$}}^{0}_{ki})_{i} \in {\mathds}{R}^{Nn_{w}},\\
{\mbox{\boldmath ${r}$}}_{ak} &\sim N_{m}(0,\sigma^{2}{\mbox{\boldmath ${S}$}}_{ak}), \hspace{3mm} {\mbox{\boldmath ${S}$}}_{ak} = \text{diag}({\mbox{\boldmath ${S}$}}_{aki})_{i} \in {\mathds}{R}^{m \times m},\\
{\mbox{\boldmath ${\epsilon}$}} &\sim N_{m}(0,\sigma^{2}{\mbox{\boldmath ${I}$}}_{m}),
\end{aligned}$$ where $g^{0}_{ki}(t) = t + w_{k}(t) + w^{0}_{ki}(t)$. $\text{diag}({\mbox{\boldmath ${B}$}}_{aki})_{ki}$ is the block diagonal matrix with the ${\mbox{\boldmath ${B}$}}_{aki}$ matrices along its diagonal, so is $\text{diag}({\mbox{\boldmath ${S}$}}_{aki})_{i}$. The derivation of the linearized model (\[eq:mod3\]) is given in \[ap:linear\]. The negative profile log likelihood function for the model (\[eq:mod3\]) is proportional to $$L(\sigma^{2}, {\mbox{\boldmath ${\rho}$}}_{s}, {\mbox{\boldmath ${\rho}$}}_{h}) = \sum_{k=1}^{K}\Big\{\sum_{a = 1}^{2}\sigma^{2}||{\mbox{\boldmath ${x}$}}_{a} - {\mbox{\boldmath ${G}$}}_{ak} + {\mbox{\boldmath ${B}$}}_{ak}{\mbox{\boldmath ${W}$}}_{k}^{0}||^{2}_{{\mbox{\boldmath ${V}$}}_{ak}} + \sum_{a=1}^{2}\text{log det}{\mbox{\boldmath ${V}$}}_{ak}\Big\} + 2mK\text{log} \sigma^{2},$$ where ${\mbox{\boldmath ${V}$}}_{ak} = \frac{\Big({\mbox{\boldmath ${S}$}}_{ak} + {\mbox{\boldmath ${B}$}}_{ak}\big({\mbox{\boldmath ${I}$}}_{N} \otimes{\mbox{\boldmath ${H}$}}_{n_{w}\times n_{w}}\big){\mbox{\boldmath ${B}$}}_{ak}^\intercal + {\mbox{\boldmath ${I}$}}_{m}\Big)}{{\mbox{\boldmath ${M}$}}_{k}} \in {\mathds}{R}^{m \times m}$, with the definition similar to $\frac{{\mbox{\boldmath ${I}$}}_{n}+{\mbox{\boldmath ${S}$}}_{ak}}{{\mbox{\boldmath ${M}$}}_{k}}$ in .
To speed up convergence, we usually repeat the above three steps several times within each iteration. **(iv) Updating $M_{ki}$ and estimating ${\boldsymbol}{\beta}$**
Fix ${\mbox{\boldmath ${\theta}$}}^{(l+1)}$ and update $$M_{ki} = \frac{\pi^{(l)}_{ki}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}^{(l+1)}_{ki})}{\sum_{j=1}^{K}\pi^{(l)}_{ji}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}^{(l+1)}_{ji})},$$ where $$\pi_{ki}^{(l)} = \frac{\text{exp}\{{\mbox{\boldmath ${v}$}}^{\intercal}_{i}{\mbox{\boldmath ${\beta}$}}_{k}^{(l)}\}}{1 + \sum_{j=1}^{K-1}\text{exp}\{{\mbox{\boldmath ${v}$}}^{\intercal}_{i}{\mbox{\boldmath ${\beta}$}}_{j}^{(l)}\}}, \hspace{5mm} k = 1,\dots, K-1,$$ and $\pi_{Ki} = 1 - \sum_{j=1}^{K-1}\pi_{ji}^{(l)}$. Then maximize $Q({\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}})$ with respect to ${\mbox{\boldmath ${\beta}$}}$, which is equivalent to maximize $$L({\mbox{\boldmath ${\beta}$}}) \overset{\Delta}{=} \sum_{i = 1}^{N}\Bigg\{\sum_{k=1}^{K-1}M_{ki}\bigg\{{\mbox{\boldmath ${v}$}}_{i}^{\intercal}{\mbox{\boldmath ${\beta}$}}_{k} - \text{log}\big[1 + \sum_{j = 1}^{K - 1}\text{exp}\{{\mbox{\boldmath ${v}$}}_{i}^{\intercal}{\mbox{\boldmath ${\beta}$}}_{j}\}\big]\bigg\} - M_{Ki}\text{log}\big[1 + \sum_{j = 1}^{K - 1}\text{exp}\{{\mbox{\boldmath ${v}$}}_{i}^{\intercal}{\mbox{\boldmath ${\beta}$}}_{j}\}\big]\Bigg\}.$$ This is very similar to the log-likelihood for a multinomial logit model ($M_{ki}$’s are corresponding to the observations) and can be maximized by iteratively re-weighted least square algorithm.
Model selection, clustering and related methods {#subsec:mod_sel}
-----------------------------------------------
There are two questions on the model selection for our proposed simultaneous registration and clustering method for multi-dimensional functional data: one is how to determine the number of knots for the splines and another is how many clusters. For the former, since our data is rather dense and insensitive, it works well using a relatively small number of equally-spaced knots. For the choice of the number of clusters, $K$, since the number of parameters, $P_{L}$, in model (\[eq:mixt\]) is relative to the number of subjects, $N$, a second-order bias correction version of AIC called $\text{AIC}_{c}$ [@Sug78; @Ken04] is utilized: $$\text{AIC}_{c} = -2L({\boldsymbol}{\hat{\Theta}}) + 2P_{L} + \frac{2P_{L}(P_{L} + 1)}{N - P_{L} - 1},$$ where $L({\mbox{\boldmath ${\hat{\Theta}}$}})$ is the maximized log-likelihood function, ${\mbox{\boldmath ${\hat{\Theta}}$}} = \{{\mbox{\boldmath ${\hat{\theta}}$}}, \hat{{\mbox{\boldmath ${\beta}$}}}\}$ in this paper.
In inference, we first choose $K$ clusters by $\text{AIC}_{c}$. Then fit the data using the method discussed in the previous subsections and denote the estimates of the parameters by $\hat{{\mbox{\boldmath ${\theta}$}}}$ and $\hat{{\mbox{\boldmath ${\beta}$}}}$. Under the framework of SRC method, the fixed-effect part of the $i$th individual curve is calculated by $$\label{eq:pred}
\hat{{\mbox{\boldmath ${x}$}}}_{ai}(t) = \sum_{k=1}^{K}\hat{\pi}_{ki}\big[\hat{\tau}_{ak}(\hat{g}_{ki}(t))\big], \hspace{2mm}a = 1, 2; \hspace{2mm} i = 1,\dots, N,$$ where $\hat{g}_{ki}(t) = t + \hat{{\mbox{\boldmath ${w}$}}}_{k}(t) + \hat{{\mbox{\boldmath ${w}$}}}_{ki}(t)$ and $\hat{\pi}_{ki} = \frac{\text{exp}\{{\mbox{\boldmath ${v}$}}_{i}\hat{{\mbox{\boldmath ${\beta}$}}}_{k}\}}{1 + \sum_{j=1}^{K-1}\text{exp}\{{\mbox{\boldmath ${v}$}}_{i}\hat{{\mbox{\boldmath ${\beta}$}}}_{j}\}}$.
For any individual data $D^{\ast} = \{({\mbox{\boldmath ${x}$}}_{1\ast},{\mbox{\boldmath ${x}$}}_{2\ast}), {\mbox{\boldmath ${v}$}}^{\ast}\}$ in $D$, the posterior distribution of the cluster membership ${\mbox{\boldmath ${z}$}}^{\ast} = (z_{1}^{\ast}, \dots, z_{K}^{\ast})^{\intercal}$ is given by $$p(z_{k}^{\ast} = 1|D^{\ast}) = \frac{\pi_{k}^{\ast}p({\mbox{\boldmath ${x}$}}_{1*}|\hat{{\mbox{\boldmath ${\theta}$}}}_{1k})p({\mbox{\boldmath ${x}$}}_{2*}|\hat{{\mbox{\boldmath ${\theta}$}}}_{2k})}{\sum_{j=1}^{K_{0}}\pi_{j}^{\ast}p({\mbox{\boldmath ${x}$}}_{1*}|\hat{{\mbox{\boldmath ${\theta}$}}}_{1j})p({\mbox{\boldmath ${x}$}}_{2*}|\hat{{\mbox{\boldmath ${\theta}$}}}_{2j})},$$ where $$\pi_{k}^{\ast} = \frac{\text{exp}\{{\mbox{\boldmath ${v}$}}^{\ast \intercal}\hat{{\mbox{\boldmath ${\beta}$}}}_{k}\}}{1 + \sum_{j=1}^{K-1}\text{exp}\{{\mbox{\boldmath ${v}$}}^{\ast \intercal}\hat{{\mbox{\boldmath ${\beta}$}}}_{j}\}}.$$ As a result, the best cluster membership for $D^{\ast}$ can be determined by $$k^{\ast} = {{\arg\!\max}}_{k = 1,\dots,K}\{p(z_{k}^{\ast} = 1|D^{\ast})\}.$$ The average mean curve for each group can be calculated from $\{\hat{{\mbox{\boldmath ${x}$}}}_{ai}(t)|_{z_{ki} = 1}\}$ for $k=1, \ldots, K$.
### Related methods {#subsubsec:rela}
Functional $k$-means method is a popular approach for clustering curves [@Chiou07], which is an extension of $k$-means cluster [@Mac67; @Lloyd82] for scalar variables. The idea can be extended to do clustering and registration simultaneously. Using the similar notation around (\[eq:mixt3\]), we can define the following objective function $$\label{eq:kmean}
F = \sum_{i=1}^{N}\sum_{k=1}^{K}z_{ki}d\big({\mbox{\boldmath ${x}$}}_{i}, {\mbox{\boldmath ${\tau}$}}_{k}(g_{ki})\big),$$ where $d$ represents one kind of distance between each curve to its assigned mean curve ${\mbox{\boldmath ${\tau}$}}_{k}(g_{ki})$ and $$z_{ki} = \left\{
\begin{array}{ll}
1, \hspace{2mm}\text{if} \hspace{2mm}k = {\arg\!\min}_{j}d\big({\mbox{\boldmath ${x}$}}_{i}, {\mbox{\boldmath ${\tau}$}}_{j}(g_{ji})\big),\\
0, \hspace{2mm}\text{otherwise}.
\end{array}
\right.$$In order to find the values $\{z_{ki}\}$ and the $\{{\mbox{\boldmath ${\tau}$}}_{k}(g_{ki})\}$ to minimize $F$, we can perform an iterative procedure in which each iteration involves two steps of the optimization with respect to $\{z_{ki}\}$ and $\{{\boldsymbol}{\tau}_{k}(g_{ki})\}$ respectively; the details are given in \[ap:imp\]. This approach is denoted by *k-means-f*.
A special case of the SRC model defined in Section \[subsec:mode\] is that the allocation model in (\[eq:allo\]) doesn’t depend on any scalar variables (denoted by *SRC-f*, i.e. use the function variable only). This special case is very similar to the above *k-means-f* approach. Actually the *k-means-f* algorithm is a special case of EM algorithm for *SRC-f*. Using the similar notation around (\[eq:mixt3\]) and assuming $
{\mbox{\boldmath ${x}$}}_{ai}|_{z_{ki} = 1} \sim N_{n_{i}}({\mbox{\boldmath ${\tau}$}}_{ak}(g_{ki}), \delta{\mbox{\boldmath ${I}$}}), a = 1, 2; i = 1, \dots, N,
$ where $\delta$ is shared by all the clusters, we have the density function of ${\mbox{\boldmath ${x}$}}_{ai}$ with the form $$p({\mbox{\boldmath ${x}$}}_{ai}|{\mbox{\boldmath ${\theta}$}}_{aki}) = (2\pi\delta)^{-\frac{n_{i}}{2}}\text{exp}\big\{-\frac{1}{2\delta}||{\mbox{\boldmath ${x}$}}_{ai} - {\mbox{\boldmath ${\tau}$}}_{ak}(g_{ki})||^{2}\big\}.$$ Let the allocation model as $p(z_{ki} = 1) = \pi_{k}, k = 1, \dots, K$ with $\sum_{k=1}^{K}\pi_{k} = 1$. Using the EM algorithm for the Gaussian mixtures described in Section \[subsec:esti\], we have $$M_{ki} = \frac{\pi_{k}\prod_{a=1}^{2}\text{exp}\big\{-||{\mbox{\boldmath ${x}$}}_{ai} - {\mbox{\boldmath ${\tau}$}}_{ak}(g_{ki})||^{2}/2\delta\big\}}{\sum_{j=1}^{K}\pi_{j}\prod_{a=1}^{2}\text{exp}\big\{-||{\mbox{\boldmath ${x}$}}_{ai} - {\mbox{\boldmath ${\tau}$}}_{aj}(g_{ji})||^{2}/2\delta\big\}}.$$ Clearly, $M_{ki} \rightarrow z_{ki}$, when $\delta \rightarrow 0$. Thus $E_{{\mbox{\boldmath ${z}$}}}[\text{log}p({\mbox{\boldmath ${x}$}},{\mbox{\boldmath ${z}$}}|{\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}})] \approx -\frac{1}{2\delta}F + \text{costant}$, when $\delta$ is small. It means the optimization problem is the same as the *k-means-f* algorithm given by (\[eq:kmean\]) (use $d = ||\cdot||^{2}$).
Numerical analysis {#sec:num_ana}
==================
We shall evaluate the performance and properties of the proposed *SRC* model in this section. We will compare it with functional $k$-means clustering (*k-means-f*) with simultaneous registration as discussed in Section \[subsubsec:rela\], the *SRC* without using an allocation model (*SRC-f*) and scalar $k$-means clustering (*k-means-s*). The *k-means-s* is a general k-means clustering method using scalar variables only. We will conduct analysis on both simulated and real data.
Simulation study {#sec:simu}
----------------
In this simulation study, we consider 2D curves coming from two groups. For each group, the corresponding observations of functional variables ${\mbox{\boldmath ${x}$}}(t)$ and scalar variables $v$ will be generated. There are $N_{k}$ batches of data in each group, where $k$ = 1 and 2. We will evaluate and compare four methods based on the simulated data $D = \{({\mbox{\boldmath ${x}$}}_{i}(t_{ij}), {\mbox{\boldmath ${v}$}}_{i}); i = 1,\dots, N; j = 1,\dots, n_{i}\}$ in different scenarios where $N=N_{1}+N_{2}$.
### Data generation
1. Generate the underlying true curves, i.e. the curves based on the internal time scale $g^{-1}(t)$. We first assume those curves in two groups share the following two slightly different true means, which is similar to two patterns of those real curves of hyoid bone’s movement $$\begin{split}
{\boldsymbol}{\mu}_{1}(t) & = \big(\mu_{11}(t), \mu_{21}(t)\big) = \big(\text{exp}\{\text{cos}(2\pi t)\}, \text{exp}\{\text{sin}(2\pi t)\}\big), \\
{\boldsymbol}{\mu}_{2}(t) & = \big(\mu_{12}(t), \mu_{22}(t)\big) = \big(\text{exp}\{\text{cos}(2\pi t^{1.05} - b_{1})\}, \text{exp}\{\text{sin}(2\pi t^{1.1} + b_{1})\}\big).
\end{split}$$
The degree of overlapping between two groups relies on the value of $b_{1}$. The smaller the value of $b_{1}$, the higher the degree of overlapping, and more difficult to cluster those curves. We use the equidistant points $t_{j} = \frac{j+1}{102}, j = 1,\dots, 100$ as input grid, i.e. $n_i=100$. The underlying true curves are generated as $${\mbox{\boldmath ${x}$}}_{ak}(g^{-1}(t)) = \mu_{ak}(t), \hspace{5mm} a = 1, 2; \hspace{2mm} k = 1, 2.$$ Figure \[fig:cluster\_true\_mean\_d010\] in \[ap:sist\] shows the shape of the true mean curves for different values of $b_{1}$.
2. Generate the original 2D curves ${\mbox{\boldmath ${x}$}}(t)$ by adding the warping function, amplitude variation and errors as follows.
1. Model the true curves ${\mbox{\boldmath ${x}$}}_{ak}(g^{-1}(t))$ using B-spline basis function with 8 knots and obtain the coefficients ${\mbox{\boldmath ${d}$}}_{ak}$.
2. For simplicity, we set $g_{ki}(t) = t + w_{ki}(t)$ and use hyman spline (monotone cubic spine using Hyman filtering) based on the anchor knots $t_{w} = (0, 0.33, 0.67, 1)$ ($n_{w} = 4$). Set ${\mbox{\boldmath ${w}$}}_{ki} \sim N_{4}(0, {\mbox{\boldmath ${T}$}}_{k}^{\intercal}{\mbox{\boldmath ${\Gamma}$}}_{i})$, where ${\mbox{\boldmath ${T}$}}_{k}^{\intercal}{\mbox{\boldmath ${T}$}}_{k} = {\mbox{\boldmath ${O}$}}_{k}$, ${\mbox{\boldmath ${O}$}}_{1} = \begin{bmatrix} 10 & 4 \\ 4 & 8\end{bmatrix}$ and ${\mbox{\boldmath ${O}$}}_{2} = \begin{bmatrix} 10 & 8 \\ 8 & 15\end{bmatrix}$, and ${\mbox{\boldmath ${\Gamma}$}}_{i} = ({\mbox{\boldmath ${\Gamma}$}}_{i1}, {\mbox{\boldmath ${\Gamma}$}}_{i2})^{\intercal}$ with ${\mbox{\boldmath ${\Gamma}$}}_{i1}, {\mbox{\boldmath ${\Gamma}$}}_{i2}$ being independent random variables $N(0, \sigma^{2}_{w})$, where $i = 1, \dots, N_{k}$.
3. Set the amplitude variation ${\mbox{\boldmath ${r}$}}_{aki} = {\mbox{\boldmath ${T}$}}_{0}^{\intercal}\cdot{\mbox{\boldmath ${\Gamma}$}}_{i0}$, where ${\mbox{\boldmath ${T}$}}_{0}^{\intercal}{\mbox{\boldmath ${T}$}}_{0} = {\mbox{\boldmath ${O}$}}_{0}$, $a = 1, 2$, $k = 1, 2$, $i =1, \dots, N_{k}$. ${\mbox{\boldmath ${O}$}}_{0}$ is created by Matern covariance function with ${\mbox{\boldmath ${\rho}$}}_{r} = (100, 0.3, 3)$, where the three elements represent the scale, range and smoothness, respectively [@Raket], and ${\boldsymbol}{\Gamma}_{j0}$ is a vector of 100 independent normal random variables $N(0, \sigma^{2}_{r})$. Set ${\boldsymbol}{\epsilon} \sim N(0, \sigma^{2}{\mathds}{I})$.
4. Generate ${\mbox{\boldmath ${x}$}}(t)$ based on the model (\[eq:mixt3\]).
3. Generate ${\mbox{\boldmath ${v}$}}$’s. We next generate those scalar variables $v$ by sampling from uniform distribution as follows: $$V_{i} \sim \left\{
\begin{array}{ll}
\text{U}(1, 2), \hspace{2mm} i = 1, 2, \dots, N_{1},\\
\text{U}(1 - b_{2}, 2 - b_{2}), \hspace{2mm} i = N_{1}+1, \dots, N_{1}+N_{2}.
\end{array}
\right.$$ Bear in mind that the larger the value of $d_{2}$, the lower the degree of overlapping and easier to carry out clustering using scalar variables.
### Results {#subsec:res}
In order to investigate how the overlapping of the observations of both scalar variables and functional variables affect the performance of clustering, we study four scenarios with the constraint $4\sigma_{w}^{2} = \sigma_{r}^{2} = \sigma^{2} = 0.01^2$ and $N_{1} = N_{2} = 30$, and with 100 replications for each scenario. We use two measures to assess the performance of each method. The first one is the Rand index (RI) [@Ran:71], having a value between 0 and 1, with 0 indicating two data clusterings disagree on any pair while 1 indicating a perfect match. And the second one is called adjusted Rand index (ARI) [@Hub:85], a modified version of Rand index (ARI). A larger value of RI or ARI indicates a higher agreement of the method and the truth.
Four methods are applied to the simulated data $D$ in Scenario 1 with $b_{1} = 0.12$, $b_{2} = 0.8$, Scenario 2 with $b_{1} = 0.10$, $b_{2} = 0.8$, Scenario 3 with $b_{1} = 0.08$, $b_{2} = 0.8$ and Scenario 4 with $b_{1} = 0.08$, $b_{2} = 0.6$. Figure \[fig:scal\] and Figure \[fig:raw\] show the raw data depending on different $b_{2}$ and $b_{1}$ respectively. First of all, we apply the $\text{AIC}_{c}$ to choose the number of clusters. The results from Figure \[fig:AICc\] in \[ap:sist\] show that $\text{AIC}_{c}$ score reaches its minimum at 2 clusters. Table \[tab:clust\] summarizes the comparisons by average ARI and RI. Overall, both measures suggest that the proposed *SRC* outperform the other three methods in all scenarios because of the use of both functional and scalar data.
[0.4]{}
[0.4]{} ![Observations of scalar variable in two cases. The ‘blue’ ones stand for those in the range of overlapping.[]{data-label="fig:scal"}](Scalar06 "fig:"){width="\textwidth"}
![The raw 2D curves in one simulation run in three cases. []{data-label="fig:raw"}](B1_X "fig:"){width="30.00000%"} ![The raw 2D curves in one simulation run in three cases. []{data-label="fig:raw"}](B1_Y "fig:"){width="30.00000%"}
![The raw 2D curves in one simulation run in three cases. []{data-label="fig:raw"}](B2_X "fig:"){width=".3\textwidth"} ![The raw 2D curves in one simulation run in three cases. []{data-label="fig:raw"}](B2_Y "fig:"){width=".3\textwidth"}
![The raw 2D curves in one simulation run in three cases. []{data-label="fig:raw"}](B3_X "fig:"){width=".3\textwidth"} ![The raw 2D curves in one simulation run in three cases. []{data-label="fig:raw"}](B3_Y "fig:"){width=".3\textwidth"}
From Table \[tab:clust\], we note that all the four methods perform best in Scenario 1 compared to the other scenarios. In this scenario, both $b_{1}$ and $b_{2}$ take the largest values, indicating that overlapping of the functional data (Figure \[fig:raw\] (c)) and scalar data (Figure \[fig:scal\] (a)) are the smallest and both greatly contribute to distinguishing those two clusters. Though the other three methods, *SRC-f*, *k-means-f* and *k-means-s* based on either functional data or scalar data, also have good performance but not as good as *SRC*.
The first three scenarios share the same value of $b_{2}$, indicating that the degree of overlapping in two clusters for scalar data does not change (Figure \[fig:scal\] (a)). It causes the performance of *k-means-s* remaining the same. The overlapping in two clusters for functional data, however, gets smaller and smaller as the value of $b_{1}$ decreases from Scenario 1 to Scenario 3. It leads to a sharp decline for the performance of *SRC-f* and *k-means–f*, both of which depend on functional data only, as opposed to a mild decrease of the performance of *SRC*, which is based on both scalar data and functional data. The scenario 4 has the smallest $b_{1}$ (Figure \[fig:raw\] (a)) and $b_{2}$ (Figure \[fig:scal\] (b)) and it is quite difficult to carry out clustering just based on functional data or scalar data only. Consequently, the values of ARI for *SRC-f*, *k-means-f* and *k-means-s* are very small. But *SRC* still performs well and are much better than the others. Other combinations with varying overlapping determined by $b_{1}$ and $b_{2}$ and with different sample sizes have also been examined. The results presented here are very typical.
------------- -- -- ------------------- ------------------- -- ------------------- ------------------- -- ------------------- ------------------- -- ------------------- -------------------
RI ARI RI ARI RI ARI RI ARI
*SRC* ${\mathbf}{0.98}$ ${\mathbf}{0.96}$ ${\mathbf}{0.95}$ ${\mathbf}{0.90}$ ${\mathbf}{0.91}$ ${\mathbf}{0.82}$ ${\mathbf}{0.80}$ ${\mathbf}{0.61}$
*SRC-f* 0.90 0.80 0.75 0.49 0.62 0.25 0.62 0.25
*k-means-f* 0.91 0.83 0.76 0.53 0.64 0.29 0.64 0.29
*k-means-s* 0.81 0.62 0.81 0.62 0.81 0.62 0.69 0.37
------------- -- -- ------------------- ------------------- -- ------------------- ------------------- -- ------------------- ------------------- -- ------------------- -------------------
: Comparison of average clustering results among four methods.[]{data-label="tab:clust"}
### Recovery of curves and cluster patterns
To understand the underlying process better, it is necessary to use the optimally alignment to estimate the entire curve, so we estimate the aligned individual curves and reconstruct the cluster pattern using equation (\[eq:pred\]).
Figure \[fig:rec1\] in Appendix displays one simulation run of $N = 100$, with $N_{1} = N_{2} = 50$ curves in each group. The top panel presents the original raw curves in the two clusters in two colors in two dimensions ($x$-axis and $y$-axis) for a new scenario with $4\sigma_{w}^{2} = \sigma_{r}^{2} = \sigma^{2} = 0.02^2, b_{1} = 0.15, b_{2} = 0.8$. The other panels respectively show the individual aligned curves resulting from *SRC*, *SRC-f* and *k-means-f*, with the value of RI (1, 0.63, 0.79) and the value of ARI (1, 0.26, 0.54) respectively. The *SRC* properly differentiates the two clusters (red and green) after curve alignment and performs better in recovering the cluster patterns.
Figure \[fig:rec2\] summarizes the result of clustering patterns. It shows the SRC method recovered the true pattens very well. As a measure of estimation error, we use the root average squared error [@Ger04] $$\textit{rase}(\hat{{\mbox{\boldmath ${\mu}$}}}) = \sqrt{\frac{\sum_{j=1}^{m}||\hat{{\mbox{\boldmath ${\mu}$}}}(t_{j}) - {\mbox{\boldmath ${\mu}$}}(t_{j})||^2}{m}},$$ where $m$ is the number of observation points, ${\mbox{\boldmath ${\mu}$}}(t)$ and $\hat{{\mbox{\boldmath ${\mu}$}}}(t)$ are the cross-sectional mean of 2D raw curves and of registered curves from the target method respectively. The values of *rase* are 2.9, 4.6 and 5.4 corresponding to three models *SRC*, *SRC-f* and *k-means-f*.
[0.45]{}
[0.45]{}
[1]{} ![(a)-(b): *solid line* and *dashed line* represent the cluster 1 and cluster 2, respectively. The *green lines* in (a), (b) and (c) are true mean functions. The *red lines*, *brown lines* and *purple lines* in (a)-(c) are respectively corresponding to *SRC*, *SRC-f* and *k-means-f*.[]{data-label="fig:rec2"}](Rec_xyc1 "fig:"){width="45.00000%"} ![(a)-(b): *solid line* and *dashed line* represent the cluster 1 and cluster 2, respectively. The *green lines* in (a), (b) and (c) are true mean functions. The *red lines*, *brown lines* and *purple lines* in (a)-(c) are respectively corresponding to *SRC*, *SRC-f* and *k-means-f*.[]{data-label="fig:rec2"}](Rec_xyc2 "fig:"){width="45.00000%"}
One extreme scenario has been discussed in \[ap:exts\], providing further evidence of the good performance of the proposed *SRC*.
Real data analysis {#subsec:real}
------------------
The application to a real data is to cluster the normal people and the patients with stroke by studying their hyoid bone motion as well as the other scalar variables. Two groups, one for normal people and the other for patients, are included. Figure (\[fig:realdata\])(a) shows one frame from a X-ray video clip. The position of hyoid bone is tracked in each frame by a semi-automatic programme developed in [@Kim17]. The raw data before being preprocessed are shown in Figure (\[fig:realdata\])(b). There are 15 subjects in each group. The scalar variable we choose is the size of Pyriform Sinus Residue (see its position in Figure \[fig:pyriform\]). Regarding to those 2D curves, we firstly carry out the preprocessing procedures like multi-dimensional shift, scaling and rotation using the package of GPA. While modeling those curves, we assume the warping function be a smooth nonlinear deformation produced by an increasing spline and the random vector ${\boldsymbol}{w}_{ki}$ be a Brownian bridge observed at discrete anchor points. B-spline basis functions are utilized for modeling the mean curves. The covariance function for the amplitude variance is the Matern covariance function.
We examine the performance of four methods *SRC*, *SRC-f*, *k-means-f* and *k-means-s* aforementioned. The values of AICc are shown in Figure \[fig:AICc\_real1\]. It shows that the two-component mixture model has the smallest value. Table \[tab:cluster\_realdata\] shows the values of RI and ARI by comparing clustering results by the four methods with the clinic outcomes. We can see that the *SRC* method outperforms the other three. As a matter of fact, both *SRC-f* and *k-means-f* with value of RI equivalent to 0.5 fail in this real data example. It is similar to the extreme example in \[ap:exts\]. More numerical results are provided in \[sec:appreal\].
Model RI ARI
------------- ----------------------- -----------------------
*SRC* ${\boldsymbol}{0.71}$ ${\boldsymbol}{0.42}$
*SRC-f* 0.50 0.02
*k-means-f* 0.50 0.02
*k-means-s* 0.67 0.33
: Results of clustering by four methods for the real data[]{data-label="tab:cluster_realdata"}
Conclusion {#sec:dis}
==========
We have proposed a methodology for simultaneous registration and clustering, SRC, for multi-dimensional functional data which considers both the curves and scalar variables. This model captures the heterogeneity from the potential time warping for curves and scalar variables corresponding to each subject while carrying out the clustering in the meantime. It can be implemented with EM algorithm. Numerical examples show that it outperforms three other related methods, *SRC-f*, *k-means-f* and *k-means-s*. The results in Section \[subsec:res\] show that in most cases the inclusion of scalar variables can improve the performance of clustering in functional data analysis.
Generally, the registration for multi-dimensional functional data is much more complicated than one dimensional case. We use the pre-processing package GPA [@Gower:1975] and a further registration via a simple warping function. The latter is one of the key parts in our model. This approach performs very well in the numerical examples presented in this article. Further research is however needed, for example, how to improve the iterative implementation for the complete registration, similar to the shape geodesic algorithm by the metric-based method proposed by [@Sri11a]. The inverse of warping function $g$ in model (\[eq:mixt\]) can also be replaced with various types of other functions depending on types of data. The success of resolving registration problem often depends on the flexibility of choosing warping function.
Model selection is an interesting but difficult issue for a mixture model, especially for the models with complex forms. [@Ken04] suggested AICc should be used unless $\frac{N}{P_{L}}>40$ for the model with the large value of $P_{L}$. In our model, the number of parameters $P_{L}$ is quite close to the number of subject $N$. Thus, we use AICc. It works well for the examples discussed in the article. It is worth a further study under a general functional data analysis framework. The results we obtained for the real data are encouraging although it is still in early stage. Research for this topic is carrying on. More features extracted from video clips along with other variables, both functional and scalar, are under investigation. Different types of models for data fitting, clustering/classification and prediction are being developed.
Bouveyron, C. and Jacques, J. (2011). Model-based clustering of time series in group-specific functional subspaces. , 5(4):281–300.
Cheng, W., Dryden, L. L., and Huang, X. Z. (2016). Bayesian registration of functions and curves. , 11(2):447–475.
Chiou, J. M. and Li, P. L. (2007). Functional clustering and identifying substructures of longitudinal data. , 69(4):679–699.
Delaigle, A. and Hall, P. (2010). Defining probability density for a distribution of random functions. , 38:1171–1193.
Ferraty, F. and Vieu, P. (2006). . Springer Series in Statistics, Springer, New York.
Gervini, D. and Gasser, T. (2004). Self-modelling warping functions. , 66(4):959–971.
Gower, J. C. (1975). Generalized procrustes analysis. , 40(1):33–51.
Green, P. J. and Richardson, S. (2000). Spatially corrrelated allocation models for count data.
Hartigan, J. A. and Wong, M. A. (1978). Algorithm as 1326: A k-means clustering algorithm. , 28:100–108.
Hubert, L. and Arabie, P. (1985). Comparing partitions. , 2:193–218.
Ieva, F., Paganoni, A. M., Pigoli, D., and Vitelli, V. (2013). Multivariate functional clustering for the analysis of ecg curves morphology. , 62(3):401–418.
James, G. M. and Sugar, C. A. (2003). Clustering for sparsely sampled functional data. , 98(462):397–408.
Kenneth, P. B. and David, R. A. (2004). Understanding aic and bic in model selection. , 33:261–304.
Kim, W.-S., Zeng, P., Shi, J. Q., Lee, Y., and Paik, N.-J. (2017). Automatic tracking of hyoid bone motion from videofluoroscopic swallowing study with automatic smoothing and segmentation. , page (to appear).
Liu, X. and Yang, M. C. K. (2009). Simultaneous curve registration and clustering for functional data. , 53:1361–1376.
Lloyd, S. P. (1982). Least squares quantization in pcm. , 28(2):129–137.
MacQueen, J. B. (1967). Some methods for classification and analysis of multivariate observations. , pages 281–297.
Marron, J., Ramsay, J. O., Sangalli, L. M., and Srivastava, A. (2015). Functional data analysis of amplitude and phase variation. , 30(4):468–484.
Raket, L. L. (2016). pavpop version 0.10. Available at <http://github.com/larslau/pavpop/>.
Raket, L. L., Grimme, B., Schoner, G., Igel, C., and Markussen, B. (2016). Separating timing, movement conditions and individual differences in the analysis of human movement. , 12(9).
Ramsay, J. O. and Silverman, B. W. (2005). . Springer-Verlag New York, USA.
Rand, W. M. (1971). Objective criteria for the evaluation of clustering methods. , 66:846–850.
Sam, A., Chamroukhi, F., Govaert, G., and Aknin, P. (2011). Model-based clustering and segmentation of time series with changes in regime. , 5(4):301–322.
Sangalli, L. M., Secchi, P., Vantini, S., and Veneziani, A. (2009). A case study in exploratory functional data analysis: geometrical features of the internal carotid artery. , 104:37–48.
Sangalli, L. M., Secchi, P., Vantini, S., and Vitelli, V. (2010). k-mean alignment for curve clustering. , 54(5):1219–1233.
Shi, J. Q. and Choi, T. (2011). . CRC Press Taylor & Francis Group.
Shi, J. Q., Murray-Smith, R., and Titterington, D. M. (2005). Hierarchical gaussian process mixtures for regression. , 15:31–41.
Shi, J. Q. and Wang, B. (2008). Curve prediction and clustering with mixtures of gaussian process functional regression models. , 18(3):267–283.
Shi, J. Q., Wang, B., Will, E. J., and West, R. M. (2012). Mixed-effect gaussian process functional regression models with application to dose-response curve prediction. , 31:3165–3177.
Srivastava, A., Klassen, E., Joshi, S. H., and Jermyn, I. H. (2011a). Shape analysis of elastic curves in euclidean spaces. , 33(7):1415–1428.
Srivastava, A., Wu, W., Kurtek, S., Klassen, E., and Marron, J. S. (2011b). Registration of functional data using the fisher-rao metric. .
Sugiura and Nariaki (1978). Further analysis of the data by akaike’s information criterion and the finite corrections. , A7:13–26.
Tang, R. and M[ü]{}ller, H. G. (1998). Pairwise curve synchronization for functional data. , 95(4):875–889.
Tarpey, T. and Kinateder, K. J. (2003). Clustering functional data. , 20(1):93–114.
Tokushige, S., Yadohisa, H., and Inada, K. (2007). Crisp and fuzzy k-means clustering algorithms for multivariate functional data. , 22:1–16.
Appendices {#appendices .unnumbered}
==========
Derivation of $M_{ki}$ {#ap:deri}
======================
Using Bayes’ theorem, the posterior distribution with respect to ${\mathbf}{z}$ has the form $$p({\mbox{\boldmath ${z}$}}|{\mbox{\boldmath ${x}$}}, {\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}}) \propto \prod_{i=1}^{N}\prod_{k=1}^{K}\pi_{ki}^{z_{ki}}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}_{ki})^{z_{ki}}.$$ By factorizing it over $i$, it is clear that the $\{{\mbox{\boldmath ${z}$}}_{i}, i = 1, \dots, N\}$ are independent under the posterior distribution. Hence, $$\begin{split}
E(z_{ki}|{\mbox{\boldmath ${x}$}}, {\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}}) &= E(z_{ki}|{\mbox{\boldmath ${x}$}}_{i}, {\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}})\\
&= p(z_{ki}=1|{\mbox{\boldmath ${x}$}}_{i}, {\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}})\\
&= \frac{p(z_{ki}=1|{\mbox{\boldmath ${\beta}$}})p({\mbox{\boldmath ${x}$}}_{i}|z_{ki} = 1, {\mbox{\boldmath ${\theta}$}})}{p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}})}\\
&= \frac{\pi_{ki}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}_{ki})}{p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}})}.
\end{split}$$ We know that $$\begin{split}
p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}}) & = \sum_{{\mbox{\boldmath ${z}$}}_{i}}p({\mbox{\boldmath ${z}$}}_{i}|{\mbox{\boldmath ${\beta}$}})p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${z}$}}_{i}, {\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}})\\
& = \sum_{z_{ki}}\prod_{k=1}^{K}\big(p(z_{ki} = 1|{\mbox{\boldmath ${\beta}$}})p({\mbox{\boldmath ${x}$}}_{i}|z_{ki}=1, {\mbox{\boldmath ${\theta}$}})\big)^{z_{ki}}\\
& = \sum_{j=1}^{K}\pi_{ji}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}_{ji}).
\end{split}$$ Thus, $$E(z_{ki}|{\mbox{\boldmath ${x}$}}, {\mbox{\boldmath ${\theta}$}}, {\mbox{\boldmath ${\beta}$}}) = \frac{\pi_{ki}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}_{ki})}{\sum_{j=1}^{K}\pi_{ji}p({\mbox{\boldmath ${x}$}}_{i}|{\mbox{\boldmath ${\theta}$}}_{ji})} .$$
Derivation of the linearized model {#ap:linear}
==================================
At the linearized level, we do the first-order Taylor approximation of model (\[eq:mixt3\]) in Section \[subsec:esti\] in the random warp ${\mbox{\boldmath ${w}$}}_{ki}$. We reconsider $g_{ki}(t)$ as a function with respect to ${\mbox{\boldmath ${w}$}}_{k}$ + ${\mbox{\boldmath ${w}$}}_{ki}$. Thus, the linearization can be carried out around the estimate of ${\mbox{\boldmath ${w}$}}_{k}$ plus ${\mbox{\boldmath ${w}$}}^{0}_{ki}$ obtained from the previous step. This results in a linear mixed-effects model as follows: $${\mbox{\boldmath ${x}$}}_{ai}|_{z_{ki=1}} \approx {\mbox{\boldmath ${x}$}}_{ai}|_{z_{ki=1}, {\mbox{\boldmath ${w}$}}_{ki} = {\mbox{\boldmath ${w}$}}^{0}_{ki}} + \nabla_{{\mbox{\boldmath ${w}$}}_{ki}}({{\mbox{\boldmath ${x}$}}_{ai}|_{z_{ki=1}}})|_{{\mbox{\boldmath ${w}$}}_{ki} = {\mbox{\boldmath ${w}$}}^{0}_{ki}}({\mbox{\boldmath ${w}$}}_{ki} - {\mbox{\boldmath ${w}$}}^{0}_{ki})$$ where $$\begin{aligned}
{\mbox{\boldmath ${x}$}}_{ai}|_{z_{ki=1}, {\mbox{\boldmath ${w}$}}_{ki} = {\mbox{\boldmath ${w}$}}^{0}_{ki}} &= {\mbox{\boldmath ${\tau_{ak}}$}}(g_{ki})|_{{\mbox{\boldmath ${w}$}}_{ki} = {\mbox{\boldmath ${w}$}}^{0}_{ki}} + {\mbox{\boldmath ${r}$}}_{aki} + {\mbox{\boldmath ${\epsilon}$}}\\
&= {\mbox{\boldmath ${\Psi}$}}_{ki}|_{g_{ki} = g_{ki}^{0}}{\mbox{\boldmath ${d}$}}_{ak} + {\mbox{\boldmath ${r}$}}_{aki} + {\mbox{\boldmath ${\epsilon}$}}\end{aligned}$$ and according to the chain rule, $$\begin{aligned}
\nabla_{{\mbox{\boldmath ${w}$}}_{ki}}({{\mbox{\boldmath ${x}$}}_{ai}|_{z_{ki=1}}})|_{{\mbox{\boldmath ${w}$}}_{ki} = {\mbox{\boldmath ${w}$}}^{0}_{ki}} & = \Bigg\{\frac{\partial{{\mbox{\boldmath ${x}$}}_{ai}|_{z_{ki=1}, t = t_{j}}}}{\partial{g_{ki}}} \bigg(\nabla_{{\mbox{\boldmath ${w}$}}_{ki}}\big(g_{ki}(t_{j})\big)\bigg)^{\intercal}\Big|_{{\mbox{\boldmath ${w}$}}_{ki} = {\mbox{\boldmath ${w}$}}^{0}_{ki}}\Bigg\}_{j} \\
& = \bigg\{\frac{\partial \Big(\tau_{ak}\big(g_{ki}(t_{j})\big)\Big)}{\partial{g_{ki}}}\Big|_{g_{ki}= g_{ki}^{0}}\Big(\nabla_{{\mbox{\boldmath ${w}$}}_{ki}}\big(g_{ki}(t_{j})\big)\Big)^\intercal\Big|_{{\mbox{\boldmath ${w}$}}_{ki} = {\mbox{\boldmath ${w}$}}_{ki}^{0}}\bigg\}_{j} \in {\mathds}{R}^{n_{i} \times n_{w}} . \end{aligned}$$ In practical, we use finite difference for calculating the derivative of $g_{ki}(t)$ with respect to ${\mbox{\boldmath ${w}$}}_{ki}$.
Implementation of functional $k$-means {#ap:imp}
======================================
We just use $d = ||\cdot||^{2}$ for the purpose of simplicity and convenience. The iterative procedures are as follows:
1. Choose the initial values for the $z_{ki}$. We can use any clustering method for scalar variables $\{{\mbox{\boldmath ${\beta}$}}_{i}\}$ corresponding to the functional variables $\{{\mbox{\boldmath ${x}$}}_{i}\}$.
2. Fix $z_{ki}$ and minimize $F$ with respect to the ${\mbox{\boldmath ${\tau}$}}_{ak}(g_{ki})$. In this phase, minimizing $F$ is equivalent to maximizing $$\sum_{k=1}^{K}\sum_{i=1}^{N}z_{ki}\bigg(\sum_{a=1}^{2}\big(\text{log}p({\mbox{\boldmath ${x}$}}_{ai}|{\mbox{\boldmath ${\theta}$}}_{aki})\big)\bigg)$$ with the assumption that the covariance matrix of ${\mbox{\boldmath ${x}$}}_{ai}$ are the same over all the subjects. The detailed estimation of $\{{\mbox{\boldmath ${\theta}$}}_{aki}\}$ have been mentioned before and the ${\mbox{\boldmath ${\tau}$}}_{ak}(g_{ki})$ can be obtained straightforwardly.
3. Fix ${\mbox{\boldmath ${\tau}$}}_{ak}(g_{ki})$ and minimize $F$ with respect to $z_{ki}$. Since the term $F$ in (\[eq:kmean\]) involving different $i$ are independent, we can optimize $F$ for each $i$ separately by choosing $z_{ki}$ as follows $$z_{ki} = \left\{
\begin{array}{ll}
1, \hspace{2mm}\text{if} \hspace{2mm}k = {\arg\!\min}_{j}||\big({\mbox{\boldmath ${x}$}}_{i} - {\mbox{\boldmath ${\tau}$}}_{j}(g_{ji})\big)||^{2};\\
0, \hspace{2mm}\text{otherwise}.
\end{array}
\right.$$
4. Repeat Step 2 and Step 3 until convergence.
Extra numerical results in simulation study {#ap:sist}
===========================================
Figures \[fig:cluster\_true\_mean\_d010\] to \[fig:rec1\] provide extra numerical results for the simulated examples discussed in Section \[sec:simu\].
An extra example is given below.
A simulation example in an extreme scenario {#ap:exts}
===========================================
It is not uncommon that sometimes the functional variables provide little information so that it fails to implement the clustering just based on those curves. However, the addition of scalar variables can make the clustering possible. We simulate a run of $N = 100$ (sample size), with $4\sigma_{w}^{2} = \sigma_{r}^{2} = \sigma^{2} = 0.02^2, b_{1} = 0.05, b_{2} = 0.8$, and $N_{1} = N_{2} = 50$ curves in each group. Figure \[fig:rec01\] displays the individual aligned curves resulting from three methods, from which no discernible clusters are visiable. The RI and ARI for *SRC*, *SRC-f* and *k-means-f* are, however, markedly different with the values of $(0.82, 0.50, 0.50)$ and $(0.64, 0, 0)$ respectively. Figure \[fig:rec02\] summarizes the mean functions of two clusters by three methods. Their values of *rase* are 1.1, 18.3 and 4.3 respectively. Those results show that the use of SRC leads to meaningful findings but the other twos are equivalent to random guess.
Extra numerical results in real data analysis {#sec:appreal}
=============================================
Figures \[fig:pyriform\] to \[fig:real1\] provide extra numerical results for the real data example discussed in Section \[subsec:real\].
Figure \[fig:real1\] shows the clustering results by SRC along with the registering effect, which are quite close to the true data shown in Figure \[fig:real0\].
![True mean curves ${\boldsymbol}{\mu}_{1}(t)$ (indicated by green lines) and ${\boldsymbol}{\mu}_{2}(t)$ (indicated by red lines) of group 1 and group 2 with $b_{1} = 0.08, 0.10, 0.12$.[]{data-label="fig:cluster_true_mean_d010"}](True_B1_X "fig:"){width="30.00000%"} ![True mean curves ${\boldsymbol}{\mu}_{1}(t)$ (indicated by green lines) and ${\boldsymbol}{\mu}_{2}(t)$ (indicated by red lines) of group 1 and group 2 with $b_{1} = 0.08, 0.10, 0.12$.[]{data-label="fig:cluster_true_mean_d010"}](True_B1_Y "fig:"){width="30.00000%"}
![True mean curves ${\boldsymbol}{\mu}_{1}(t)$ (indicated by green lines) and ${\boldsymbol}{\mu}_{2}(t)$ (indicated by red lines) of group 1 and group 2 with $b_{1} = 0.08, 0.10, 0.12$.[]{data-label="fig:cluster_true_mean_d010"}](True_B2_X "fig:"){width=".3\textwidth"} ![True mean curves ${\boldsymbol}{\mu}_{1}(t)$ (indicated by green lines) and ${\boldsymbol}{\mu}_{2}(t)$ (indicated by red lines) of group 1 and group 2 with $b_{1} = 0.08, 0.10, 0.12$.[]{data-label="fig:cluster_true_mean_d010"}](True_B2_Y "fig:"){width=".3\textwidth"}
![True mean curves ${\boldsymbol}{\mu}_{1}(t)$ (indicated by green lines) and ${\boldsymbol}{\mu}_{2}(t)$ (indicated by red lines) of group 1 and group 2 with $b_{1} = 0.08, 0.10, 0.12$.[]{data-label="fig:cluster_true_mean_d010"}](True_B3_X "fig:"){width=".3\textwidth"} ![True mean curves ${\boldsymbol}{\mu}_{1}(t)$ (indicated by green lines) and ${\boldsymbol}{\mu}_{2}(t)$ (indicated by red lines) of group 1 and group 2 with $b_{1} = 0.08, 0.10, 0.12$.[]{data-label="fig:cluster_true_mean_d010"}](True_B3_Y "fig:"){width=".3\textwidth"}
[0.45]{}
[0.45]{} ![The value of AICc calculated from one replication in each scenario for the method *SRC*.[]{data-label="fig:AICc"}](AICc2 "fig:"){width="\textwidth"}
[0.45]{} ![The value of AICc calculated from one replication in each scenario for the method *SRC*.[]{data-label="fig:AICc"}](AICc3 "fig:"){width="\textwidth"}
[0.45]{} ![The value of AICc calculated from one replication in each scenario for the method *SRC*.[]{data-label="fig:AICc"}](AICc4 "fig:"){width="\textwidth"}
[0.3]{}
[0.3]{} ![The value of AICc for the method *SRC* for one replication in three scenarios.[]{data-label="fig:AICc"}](AICc_B2 "fig:"){width="\textwidth"}
[0.3]{} ![The value of AICc for the method *SRC* for one replication in three scenarios.[]{data-label="fig:AICc"}](AICc_B3 "fig:"){width="\textwidth"}
[0.32]{}
[0.32]{}
[0.32]{}
[0.32]{}
[0.32]{}
[0.32]{}
[0.32]{}
[0.32]{}
[0.32]{}
[0.32]{}
[0.32]{}
[0.32]{}
[0.32]{}
[0.32]{}
[0.32]{}
[0.32]{}
[0.45]{}
[0.45]{}
![(a)-(c):*solid line* and *dashed line* represent the cluster 1 and cluster 2, respectively. The *green lines* in (a), (b) and (c) are true mean functions. The *red lines*, *brown lines* and *purple lines* in (a)-(c) are respectively corresponding to *SRC*, *SRC-f* and *k-means-f*.[]{data-label="fig:rec02"}](Rec_xymean0){width="50.00000%"}
[0.45]{} ![(a)-(c):*solid line* and *dashed line* represent the cluster 1 and cluster 2, respectively. The *green lines* in (a), (b) and (c) are true mean functions. The *red lines*, *brown lines* and *purple lines* in (a)-(c) are respectively corresponding to *SRC*, *SRC-f* and *k-means-f*.[]{data-label="fig:rec02"}](Rec_xmean0 "fig:"){width="\textwidth"}
[0.45]{} ![(a)-(c):*solid line* and *dashed line* represent the cluster 1 and cluster 2, respectively. The *green lines* in (a), (b) and (c) are true mean functions. The *red lines*, *brown lines* and *purple lines* in (a)-(c) are respectively corresponding to *SRC*, *SRC-f* and *k-means-f*.[]{data-label="fig:rec02"}](Rec_ymean0 "fig:"){width="\textwidth"}
\[ap:rean\]
![Curves of hyoid bone motion for two true groups, where the bold curves in green (upper panel) and in red (lower panel) are the average mean curve for each group[]{data-label="fig:real0"}](Real_C1_X "fig:"){width="30.00000%"} ![Curves of hyoid bone motion for two true groups, where the bold curves in green (upper panel) and in red (lower panel) are the average mean curve for each group[]{data-label="fig:real0"}](Real_C1_Y "fig:"){width="30.00000%"}
![Curves of hyoid bone motion for two true groups, where the bold curves in green (upper panel) and in red (lower panel) are the average mean curve for each group[]{data-label="fig:real0"}](Real_C2_X "fig:"){width=".3\textwidth"} ![Curves of hyoid bone motion for two true groups, where the bold curves in green (upper panel) and in red (lower panel) are the average mean curve for each group[]{data-label="fig:real0"}](Real_C2_Y "fig:"){width=".3\textwidth"}
![Curves of hyoid bone motion for two groups clustered by *SRC*, where the bold curves in green (upper panel) and in red (lower panel) are the average mean curve for each group[]{data-label="fig:real1"}](Mod_C1_X "fig:"){width="30.00000%"} ![Curves of hyoid bone motion for two groups clustered by *SRC*, where the bold curves in green (upper panel) and in red (lower panel) are the average mean curve for each group[]{data-label="fig:real1"}](Mod_C1_Y "fig:"){width="30.00000%"}
![Curves of hyoid bone motion for two groups clustered by *SRC*, where the bold curves in green (upper panel) and in red (lower panel) are the average mean curve for each group[]{data-label="fig:real1"}](Mod_C2_X "fig:"){width=".3\textwidth"} ![Curves of hyoid bone motion for two groups clustered by *SRC*, where the bold curves in green (upper panel) and in red (lower panel) are the average mean curve for each group[]{data-label="fig:real1"}](Mod_C2_Y "fig:"){width=".3\textwidth"}
[^1]: Corresponding author, email: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Antenna selection in Massive MIMO (Multiple Input Multiple Output) communication systems enables reduction of complexity, cost and power while keeping the channel capacity high and retaining the diversity, interference reduction, spatial multiplexity and array gains of Massive MIMO. We investigate the possibility of decentralised antenna selection both to parallelise the optimisation process and put the environment awareness to use. Results of experiments with two different power control rules and varying number of users show that a simple and computationally inexpensive algorithm can be used in real time. The algorithm we propose draws its foundations from self-organisation, environment awareness and randomness.'
author:
- 'Harun Siljak, Irene Macaluso, and Nicola Marchetti[^1][^2]'
title: 'Distributing Complexity: A New Approach to Antenna Selection for Distributed Massive MIMO'
---
Distributed Massive MIMO, antenna selection, optimisation, self-organisation.
Introduction
============
the classical idea of antennas distributed in space [@kerpez1996radio] and the new concept of massive MIMO antenna systems [@marzetta_noncooperative_2010], distributed massive MIMO offers diversity, spatial multiplexing opportunities, interference suppression and redundance [@ozgur2013spatial]. The question of redundance and the number of antennas needed to operate in a certain environment can be answered under certain conditions [@hoydis_massive_2013], and it reinforces the importance of antenna selection. Antenna selection in Massive MIMO system can help with power optimisation, complexity reduction and provide a set of antennas available for other purposes, e.g. nulling [@geraci_operating_2017]. Of course, a trade-off between these benefits exists [@liu_energy_2017] and different objectives and performance measures are used, from spectral efficiency and constructive interference [@amadori_interference-driven_2016] to fairness measures [@makki_genetic_2017]. Other factors are also taken into account, such as the circuit power consumption [@hamdi_resource_2016]. Another aspect investigated is the difference between FDD (frequency division duplex) and TDD (time division duplex) in terms of CSI (channel state information) collection, which is essential for antenna selection [@benmimoune_novel_2017]. These algorithms are most often centralised and based on co-located antenna systems.
A large number of antenna selection algorithms for MIMO (and recently massive MIMO) has been proposed over the last 15 years, one of the first being the removal of antennas highly correlated with other antennas in the selected set [@molisch_capacity_2005]. Several approaches were based on the greedy principle: in [@gharavi-alkhansari_fast_2004], the authors proposed an iterative algorithm that starts from an empty set of selected antennas and in each turn picks the antenna that contributes the most to the capacity of the selected antennas set. Its dual, i.e. the algorithm that starts with all the antennas and removes those that contribute the least iteratively, has been presented in [@gorokhov_receive_2003]. Both algorithms terminate once the desired number of selected antennas is reached. A case for environment-ignoring random selection was made as well, pointing out that for a large number of selected antennas it is comparable to other selection procedures [@lee_energy_2013], which has been observed in practice for planar co-located massive MIMO [@gao_massive_2015]. To exemplify the various approaches used, we note that methods using convex optimisation [@mahboob_transmit_2012], combinatorial optimisation [@michalopoulos_distributed_2008], genetic algorithm [@makki_genetic_2017] have been proposed.
In this letter we propose a novel distributed, local environment-aware antenna selection algorithm based on sum-capacity maximisation. The aim is to achieve sum rates comparable to those achievable through the use of more computationally complex centralised algorithms and allow flexibility and adaptability of the scheme. By distributing the computation over the nodes, we reduce computational complexity and allow the systemic complexity to enhance the performance.
The Antenna Selection Algorithm
===============================
We consider the scenario of downlink (transmit) antenna selection at the distributed massive MIMO base station with $N_{T}$ antennas. In the cell there are $N_{R}$ single antenna users and we aim at maximising the sum-capacity
$$\mathcal{C}=\max_{\mathbf{P},\mathbf{H_{c}}}\log_{2}\det\left(\mathbf{I}+f(\rho,\,N_{TS},\,N_{R})\mathbf{H_{c}}\mathbf{P}\mathbf{H_{c}}^{H}\right)$$
where $\mathbf{I}$ is $N_{TS}\times N_{TS}$ identity matrix, $\mathbf{P}$ is a diagonal $N_{R}\times N_{R}$ matrix describing the power distribution and $\mathbf{H_{c}}$ is the $N_{TS}\times N_{R}$ channel matrix representing a selected subset of antennas from a set of $N_{T}$ antennas ($N_{T}\ge N_{TS}$) represented in the channel matrix $\mathbf{H}$, sized $N_{T}\times N_{R}$ [@gao_massive_2015]. The term $f(\rho,\,N_{TS},\,N_{R})$ represents the transmission power factor from the downlink channel model $$\mathbf{y}=\sqrt{f(\rho,\,N_{TS},\,N_{R})}\mathbf{H_{c}z}+\mathbf{n}$$ with $\mathbf{y}$ being the $N_{R}\times1$ received vector, $\mathbf{z}$ being the $N_{T}\times1$ transmit vector and $\mathbf{n}$ representing the noise vector; $\rho$ is the signal to noise ratio (SNR) at each user.
There are two ways the power can be managed in downlink, and we address them both. One is harvesting the array gain by improving the SNR at the user side by taking $f(\rho,\,N_{TS},\,N_{R})=\rho N_{R}$ (power control A), while the other harvests the array gain as a means of reducing transmit power, achieved through taking $f(\rho,\,N_{TS},\:N_{R})=\rho N_{R}/N_{TS}$ (power control B). In both cases, the increase in the number of users increases the transmit power. These two optimisation problems are fundamentally different. The problem using the power control A is akin to the receiver antenna selection. This problem can be solved using greedy algorithms with a guaranteed (suboptimal) performance bound, simply adding antennas in an initially empty set of selected antennas based on the contribution to the sum rate they bring in. The problem with model B is that it is not submodular (in particular, not monotonic) [@vaze_sub-modularity_2012]. This means that the addition of an antenna in the selected set of antennas can decrease channel capacity, and greedy algorithms cannot provide performance guarantees.
The optimisation problem we are solving is twofold. We are looking for both the subset of the total set of available antennas and for the optimal power distribution over them. Following the practice from [@gao_massive_2015], we initially assume all diagonal elements of $\mathbf{P}$ equal to $1/N_{R}$ (their sum is unity, making the total power equal to $f(\rho,\,N_{TS},\,N_{R})$), perform the antenna selection and then perform the selection of matrix $\mathbf{P}$ using water filling for zero forcing. The choice of zero forcing for precoding was a matter of practicality, the antenna selection algorithm we propose is independent of the channel model or the precoding scheme.
In the following explanation of the algorithm we will use the term neighbourhood for the set of $k$ elements/antenna indices denoting the $k$ nearest neighbours of an antenna ($D_{i}$: the $i$th antenna’s neighbourhood). Neighbourhood of an antenna $A$ is shown in Fig. \[antennas\]. Flag bit $f_{j}$ represents the on/off state our algorithm proposes for the $j$th antenna in an iteration.
Our local algorithm is motivated by a simple model in which every antenna communicates with its neighbourhood to determine whether it should participate in the set of selected antennas $S_{i}$ from the neighbourhood $D_{i}$ or not, i.e. $S_{i}=\{j\in D_{i}|f_{j}=1\}$. Each antenna element node calculates the sum-capacity with power model B for the currently selected set of antennas from its neighbourhood, and then for this set augmented with its antenna.[^3] If the latter is larger than the former, the node sets its flag to one (else, it resets to zero).
Starting from a random selection of antennas, this simple rule in principle organises the antennas either into a stable configuration or an oscillation between two configurations. As these state(s) may be a local but not a global maximum of the achievable sum rates, we introduce a mutation flipping every ($1/p_{M}$)th flag on average. In case of long coherence intervals, the mutation happens often as the algorithm gets a chance to run longer on the same CSI. In our case, it was a rare event as we assumed a short coherence period.[^4]
The algorithm runs on the same CSI for a prescribed number of iterations $N_{i}$. After the last iteration, antennas turn on and off to form the configuration from the iteration that had the highest total capacity over all antennas. Changing the physical state of the antennas after each iteration is inefficient: this is why we work with flag bits. Size of $N_{i}$ depends on the coherence interval.
Algorithm 1 presents the self-organising behaviour we described. We dub the algorithm *local* to emphasise the locality of computation and perspective.
1. [Start from a random seed of $n$ flags “on”, the rest being “off”.]{}
2. [For each antenna $i$, $1\le i\le N_{T}$ perform the following:]{}
1. [Calculate the sum-capacity $\mathcal{C}_{i-}$ of the system with the channel matrix $H_{S_{i}}$ (the antennas in $S_{i}$)]{}
2. [Calculate the sum-capacity $\mathcal{C}_{i+}$ of the system $H_{S_{i}\cup i}$]{}
3. [Compare the two values: if $\mathcal{C}_{i+}>\mathcal{C}_{i-}$, the antenna in question should be selected (i.e. on). Otherwise it should be off. ]{}
3. [Update the flags $f_{i}$ of each of the $N_{T}$ antennas to the result of 2(c) or its opposite, if a random mutation occurs (with a probability of $p_{M}$). Store the current configuration.]{}
4. [Repeat $N_{i}$ times steps 2 and 3 before updating the physical state of the antennas (on and off) to the flag state that resulted in the highest total capacity.]{}
5. [Repeat steps 2-4.]{}
Sum rates: Local vs. Greedy Algorithm
=====================================
The algorithm was tested using raytracing Matlab tool Ilmprop [@galdo2005geometry] on a system composed by 64 antennas randomly distributed in space at the same height and shown in Fig. \[antennas\]. In all computations, CSI was normalised to unit average energy over all antennas, users and subcarriers [@gao_massive_2015].
{width="1\columnwidth"}
\[antennas\]
The two power control scenarios described before are tested and the results are shown in Fig. \[locvsrand\] (16 users case omitted for clarity in the second scenario plot). A clear difference between the results of local antenna selection and random selection of antennas was expected: while such a difference is often not detectable in the case of planar co-located massive MIMO, the distributed case is closer to cylindrical arrays in this sense, antennas being more far apart [@gao_massive_2015]. The curve representing the local antenna selection shows the highest sum rate values obtained for different numbers of selected antennas. It is compared with the sum rates obtained for the average random case of choosing the same numbers of antennas and the results of the previously described greedy algorithm (the one starting with an empty set of antennas [@gharavi-alkhansari_fast_2004]). The sum rates obtained through the proposed local algorithm are on par (even marginally better) as those obtained through the greedy algorithm (very low antenna counts for which our algorithm gives zero rates are not relevant, as they are within $N_{TS}\lesssim N_{R}$ range, not enabling proper beamforming for all users). The proposed local selection mechanism achieves a comparable performance to the greedy selection, while reducing the computational complexity.
{width="1\columnwidth"}
\[locvsrand\]
{width="1\columnwidth"}
\[12comp\]
Computational and Systemic Complexity
=====================================
Computational Complexity
------------------------
The worst case complexity of the proposed algorithm is $\mathcal{O}(N_{T}^{\omega})$, where $\omega$, $2<\omega<3$ is the exponent in the employed matrix multiplication algorithm complexity.[^5].
We first note that $\det(\mathbf{I}+\mathbf{AB})=\det(\mathbf{I}+\mathbf{BA})$ for properly sized identity matrices $\mathbf{I}$ on both sides (Sylvester identity), so we can shift from $N_{T}\times N_{T}$ matrix $\mathbf{H}\mathbf{H}^{H}$ to a smaller $N_{R}\times N_{R}$ matrix $\mathbf{H}^{H}\mathbf{H}$. The total number of matrix multiplications is $2N_{i}$ and they would be conducted using some of the standard algorithms with complexity $\mathcal{O}(N_{R}^{\omega-1}N_{T})$, $2<\omega<3$ [@knight1995fast][^6]. This brings the multiplication complexity of our algorithm to $\mathcal{O}(N_{i}N_{R}^{\omega-1}N_{T})$. In the worst case $N_{R}=N_{T}$, and since $N_{i}=\text{const}$ we obtain the complexity of $\mathcal{O}(N_{T}^{\omega})$. The number of determinant calculations is constant ($\mathcal{O}(N_{i})$), so the overall complexity is $\mathcal{O}(N_{T}^{\omega})$.
Comparing this to the case of the greedy algorithms with complexity $\mathcal{O}(N_{T}^{2}N_{R}^{2})$ [@gorokhov_receive_2003] and $\mathcal{O}(N_{T}N_{R}N_{TS})$ [@gharavi-alkhansari_fast_2004], we see that the complexity is reduced: the effect is best seen for a large number of antennas (massive MIMO) where the constant nature of $N_{i}$ enables efficient scaling. We also note that $N_{R}$ and $N_{T}$ in our algorithm consideration are not the entire set of transmitters and receivers as in the greedy algorithms, but just a $k$-neighbourhood of transmitters and the receivers in their vicinity. Hence, the computational complexity decreases even more in the practical implementation.
The main computational burden of the single iteration of both algorithms is the fact that it is repeated for each of the $c$ OFDM subcarriers. A natural question to ask is whether we need to optimise over the whole set of subcarriers, and if a significantly smaller subset could be selected to represent the channel appropriately. In our study we propose a random selection of a subcarrier subset and argue that 5% of the whole set is enough for practical use, based on the results. This approach gives a good representation of the channel for the algorithm as the procedure is repeated $N_{i}$ times, allowing most of the subcarriers to appear in different iterations and influence the antenna selection. The alternative is selecting a fixed number of the subcarriers with the largest average power over all users. Using any subset of subcarriers could also speed up the greedy algorithm, but our local selection algorithm still has lower complexity. Fig. \[12comp\] represents the gain of sum capacity over random antenna selection for eight users scenario where we compare the results of greedy selection and three local algorithm variants: the original proposal using all 300 subcarriers, one with 15 randomly selected subcarriers and one with 60 strongest subcarriers selected. The random subcarrier selection variant is computationally superior to both alternatives. It is marginally better in terms of sum rates and we can observe from Fig. \[12comp\] that its minimum selection sets are in general smaller than those of the full 300 subcarrier algorithm. The strongest subset-based variant a limited applicability due to a large minimum number of selected antennas.
Another issue with the growing complexity is the collection of CSI. It is unrealistic to expect perfect knowledge of CSI, but as Fig. \[12comp\] shows, it is not necessary: a 30% uncertainty in CSI hs negligible impact on the result of antenna selection.
{width="1\columnwidth"}
\[neighsize\]
Systemic Complexity: Self-Organisation
--------------------------------------
Our algorithm does not start with a predefined number of antennas to be selected, but converges to a subset whose size depends on the size of the neighbourhood. This means that for the comparison with the greedy algorithm shown in Figs. \[locvsrand\] and \[12comp\] we have run the algorithm for varying numbers of neighbours considered. Fig. \[neighsize\] shows the size of the neighbourhood observed and the number of selected antennas in each of those cases for scenarios with 4 and 12 users (the other two scenarios omitted for clarity) in the power control scheme A. The variable parameter is placed on the ordinate axis to align the graph with Fig. \[locvsrand\].
The figure demonstrates that the smallest number of antennas is selected in the case when a single antenna sees most of the other antennas as its neighbours, and vice versa, the largest number of antennas is selected in correspondence to the smallest neighbourhoods. Small neighbourhoods make large effort to support all users, hence turning on most of their antennas, leading to a high total count of selected antennas. In large neighbourhoods suboptimal antennas keep themselves out of the selected subset, seeing the better antennas already in. We also note the characteristic bimodal shape, implying that in large neighbourhoods the algorithm switches (oscillates) between small number of selected antennas and roughly 50% of the total number. This is a consequence of the power control B we use (cf. the location of maxima of the sum rates in Fig. \[locvsrand\](b)) as roughly half of the antennas do not contribute anything new once the other antennas are included in the selected set. In large neighbourhoods, entropy is low as all antennas know how they fare against other antennas, making them more aware of the environment and the rest of the system.
Figs. \[locvsrand\] and \[neighsize\] also show that the minimal number of selected antennas is not always the number of users: the algorithm sometimes adds more antennas for better beamforming.
Conclusions
===========
We have presented a novel local antenna selection method for distributed massive MIMO. Its local nature allows it to be environment-aware and enables distributed computing at every node. While reducing the complexity of matrix operations and distributing it over all antenna nodes, we have additionaly reduced computational complexity by using a very small subset of subcarriers for optimisation, reducing the time cost by 20 times. This reduction resulted in both enabling the real-time application of the algorithm in dynamic environments with short coherence time and retaining sum-rates on par with other antenna selection algorithms.
Relying on self-organisation, this algorithm emphasises the local properties of distributed massive MIMO and supports its modularity, namely the option of cluster separation and distributed control. The distributed control aspect allows user selection aided by our algorithm for users close to a neighbourhood cluster, and also the service consolidation in case of device failures within a cluster. The neighbouring antennas are aware of local faults and organise themselves accordingly in an emergent manner, building up on the inherent systemic complexity of the antenna selection algorithm. The clusters may operate on their own and/or interact with other clusters, depending on the set neighbourhood size.
The local selection algorithm relies on randomness in two ways. Randomly selecting subcarriers to do the optimisation on, it keeps the diversity of the full subcarrier set while reducing computation time. Randomly performing mutations on state transitions, it allows leaving local maxima.
The reduced computational complexity and environment awareness enable a flexible real time application. Being independent from precoding choice, channel model and the form of power control, the algorithm has been shown to perform well in two different variants of transmit antenna selection.
[10]{}
K. J. Kerpez. A radio access system with distributed antennas. , 45(2):265–275, 1996.
T. L. Marzetta. Noncooperative [Cellular]{} [Wireless]{} with [Unlimited]{} [Numbers]{} of [Base]{} [Station]{} [Antennas]{}. , 9(11):3590–3600, November 2010.
A. Ozgur, O. L[é]{}v[ê]{}que, and D. Tse. Spatial degrees of freedom of large distributed mimo systems and wireless ad hoc networks. , 31(2):202–214, 2013.
J. Hoydis, S. ten Brink, and M. Debbah. Massive [MIMO]{} in the [UL]{}/[DL]{} of [Cellular]{} [Networks]{}: [How]{} [Many]{} [Antennas]{} [Do]{} [We]{} [Need]{}? , 31(2):160–171, February 2013.
G. Geraci, A. Garcia-Rodriguez, D. Lopez-Perez, A. Bonfante, L. Galati Giordano, and H. Claussen. Operating [Massive]{} [MIMO]{} in [Unlicensed]{} [Bands]{} for [Enhanced]{} [Coexistence]{} and [Spatial]{} [Reuse]{}. , 35(6):1282–1293, June 2017.
Z. Liu, W. Du, and D. Sun. Energy and [Spectral]{} [Efficiency]{} [Tradeoff]{} for [Massive]{} [MIMO]{} [Systems]{} [With]{} [Transmit]{} [Antenna]{} [Selection]{}. , 66(5):4453–4457, May 2017.
P. V. Amadori and C. Masouros. Interference-[Driven]{} [Antenna]{} [Selection]{} for [Massive]{} [Multiuser]{} [MIMO]{}. , 65(8):5944–5958, August 2016.
B. Makki, A. Ide, T. Svensson, T. Eriksson, and M. S. Alouini. A [Genetic]{} [Algorithm]{}-[Based]{} [Antenna]{} [Selection]{} [Approach]{} for [Large]{}-but-[Finite]{} [MIMO]{} [Networks]{}. , 66(7):6591–6595, July 2017.
R. Hamdi, E. Driouch, and W. Ajib. Resource [Allocation]{} in [Downlink]{} [Large]{}-[Scale]{} [MIMO]{} [Systems]{}. , 4:8303–8316, 2016.
M. Benmimoune, E. Driouch, W. Ajib, and D. Massicotte. Novel transmit antenna selection strategy for massive [MIMO]{} downlink channel. , 23(8):2473–2484, November 2017.
A. F. Molisch, M. Z. Win, Yang-Seok Choi, and J. H. Winters. Capacity of [MIMO]{} systems with antenna selection. , 4(4):1759–1772, July 2005.
M. Gharavi-Alkhansari and A. B. Gershman. Fast antenna subset selection in [MIMO]{} systems. , 52(2):339–347, February 2004.
A. Gorokhov, D. A. Gore, and A. J. Paulraj. Receive antenna selection for [MIMO]{} spatial multiplexing: theory and algorithms. , 51(11):2796–2807, November 2003.
B. M. Lee, J. Choi, J. Bang, and B. C. Kang. An energy efficient antenna selection for large scale green [MIMO]{} systems. In [*2013 [IEEE]{} [International]{} [Symposium]{} on [Circuits]{} and [Systems]{} ([ISCAS]{}2013)*]{}, pages 950–953, May 2013.
X. Gao, O. Edfors, F. Tufvesson, and E. G. Larsson. Massive [MIMO]{} in [Real]{} [Propagation]{} [Environments]{}: [Do]{} [All]{} [Antennas]{} [Contribute]{} [Equally]{}? , 63(11):3917–3928, 2015.
S. Mahboob, R. Ruby, and V. C. M. Leung. Transmit [Antenna]{} [Selection]{} for [Downlink]{} [Transmission]{} in a [Massively]{} [Distributed]{} [Antenna]{} [System]{} [Using]{} [Convex]{} [Optimization]{}. In [*2012 [Seventh]{} [International]{} [Conference]{} on [Broadband]{}, [Wireless]{} [Computing]{}, [Communication]{} and [Applications]{}*]{}, pages 228–233, November 2012.
D. S. Michalopoulos, G. K. Karagiannidis, T. A. Tsiftsis, and R. K. Mallik. Distributed [Transmit]{} [Antenna]{} [Selection]{} ([DTAS]{}) [Under]{} [Performance]{} or [Energy]{} [Consumption]{} [Constraints]{}. , 7(4):1168–1173, April 2008.
R. Vaze and H. Ganapathy. Sub-[Modularity]{} and [Antenna]{} [Selection]{} in [MIMO]{} [Systems]{}. , 16(9):1446–1449, September 2012.
G Del Galdo, M Haardt, and C Schneider. Geometry-based channel modelling of mimo channels in comparison with channel sounder measurements. , 2(BC):117–126, 2005.
P. A. Knight. Fast rectangular matrix multiplication and qr decomposition. , 221:69–81, 1995.
[^1]: The authors are with the CONNECT Centre, Trinity College Dublin, Ireland (emails: {siljakh, macalusi, marchetn}@tcd.ie)
[^2]: This publication has emanated from research supported in part by a research grant from Science Foundation Ireland (SFI) and is co-funded under the European Regional Development Fund under Grant Number 13/RC/2077. The project has received funding from the European Unions Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 713567.
[^3]: The calculation of local sum capacity using power control B is important for the algorithm as the case of control A would allow every antenna to join and improve the selected set by just adding more power to it. Control B allows only the antennas improving the information content to join in.
[^4]: While the mutation bears resemblance to the genetic algorithm approach in [@makki_genetic_2017], the core process is different. In our algorithm, we converge to the best antenna selection by a simple but *directed* search strategy. In addition to that, we use the *local* measure of the capacity as the selection criterion.
[^5]: It is not always feasible to use the multiplication algorithm with the lowest complexity due to large constant factors making it hard to do even a single iteration: the current lowest complexity algorithms ($\omega=2.373$) cannot be implemented in technology at all.
[^6]: If two $n\times n$ square matrices are multiplied with complexity$\mathcal{O}(n^{\omega}),$ rectangular matrices $a\times b$ and $b\times c$ are multiplied with complexity $\mathcal{O}(n_{1}^{\omega-2}n_{2}n_{3})$ where $n_{1}=\min(a,b,c)$ and $n_{1},\ n_{2}$ are the other two dimensions.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Buffer overflow detection and mitigation for C programs has been an important concern for a long time. This paper defines a string buffer overflow analysis for C. The key ideas of our formulation are $(a)$ separating buffers from the pointers that point to them, $(b)$ modelling buffers in terms of sizes and sets of positions of null characters, and $(c)$ defining stateless functions to compute the sets of null positions and mappings between buffers and pointers.
This exercise has been carried out to test the feasibility of describing such an analysis in terms of lattice valued functions and relations to facilitate automatic construction of an analyser without the user having to write C/C++/Java code. This is facilitated by avoiding stateful formulations because they combine effects through side effects in states raising a natural requirement of C/C++/Java code to be written to describe them. Given the above motivation, the focus of this paper is not to build good static approximations for buffer overflow analysis but to show how given static approximations could be formalized in terms of stateless formulations so that they become amenable to automatic construction of analysers.
author:
- '[Uday P. Khedker]{}'
bibliography:
- 'bof.bib'
title: '**Buffer Overflow Analysis for C**'
---
Introduction
============
Low level modelling of strings in C and associated unchecked operations potentially lead to the possibility of buffer overflows. Given the possibility of a potentially fraudulent use of these loop holes in C programs, detection and mitigation of buffer overflows is critical [@citeulike:6345464; @buf.study; @Lhee:2003:BOF:781669.781672; @Shahriar:2010:MBO:2441114.2441116; @Zitser:2004:TSA:1029894.1029911].
This paper proposes an analysis to discover buffer overflows. The key ideas of our formulation are $(a)$ separating buffers from the pointers that point to them, $(b)$ modelling buffers in terms of sizes and sets of positions of null characters, and $(c)$ defining stateless functions to compute the sets of null positions and mappings between buffers and pointer. The first idea is not new; the novelty of our work lies in modelling the computations of null position sets and an insistence on stateless formulations. As is customary, we present our formulation in an intraprocedural setting. It should be easy to lift it to an interprocedural setting using standard techniques of interprocedural data flow analysis such as the method of value contexts [@Khedker.UP.Karkare.B:2008:Efficiency-Precision-Simplicity; @Khedker.UP.Sanyal.A.Karkare.B:2009:Data-Flow-Analysis; @Padhye:2013:IDF:2487568.2487569].
Our goal is not to device the best possible static approximations for buffer overflow analysis but to show how given static approximations could be formalized to devise a mathematical formulation which can be transcribed into a declarative specification of the analysis so that it becomes amenable to automatic construction of analysers.
The rest of the paper is organized as follows: Section \[sec:requirements\] describes the requirements of buffer overflow analysis by defining the core statements for analysis, the soundness criterion, and our assumptions. Section \[sec:formulation\] describes our modelling and defines the analysis in terms of lattices and lattice valued functions and relations. For simplicity of exposition, it assumes that a pointer points to a single buffer at a program point. Section \[sec:exmp\] shows the analysis of our running example. Section \[sec:extensions\] shows how the model can be extended to allow a pointer to point to multiple buffers at a program point. Section \[sec:related\] briefly describes the related work to highlight the trends while Section \[sec:conclusions\] concludes the paper. Appendix \[app:stmt.modelling\] describes how other statements can be modelled in terms of core statements.
Requirements of Buffer Overflow Analysis {#sec:requirements}
========================================
Our formulation is guided by the following requirements and assumptions.
Program Model
-------------
We assume the following model of programs.
- A buffer is an array of `char` or pointers to `char` storing C-style strings. It could contain multiple strings and hence multiple null characters (`'\0'`). A pointer may point to any location within a buffer.
- The list of core buffer manipulation statements to be considered for this analysis is as follows.
- Buffer creation: and where $i$ is a compile time constant, and .
- Buffer assignment:
- and where $i$ is a compile time constant.
- which copies a $m$ character long block of memory pointed to by $y$ to the memory pointed to by $x$.
- Buffer modification:
- Direct modification. and where $i$ is a compile time constant and c is a non-null character.
- Modification through string functions. and their length limited versions ( and ).
- Buffer reading: Any statement using $x$, , or as an rvalue or calls to .
- A program is viewed as a convention control flow graph with each node representing a single statement.
Appendix \[app:stmt.modelling\] explains how other statements are modelled in terms of these statements.
Soundness Criterion, Required Approximations, and Assumptions {#sec:assumptions}
-------------------------------------------------------------
We assume the following soundness criterion: no buffer overflow should go undetected; false positives about buffer overflow can be tolerated.
In order the ensure soundness, we introduce the following approximation: A buffer has a single set of null positions associated with it. If the sets of a buffer differ along different execution paths reaching a program point, we create an approximate buffer such that regardless of the execution path, every string contained in the original buffer is a substring of some string present in the approximate buffer. This approximation may cause some imprecision in that our analysis may consider longer strings than are actually present in the buffer leading to false positives.
This approximation is implemented in the following manner:
- At a given program point, a buffer could have different null position sets along different control flow paths reaching the program point. Hence, at the join points in the program, the null position sets of a buffer reaching along different control flow paths are intersected.
- We assume that the memory allocated using does not contain a null character.
- We assume that the values of integer variables appearing in , , or are known at compile time or a range analysis has been performed. If range information is available, we choose the low limit of the range of $i$ for and the the high limit of the range of $i$ for and . If $i$ is not a compile time constant and its range is not known, we assume its value to be $\infty$.
Formulating Buffer Overflow Analysis {#sec:formulation}
====================================
In this section, we model buffers and the relations of pointers holding the addresses of buffers. This is then followed by formulating data flow equations that compute them. As is customary, we present our formulation in an intraprocedural setting. It should be easy to lift it to an interprocedural setting using standard techniques of interprocedural data flow analysis such as the method of value contexts [@Khedker.UP.Karkare.B:2008:Efficiency-Precision-Simplicity; @Khedker.UP.Sanyal.A.Karkare.B:2009:Data-Flow-Analysis; @Padhye:2013:IDF:2487568.2487569].
Assumptions for Simplifying the Formulation
-------------------------------------------
For simplicity of exposition, we make the following assumptions for the purpose of presenting the formulation. Section \[sec:extensions\] extends the formulation by relaxing these assumptions.
- At a given program point, in general, a pointer could point to different buffers along different control flow paths reaching the program point. For simplicity of modelling, we assume that a pointer points to a single buffer.
- Under the assumption that range information is available, we ignore all back edges in the program and view it as a directed acyclic graph. This allows us to handle the situation where pointers to a buffer are advanced in a loop and hence at the loop entry, such a pointer points to two different positions in a buffer. If we do not ignore back edges, such common case usage of pointers will lead to over-approximation of null position sets leading to proliferation of false negatives.
Loops basically contribute to advancement of pointers or increments of indices. Range information captures these effect. Hence we can reduce false negatives by restricting the merge points to those resulting from forward edges in the control flow graph of the program.
Note that a single pointee assumption does not preclude the possibility that a buffer may be pointed to by multiple pointers. Such a situation is easily handled by our formulation. Indeed, our running example of Figure \[fig:exmp.1\] has pointers $x$, $y$, and $z$ all pointing to the same buffer $b_1$.
Modelling Buffers and Buffer Pointer Relations {#sec:model}
----------------------------------------------
We identify a buffer by its allocation site name and record its size and the set of positions in the buffer that hold the null character $'\backslash 0'$.
Let be the set of nodes in the control flow graph of the program being analysed, be the set of positive integers (including 0) and be the set of pointers to buffers. Using these sets, we define as a set of buffer identities, as a set of buffer sizes (or offsets into buffers), and as the set of positions of null characters in a buffer. Buffers are described using the following functions.
- maps a buffer identity to buffer information.
A buffer is identified by the program point of its creation [e.g.]{} $b_n$ denotes the buffer created in statement . Each buffer has exactly one size and exactly one set of null positions at any given program point. also contains a fictitious “undefined” buffer .
Buffer size $\infty$ may also be associated with a valid buffer indicating that at that program point, the buffer size is not a compile time constant. This could be either because the buffer has not been created along some execution path, or the buffer has been created using a size that is not known at compile time, or the memory has been freed along some execution path.[^1]
If the null position set is empty, it indicates that a read will cause a buffer overflow. The presence of $\infty$ in a null position set indicates that a write has already caused buffer overflow.
- relates a buffer pointer to its pointee buffer and an offset into the buffer (because a pointer may point to some position in the middle of a buffer).
[@c@|@c@]{}
[@c@]{}
/* Initial situation:
w points to buffer b0
x points to buffer b1*/
z = x+6;
/* After this, z points
to offset 6 in b1 */
y = x+4;
/* After this, y points
to offset 4 in b1 */
/* Call 1. No overflow */
strcat(z,y);
/* Call 2. b1 overflows */
strcat(z,y);
/* Call 3. No overflow */
strcpy(y,w);
\
\
(a) Program
&
[@c@]{}
(0,-2)(110,38)
\
(b) Memory before call 1. “??” indicates garbage value.\
\
\
(0,-2)(110,38)
\
(c) Memory before call 2. “??” indicates garbage value.\
\
\
\
We assume that both these functions are total functions. This simplifies our formulations.
We use the following notational conventions:
- ranges over the set $A=2^{\text{\sf\em buf\/}\xspace}$ (and thus represent buffer mappings).
- ranges over the set $B=2^{\text{\sf\em bpt\/}\xspace}$ (and thus represent pointer to buffer mappings).
- ranges over buffer identities in ${\text{{\sffamily N}$_b^\infty$}\xspace}$.
- range over the set of pointers ${\text{\sffamily P}\xspace}$.
- range over the set .
- range over some set (the types of their elements are evident from the context).
For a buffer map ${\text{$\alpha$}\xspace}$, the extractor functions [*size$({\text{$\alpha$}\xspace},b)$*]{} and compute the size and the null position set of a given buffer $b$ in . $$\begin{aligned}
(b,k,X) \in {\text{$\alpha$}\xspace}\Rightarrow {\text{\sf\em size\text{$({\text{$\alpha$}\xspace}, b)$}\/}\xspace} =k \wedge {\text{\sf\em nps\/}\xspace}({\text{$\alpha$}\xspace}, b) =X
\label{eq:fun.size}\end{aligned}$$ In other words, .
For a buffer pointer map ${\text{$\beta$}\xspace}$, the extractor functions and return the pointee buffer and the start position of $x$ in a given buffer $b$. $$\begin{aligned}
(x,b,i) \in {\text{$\beta$}\xspace}\Rightarrow {\text{\sf\em pt\/}\xspace}({\text{$\beta$}\xspace}, x) = b \wedge {\text{\sf\em start\/}\xspace}({\text{$\beta$}\xspace}, x, b) =i
\label{eq:fun.start}\end{aligned}$$ We compute the positions of null characters and start positions using compile time saturated addition of integers that limits the result to the given (compile time) constant $k$ as defined below. $$\begin{aligned}
\forall\, i, j \in {\text{\sffamily Z$_b$}\xspace}\!:\;\;
i \oplus_k j &=
\begin{cases}
i + j & i + j \leq k \wedge k \neq \infty
\\
\infty & \text{otherwise}
\end{cases}\end{aligned}$$
### Running Example {#running-example .unnumbered}
Figure \[fig:exmp.1\] illustrates our modelling. The details of these computations are explained in Section \[sec:exmp\].
Data Flow Equations
-------------------
This section describes the lattices of data flow values, the data flow equations, and the flow functions used in the data flow equations.
### Lattices
Our analysis computes subsets of and flow sensitively at each node . The lattices of these values are and where and . The overall lattice $L$ is the product lattice . Its meet operation is defined in terms of the meet operations and ${\text{$\sqcap_B\,$}\xspace}$ of the constituent lattices $A$ and $B$. $$\begin{aligned}
\forall \, {\text{$\alpha$}\xspace},{\text{$\alpha$}\xspace}' \in A, \;
\forall \, {\text{$\beta$}\xspace},{\text{$\beta$}\xspace}' \in B\!: \;\;
({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace}) \,{\text{$\sqcap_{AB}\,$}\xspace}({\text{$\alpha$}\xspace}',{\text{$\beta$}\xspace}')
&= ({\text{$\alpha$}\xspace}\,{\text{$\sqcap_A\,$}\xspace}{\text{$\alpha$}\xspace}', {\text{$\beta$}\xspace}\,{\text{$\sqcap_B\,$}\xspace}{\text{$\beta$}\xspace}') \end{aligned}$$
We first define the meet for merging the buffers (equations \[eq:meet.A\] to \[eq:meet.NP\]) and then define the meet for merging the pointer buffer mappings (equation \[eq:meet.B\]).
### Merging Buffers {#merging-buffers .unnumbered}
Recall that lattice where is defined in terms of lattices and . Hence is defined in terms of and . $$\begin{aligned}
\forall \, {\text{$\alpha$}\xspace},{\text{$\alpha$}\xspace}' \in A\!: \;\;
{\text{$\alpha$}\xspace}\,{\text{$\sqcap_A\,$}\xspace}{\text{$\alpha$}\xspace}' &= \left\{ \left(b, \;
\left(
\left(i \,{\text{$\sqcap_S\,$}\xspace}j , \; X \,{\text{$\sqcap_{Z}\,$}\xspace}Y \right)
\right)
\right)
\; \middle| \; \left(b,i, X\right) \in {\text{$\alpha$}\xspace}, \left(b,j, Y\right) \in {\text{$\alpha$}\xspace}'
\right\}
\label{eq:meet.A}\end{aligned}$$ The definitions of and warrant some explanation.
- The meet enforces a buffer to have the same size along all paths. In case of inconsistent sizes, the buffer size is recorded as $\infty$ which indicates that the buffer size is statically undefined (see Section \[sec:model\] for the semantics of $\infty$). $$\begin{aligned}
\forall \, i,j \in {\text{\sffamily S$_b$}\xspace}\!: \;\; i\, {\text{$\sqcap_S\,$}\xspace}j
&=
\begin{cases}
i & i = j
\\
\infty & \text{otherwise}
\end{cases}
\label{eq:meet.SZ}
\end{aligned}$$ The definition of indicates that $\infty$ is the $\bot$ element of . However, it does not contain a $\top$ and hence is only a meet semilattice.
- Given the soundness requirement described in Section \[sec:assumptions\], the should approximate the null position sets of buffers by intersecting them. This effectively lengthens the strings in the buffer guaranteeing the soundness of this approximation. However, this idea has a minor flaw: Assume that for a given buffer, we get the set along one path and along the other path. We would like to assume that we have a null character at position 1 alone and intersection indeed achieves this. However, we cannot ignore the buffer overflow that has occurred along the second control flow path. An intersection would remove $\infty$ from the resulting set because it appears in only one set.
We overcome this problem by observing that a set of null positions $X$ is a subset of . Hence we define two function ${\text{\sf\em fin\/}\xspace}(X)$ and ${\text{\sf\em inf\/}\xspace}(X)$ that partition $X$ into subsets of and : ${\text{\sf\em fin\/}\xspace}(X)$ computes the maximal subset of $X$ containing finite numbers, and ${\text{\sf\em inf\/}\xspace}(X)$ computes the minimal subset of $X$ containing $\infty$.
Consider the following sets , , , and . Then,
- , and .
- , and .
- , and .
- .
This distinction allows us to use intersection for ${\text{\sf\em fin\/}\xspace}(X)$ and union for ${\text{\sf\em inf\/}\xspace}(X)$ in the definition of . $$\begin{aligned}
\forall \, X,Y \subseteq {\text{\sffamily Z$_b$}\xspace}\!: \;\;
X \,{\text{$\sqcap_{Z}\,$}\xspace}Y
&=
\left({\text{\sf\em fin\/}\xspace}(X) \cap {\text{\sf\em fin\/}\xspace}(Y)\right)\cup {\text{\sf\em inf\/}\xspace}(X) \cup {\text{\sf\em inf\/}\xspace}(Y)
\label{eq:meet.NP}
\end{aligned}$$
### Merging Pointer Buffer Mappings {#merging-pointer-buffer-mappings .unnumbered}
Buffer mappings are merged using . Since where , is defined in terms of and : $$\begin{aligned}
\forall \, {\text{$\beta$}\xspace},{\text{$\beta$}\xspace}' \in B\!: \;\;
{\text{$\beta$}\xspace}\,{\text{$\sqcap_B\,$}\xspace}{\text{$\beta$}\xspace}'
& =
\left\{
\left( x, b \,{\text{$\sqcap_N\,$}\xspace}b', k \,{\text{$\sqcap_S\,$}\xspace}k' \right)
\;\middle| \;
\left( x, b, k \right) \in {\text{$\beta$}\xspace},
\left( x, b', k' \right) \in {\text{$\beta$}\xspace}'
\right\}
\label{eq:meet.B}\end{aligned}$$ The meet operation is defined in equation (\[eq:meet.SZ\]) earlier; is similarly defined below. $$\begin{aligned}
\forall \, b,b' \in {\text{{\sffamily N}$_b^\infty$}\xspace}\!: \;\;
b \,{\text{$\sqcap_N\,$}\xspace}b'
&=
\begin{cases}
b & b = b'
\\
b_\infty & \text{otherwise}
\end{cases}
\end{aligned}$$ imposes the restriction that a pointer points to the same buffer along all paths reaching a program point; ensures that the offsets are also identical.
Since and are meet semilattice, their product is also a meet semilattice. It does not have a $\top$ element and its $\bot$ element is .
$$\begin{array}{|l|c|c|l@{}|}
\hline
\multicolumn{4}{|l|}{ \begin{tabular}{@{}l@{}}
For brevity, we use the following short forms
\\
$\bullet$ \; Destination buffer: $D_n$, New size: $K_n({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace})$, New relevant null position set: $R_n({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace})$
\\
$\bullet$ \; ${\text{\sf\em pt\/}\xspace}({\text{$\beta$}\xspace},z) = pt_z, {\text{\sf\em size\text{$($}\/}\xspace}{\text{$\alpha$}\xspace},{\text{\sf\em pt\/}\xspace}_z) = {\text{\sf\em size\text{$$}\/}\xspace}_z, {\text{\sf\em nps\/}\xspace}({\text{\sf\em pt\/}\xspace}_z) = {\text{\sf\em nps\/}\xspace}_z$,
$ {\text{\sf\em start\/}\xspace}({\text{$\beta$}\xspace},z,{\text{\sf\em pt\/}\xspace}_z) = {\text{\sf\em start\/}\xspace}_z, {\text{\sf\em end\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},z,{\text{\sf\em pt\/}\xspace}_z) = {\text{\sf\em end\/}\xspace}_z $
\\
\end{tabular}
}
\\ \hline
\text{Statement } n
& D_n
& K_n({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace})
& R_n({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace})
\\ \hline\hline
x = {\text{\sf\em malloc\/}\xspace}(m)
& b_n
& m
& \emptyset
\\\hline
x = {\text{\sf\em calloc\/}\xspace}(m)
& b_n
& m
& \{ i \mid i \leq m \}
\\\hline
{\text{\sf\em free\/}\xspace}(x)
& {\text{\sf\em pt\/}\xspace}_x
& 0
& {\text{\sf\em nps\/}\xspace}_x
\\\hline
x[i] = '\backslash 0'
& {\text{\sf\em pt\/}\xspace}_x
& {\text{\sf\em size\text{$$}\/}\xspace}_x
& {\text{\sf\em nps\/}\xspace}_x \cup \{ i \oplus_{{\text{\sf\em size\text{$$}\/}\xspace}_x} {\text{\sf\em start\/}\xspace}_x \}
\\\hline
x[i] = c
& {\text{\sf\em pt\/}\xspace}_x
& {\text{\sf\em size\text{$$}\/}\xspace}_x
& {\text{\sf\em nps\/}\xspace}_x - \{ i \oplus_{{\text{\sf\em size\text{$$}\/}\xspace}_x} {\text{\sf\em start\/}\xspace}_x \}
\\\hline
\multirow{7}{*}{\begin{tabular}{@{}l@{}}
{\text{\sf\em strcpy\/}\xspace}$(x,y)$
\\
{\text{\sf\em strcat\/}\xspace}$(x,y)$
\\
{\text{\sf\em strncpy\/}\xspace}$(x,y,m)$
\\
{\text{\sf\em strncat\/}\xspace}$(x,y,m)$
\end{tabular}
} &
\multirow{7}{*}{$
{\text{\sf\em pt\/}\xspace}_x
$}
&
\multirow{7}{*}{$
{\text{\sf\em size\text{$$}\/}\xspace}_x
$}
&
\begin{array}{@{}ll@{}}
\text{\em let}
\\
& \begin{array}{@{}r@{\ }l}
{\text{\sf\em length\/}\xspace}_y & = \begin{cases}
{\text{\sf\em end\/}\xspace}_y - {\text{\sf\em start\/}\xspace}{}_y
& {\text{\sf\em strcpy\/}\xspace}\text{ or } {\text{\sf\em strcat\/}\xspace}\\
\min(m\!+\!1, {\text{\sf\em end\/}\xspace}_y - {\text{\sf\em start\/}\xspace}{}_y)
& {\text{\sf\em strncpy\/}\xspace}\text{ or } {\text{\sf\em strncat\/}\xspace}\end{cases}
\\
\rule{0em}{2.em} {\text{\sf\em copy\_pos\/}\xspace}_x & = \begin{cases}
{\text{\sf\em start\/}\xspace}{}_x \hspace*{30mm}
& {\text{\sf\em strcpy\/}\xspace}\text{ or } {\text{\sf\em strncpy\/}\xspace}\\
{\text{\sf\em end\/}\xspace}_x
& {\text{\sf\em strcat\/}\xspace}\text{ or } {\text{\sf\em strncat\/}\xspace}\end{cases}
\\
{\text{\sf\em np\_src\/}\xspace}_{xy} & = {\text{\sf\em length\/}\xspace}_y \oplus_{{\text{\sf\em size\text{$$}\/}\xspace}_x} {\text{\sf\em copy\_pos\/}\xspace}_x
\\
{\text{\sf\em nps\_before\/}\xspace}_{xy} & = {\text{\sf\em rnps\/}\xspace}({\text{\sf\em nps\/}\xspace}_x,{\text{\sf\em size\text{$$}\/}\xspace}_x, 0, {\text{\sf\em start\/}\xspace}_x, < )
\\
{\text{\sf\em np\_at\/}\xspace}_{xy} &= \{{\text{\sf\em np\_src\/}\xspace}_{xy}\}
\\
{\text{\sf\em nps\_after\/}\xspace}_{xy} &= {\text{\sf\em rnps\/}\xspace}({\text{\sf\em nps\/}\xspace}_x,{\text{\sf\em size\text{$$}\/}\xspace}_x, 0, {\text{\sf\em np\_src\/}\xspace}_{xy}, \geq )
\end{array}
\\
\text{\em in}
\\
& {\text{\sf\em nps\_before\/}\xspace}_{xy} \; \cup \; {\text{\sf\em np\_at\/}\xspace}_{xy} \;\cup \; {\text{\sf\em nps\_after\/}\xspace}_{xy}
\end{array}
\\\hline
{\text{\sf\em memcpy\/}\xspace}(x,y,m)
& {\text{\sf\em pt\/}\xspace}_x
& {\text{\sf\em size\text{$$}\/}\xspace}_x
&
\begin{array}{@{}ll@{}}
\text{\em let}
\\
&
\begin{array}{r@{\ }l}
{\text{\sf\em shift\_dist\/}\xspace}_{xy} & = {\text{\sf\em start\/}\xspace}_x-{\text{\sf\em start\/}\xspace}_y
\\
{\text{\sf\em nps\_src\/}\xspace}_{xy}& = {\text{\sf\em rnps\/}\xspace}({\text{\sf\em nps\/}\xspace}_y,{\text{\sf\em size\text{$$}\/}\xspace}_x,{\text{\sf\em shift\_dist\/}\xspace}_{xy}, {\text{\sf\em start\/}\xspace}_y, \geq ) \cap
\\
& \hspace*{4mm}{\text{\sf\em rnps\/}\xspace}({\text{\sf\em nps\/}\xspace}_y,{\text{\sf\em size\text{$$}\/}\xspace}_x,{\text{\sf\em shift\_dist\/}\xspace}_{xy}, {\text{\sf\em start\/}\xspace}_y + m, \leq )
\\
{\text{\sf\em nps\_before\/}\xspace}_{xy} & = {\text{\sf\em rnps\/}\xspace}({\text{\sf\em nps\/}\xspace}_x,{\text{\sf\em size\text{$$}\/}\xspace}_x, 0, {\text{\sf\em start\/}\xspace}_x, < )
\\
{\text{\sf\em nps\_after\/}\xspace}_{xy} &= {\text{\sf\em rnps\/}\xspace}({\text{\sf\em nps\/}\xspace}_x,{\text{\sf\em size\text{$$}\/}\xspace}_x, 0, {\text{\sf\em start\/}\xspace}_x + m, \geq )
\\
{\text{\sf\em o\_flow\/}\xspace}_{xy} & = \{m\oplus_{{\text{\sf\em size\text{$$}\/}\xspace}_x}{\text{\sf\em start\/}\xspace}_x \} - \{m+{\text{\sf\em start\/}\xspace}_x\} \; \cup
\\
& \hspace*{4mm} \{m\oplus_{{\text{\sf\em size\text{$$}\/}\xspace}_y}{\text{\sf\em start\/}\xspace}_y \} - \{m+{\text{\sf\em start\/}\xspace}_y\}
\end{array}
\\
in
\\
& \begin{array}{@{}l@{}}
{\text{\sf\em nps\_before\/}\xspace}_{xy} \; \cup {\text{\sf\em nps\_after\/}\xspace}_{xy} \; \cup \; {\text{\sf\em nps\_src\/}\xspace}_{xy}
\; \cup \; {\text{\sf\em o\_flow\/}\xspace}_{xy}
\end{array}
\end{array}
\\ \hline
\end{array}$$
### Data Flow Equations
Recall that our data flow values are pairs that record the buffer mappings and pointer to buffer mappings at each program point.
The data flow equations are as follows: $$\begin{aligned}
{\text{\sf\em bIn\/}\xspace}_n & =
\begin{cases}
\left(\left\{ \left(b_n, \infty, \emptyset\right) \;\middle|\; n \in {\text{\sffamily N}\xspace}\right\},
\left\{ \left(x, b_\infty, \infty \right) \;\middle|\; x \in {\text{\sffamily P}\xspace}\right\}
\right)
& n = {\text{{\sf\em StartNode}$_{}$}\xspace}
\\
\displaystyle{\bigsqcap_{p \in pred(n)}}{\hspace*{-4mm}\raisebox{-2mm}{$\scriptstyle AB$}}
{\text{\sf\em bOut\/}\xspace}_p
& \text{otherwise}
\end{cases}
\\
{\text{\sf\em bOut\/}\xspace}_n & = {\text{\sf\em update\_maps\/}\xspace}({\text{\sf\em bIn\/}\xspace}_n,n)\end{aligned}$$ At the start of the program, all buffers are undefined (i.e. their sizes are $\infty$ and their null pointer sets are $\emptyset$) and all pointers point to undefined buffer . We need this boundary information instead of empty mappings $\emptyset$ because we need to maintain the invariant that our mappings are total functions.
Flow function updates the buffer mappings and pointer to buffer mappings. $$\begin{aligned}
{\text{\sf\em update\_maps\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},n)
& =
(
{\text{\sf\em update\_buf\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},n),
{\text{\sf\em update\_bpt\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},n)
)\end{aligned}$$ For statement $n$, the destination buffer $D_n$ in buffer mapping is updated using auxiliary extractor functions $K_n({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace})$ and $R_n({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace})$; the three terms are defined for relevant statements Figure \[fig:flow.fun\] and are explained Section \[sec:flow.func\]. $$\begin{aligned}
{\text{\sf\em update\_buf\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},n)
&= \begin{cases}
{\text{$\alpha$}\xspace}\left[D_n \mapsto \left(
K_n({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace}),
R_n({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace})\right)\right]
& \renewcommand{\arraystretch}{.9}
\begin{array}{@{}l@{}}
\text{Statement $n$ involves a string }
\\
\text{function listed in Figure \ref{fig:flow.fun}}
\end{array}
\\
{\text{$\alpha$}\xspace}& \text{Otherwise}
\end{cases}\end{aligned}$$ The pointer to buffer maps are updated only for the statements which involve a pointer assignment. $$\begin{aligned}
{\text{\sf\em update\_bpt\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},n)
&=
\begin{cases}
{\text{$\beta$}\xspace}[x \mapsto \{ (b_n,0) \}]
& \renewcommand{\arraystretch}{.9}
\begin{array}{@{}l@{}}
\text{Statement $n$ is } x = {\text{\sf\em malloc\/}\xspace}(k) \text{ or } x = {\text{\sf\em calloc\/}\xspace}(k)
\\
\text{or } x = {\text{\sf\em string\/}\xspace}(k)
\end{array}
\\
{\text{$\beta$}\xspace}[x \mapsto {\text{$\beta$}\xspace}(y)]
& \text{Statement $n$ is } x = y
\\
{\text{$\beta$}\xspace}[x \mapsto {\text{\sf\em shift\_offset\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},y,i)]
& \text{Statement $n$ is } x = y + i
\\
{\text{$\beta$}\xspace}& \text{Otherwise}
\end{cases}\end{aligned}$$ where is used to compute the offset of $y$ in the buffer by shifting it by $i$. $$\begin{aligned}
{\text{\sf\em shift\_offset\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},y,i) &=
\{ (b, j \oplus_k i) \mid b = {\text{\sf\em pt\/}\xspace}({\text{$\beta$}\xspace},y), j = {\text{\sf\em start\/}\xspace}({\text{$\beta$}\xspace},y,b), k = {\text{\sf\em size\text{$({\text{$\alpha$}\xspace},b)$}\/}\xspace})\} \end{aligned}$$
### Flow Functions for Computing the Set of Null Positions {#sec:flow.func}
The heart of this analysis lies in computing the set of positions of null characters. We use the following functions to achieve this:
1. Functions , , , and introduced in equations (\[eq:fun.size\]) and (\[eq:fun.start\]).
2. Function which performs arithmetic on null positions to compute the relevant null positions. $$\begin{aligned}
{\text{\sf\em rnps\/}\xspace}(X, {\text{\sf\em sat\/}\xspace}, {\text{\sf\em shift\/}\xspace}, {\text{\sf\em pivot\/}\xspace},{\text{\pscirclebox[framesep=.1,linewidth=.15]{\scriptsize\bf?}}})
&=
\{ i \oplus_{\text{\sf\em sat\/}\xspace}{\text{\sf\em shift\/}\xspace}\mid i \in X, i \;{\text{\pscirclebox[framesep=.1,linewidth=.15]{\scriptsize\bf?}}}\,{\text{\sf\em pivot\/}\xspace}\}
\end{aligned}$$ It is defined in terms of
- a set of null positions ($X$),
- a saturation limit (),
- an offset () in the destination buffer using which the null positions of the source buffer may have to be shifted,
- null positions in the source buffer described in terms of
- a pivot () around which the null positions are to be examined, and
- a relational operator () used for comparison with the pivot.
3. Function which computes the end of a string in a buffer using by finding out the null positions that lie beyond the of the string and then computing their greatest lower bound () under $\leq$ as the partial order. This allows us to restrict the set in a buffer to the first null position after the start position of a given string. $$\begin{aligned}
{\text{\sf\em end\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},x, b)
&=
{\text{\sf\em glb\_le\/}\xspace}({\text{\sf\em rnps\/}\xspace}({\text{\sf\em nps\/}\xspace}(b),{\text{\sf\em size\text{$($}\/}\xspace}{\text{$\alpha$}\xspace}, b),0,{\text{\sf\em start\/}\xspace}({\text{$\beta$}\xspace},x,b), \geq))
\end{aligned}$$ Although our primary interest is in computing the minimum among the null positions, is more convenient. It inherently models the situation when no null position is found because ${\text{\sf\em glb\_le\/}\xspace}(\emptyset) = \infty$ indicating that the string is infinitely long (i.e. the string overflows the buffer).[^2]
### Set of Null Positions for String Functions {#set-of-null-positions-for-string-functions .unnumbered}
We use the above functions by dividing the null position sets resulting from a string operation into three categories based on the start position $s$ and the end position $e$ of the copied string in the destination buffer. Given $x$ as the destination pointer and $y$ as the source pointer:
1. . The null positions that remain unchanged in the destination buffer because they appear before $s$. This is defined by in Figure \[fig:flow.fun\].
2. . The null position appearing at $e$. This is defined by in Figure \[fig:flow.fun\]. It represents the null position imported from the source buffer.
3. . The null positions that remain unchanged in the destination buffer because they appear after $e$. This is defined by in Figure \[fig:flow.fun\].
The difference between the and functions is evident from the definition in Figure \[fig:flow.fun\]—for , the starting point of copying is ${\text{\sf\em start\/}\xspace}_{x}$ whereas for it is ${\text{\sf\em end\/}\xspace}_{x}$. Further, their length limited versions choose the minimum between the null distance and the length provided.
### Set of Null Positions for Memory Copy {#set-of-null-positions-for-memory-copy .unnumbered}
We view as a special case of . The main difference is that copies from $y$ only upto the first null character whereas copies $m$ characters from $y$. Thus the main change is in computing instead of . Former includes multiple position whereas the latter computes a single position. Thus we compute
1. . This identifies all null character positions that lie between and in the source buffer. These should be shifted by the start position in the destination buffer. This is computed by the intersection of with where represents the distance by which the null positions should be shifted.
2. . We must include $\infty$ to indicate overflow if
- (the write operation crosses the destination boundary), or
- (the read operation crosses the source boundary).
This is easily achieved by computing the following set differences and .
Overflow Detection {#sec:oflow.detection}
------------------
Computation of data flow values ${\text{\sf\em bIn\/}\xspace}_n/{\text{\sf\em bOut\/}\xspace}_n$ detects a buffer overflow at a program point at which a buffer is written into by introducing $\infty$ in the set. Note that this is a conservative conclusion and may well be a false positive. Our analysis is sound when it concludes the absence of buffer overflow.
For a program statement that merely reads a buffer (eg. call to ), the ${\text{\sf\em bIn\/}\xspace}_n/{\text{\sf\em bOut\/}\xspace}_n$ values remain unmodified. Detecting a potential buffer overflow for a read using a pointer is trivially achieved by
- Checking the buffer identity. If it is $b_\infty$, there is a potential buffer overflow.
- Checking the buffer size. If it is $\infty$, there is a potential buffer overflow.
- Checking the offset of the pointer. If it is $\infty$, there is a potential buffer overflow.
- Otherwise, we compute the of the pointer being read. If it is $\infty$, there is a potential buffer overflow.
Running Example Revisited {#sec:exmp}
=========================
This section shows the application of our analysis to the example in Figure \[fig:exmp.1\] and describes how buffer overflows can be detected.
Computing the Set of Null Positions
-----------------------------------
It is clear from Figure \[fig:flow.fun\] that the only non-obvious computation in this analysis is the computation of $R_n$. We illustrate it for our running example. The relevant maps and the values of the default extractor functions have already been provided in Figure \[fig:exmp.1\]. Observe that for the first two calls, buffer $b_1$ serves both as source and destination whereas for the third call, buffer $b_0$ is the source and $b_1$ is the destination.
- After the first call to in the example in Figure \[fig:exmp.1\], we have $$\begin{aligned}
{\text{\sf\em np\_src\/}\xspace}&= {\text{\sf\em length\/}\xspace}\oplus_{14} {\text{\sf\em copy\_pos\/}\xspace}\\ & = ({\text{\sf\em end\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},y,b_1) - {\text{\sf\em start\/}\xspace}({\text{$\beta$}\xspace},y,b_1)) \oplus_{14} {\text{\sf\em end\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace}, z, b_1)
\\
&= \left({\text{\sf\em glb\_le\/}\xspace}({\text{\sf\em rnps\/}\xspace}(\{3,7,13\},14,0,4, \geq)) - {\text{\sf\em start\/}\xspace}({\text{$\beta$}\xspace},y,b_1)\right)
\\
&\phantom{=}\;\; \oplus_{14}
\\
&\phantom{=}\;\; ({\text{\sf\em glb\_le\/}\xspace}({\text{\sf\em rnps\/}\xspace}(\{3,7,13\},14,0,6, \geq))
\\
& = ({\text{\sf\em glb\_le\/}\xspace}(\{ 7\oplus_{14} 0, 13 \oplus_{14} 0 \}) - 4) \oplus_{14}
({\text{\sf\em glb\_le\/}\xspace}(\{ 7\oplus_{14} 0, 13 \oplus_{14} 0 \}))
\\
& = ({\text{\sf\em glb\_le\/}\xspace}(\{ 7, 13 \}) - 4) \oplus_{14}
({\text{\sf\em glb\_le\/}\xspace}(\{ 7, 13 \}) )
\\
& = (7 - 4) \oplus_{14} 7 = 10\end{aligned}$$ The computation of relevant null positions (i.e. $R_n$) is as follows: $$\begin{aligned}
R_n({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},b_1) & =
{\text{\sf\em rnps\/}\xspace}(\{3,7,13\}, 14, 0, 6, <)
\cup \{ 10 \}
\cup
{\text{\sf\em rnps\/}\xspace}(\{3,7,13\}, 14, 0, 10, \geq)
\\
&=
\{ 3 \oplus_{14} 0 \}
\cup \{ 10 \}
\cup \{ 13 \oplus_{14} 0 \}
\\
&=
\{ 3 \} \cup \{ 10 \} \cup \{ 13 \} = \{ 3, 10, 13 \} \end{aligned}$$
- After the second call to in the example, we have $$\begin{aligned}
{\text{\sf\em np\_src\/}\xspace}&= {\text{\sf\em length\/}\xspace}\oplus_{14} {\text{\sf\em copy\_pos\/}\xspace}\\
& = ({\text{\sf\em end\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},y,b_1) - {\text{\sf\em start\/}\xspace}({\text{$\beta$}\xspace},y,b_1)) \oplus_{14} {\text{\sf\em end\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace}, z, b_1)
\\
&= \left({\text{\sf\em glb\_le\/}\xspace}({\text{\sf\em rnps\/}\xspace}(\{3,10,13\},14,0,4, \geq)) - {\text{\sf\em start\/}\xspace}({\text{$\beta$}\xspace},y,b_1)\right)
\\
&\phantom{=}\;\; \oplus_{14}
\\
&\phantom{=}\;\; ({\text{\sf\em glb\_le\/}\xspace}({\text{\sf\em rnps\/}\xspace}(\{3,10,13\},14,0,6, \geq))
\\
& = ({\text{\sf\em glb\_le\/}\xspace}(\{ 10\oplus_{14} 0, 13 \oplus_{14} 0 \}) - 4) \oplus_{14}
({\text{\sf\em glb\_le\/}\xspace}(\{ 10\oplus_{14} 0, 13 \oplus_{14} 0 \}))
\\
& = ({\text{\sf\em glb\_le\/}\xspace}(\{ 10, 13 \}) - 4) \oplus_{14}
({\text{\sf\em glb\_le\/}\xspace}(\{ 10, 13 \}) )
\\
& = (10 - 4) \oplus_{14} 10 = \infty\end{aligned}$$ $R_n$ computation is as follows. $$\begin{aligned}
R_n({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},b_1) & =
{\text{\sf\em rnps\/}\xspace}(\{3,7,13\}, 14, 0, 6, <) \cup \{ \infty \}
\cup {\text{\sf\em rnps\/}\xspace}(\{3,7,13\}, 14, 0, \infty, \geq)
\\
&=
\{ 3 \oplus_{14} 0 \}
\cup \{ \infty \}
\cup \emptyset
\\
&=
\{ 3 \} \cup \{ \infty \} = \{ 3, \infty \} \end{aligned}$$
- The third call involves $b_0$ as the source buffer and $b_1$ as the destination buffer. $$\begin{aligned}
{\text{\sf\em np\_src\/}\xspace}&= {\text{\sf\em length\/}\xspace}\oplus_{14} {\text{\sf\em copy\_pos\/}\xspace}\\
& = ({\text{\sf\em end\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},w,b_0) - {\text{\sf\em start\/}\xspace}({\text{$\beta$}\xspace},w,b_0)) \oplus_{14} {\text{\sf\em start\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace}, y, b_1)
\\
&= \left({\text{\sf\em glb\_le\/}\xspace}({\text{\sf\em rnps\/}\xspace}(\{10\},10,0,0, \geq)) - 0\right) \oplus_{14} 4
\\
& = {\text{\sf\em glb\_le\/}\xspace}(\{ 10\oplus_{10} 0 \}) \oplus_{14} 4
\\
& = {\text{\sf\em glb\_le\/}\xspace}(\{ 10 \}) \oplus_{14} 4
\\
& = 10 \oplus_{14} 4 = 14\end{aligned}$$ $R_n$ computation is as follows. $$\begin{aligned}
R_n({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},b_1) & =
{\text{\sf\em rnps\/}\xspace}(\{3,\infty\}, 14, 0, 4, <) \cup \{ 14 \}
\cup {\text{\sf\em rnps\/}\xspace}(\{3,\infty\}, 14, 0, 14, \geq)
\\
&=
\{ 3 \oplus_{14} 0 \}
\cup \{ 14 \}
\cup \{ \infty \}
\\
&=
\{ 3 \} \cup \{14 \}\cup \{ \infty \} = \{ 3, 14, \infty \} \end{aligned}$$ Note that the $\infty$ has not been generated in the computation of , it has been carried forward from the previous set. In other words, this call does not cause an overflow, the presence of $\infty$ in $R_n$ indicates that an overflow has occurred in the buffer earlier.
Overflow Detection {#overflow-detection}
------------------
It is clear from Section \[sec:oflow.detection\] that checking overflow is trivial except in the case where is to be used. We illustrate such uses for our running example in the following cases.
1. Assume that a statement is added to our running example. Then . For simplicity, assume that a call to ${\text{\sf\em strlen\/}\xspace}(v)$ occurs before the first call to . We compute $$\begin{aligned}
{\text{\sf\em end\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},v,b_1) & =
{\text{\sf\em glb\_le\/}\xspace}({\text{\sf\em rnps\/}\xspace}(\{3,7,13\},14,0,14,\geq))
\\
&= {\text{\sf\em glb\_le\/}\xspace}(\emptyset) = \infty
\end{aligned}$$ Thus ${\text{\sf\em strlen\/}\xspace}(v)$ may cause a buffer overflow. In this case, this is not a false negative but a certain buffer overflow. However, our analysis does not have any means of distinguishing between a certain overflow and a false negative.
2. For ${\text{\sf\em strlen\/}\xspace}(y)$ (before the first call to in the same example), we compute $$\begin{aligned}
{\text{\sf\em end\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},y,b_1) & =
{\text{\sf\em glb\_le\/}\xspace}({\text{\sf\em rnps\/}\xspace}(\{3,7,13\},14,0,4,\geq))
\\
&= {\text{\sf\em glb\_le\/}\xspace}(\{7 \oplus_{14} 0, 13\oplus_{14} 0\})
\\
&= {\text{\sf\em glb\_le\/}\xspace}(\{7, 13\}) = 7
\end{aligned}$$ The result indicates that the read will encounter a null character in the buffer at position 7 and hence there is no buffer overflow.
3. For ${\text{\sf\em strlen\/}\xspace}(y)$ (after the first call to in the same example), we compute $$\begin{aligned}
{\text{\sf\em end\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},y,b_1) & =
{\text{\sf\em glb\_le\/}\xspace}({\text{\sf\em rnps\/}\xspace}(\{3,10,13\},14,0,4,\geq))
\\
&= {\text{\sf\em glb\_le\/}\xspace}(\{10 \oplus_{14} 0, 13\oplus_{14} 0\})
\\
&= {\text{\sf\em glb\_le\/}\xspace}(\{10, 13\}) = 10
\end{aligned}$$ The result indicates that the read will encounter a null character in the buffer at position 10 and hence there is no buffer overflow.
4. For ${\text{\sf\em strlen\/}\xspace}(y)$ (after the second call to in the same example), we compute $$\begin{aligned}
{\text{\sf\em end\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},y,b_1) & =
{\text{\sf\em glb\_le\/}\xspace}({\text{\sf\em rnps\/}\xspace}(\{3,\infty\},14,0,4,\geq))
\\
&= {\text{\sf\em glb\_le\/}\xspace}(\emptyset) = \infty
\end{aligned}$$ The result indicates that the read may not find any null character in the buffer and hence there is a potential buffer overflow. In this case also, this is a certain buffer overflow but we have no way of concluding so.
Extensions for Handling Multiple Pointee Buffers of a Pointer {#sec:extensions}
=============================================================
In this section we show how our formulation can be extended to allow a pointer to point to multiple buffers at a program point. This requires a change in the lattice $B$, in the flow function , and in the computation of extractor function $R_n$ of Figure \[fig:flow.fun\]. These changes allow some relaxation in the assumptions about program model.
Changes in the Lattices
-----------------------
Given and , we allow the coexistence of triples and in a . However, multiple triples of the kind and are still prohibited.
This is achieved by redefining for the lattice as . Observe that now we use and not because now we do not need the fictitious “undefined” buffer $b_\infty$. Also, because of the cross product , now returns a set of buffers rather than a single buffer.
The meet operation is simplified as follows: $$\begin{aligned}
\forall \, {\text{$\beta$}\xspace},{\text{$\beta$}\xspace}' \in B\!: \;\;
{\text{$\beta$}\xspace}\,{\text{$\sqcap_B\,$}\xspace}{\text{$\beta$}\xspace}'
& =
\left\{
\left( x, b, k \,{\text{$\sqcap_S\,$}\xspace}k' \right)
\;\middle| \;
\left( x, b, k \right) \in {\text{$\beta$}\xspace},
\left( x, b, k' \right) \in {\text{$\beta$}\xspace}'
\right\}\end{aligned}$$ Exclusion of $b_\infty$ also leads to a change in the boundary information in the data flow equation for in which the triple is replaced by the triple for .
Changes in Extractor Function $R_n$
-----------------------------------
When we have multiple source destination buffers in a string operation, we use the following approximation by combining the effects of all these buffers: We use the longest string among all destination buffers, the smallest size among all source buffers, and the farthest position of copying among all source buffers. These three changes are reflected in our formulation by computing a single approximate $R_n$ by changing the terms appearing in Figure \[fig:flow.fun\].
We describe the changes in $R_n$ computation for string operations.
- [*Computing the longest string among all destination buffers.*]{} We compute ${\text{\sf\em length\/}\xspace}_y$ separately for each pointee buffer of $y$ and take the largest value.
- [*Computing the smallest size among all source buffers.*]{} We take the smallest value of ${\text{\sf\em size\text{$$}\/}\xspace}_x$ among all pointee buffers of $x$ for computing ${\text{\sf\em np\_src\/}\xspace}_{xy}$.
- [*Computing the farthest position of copying among all source buffers.*]{} We take the largest value of ${\text{\sf\em copy\_pos\/}\xspace}_{x}$ among all pointee buffers of $x$.
- [*Computing $R_n$.*]{} This involves the following changes.
- Computing ${\text{\sf\em nps\_before\/}\xspace}_{xy}$ and ${\text{\sf\em nps\_after\/}\xspace}_{xy}$. We take the smallest value of ${\text{\sf\em size\text{$$}\/}\xspace}_x$ and the largest value of ${\text{\sf\em np\_src\/}\xspace}_{xy}$.
- Computing ${\text{\sf\em np\_at\/}\xspace}_{xy}$. We take the largest value of ${\text{\sf\em np\_src\/}\xspace}_{xy}$.
Observe that these changes are declarative in the sense that they basically involve ranging over the buffers and computing the maximum or the minimum. Hence a stateless formulation of these changes is easy to write.
Similar changes can be defined for the operation.
Changes in Flow Function
-------------------------
With the possibility of a pointer pointing to multiple buffers, $D_n$ now becomes a set of buffers and $K_n$ is computed separately for each buffer. However, $R_n$ is common because it is an approximation of all source buffers. A buffer mapping should accumulate updates in all source buffers in $D_n$. In other words, when a buffer $b \in D_n$ is updated, the resulting mapping ${\text{$\alpha$}\xspace}'$ should be passed on to for updating some other buffer $b' \in D_n$. This is achieved by defining recursively and passing $D_n$ as the value of argument $X$ for the top level call to . $$\begin{aligned}
\!\!\!
\!\!\!
\!
{\text{\sf\em update\_buf\/}\xspace}({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},n,X)
&=
\begin{cases}
{\text{\sf\em update\_buf\/}\xspace}\left(
{\text{$\alpha$}\xspace}\left[b \mapsto (
K_n({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace},b),
R_n({\text{$\alpha$}\xspace},{\text{$\beta$}\xspace})
)\right], \;
{\text{\sf\em bpt\/}\xspace}, \;
n, X \!- \!\{ b \}
\right)
& \renewcommand{\arraystretch}{.9}
\begin{array}{@{}l}
b \in X
\end{array}
\\
{\text{$\alpha$}\xspace}& \text{Otherwise}
\rule{0em}{1.5em}
\end{cases}\end{aligned}$$ In each recursive call, we accumulate the updates and the set $X$ becomes smaller. When all updates are accumulated in , $X$ becomes $\emptyset$.
Relaxation in Program Model
---------------------------
With the extensions for handling multiple buffers of a pointer at a program point in place, now we do not need to ignore the back edges and we may be able to handle some cases where range information is not available. However, this computation may be expensive as a loop may be iterated many times depending upon the increment in the values of the pointers or array indices.
Related Work {#sec:related}
============
Buffer overflow detection and mitigation has been an important concern for a long time. The problem is compounded by the idiosyncrasies of C string operations [@Wagner00afirst] and non-trivial semantics [@Dor:2003:CTR:781131.781149]. There is an abundance of literature on the topic and many tools have been created to address the concern; we cite only a few references as a starting point [@citeulike:6345464; @buf.study; @Lhee:2003:BOF:781669.781672; @Shahriar:2010:MBO:2441114.2441116; @Zitser:2004:TSA:1029894.1029911].
A popular approach of addressing this concern has been to perform run time checks using code instrumentation, or expanded representation of pointers and buffers to store bookkeeping information (effected through compilers and changed library function) [@Dhurjati:2006:BAB:1134285.1134309; @Jones97backwards-compatiblebounds; @Ruwase04apractical]. The known static analysis methods are characterized by some combination of the following features:
- User annotations in the program [@Dor:2003:CTR:781131.781149; @Larochelle:2001:SDL:1251327.1251341] in the form of contracts or assertions.
- Range analysis [@Li:2010:PES:1882291.1882338; @Wagner00afirst] storing the lower and the upper limits of the buffers.
- Integer linear programming [@Dor:2003:CTR:781131.781149; @Ganapathy:2003:BOD:948109.948155; @Wagner00afirst] to solve constraints to compute the ranges.
- Symbolic computation [@Li:2010:PES:1882291.1882338] to store the ranges in terms of expressions rather than in terms of integers.
Most approaches set up constraints flow insensitively (or use flow insensitive pointer analysis) except [@Li:2010:PES:1882291.1882338] which does path sensitive analysis by storing relevant path predicates.[^3] What is common to all these approaches is that their formulations are stateful algorithms; none of them have a stateless formalization amenable to reasoning and automatic construction of analysers.
In terms of modelling, most approaches do not separate pointers and buffers as we do; they need to explicitly discover aliasing between buffers to record the effects of one change through a pointer, into the buffer of another pointer. In this sense our modelling is closer to the modelling for “intended referent” [@Jones97backwards-compatiblebounds]. However they store the information for run time checking, we use it for static analysis. Besides, unlike them, we do not treat a string as a buffer and instead store a set of null positions within a buffer. Most approaches treat a string as a buffer and hence store a single length with a buffer.
Conclusions {#sec:conclusions}
===========
Modelling buffers and buffer overflows in terms of
- mappings between pointers and buffers whose addresses they hold,
- buffer information storing size and sets of positions of null characters, and
- defining functions to compute these models
makes it possible to formulate buffer overflow analysis as a data flow analysis of C programs. A key facilitator of this is an emphasis on stateless formalizations of analyses in terms of lattice valued functions and relations. A stateful formulation combines features through side effects recorded in states thereby raising a natural requirement of C/C++/Java code to be written to complement a partially high level specification so that a generator can generate the analyser.
Banishing states from formulations enable higher levels of abstraction. The resulting conciseness, together with higher levels of abstraction, makes the formulations amenable to human reasoning. Further, such formulations allow a generator to check the specifications, combine their features freely, and decide how and where to introduce states in the generated code.
This paper does not claim to define the best buffer overflow analysis; it may well be possible to devise better static approximations. In particular, storing “out of bound” null positions [@Ruwase04apractical] may eliminate many false positives. It may also be interesting to explore the possibility of using symbolic expressions [@Li:2010:PES:1882291.1882338] for modelling the set of null character positions.
So long as we have a well defined static approximation, we believe that a stateless mathematical formulation leading to high level specifications amenable to automatic construction of analysers is possible.
Acknowledgements {#acknowledgements .unnumbered}
================
Dipali Bhutada, Kalyani Zope, Swati Jaiswal, and Vini Kanvar provided useful feedback on the drafts of this paper.
Modelling Other Statements {#app:stmt.modelling}
==========================
We show how most other statements can be modelled in terms of the core statements of our formulation.
- Assignment of string literals (eg. a statement ) is modelled as a sequence of two statements where $k$ represents the length of the string literal.
- Reallocation of memory to resize a buffer using is modelled as a sequence of two statements .
- Statements and are equivalent to and respectively.
- A declaration of an uninitialized array (eg. ) is modelled as statement.
- A declaration of an initialized array (eg. ) is modelled as a sequence of two statements where $k$ represents the number of elements in the array.
- The following string functions are modelled in terms of for the purpose of buffer overflow analysis: , , , and . The function finds the first occurrence of string $y$ in string $x$ and is viewed as a combination of and for this analysis. A call to is similarly modelled in terms of and .
Some of these functions either return a buffer pointer containing a possible substring (or a collection of substrings). Their lengths are dynamic and hence are not amenable to static analysis. Hence, we view the functions returning pointers to buffers with dynamic lengths, as functions creating buffers of size $\infty$.
- The remaining memory handling functions can also be similarly modelled and statically approximated in case their result depends on a dynamic value.
We do not rule out the possibility of better static approximations and expect them to be handled in a similar manner provided they can be defined precisely in terms of stateless functions.
[^1]: Since a buffer is identified by its program point of creation, known but dissimilar buffer sizes along different execution paths reaching a program point are not possible.
[^2]: Given that computes the greatest number that is smaller than any number in the given set, if the given set does not contain any number, the greatest number that is smaller than no number is $\infty$.
[^3]: It avoids a combinatorial explosion by a sparse demand driven computation that uses transitive data and control dependences.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In the information loss problem in black hole evaporation, the investigation of the purification partner of a Hawking particle is crucial. It is a well-known fact that the 3+1 dimensional spherically symmetric gravitational collapse can be mimicked by 1+1 dimensional moving mirror models. Since a detected particle in field theory is defined by what a particle detector observes, the diversity of detector designs yields a variety of particles and their partners. We provide a formula of generalized partners of detected particles emitted out of mirrors in an arbitrary motion for any Gaussian state in a free massless scalar field theory. Using our formula, we directly demonstrate information storage about initial phase information in a pure state of a detected particle and its partner. The form of the partner drastically changes depending on the detailed designs of particle detectors for Hawking radiation. In the case of a detected particle and its partner sensitive to information about initial phase, spatial configurations of the partner has long tails in a stage where only zero-point fluctuation is emitted out of the mirror.'
author:
- Takeshi Tomitsuka
- Koji Yamaguchi
- Masahiro Hotta
title: Partner formula for an arbitrary moving mirror in $1+1$ dimensions
---
Introduction
============
The information loss problem in black hole evaporation was posed first by Hawking in 1976 [@H]. Suppose that a black hole is formed by a collapsing star. The black hole evaporates by emitting the Hawking radiation with a thermal spectrum, which does not depend on the details of the initial phase of black hole formation. Therefore, it seems that some information about the initial phase cannot be reconstructed only from the Hawking radiation, and cannot be lost during the black hole evaporation process. This problem is a longstanding issue of fundamental physics during more than four decades. From the viewpoint of quantum information theory, the problem can be recast into a quest of a purification partner of the Hawking radiation if quantum gravity maintains unitarity. Since quantum gravity theory has not been completed yet, nobody knows the correct partner. Many possible candidates have been proposed including baby universe [@D][@Z], massive remnant[@A][@B][@G], and the Hawking radiation itself proposed by Page [@P], zero-point fluctuation of quantum fields [@W][@HSU][@HS], and soft hairs [@HSS] [@HPS].
Just like in the Page scenario, we assume that quantum gravity effects become negligibly small in null future infinity. Hawking particles in the radiation of free fields should have partners with degrees of freedom of the same fields. The composite system of a Hawking particle and its partner is in a pure state and carries the initial phase information of the gravitational collapse. The partner quest in realistic 3+1 dimensional cases encounters difficulties of unknown quantum gravity dynamics and cannot be achieved.
In this paper, we adopt a detour route. Similar to black holes, quantum fields scattered by some accelerated mirrors have a thermal spectrum [@MM][@F][@M2][@CW]. The radiation can be approximately interpreted as the Hawking radiation emitted by spherically symmetric black hole formation in 3+1 dimensions [@W]. The mirror trajectory shape reflects a part of the initial parameters of a 3+1 dimensional gravitational collapse. The emitted Hawking particle is observed by an Unruh-De Witt detector [@Unruh][@Dewitt]. Note that a wide variety of detector designs are possible even if we use a localized detector . Since a particle in field theory is defined by what a particle detector observes [@Unruh], the diversity of detector designs yields a variety of particles and their partners [@TYH][@TYH2]. The information of the mirror trajectory is stored in each pair of the detected particle and its partner. We provide a formula of generalized partners in an arbitrary mirror motion for any Gaussian state in a free massless scalar field. Using this formula, we directly demonstrate information storage about initial phase information. The spatial form of the partner is specified by its weighting functions introduced in Sec. \[section\_3\]. We find two distinct cases of the partners. In one case, the weighting functions of the detector modes are well localized in the Hawking radiation stage. The weighting functions of the partner modes have almost no dependence on the initial phase information. In the other case, the weighting functions of the detector modes have non vanishing tails in the regime of the initial phase. The weighting functions of the partner modes change drastically depending on the initial phase information.
In Sec. \[section\_2\], we give a short review on a moving mirror model. In Sec. \[section\_3\], we provide a partner formula for an arbitrary moving mirror in $1+1$ dimension. In Sec. \[section\_4\], we demonstrate information storage about initial phase information by using our formula. Finally, In Sec. \[section\_5\], a summary is provided.
In this paper, we adopt natural units, $c=\hbar=k_{B}=1$.
Moving mirror model\[section\_2\]
=================================
First, we review a moving mirror model which mimics the Hawking radiation and black hole evaporation. Consider a flat $1+1$ dimensional spacetime $$ds^2=-dt^2+dx^2=-dudv,$$ where we introduce light-cone coordinates $$u=t-x\ ,\ v=t+x.$$ A massless scaler field $\phi(u,v)$ satisfies the following Klein-Gordon equation $$\partial_{v}\partial_{u}\phi(u,v)=0$$ and vanishes at the location of the mirror $v=p(u)$ $$\label{boundary}
\phi(u,p(u))=0.$$ General solutions are written by $$\phi(u,v)=\phi_{in}(v)+\phi_{out}(u),$$ where $\phi_{in}(v)(\phi_{out}(u))$ is an arbitrary function of $v(u)$ and satisfies the following relation $$\label{boundary_condition}
\phi_{out}(u)=-\phi_{in}(p(u))$$ due to Eq.. The incoming field $\phi_{in}(v)$ can be quantized as $$\hat{\phi}_{in}(v)=\int_{0}^{\infty}\frac{d\omega}{\sqrt{4\pi\omega}}\left(\hat{a}_{\omega}e^{-i\omega v}+\hat{a}_{\omega}^{\dagger}e^{i\omega v}\right),$$ where creation and annihilation operators $\hat{a}_{\omega}$ and $\hat{a}_{\omega}^{\dagger}$ satisfy $[\hat{a}_{\omega},\hat{a}_{\omega'}^{\dagger}]=\delta(\omega-\omega')$. The in-vacuum state $\ket{0_{in}}$ is defined by $\hat{a}_{\omega}\ket{0_{in}}=0$. The outgoing field operator $\phi_{out}(u)$ can also be quantized as $$\hat{\phi}_{out}(u)=\int_{0}^{\infty}\frac{d\omega}{\sqrt{4\pi\omega}}\left(\hat{b}_{\omega}e^{-i\omega u}+\hat{b}_{\omega}^{\dagger}e^{i\omega u}\right),$$ where creation and annihilation operators $\hat{b}_{\omega}$ and $\hat{b}_{\omega}^{\dagger}$ satisfy $[\hat{b}_{\omega},\hat{b}_{\omega'}^{\dagger}]=\delta(\omega-\omega')$. The two sets of operator $\{\hat{a}_{\omega},\hat{a}_{\omega}^{\dagger}\}$ and $\{\hat{b}_{\omega},\hat{b}_{\omega}^{\dagger}\}$ are related by $$\hat{b}_{\omega}=\int_{0}^{\infty}d\omega'\left(\alpha_{\omega\omega'}\hat{a}_{\omega'}+\beta_{\omega\omega'}\hat{a}_{\omega'}^{\dagger}\right),$$ where $\alpha_{\omega\omega'}$ and $\beta_{\omega\omega'}$ are Bogoliubov coefficients. If $\beta_{\omega\omega'}\neq 0$, particle creation occurs with a distribution given by $$\bra{0_{in}}\hat{b}_{\omega}^{\dagger}\hat{b}_{\omega}\ket{0_{in}}=\int_{0}^{\infty}|{\beta_{\omega\omega'}}|^{2}d\omega'$$ and the energy flux observed in the future null infinity is given in terms of the mirror trajectory $p(u)$ by $$\label{energy_flux}
\bra{0_{in}}\hat{T}_{uu}(u)\ket{0_{in}}=-\frac{1}{24\pi}\left[\frac{\partial_{u}^{3}p(u)}{\partial_{u}p(u)}-\frac{3}{2}\left(\frac{\partial_{u}^{2}p(u)}{\partial_{u}p(u)}\right)^{2}\right].$$ For example, consider the following trajectory $$\label{p(u)}
p(u)=v_{H}-\frac{1}{\kappa}e^{-\kappa u}.$$ This trajectory mimics an eternal black hole without backreaction of radiation, that is, this black hole never evaporates. $v=v_{H}$ corresponds to the event horizon (Fig.\[p\]). The incoming field $\hat{\phi}_{in}(v<v_{H})$ is reflected into $\hat{\phi}_{out}(u)$ while $\hat{\phi}_{in}(v \geq v_{H})$ goes through and never comes back. Figure \[Penrose\_1\] represents the Penrose diagram of the corresponding $3+1$ dimensional spherically symmetric gravitational shell collapse. The energy flux is constant which can be calculated by Eq. as $$\bra{0_{in}}\hat{T}_{uu}(u)\ket{0_{in}}=\frac{\kappa^{2}}{48\pi},$$ with its Hawking temperature $T_{H}=\frac{\kappa}{2\pi}$.
![Penrose diagram of $3+1$ dimensional spherically symmetric gravitational shell collapse mimics trajectory . Mirror trajectory corresponds to the origin $(r=0)$. Red line represents collapsing the null shell and dashed line denotes the event horizon.[]{data-label="Penrose_1"}](p.pdf){width="70mm"}
![Penrose diagram of $3+1$ dimensional spherically symmetric gravitational shell collapse mimics trajectory . Mirror trajectory corresponds to the origin $(r=0)$. Red line represents collapsing the null shell and dashed line denotes the event horizon.[]{data-label="Penrose_1"}](Penrose_1.png){height="70mm"}
The second example is the following trajectory $$\label{p5(u)}
p(u)=-\frac{1}{\kappa}\ln{\left(\frac{1+e^{-\kappa u}}{1+e^{\kappa(u-h)}}\right)},$$ which mimics a black hole evaporation process with a backreaction of radiation. Parameter $h$ controls the lifetime of the black hole. This mirror is at rest in $u \sim -\infty$, accelerated in $0<u<h$ and finally stopped (Fig.\[p5\]). Almost constant energy flux $\bra{0_{in}}\hat{T}_{uu}(u)\ket{0_{in}}\sim \frac{\kappa^{2}}{48\pi}$ is emitted during evaporation and no energy flux exists initially and finally. In this case, all incoming fields $\hat{\phi}_{in}(v)$ come back to the future null infinity as $\hat{\phi}_{out}(u)$.
![The trajectory $(\kappa=1,\ h=100)$.[]{data-label="p5"}](p5sub.pdf){width="80mm"}
Partner formula for an arbitrary mirror motion\[section\_3\]
============================================================
In this section, we provide a formula of generalized partners of detected particles emitted out of mirrors in arbitrary motion for any Gaussian state in a free massless scalar field. In the following, we assume that the trajectory $p(u)$ is a monotonically increasing function and its range is $(-\infty,\infty)$, that is, we consider an evaporation case. Let us introduce the incoming momentum operator in the past null infinity ${\mathscr S^{-}}$ $$\hat{\Pi}_{in}(v) \equiv \partial_{v}\hat{\phi}_{in}(v)=-i\int_{0}^{\infty}\sqrt{\frac{\omega}{4\pi}}\left(\hat{a}_{\omega}e^{-i\omega v}-\hat{a}_{\omega}^{\dagger}e^{i\omega v}\right)d\omega.$$ We fix the detected particle mode $A$ by a set of canonical operators using $\hat{\Pi}_{in}(v)$ $$(\hat{q}_{A},\hat{p}_{A})=\left(\int q_{A}^{in}(v)\hat{\Pi}_{in}(v)dv,\int p_{A}^{in}(v)\hat{\Pi}_{in}(v)dv \right),$$ where $q_{A}^{in}(v)$ and $p_{A}^{in}(v)$ are real functions of $v$ satisfying $$\frac{1}{2}\int_{-\infty}^{\infty}q^{in}_{A}(v)\partial_{v}p_{A}^{in}(v)dv=1$$ due to the canonical commutation relation $[\hat{q}_{A},\hat{p}_{A}]=i$ and call them weighting functions. We make a local symplectic transformation $S_{A}$ $$\begin{aligned}
\left(
\begin{array}{c}
\hat{Q}_{A} \\
\hat{P}_{A}
\end{array}\right)\equiv
S_{A}\left(
\begin{array}{c}
\hat{q}_{A} \\
\hat{p}_{A}
\end{array}\right)=\left(
\begin{array}{c}
\int_{-\infty}^{\infty}Q_{A}^{in}(v)\hat{\Pi}_{in}(v)dv \\
\int_{-\infty}^{\infty}P_{A}^{in}(v)\hat{\Pi}_{in}(v)dv
\end{array}\right)\end{aligned}$$ such that the operators $(\hat{Q}_{A},\hat{P}_{A})$ are in the standard form (Appendix \[symplectic\]). Here we defined the functions $(Q_{A}^{in}(v),P_{A}^{in}(v))$ by $$\begin{aligned}
\left(
\begin{array}{c}
Q_{A}^{in}(v) \\
P_{A}^{in}(v)
\end{array}\right)\equiv S_{A}\left(
\begin{array}{c}
q_{A}^{in}(v) \\
p_{A}^{in}(v)
\end{array}\right).\end{aligned}$$ Then, its partner mode B, described as a set of canonical operators $(\hat{Q}_{B},\hat{P}_{B})$, is given by the following: (the derivation is shown in Appendix \[derivation\_A\]) $$\begin{aligned}
\label{in_formula_Q}
\hat{Q}_{B}&=\frac{\sqrt{1+g^{2}}}{g}\hat{Q}_{A}+\frac{2}{g}\int_{-\infty}^{\infty}\int_{-\infty}^{\infty}{\rm sgn}(v-v'){\rm Re}\left(\bra{\Psi}\hat{P}_{A}\hat{\Pi}_{in}(v')\ket{\Psi}\right)\hat{\Pi}_{in}(v)dv'dv \\
\label{in_formula_P}
\hat{P}_{B}&=-\frac{\sqrt{1+g^{2}}}{g}\hat{P}_{A}+\frac{2}{g}\int_{-\infty}^{\infty}\int_{-\infty}^{\infty}{\rm sgn}(v-v'){\rm Re}\left(\bra{\Psi}\hat{Q}_{A}\hat{\Pi}_{in}(v')\ket{\Psi}\right)\hat{\Pi}_{in}(v)dv'dv,\end{aligned}$$ where $g\equiv \sqrt{4(\bra{\Psi}\hat{q}_{A}^{2}\ket{\Psi}\bra{\Psi}\hat{p}_{A}^{2}\ket{\Psi}-({\rm Re}\bra{\Psi}\hat{q}_{A}\hat{p}_{A}\ket{\Psi})^{2})-1}$, $\ket{\Psi}$ is an arbitrary Gaussian state, and $$\begin{aligned}
{\rm sgn}(x)\equiv
\begin{cases}
1 & : x>0 \\
0 & : x=0\\
-1 & : x<0
\end{cases}.\end{aligned}$$ In the future null infinity ${\mathscr S}^{+}$, we define the outgoing momentum operator $$\hat{\Pi}_{out}(u)\equiv \partial_{u}\hat{\phi}_{out}(u)=-i\int_{0}^{\infty}\sqrt{\frac{\omega}{4\pi}}d\omega(\hat{b}_{\omega}e^{-i\omega u}-\hat{b}_{\omega}^{\dagger}e^{i\omega u}).$$ Due to the boundary condition , the relation $$\label{Pi_relation}
\hat{\Pi}_{out}(u)=-\partial_{u}p(u)\hat{\Pi}_{in}(p(u))$$ holds. The operators of mode $A$ can also be written in terms of $\hat{\Pi}_{out}(u)$ $$\begin{aligned}
\hat{Q}_{A}&=-\int_{-\infty}^{\infty}Q_{A}^{in}(p(u))\hat{\Pi}_{out}(u)du=\int_{-\infty}^{\infty}Q_{A}(u)\hat{\Pi}_{out}(u)du \\
\hat{P}_{A}&=-\int_{-\infty}^{\infty}P_{A}^{in}(p(u))\hat{\Pi}_{out}(u)du=\int_{-\infty}^{\infty}P_{A}(u)\hat{\Pi}_{out}(u)du,\end{aligned}$$ where $$Q_{A}(u)\equiv -Q_{A}^{in}(p(u)),\ P_{A}(u)\equiv -P_{A}^{in}(p(u)).$$ These are the weighting functions in the future null infinity ${\mathscr S}^{+}$. We can get the formula in terms of $\hat{\Pi}_{out}(u)$ from Eqs., using $v=p(u)$ and the relation $$\begin{aligned}
\label{out_formula_Q}
\hat{Q}_{B}&=\frac{\sqrt{1+g^{2}}}{g}\hat{Q}_{A}+\frac{2}{g}\int_{-\infty}^{\infty}\int_{-\infty}^{\infty}{\rm sgn}\left(p(u)-p(u')\right){\rm Re}\left(\bra{\Psi}\hat{P}_{A}\hat{\Pi}_{out}(u')\ket{\Psi}\right)\hat{\Pi}_{out}(u)du'du \\
\label{out_formula_P}
\hat{P}_{B}&=-\frac{\sqrt{1+g^{2}}}{g}\hat{P}_{A}+\frac{2}{g}\int_{-\infty}^{\infty}\int_{-\infty}^{\infty}{\rm sgn}\left(p(u)-p(u')\right){\rm Re}\left(\bra{\Psi}\hat{Q}_{A}\hat{\Pi}_{out}(u')\ket{\Psi}\right)\hat{\Pi}_{out}(u)du'du.\end{aligned}$$ This is the partner formula for an arbitrary moving mirror in $1+1$ dimension and the main result in this paper. Entanglement entropy between the mode $A$ and $B$ is given by $$S(g)=\sqrt{1+g^{2}}\ln\left(\frac{\sqrt{1+g^{2}}+1}{g}\right)+\ln\left(\frac{g}{2}\right).$$ For the vacuum Gaussian state $\ket{\Psi}=\ket{0_{in}}$, this formula is more simplified. The weighting functions are given by $$\begin{aligned}
\label{out_formula_Qw}
Q_{B}(u)&=\frac{\sqrt{1+g^2}}{g}Q_{A}(u)+\frac{1}{g}\int_{-\infty}^{\infty}\Delta(p(u)-p(u'))P_{A}(u')\partial_{u'}p(u')du' \\
\label{out_formula_Pw}
P_{B}(u)&=-\frac{\sqrt{1+g^2}}{g}P_{A}(u)+\frac{1}{g}\int_{-\infty}^{\infty}\Delta(p(u)-p(u'))Q_{A}(u')\partial_{u'}p(u')du',\end{aligned}$$ where we have defined $$\Delta(x-x')\equiv \frac{i}{2\pi}\int_{-\infty}^{\infty}d\omega\ {\rm sgn}(\omega)e^{-i\omega(x-x')}.$$ (the derivation is shown in Appendix \[derivation\_B\])
Application to the simple evaporation mirror model\[section\_4\]
================================================================
We want to investigate how the initial phase information of a moving mirror is stored in a pure state of a pair of detected particle and its partner. To do this, we consider the following simple trajectory $$\begin{aligned}
\label{p4(u)}
p(u)=
\begin{cases}
-\frac{e^{\kappa_{2}h}}{\kappa_{1}}e^{-\kappa_{1}u}+\left(\frac{1}{\kappa_{1}}-\frac{1}{\kappa_{2}}\right)e^{\kappa_{2}h}+h+\frac{1}{\kappa_{2}} & (u\leq 0) \\
-\frac{e^{\kappa_{2}h}}{\kappa_{2}}e^{-\kappa_{2}u}+h+\frac{1}{\kappa_{2}} & (0 < u \leq h) \\
\ \ u & (h<u)
\end{cases}\end{aligned}$$ which has three stages, namely, the initial phase $(u\leq 0)$, the Hawking radiation stage $(0<u\leq h)$ and the no radiation stage $(h < u)$. Figure \[p4\] shows the trajectory . Note that dynamics of realistic black hole formation is different from that of the moving mirror. Thus, the causal structure of the black hole case cannot be correctly described by the moving mirror model. However, there exists an apparent information loss in the moving mirror model since it seems that Hawking radiation does not depend on the initial phase information, as already stressed above. It is significant to address the question where the information is stored in the final quantum states. This trajectory mimics the two-step spherically symmetric gravitational shell collapse and evaporation process (Fig.\[collapse\]). The initial phase corresponds to the first shell collapse, the Hawking radiation stage to the second shell collapse, and the no radiation stage to the region where only zero-point fluctuation exists, respectively.
![Two-step spherically symmetric gravitational shell collapse. First shell collapse (red circle) is described by red line in Fig.\[p4\] and second one (blue circle) is described by blue line in Fig.\[p4\].[]{data-label="collapse"}](p4.pdf){width="70mm"}
![Two-step spherically symmetric gravitational shell collapse. First shell collapse (red circle) is described by red line in Fig.\[p4\] and second one (blue circle) is described by blue line in Fig.\[p4\].[]{data-label="collapse"}](collapse.png){width="70mm"}
Figure \[T4\] shows the energy flux for the trajectory. The Hawking temperature in the initial phase $(0\leq u)$ and in the Hawking radiation stage $(0 < u \leq h)$ are $T_{1}=\frac{\kappa_{1}}{2\pi}$ and $T_{2}=\frac{\kappa_{2}}{2\pi}$, respectively.
![Energy flux $\bra{0_{in}}\hat{T}_{uu}(u)\ket{0_{in}}$ for the trajectory $(\kappa_{1}=3,\ \kappa_{2}=1,\ h=2)$. In the initial phase $(u \leq 0)$, the Hawking temperature $T_{1}=\frac{3}{2\pi}$, in the Hawking radiation stage $(0 < u \leq 2)$, $T_{2}=\frac{1}{2\pi}$. Red dashed lines represent energy fluxes proportional to delta function at the junctions of the trajectory. Blue solid line represents energy flux without their contributions.[]{data-label="T4"}](T4.pdf){width="80mm"}
In this setup, we assume that $\ket{\Psi}=\ket{0_{in}}$ in the past null infinity ${\mathscr S}^{-}$ and it evolves under the trajectory . Fixing parameters $\kappa_{2}=1,\ h=2$ and setting the detected particle mode $A$ in the future null infinity ${\mathscr S}^{+}$, we investigate how its partner mode $B$ changes with respect to $\kappa_{1}$ which is initial phase information. Let us set the mode $A$ by the following form: $$\begin{aligned}
\hat{q}_{A}&=\int_{-\infty}^{\infty}q_{A}(u)\hat{\Pi}_{out}(u)du\\
\hat{p}_{A}&=\int_{-\infty}^{\infty}p_{A}(u)\hat{\Pi}_{out}(u)du \\
q_{A}(u)&=e^{-C^{2}(u-\frac{h}{2})^{2}} \\
p_{A}(u)&=4C\sqrt{\frac{2}{\pi}}\left(u-\frac{h}{2}\right)e^{-C^{2}(u-\frac{h}{2})^{2}}\end{aligned}$$ satisfying $\frac{1}{2}\int_{-\infty}^{\infty}q_{A}(u)\partial_{u}p_{A}(u)du=1$ (which corresponds to $[\hat{q}_{A},\hat{p}_{A}]=i$). A positive parameter $C$ controls the degree of localization in the Hawking radiation stage. First, we set them well-localized in the Hawking radiation stage $(C=6)$ (Fig.\[qA\],\[pA\]). The schematic picture of this setup is shown in Fig.\[mirror\_1\].
![The weighting function of mode $A$ $p_{A}(u)$ $(C=6)$ well-localized in the Hawking radiation stage $(0<u\leq 2)$.[]{data-label="pA"}](qA.pdf){width="70mm"}
![The weighting function of mode $A$ $p_{A}(u)$ $(C=6)$ well-localized in the Hawking radiation stage $(0<u\leq 2)$.[]{data-label="pA"}](pA.pdf){width="70mm"}
![The schematic picture of this setup. The weighting functions of mode $A$ are well localized in the Hawking radiation stage.[]{data-label="mirror_1"}](mirror_1.png){width="70mm"}
For this mode $A$, Figs.\[QB\] and \[PB\] show the numerical result of the weighting functions of its partner mode $B$.
![The weighting function of partner mode $B$ $P_{B}(u)$ for the localized mode $A$.[]{data-label="PB"}](QB.pdf){width="70mm"}
![The weighting function of partner mode $B$ $P_{B}(u)$ for the localized mode $A$.[]{data-label="PB"}](PB.pdf){width="70mm"}
Its partner mode $B$ changes little in the initial phase with respect to $\kappa_{1}$ (Figs.\[QBsub\] and \[PBsub\]). So we can hardly determine what $\kappa_{1}$ is from this pair of partners, that is, this pair has little initial phase information. The result that there is no change in the thermal and the no radiation stage is clear by the partner formula. When the weighting functions of mode $A$ are well localized in the Hawking radiation stage, the dependence of $\kappa_{1}$ emerges only in the initial phase of the second terms of the right-hand sides in Eqs. and .
![The weighting function of partner mode $B$ in the initial phase $P_{B}(u<0)$.[]{data-label="PBsub"}](QBsub.pdf){width="70mm"}
![The weighting function of partner mode $B$ in the initial phase $P_{B}(u<0)$.[]{data-label="PBsub"}](PBsub.pdf){width="70mm"}
Next, we let the weighting functions of mode $A$ nonlocalized a little $(C=1)$ which have nonvanishing tails in the initial phase, as depicted in Figs.\[qA2\] and \[pA2\]. The schematic picture of this setup is shown in Fig.\[mirror\_2\].
![The weighting function of mode $A$ $p_{A}(u)$ $(C=1)$ nonlocalized in the Hawking stage $(0 < u \leq 2)$.[]{data-label="pA2"}](qA2.pdf){width="70mm"}
![The weighting function of mode $A$ $p_{A}(u)$ $(C=1)$ nonlocalized in the Hawking stage $(0 < u \leq 2)$.[]{data-label="pA2"}](pA2.pdf){width="70mm"}
![The schematic picture of this setup. The weighting functions of mode $A$ are nonlocalized in the Hawking radiation stage which have the nonvanishing tails in the initial phase.[]{data-label="mirror_2"}](mirror_2.png){width="70mm"}
For this mode $A$, the weighting functions of its partner mode $B$ are shown in Figs.\[QB2\] and \[PB2\].
![The weighting function of partner mode $B$ $P_{B}(u)$ with respect to $\kappa_{1}$ for the nonlocalized mode $A$.[]{data-label="PB2"}](QB2.pdf){width="70mm"}
![The weighting function of partner mode $B$ $P_{B}(u)$ with respect to $\kappa_{1}$ for the nonlocalized mode $A$.[]{data-label="PB2"}](PB2.pdf){width="70mm"}
Its partner mode $B$ changes dramatically with respect to $\kappa_{1}$ in all stages due to the second terms of the right-hand sides in Eqs. and . Then we can determine what $\kappa_{1}$ is from this pair of partners, that is, this pair has the initial phase information. Note that the partner mode $B$ has a long tail in the no radiation stage $(u>2)$, so initial phase information is also stored in the correlation between the mode $A$ and zero-point fluctuation, pointed out in [@HSU].
Summary\[section\_5\]
=====================
For an arbitrary Gaussian state, we provide a formula of generalized partners for an arbitrary moving mirror in $1+1$ dimensions \[Eqs.,\]. Next, we demonstrate information storage about initial phase information in a pair of partners by applying this formula to a simple mirror trajectory which mimics two-step spherically symmetric gravitational shell collapse. The sensitivity of a pair of partners to initial phase information depends on the design of the detected particle mode. In particular, for the detected particle mode well-localized in the Hawking radiation stage (Figs.\[qA\],\[pA\]), little initial phase information is stored in that pair of partners (Figs.\[QB\],\[PB\]). On the other hand, for a little nonlocalized one (Figs.\[qA2\],\[pA2\]), initial phase information is stored in it including the correlation with zero-point fluctuation (Figs.\[QB2\],\[PB2\]). It would be interesting to extend our formula to the higher dimensional case and two mirror case. The latter could be applicable in a dynamical Casimir effect experiment where two imperfect reflectional mirrors are used.\
Recently, Wald wrote an interesting paper [@Wald] and proposed a Milne partner of a Hawking particle. Although ordinary entanglement of two particles is spacelike correlation, the entanglement between a Milne particle and a Hawking particle is timelike correlation. Our results in this paper will be useful to explore the above possibility.
Acknowledgment {#acknowledgment .unnumbered}
==============
We thank the participants in “Relativistic Quantum Information - North 2019” at National Cheng Kung University for useful discussion and Michael R.R. Good for useful comments. We appreciate a discussion with Robert M. Wald. This research is partially supported by JSPS KAKENHI Grant Number JP19K03838 (M.H.) and JP18J20057 (K.Y.), and by the Graduate Program on Physics for the Universe (GP-PU), Tohoku University (T.T and K.Y.).
[99]{}
S. W. Hawking, Phys. Rev. D [**14**]{}, 2460 (1976).
F. Dyson, Institute for Advanced Study Preprint (unpublished).
Y. B. Zeldovich, Zh. Eksp. Teor. Fiz. [**79**]{}, 18 (1977) \[Sov. Phys. JETP [**45**]{}, 9 (1977)\].
Y. Aharonov, A. Casher and S. Nussinov, Phys. Lett. [**191**]{} B, 51 (1987).
T. Banks, A. Dabholkar, M. R. Douglas and M. O’Loughlin, Phys. Rev. D [**45**]{} 3607 (1992).
S. B. Giddings, Phys. Rev. D [**49**]{}, 947 (1994).
D. N. Page, Phys. Rev. Lett. [**71**]{}, 3743 (1993).
F. Wilczek, in *Proceedings of the International Symposium on Black Holes, Membranes, Wormholes, and Superstrings: Houston Advanced Research Center, USA, 1992*, edited by S. Kalara and D. Nanopoulos (World Scientific, Singapore, 1993).
M. Hotta, R. Schützhold, and W. G. Unruh, Phys. Rev. D [**91**]{},124060 (2015).
M. Hotta, and A. Sugita, Prog. Theor. Exp. Phys. (2015), 123B04.
M. Hotta, K. Sasaki, and T. Sasaki, Classical Quantum Gravity [**18**]{}, 1823 (2001).
S.W. Hawking, M. J. Perry, and A. Strominger, Phys. Rev.Lett. [**116**]{}, 231301 (2016).
B. S. De Witt, Phys. Rep. [**19C**]{}, 297 (1975).
P. C. Davies, and S. A. Fulling, Proc. R. Soc. A [**356**]{}, 237 (1977).
W. R. Walker, Phys. Rev. D [**31**]{}, 767 (1985).
R. D. Carlitz and R. S. Willey, Phys. Rev. D [**36**]{}, 2336 (1987); [**36**]{}, 2327 (1987).
W. Unruh, Phys. Rev. D [**14**]{}, 870 (1976).
B. De Witt, General Relativity: An Einstein Centenary Survey, edited by S.W. Hawking and W. Israel (Cambridge University Press, Cambridge, England, 1979), p. 680.
J. Trevison, K. Yamaguchi and M. Hotta, Prog. Theor. Exp. Phys. ([**2018**]{}) 103A03.
J. Trevison, K. Yamaguchi, and M. Hotta, J. Phys. A [**52**]{}, 125402 (2019).
R. M. Wald, Phys. Rev. D [**100**]{}, 065019 (2019).
THE LOCAL SYMPLECTIC TRANSFORMATION TO THE STANDARD FORM\[symplectic\]
======================================================================
By the local symplectic transformation of particle mode A $$\begin{aligned}
\left(
\begin{array}{c}
\hat{Q}_{A} \\
\hat{P}_{A}
\end{array}
\right)=S_{A}\left(
\begin{array}{c}
\hat{q}_{A} \\
\hat{p}_{A}
\end{array}
\right)=\left(
\begin{array}{cc}
\cos\theta_{A}' & \sin\theta_{A}' \\
-\sin\theta_{A}' & \cos\theta_{A}'
\end{array}\right)\left(
\begin{array}{cc}
e^{\sigma_{A}} & 0 \\
0 & e^{-\sigma_{A}}
\end{array}\right)\left(
\begin{array}{cc}
\cos\theta_{A} & \sin\theta_{A} \\
-\sin\theta_{A} & \cos\theta_{A}
\end{array}\right)\left(
\begin{array}{c}
\hat{q}_{A} \\
\hat{p}_{A}
\end{array}\right)\end{aligned}$$ , we can transform the corresponding canonical operators $(\hat{q}_{A}, \hat{p}_{A})^{T} \rightarrow (\hat{Q}_{A},\hat{P}_{A})^{T}$ such that $$\begin{aligned}
\left(
\begin{array}{cc}
\bra{\Psi}\hat{Q}_{A}^{2}\ket{\Psi} & {\rm Re}\left(\bra{\Psi}\hat{Q}_{A}\hat{P}_{A}\ket{\Psi}\right) \\
{\rm Re}\left(\bra{\Psi}\hat{P}_{A}\hat{Q}_{A}\ket{\Psi}\right) & \bra{\Psi}\hat{P}_{A}^{2}\ket{\Psi}
\end{array}
\right)=\frac{\sqrt{1+g^{2}}}{2}\left(
\begin{array}{cc}
1 & 0 \\
0 & 1
\end{array}
\right),\end{aligned}$$ where $g\equiv \sqrt{4(\bra{\Psi}\hat{q}_{A}^{2}\ket{\Psi}\bra{\Psi}\hat{p}_{A}^{2}\ket{\Psi}-({\rm Re}\bra{\Psi}\hat{q}_{A}\hat{p}_{A}\ket{\Psi})^{2})-1}$, $\ket{\Psi}$ is an arbitrary Gaussian state.
DERIVATION OF EQS. AND \[derivation\_A\]
========================================
We define the canonical operators $$\hat{Q}_{\omega}\equiv \frac{1}{\sqrt{2}}(\hat{a}_{\omega}+\hat{a}_{\omega}^{\dagger}),\ \ \ \hat{P}_{\omega}\equiv \frac{1}{\sqrt{2}i}(\hat{a}_{\omega}-\hat{a}_{\omega}^{\dagger})$$ and $$\hat{\boldsymbol{R}}_{\omega}\equiv(\hat{Q}_{\omega},\hat{P}_{\omega})^{T}$$ and assume that $\bra{\Psi}\hat{\boldsymbol R}_{\omega}\ket{\Psi}=0$ holds for an arbitrary state $\ket{\Psi}$ without loss of generality by shifting $$\hat{Q}_{\omega}\rightarrow \hat{Q}_{\omega}-\bra{\Psi}\hat{Q}_{\omega}\ket{\Psi},\ \hat{P}_{\omega}\rightarrow \hat{P}_{\omega}-\bra{\Psi}\hat{P}_{\omega}\ket{\Psi}.$$ Then $\hat{{\boldsymbol R}}_{\omega}$ satisfies $$[\hat{\boldsymbol R}_{\omega},\hat{\boldsymbol R}_{\omega'}^{T}]=i\Omega(\omega,\omega')$$ with $$\begin{aligned}
\Omega(\omega,\omega')\equiv \left(
\begin{array}{cc}
0 & \delta(\omega-\omega') \\
-\delta(\omega-\omega') & 0
\end{array}\right)\end{aligned}$$ The covariance matrix is defined by $$M(\omega,\omega')\equiv {\rm Re}(\bra{\Psi}\hat{\boldsymbol R}_{\omega}\hat{\boldsymbol R}_{\omega'}^{T}\ket{\Psi})$$ If $\ket{\Psi}$ is a Gaussian pure sate, the relation $$\int_{0}^{\infty}d\omega_{1}d\omega_{2}M(\omega,\omega_{1})\Omega(\omega_{1},\omega_{2})M(\omega_{2},\omega')=\frac{1}{4}\Omega(\omega,\omega')$$ is satisfied. An incoming momentum operator can be expressed in terms of $\hat{Q}_{\omega}$ and $\hat{P}_{\omega}$ as $$\begin{aligned}
\hat{\Pi}_{in}(v)&=-i\int_{0}^{\infty}\sqrt{\frac{\omega}{4\pi}}\left(\frac{\hat{Q}_{\omega}+i\hat{P}_{\omega}}{\sqrt{2}}e^{-i\omega v}-\frac{\hat{Q}_{\omega}-i\hat{P}_{\omega}}{\sqrt{2}}e^{i\omega v}\right)d\omega \nonumber \\
&=\int_{0}^{\infty}\sqrt{\frac{\omega}{2\pi}}\left(-\sin{(\omega v)}\hat{Q}_{\omega}+\cos{(\omega v)}\hat{P}_{\omega}\right)d\omega\end{aligned}$$ So, the mode $A$ operators are also written by $$\begin{aligned}
\hat{q}_{A}&=\int_{-\infty}^{\infty}q_{A}^{in}(v)\hat{\Pi}_{in}(v)dv=\int_{0}^{\infty}d\omega {\boldsymbol v}_{A}(\omega)^{T}\hat{\boldsymbol R}_{\omega}\\
\hat{p}_{A}&=\int_{-\infty}^{\infty}p_{A}^{in}(v)\hat{\Pi}_{in}(v)dv=\int_{0}^{\infty}d\omega {\boldsymbol u}_{A}(\omega)^{T}\hat{\boldsymbol R}_{\omega}\end{aligned}$$ where $$\begin{aligned}
{\boldsymbol v}_{A}(\omega)&\equiv \int_{-\infty}^{\infty}dv \sqrt{\frac{\omega}{2\pi}}q_{A}(v)(-\sin{(\omega v)},\cos{(\omega v)})^{T} \\
{\boldsymbol u}_{A}(\omega)&\equiv \int_{-\infty}^{\infty}dv \sqrt{\frac{\omega}{2\pi}}p_{A}(v)(-\sin{(\omega v)},\cos{(\omega v)})^{T} \end{aligned}$$ We make a symplectic transformation $$\begin{aligned}
\left(
\begin{array}{c}
\hat{Q}_{A} \\
\hat{P}_{A}
\end{array}\right)\equiv
S_{A}\left(
\begin{array}{c}
\hat{q}_{A} \\
\hat{p}_{A}
\end{array}\right)=\left(
\begin{array}{c}
\int_{0}^{\infty}d\omega{\boldsymbol V}_{A}(\omega)^{T}\hat{{\boldsymbol R}}_{\omega} \\
\int_{0}^{\infty}d\omega{\boldsymbol U}_{A}(\omega)^{T}\hat{{\boldsymbol R}}_{\omega}
\end{array}\right)\end{aligned}$$ such that the operators $(\hat{Q}_{A},\hat{P}_{A})$ are in the standard form where $$\begin{aligned}
\left(
\begin{array}{c}
{\boldsymbol V}_{A}(\omega) \\
{\boldsymbol U}_{A}(\omega)
\end{array}\right)\equiv S_{A}\left(
\begin{array}{c}
{\boldsymbol v}_{A}(\omega) \\
{\boldsymbol u}_{A}(\omega)
\end{array}\right).\end{aligned}$$ Then, its partner mode $B$ can be given by $$\begin{aligned}
\label{QBin}
\hat{Q}_{B}&=\int_{0}^{\infty}d\omega{\boldsymbol V}_{B}^{T}(\omega)\hat{{\boldsymbol R}}_{\omega} \\
\label{PBin}
\hat{P}_{B}&=\int_{0}^{\infty}d\omega{\boldsymbol U}_{B}^{T}(\omega)\hat{{\boldsymbol R}}_{\omega}\end{aligned}$$ where the functions ${\boldsymbol V}_{B}(\omega),\ {\boldsymbol U}_{B}(\omega)$ are determined from the covariance matrix [@TYH] as $$\begin{aligned}
\label{partner V}
{\boldsymbol V}_{B}(\omega)&=\frac{\sqrt{1+g^2}}{g}{\boldsymbol V}_{A}(\omega)-\frac{2}{g}\int_{0}^{\infty}d\omega_{1}d\omega_{2}\Omega(\omega,\omega_{1})M(\omega_{1},\omega_{2}){\boldsymbol U}_{A}(\omega_{2})\\
\label{partner U}
{\boldsymbol U}_{B}(\omega)&=-\frac{\sqrt{1+g^2}}{g}{\boldsymbol U}_{A}(\omega)-\frac{2}{g}\int_{0}^{\infty}d\omega_{1}d\omega_{2}\Omega(\omega,\omega_{1})M(\omega_{1},\omega_{2}){\boldsymbol V}_{A}(\omega_{2}).\end{aligned}$$ At this time, the covariance matrix of mode $A$ and $B$ including the correlations between $A$ and $B$ is $$\begin{aligned}
M_{AB}&\equiv\left(
\begin{array}{cccc}
\bra{\Psi}\hat{Q}_{A}^{2}\ket{\Psi} & {\rm Re}\bra{\Psi}\hat{Q}_{A}\hat{P}_{A}\ket{\Psi} & \bra{\Psi}\hat{Q}_{A}\hat{Q}_{B}\ket{\Psi} & \bra{\Psi}\hat{Q}_{A}\hat{P}_{B}\ket{\Psi} \\
{\rm Re}\bra{\Psi}\hat{P}_{A}\hat{Q}_{A}\ket{\Psi} & \bra{\Psi}\hat{P}_{A}^{2}\ket{\Psi} & \bra{\Psi}\hat{P}_{A}\hat{Q}_{B}\ket{\Psi} & \bra{\Psi}\hat{P}_{A}\hat{P}_{B}\ket{\Psi} \\
\bra{\Psi}\hat{Q}_{B}\hat{Q}_{A}\ket{\Psi} & \bra{\Psi}\hat{Q}_{B}\hat{P}_{A}\ket{\Psi} & \bra{\Psi}\hat{Q}_{B}^{2}\ket{\Psi} & {\rm Re}\bra{\Psi}\hat{Q}_{B}\hat{P}_{B}\ket{\Psi} \\
\bra{\Psi}\hat{P}_{B}\hat{Q}_{A}\ket{\Psi} &\bra{\Psi}\hat{P}_{B}\hat{P}_{A}\ket{\Psi} &{\rm Re}\bra{\Psi}\hat{P}_{B}\hat{Q}_{B}\ket{\Psi} & \bra{\Psi}\hat{P}_{B}^{2}\ket{\Psi}
\end{array}\right) \\
&=\left(
\begin{array}{cccc}
\frac{\sqrt{1+g^{2}}}{2} & 0 & \frac{g}{2} & 0 \\
0 & \frac{\sqrt{1+g^{2}}}{2} &0 & -\frac{g}{2} \\
\frac{g}{2} & 0 & \frac{\sqrt{1+g^{2}}}{2} & 0 \\
0 & -\frac{g}{2} & 0 & \frac{\sqrt{1+g^{2}}}{2}
\end{array}\right)\end{aligned}$$
Substituting and for Eqs. and , we can get the formula and after simple calculation using the relation $$\begin{aligned}
\hat{\boldsymbol R}_{\omega}=\sqrt{\frac{2}{\omega \pi}}\int_{-\infty}^{\infty}dv\left(
\begin{array}{c}
-\sin{(\omega v)} \\
\cos{(\omega v)}
\end{array}\right)\hat{\Pi}_{in}(v).\end{aligned}$$
DERIVATION OF EQS. AND \[derivation\_B\]
========================================
For the vacuum Gaussian state $\ket{0_{in}}$ which is defined by $\hat{a}_{\omega}\ket{0_{in}}=0\ (\forall \omega)$, the covariance matrix is simplified as $$\begin{aligned}
M(\omega_{1},\omega_{2})=\frac{1}{2}\left(
\begin{array}{cc}
\delta(\omega_{1}-\omega_{2}) & 0 \\
0 & \delta(\omega_{1}-\omega_{2}) \\
\end{array}\right).\end{aligned}$$ Equations and become $$\begin{aligned}
\label{Vb}
{\boldsymbol V}_{B}(\omega)&=\frac{\sqrt{1+g^2}}{g}{\boldsymbol V}_{A}(\omega)-\frac{1}{g}\Omega {\boldsymbol U}_{A}(\omega) \\
\label{Ub}
{\boldsymbol U}_{B}(\omega)&=\frac{\sqrt{1+g^2}}{g}{\boldsymbol U}_{A}(\omega)-\frac{1}{g}\Omega {\boldsymbol V}_{A}(\omega)\end{aligned}$$ where $$\Omega\equiv\left(
\begin{array}{cc}
0 & 1 \\
-1 & 0
\end{array}\right).$$ Substituting and for Eqs. and , $$\begin{aligned}
\int_{0}^{\infty}d\omega (\Omega {\boldsymbol U}_{A}(\omega))^{T}\hat{\boldsymbol R}_{\omega}&=\int_{0}^{\infty}d\omega \sqrt{\frac{2}{\pi\omega}}(\Omega {\boldsymbol U}_{A}(\omega))^{T}\left(
\begin{array}{c}
-\sin(\omega v) \\
\cos (\omega v)
\end{array}\right)\hat{\Pi}_{in}(v) \\
&=\frac{1}{\pi}\int_{-\infty}^{\infty}dv\int_{-\infty}^{\infty}dv'\int_{0}^{\infty}d\omega P_{A}^{in}(v')\sin(\omega(v'-v))\hat{\Pi}_{in}(v) \nonumber \\
&=\frac{1}{2\pi i}\int_{-\infty}^{\infty}dv\int_{-\infty}^{\infty}dv'P_{A}^{in}(v')\int_{0}^{\infty}d\omega(e^{-i\omega(v-v')}-e^{i\omega(v-v')})\hat{\Pi}_{in}(v)\nonumber \\
&=-\int_{-\infty}^{\infty}dv\int_{-\infty}^{\infty}dv'\left(\frac{i}{2\pi}\int_{-\infty}^{\infty}d\omega {\rm sgn}(\omega)e^{-i\omega(v-v')}\right)P_{A}^{in}(v')\hat{\Pi}_{in}(v)\nonumber \\
&=-\int_{-\infty}^{\infty}dv\int_{-\infty}^{\infty}dv'\Delta(v-v')P_{A}^{in}(v')\hat{\Pi}_{in}(v)\end{aligned}$$ where we have defined $$\Delta(v-v')\equiv \frac{i}{2\pi}\int_{-\infty}^{\infty}d\omega\ {\rm sgn}(\omega)e^{-i\omega(v-v')}$$ Therefore, we get $$\hat{Q}_{B}=\int_{-\infty}^{\infty} Q_{B}^{in}(v)\hat{\Pi}_{in}(v)dv,\ \hat{P}_{B}=\int_{-\infty}^{\infty} P_{B}^{in}(v)\hat{\Pi}_{in}(v)dv$$ where the weighting functions are given by $$\begin{aligned}
Q_{B}^{in}(v)&=\frac{\sqrt{1+g^2}}{g}Q_{A}^{in}(v)+\frac{1}{g}\int_{-\infty}^{\infty}dv' \Delta(v-v')P_{A}^{in}(v') \\
P_{B}^{in}(v)&=-\frac{\sqrt{1+g^2}}{g}P_{A}^{in}(v)+\frac{1}{g}\int_{-\infty}^{\infty}dv' \Delta(v-v')Q_{A}^{in}(v').\end{aligned}$$ Finally, using $v=p(u)$ and the relation , we can obtain $$\hat{Q}_{B}=\int_{-\infty}^{\infty}Q_{B}(u)\hat{\Pi}_{out}(u)du,\ \hat{P}_{B}=\int_{-\infty}^{\infty}P_{B}(u)\hat{\Pi}_{out}(u)du$$ where $Q_{B}(u)$ and $P_{B}(u)$ are given by Eqs. and .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that a violation of the Clauser-Horne-Shimony-Holt (CHSH) inequality can be demonstrated in a certain kind of Bell experiment if and only if the state is entangled. Our protocol allows local filtering measurements and involves shared ancilla states that do not themselves violate CHSH. Our result follows from two main steps. We first provide a simple characterization of the states that violate the CHSH-inequality after local filtering operations in terms of witness-like operators. Second, we prove that for each entangled state $\sigma$, there exists another state $\rho$ not violating CHSH, such that $\rho\otimes\sigma$ violates CHSH. Hence, in this scenario, $\sigma$ cannot be substituted by classical correlations without changing the statistics of the experiment; we say that $\sigma$ is not [*simulable*]{} by classical correlations and our result is that entanglement is equivalent to non-simulability.'
author:
- Lluís Masanes
- 'Yeong-Cherng Liang'
- 'Andrew C. Doherty'
title: All bipartite entangled states display some hidden nonlocality
---
In 1964 Bell ruled out the possibility that a local-realistic theory could reproduce all the experimental predictions given by Quantum Mechanics [@Be]. In a local-realistic theory the outcomes of measurements are determined in advance by unknown (or hidden) variables, and do not depend on the choice of measurements made by distant observers. Bell’s theorem states that the quantum mechanical probabilities for outcomes of measurements distributed in space cannot, in general, be replicated in [*any*]{} local-realistic theory. This fact is demonstrated for particular states and measurements by the violation of a Bell inequality. Entanglement is in some way responsible for this phenomenon since entangled states are required to demonstrate the violation of Bell inequalities. However since it has been known for some time that there are entangled states that do not violate any Bell inequality [@W; @B] the precise relationship between entanglement and Bell inequality violation has remained poorly understood.
The definition of entangled state is made in terms of the physical resources needed for the [*preparation*]{} of the state: a multipartite state is said to be entangled if it cannot be prepared from classical correlations using local quantum operations [@W]. But this definition tells us nothing about the “behavior” of the state. For example, does the state violate a Bell inequality, or is it useful in some quantum protocol such as teleportation?
It is known that every pure entangled state violates a Bell inequality [@Gi; @Po] and that no separable state does [@W], but the situation gets more complicated for mixed entangled states. There are bipartite mixed states that, though being entangled, possess a local hidden variables model (LHVM) whenever measurements are made on a single copy of the state (see for example [@B; @W]). But some of these states violate Bell inequalities if, prior to the measurement, the state is processed by local operations and classical communication (LOCC) [@SP; @G]. Moreover, by jointly measuring more than one copy of these states after some LOCC preprocessing, it was shown that an even larger set of entangled states could be detected through their violation of a Bell inequality [@P].
Generalizing this idea one can get a strong test of the nonlocality “hidden” in a state by combining local filtering operations and collective measurements: Perform joint local filtering operations on an arbitrarily large number of copies of the state and then a Bell inequality test on the resulting state. If the resulting probabilities violate a Bell inequality, we say that the original state violates this inequality asymptotically. In Ref. [@Masymp] it is shown that a bipartite state violates the Clauser-Horne-Shimony-Holt (CHSH) inequality [@CHSH] asymptotically if, and only if, it is distillable. This result suggests that undistillable entangled states may admit a LHVM description even when experiments are performed on many copies of the state.
Given these negative results, it seems necessary to allow still more general protocols for the nonlocality “hidden" in arbitrary entangled states to manifest itself. One natural possibility is to allow joint processing with auxiliary states (that do not themselves violate the Bell inequality) rather than just with more copies of the state in question. This idea has been fruitful to show that useful entanglement can be extracted from all non-separable states [@Mtele] and in this Letter we use it to show that there is indeed a kind of hidden nonlocality that is possessed by all entangled states.
In order to investigate the possibilities of this more general kind of hidden nonlocality we introduce the concept of a [*simulable*]{} state. We say that a bipartite state $\sigma$ is simulable by classical correlations, or just [*simulable*]{}, if in any protocol (possibly involving other resources such as shared quantum states) two separated parties sharing classical correlations instead of $\sigma$ can obtain the same statistics for the outcomes of the protocol. In this sense, simulable states have a completely classical “behavior". Of course we are most interested in this Letter in the case where the protocol concerned is a test of nonlocality. Clearly, all separable states are simulable. A possible way to simulate a separable state is by just preparing it from classical correlations [@W].
The scenario that we consider is the typical Bell-like experiment, where two parties share a bipartite system and perform local measurements on it. Alice chooses between the observables $x=0,1$ and obtains the outcomes $a=0,1$, and analogously for Bob, $y$ and $b$. All the relevant experimental information is contained in the joint probability distribution for the outcomes conditioned on the choice of observables $P(a,b|x,y)$. It is convenient to define the correlation functions $$\label{correlators}
C_{xy}\equiv P(a=b|x,y)-P(a\neq b|x,y)\ .$$ By local relabeling of $(a,b,x,y)$ it is always possible to make $C_{00}, C_{01}, C_{10} \geq 0$. With this convention, the distribution $P(a,b|x,y)$ admits a LHVM if [@A.Fine:PRL:1982], and only if, it satisfies the CHSH-inequality [@CHSH] $$\label{chsh}
C_{00} +C_{01} +C_{10} -C_{11} \leq 2\ .$$ Let us characterize the set of bipartite states that violate the CHSH-inequality after preprocessing.
Denote by $\c$ the set of bipartite states that do not violate the CHSH-inequality, even after stochastic local operations without communication.
By stochastic we mean that the operation can fail, and we do not care about the probability of failure, as long as it is strictly smaller than one. Up to normalization, these operations allow the transformations $$\label{sepmap}
\rho
\rightarrow \Omega(\rho) = \sum_i \left(A_i\otimes B_i \right) \rho \left(A_i\otimes
B_i \right)^\dagger,$$ where $A_i$ and $B_i$ are, respectively, Kraus operators acting on the first and second system. This class of maps is known as the [ *separable maps*]{}.
In Ref. [@Masymp] it is shown that the states in $\c$ do not violate CHSH even after stochastic local operations [*with*]{} communication. So the exact nature of the local operations allowed in the definition of $\c$ is not important. Clearly, states that do not violate the CHSH-inequality asymptotically are in $\c$. Thus, $\c$ contains all undistillable states [@Masymp]. We are now able to state precisely the central result of this Letter.
A bipartite state $\sigma$ is entangled if, and only if, there exists a state $\rho \in \c$ such that $\rho\otimes\sigma$ is not in $\c$.
The consequences of this theorem are dramatic. If $\rho$ belongs to $\c$, no matter how much additional classical correlation (which can always be represented by a separable state $\eta_{\mbox{\scriptsize sep}}$) we supply to it, the result $\rho
\otimes \eta_{\mbox{\scriptsize sep}}$ is still in $\c$. Contrary, the state $\rho \otimes \sigma$ is not in $\c$ even if both $\rho$ and $\sigma$ are in $\c$. The violation of CHSH manifests the qualitatively different behavior between $\rho \otimes \sigma$ and $\rho \otimes \eta_{\mbox{\scriptsize sep}}$, where $\eta_{\mbox{\scriptsize sep}}$ is any separable state, and $\sigma$ is any entangled state. Summarizing, for each entangled state $\sigma$ there exists a protocol (which also involves the auxiliary state $\rho$ associated with the theorem) in which $\sigma$ cannot be substituted by an arbitrarily large amount of classical correlations without changing the result: [*Entangled states are the ones that cannot be simulated by classical correlations.*]{}
The proof of the above theorem has two main ingredients. Firstly we note that $\c$ is a convex set and provide a characterization of $\c$ in terms of witness-like operators that detect CHSH-violation. Secondly we use convexity arguments similar to those in Ref. [@Mtele] to prove by contradiction that there exists some $\rho\in \c$ such that one of these witnesses may be constructed for $\rho\otimes
\sigma$ whenever $\sigma$ is entangled. To carry this argument through we require a characterization of the separable completely positive maps between Bell diagonal states that will be explained further elsewhere [@DLM]. Firstly we describe the witnesses for CHSH-violation.
A bipartite state $\rho$ acting on $\H_{\A}\otimes\H_{\B}$ belongs to $\c$ if, and only if, it satisfies $$\label{C}
\tr\!\left[\rho\, (A\otimes B) H_\theta
(A\otimes B)^\dag \right]
\geq 0\ ,$$ for all matrices of the form $A\! :{{\mathbb{C}}}^2\rightarrow\H_{\A}$, $B\!
:{{\mathbb{C}}}^2\rightarrow\H_{\B}$ and all numbers $\theta \in [0,\pi/4]$, where $$\label{H}
H_\theta\equiv \id\otimes\id-\cos\theta\,
\sigma_x\otimes\sigma_x - \sin\theta\,
\sigma_z\otimes \sigma_z,$$ $\id$ being the $2\times 2$ identity matrix and $\{\sigma_i\}_{i=x,y,z}$ the Pauli matrices.
To start off, we recall that, without preprocessing, a two-qubit state $\varrho$ violates the CHSH inequality iff $\mu_1^2+\mu_2^2 > 1$, where $\mu_1$ and $\mu_2$ are the two largest singular values of the $3\times 3$ real matrix $R_{ij}= \tr[\varrho\,
\sigma_i\otimes \sigma_j]$, with indices $i,j=x,y,z$ [@VW; @Horodecki]. Equivalently, $(\mu_1,\mu_2)$ derived from $\varrho$ must lie outside the unit circle $\mu_1^2+\mu_2^2=1$, which is true if and only if there exists $\theta\in [0,2\pi]$ such that $$\label{Eq:Linear:CHSH}
\mu_1\cos\theta+\mu_2\sin\theta >1.$$ Now, it is also well known that by appropriate local unitary transformations $U$, $V$, it is always possible to arrive at a local basis such that $R$ is diagonal with $\mu_1=R_{xx}$ and $\mu_2=R_{zz}$. From the definition of $R$ it follows that $$\mu_1\cos\theta = \tr\!\left[(U \otimes V)
\varrho\, (U \otimes V)^\dag (\cos\theta\,
\sigma_x\otimes\sigma_x)
\right],$$ with the expression for $\mu_2 \sin \theta$ involving obvious modifications. Since singular values are non-negative, it thus follows that if $\varrho$ violates CHSH then there exist $U,V\in\rm SU(2)$, $\theta\in [0,\pi/4]$ such that $$\label{Cqubits}
\tr\!\left[\varrho\, (U \otimes V)^\dag H_\theta
(U \otimes V) \right]
< 0\ .$$ On the other hand suppose that there exists some $(U,V,\theta)$ satisfying (\[Cqubits\]). Thus we have $R_{xx}\cos \theta
+R_{zz}\sin \theta > 1$. If we assume $\mu_1 \geq \mu_2$, the inequalities $|R_{xx}|,|R_{zz}| \leq
\mu_1 \leq 1$ follow from the definition of singular values and the well known fact that all singular values of $R$ are less than one. Since $0
\leq \theta \leq \pi/4$ both $R_{xx}$ and $R_{zz}$ must be positive and since $\cos \theta \geq \sin \theta$ we may assume without loss of generality that $R_{xx}\geq R_{zz}$. The singular values of $R$ obey the inequality $|R_{xx}+R_{zz}|\leq
\mu_1+\mu_2$ [@Bhat] and as a result we find $ \mu_1\cos\theta+\mu_2\sin\theta > 1$ so $\rho$ violates the CHSH-inequality. Thus $\rho$ violates CHSH iff (\[Cqubits\]) holds.
Let us come back to the question of CHSH violation after local filtering operations. Assume that $\rho$ violates CHSH after stochastic local operations. Let us show that it must violate (\[C\]) for some $(A,B,\theta)$. In Ref. [@Masymp] it is proven that, if a state violates CHSH then it can be transformed by stochastic local operations into a two-qubit state which also violates CHSH. Therefore, there must exist a separable map $\Omega$ with two-qubit output, such that the state $\Omega(\rho)$ satisfies condition (\[Cqubits\]) for some $(U,V,\theta)$, denote them by $(U_0,V_0,\theta_0)$. Clearly, if $\Omega(\rho)$ satisfies (\[Cqubits\]) there must exist at least one value of $i$ in the Kraus decomposition of (\[sepmap\]) such that $(A_i\otimes B_i)~\rho~(A_i \otimes
B_i)^\dag$ also satisfies (\[Cqubits\]). This implies that $\rho$ violates (\[C\]) for $A= A_i^\dag U_0^\dag$, $B= B_i^\dag V_0^\dag$ and $\theta=\theta_0$. This proves one direction of the lemma, let us show the other.
Assume that $\rho$ violates (\[C\]) for $(A_0,B_0,\theta_0)$. It is straightforward to see that $\rho$ violates CHSH after stochastic LOCC. Consider operation that transforms $\rho$ into $(A_0\otimes
B_0)^\dag\, \rho\, (A_0\otimes B_0)$. By assumption, the final state satisfies (\[Cqubits\]) with $U=V=\id$ and $\theta=\theta_0$, which implies that it violates CHSH. $\Box$
The above characterization is interesting on its own. Here we use it to prove our main result.
If $\sigma$ is separable then, $\rho
\in\c$ implies $\rho\otimes\sigma \in \c$. This is so because the preprocessing by LOCC of $\rho$, before the Bell experiment, can include the preparation of the state $\sigma$. Let us prove the other direction of the theorem.
>From now on $\sigma$ is an arbitrary entangled state acting on $\H=\H_{\A} \otimes\H_{\B}$. Let us show that there always exists an ancilla state $\rho\in\c$ such that $\rho\otimes\sigma \not\in \c$. Fix $\rho$ to act on the bipartite Hilbert space $\left[\H_{\A'}\otimes\H_{\A''}\right] \otimes
\left[\H_{\B'}\otimes\H_{\B''}\right]$, where $\H_{\A'}=\H_\A$, $\H_{\B'}=\H_{\B}$ and $\H_{\A''} =\H_{\B''} ={{\mathbb{C}}}^2$ (see Fig. \[Fig:Protocol\]).
Our aim is to prove that the state $\rho\otimes\sigma$ violates (\[C\]) for some choice of $A$, $B$, and $\theta$. In particular, let $$\label{tildeM}
\tilde{A} = {\mbox{$| \Phi_{\A \A'} \rangle$}}\otimes
\id_{\A''}\ ,\quad
\tilde{B} = {\mbox{$| \Phi_{\B \B'} \rangle$}}\otimes
\id_{\B''}\ , \quad \theta = \pi/4,$$ where ${\mbox{$| \Phi_{\A \A'} \rangle$}}$ is the maximally-entangled state between the spaces $\H_{\A}$ and $\H_{\A'}$ (which have the same dimension), and $\id_{\A''}$ is the identity matrix acting on ${{\mathbb{C}}}^2$ (analogously for Bob). A little calculation shows that for any $\rho$ $$\label{eq}
\tr\!\left[\rho\otimes\sigma\,
(\tilde{A}\otimes\tilde{B}) H_{\pi/4}
(\tilde{A}\otimes \tilde{B})^\dag\right]
= \nu\, \tr\!\left[\rho\, (\sigma\t\otimes H_{\pi/4})\right]$$ where $\nu$ is a positive constant and $\sigma\t$ stands for the transpose of $\sigma$. The inequality (\[C\]) with $\theta=\pi/4$, $A=\tilde{A}$, $B=\tilde{B}$ becomes $$\label{cond}
\tr\left[\rho \left(\sigma\t
\otimes H_{\pi/4}\right) \right]
<0\ .$$ For convenience, in the rest of the proof we allow $\rho$ to be unnormalized. The only constraints on the matrices $\rho \in \c$ are: positive semi-definiteness ($\rho \in \s$), and satisfiability of all the inequalities (\[C\]) in Lemma 1. $\c$ is now a convex cone, and its dual cone is defined as $$\label{dual}
\c^*=\{X : \tr[\rho\, X]\geq 0,\
\forall \rho\in\c\}\ ,$$ where $X$ are Hermitian matrices. Farkas’ Lemma [@farkas] states that all matrices in $\c^*$ can be written as non-negative linear combinations of matrices $P \in \s$ and matrices $(A\otimes
B) H_\theta (A\otimes B)^\dag$ with $A:{{\mathbb{C}}}^2\rightarrow
\H_{\A'}\otimes\H_{\A''}$ and $B:{{\mathbb{C}}}^2\rightarrow
\H_{\B'} \otimes\H_{\B''}$.
We now show that there always exists $\rho\in\c$ satisfying (\[cond\]) by supposing otherwise and arriving at a contradiction. Suppose that for all $\rho \in \c$ the inequality $\tr[\rho\,
(\sigma\t\otimes H_{\pi/4})]\geq 0$ holds, and thus the matrix $\sigma\t\!\otimes H_{\pi/4}$ belongs to $\c^*$. Applying Farkas’ Lemma [@farkas] we can write $$\label{pl}
\sigma\t\! \otimes H_{\pi/4}=
\int\!\! dx\, (A_x\otimes B_x) H_{\theta\!_x} (A_x\otimes
B_x)^\dag + \int\!\! dy\, P_y\ ,$$ which is equivalent to $$\label{principal0}
\sigma\t\! \otimes H_{\pi/4} - \int\!\! dx
\ \Omega_x\!\left( H_{\theta\!_x} \right) \geq 0\ ,$$ where each $\Omega_x$ is a separable map (\[sepmap\]). We prove in Lemma 2 that (\[principal0\]) requires that $\sigma$ is separable, which gives the desired contradiction. Thus the result is proven. $\Box$
In order to arrive at a contradiction from it is necessary to use the constraint that the maps $\Omega_x$ are separable. The problem of characterizing the separable maps is hard in general since it maps onto the separability problem for bipartite states. However it turns out only to be necessary to determine the set of separable maps that take Bell diagonal states to Bell diagonal states and this can be done exactly [@DLM]. This characterization may be used to prove the following lemma and thereby our theorem.
Let $\Omega_\theta: [{{\mathbb{C}}}^2] \otimes [{{\mathbb{C}}}^2]
\rightarrow [\H_{\A}\otimes {{\mathbb{C}}}^2]\otimes[\H_{\B} \otimes {{\mathbb{C}}}^2]$ be a family of maps, separable with respect to the partition denoted by the brackets. Let $\mu$ be a unit-trace, positive semi-definite matrix acting on $[\H_{\A}]\otimes[\H_{\B}]$ such that $$\label{principal}
\mu\t\! \otimes H_{\pi/4} - \int\!\! dx
\ \Omega_x\!\left( H_{\theta_x} \right)
\geq 0\ ,$$ where $H_\theta$ is defined in (\[H\]), then $\mu$ has to be separable.
Now, let us characterize the solutions $\Omega_x$ of (\[principal\]). The Bell basis is defined as $$\begin{aligned}
\label{bell}
{\mbox{$| \Phi_{^1_2} \rangle$}} &=& 2^{-1/2}\left( {\mbox{$| 00 \rangle$}} \pm {\mbox{$| 11 \rangle$}} \right)\ , \\
{\mbox{$| \Phi_{^3_4} \rangle$}} &=& 2^{-1/2}\left( {\mbox{$| 01 \rangle$}} \pm {\mbox{$| 10 \rangle$}} \right)\ .\end{aligned}$$ The matrices $H_\theta$ are diagonal in this basis, $H_\theta=
\sum_{i=1}^4 N_\theta^i\, \Pi_i$, where $\Pi_i\equiv{\mbox{$| \Phi_i \rangle\!\langle \Phi_i |$}}$ are the Bell projectors and $N_\theta^i$ are the components of the vector $$\label{N}
N_\theta = \left[
\begin{array}{c}
1-\cos\theta-\sin\theta \\
1+\cos\theta-\sin\theta \\
1-\cos\theta+\sin\theta \\
1+\cos\theta+\sin\theta \\
\end{array}
\right].$$ For each value of $x$ define the sixteen matrices $$\label{omega}
\omega_x^{ij} \equiv \tr_{\A'' \B''}\!\left[\left(\id\o\Pi_i\right)\,
\Omega_x(\Pi_j)\right]\ ,$$ for $i,j=1,2,3,4$, where the identity matrix $\id$ acts on $\H_{\A'}\otimes\H_{\B'}$ and $\Pi_i$ acts on $\H_{\A''}\otimes\H_{\B''}$. Each $\omega_x^{ij}$ is the result of a physical operation, and hence positive. One can see $\omega_x^{ij}$ as the Jamio[ł]{}kowski state corresponding to the map $\Omega_x$, after “twirling" the input and output subsystems with the group of unitaries that leaves Bell-diagonal states invariant. Projecting the left hand side of onto the four Bell projectors $\Pi_i$, and taking the relevant partial trace, we get $$\label{uk}
\mu\t N_{\pi/4}^i - \int\!\! dx
\sum_{j=1}^4\ \omega_x^{ij} N_{\theta\!_x}^j \geq 0
\ ,$$ for $i=1,2,3,4$. Denote by $M_x$ the $4\times4$ matrix with components $M_x^{ij}=\tr[\omega_x^{ij}]$. Performing the trace on the left hand side of (\[uk\]) we obtain the four inequalities $$\label{io}
N_{\pi/4} - \int\!\! dx
\ M_x \cdot N_{\theta\!_x} \succeq \nv\ ,$$ where $\nv$ is the 4-dimensional null vector, and the symbols $\cdot$ and $\succeq$ mean, respectively, standard matrix multiplication and component-wise inequality. Consider the set of matrices $M$ that are generated by tracing the left-hand side of (\[omega\]) when $\Omega_x$ is any separable map. The characterization of this set of matrices is obtained in Ref. [@DLM], and goes as follows. Denote by $\mathcal{D}$ the set of $4\times 4$ doubly-stochastic matrices, that is, the convex hull of the permutation matrices [@Bhat]. Denote by $\mathcal{G}$ the convex hull of all matrices obtained when independently permuting the rows and columns of $$\label{e2}
G_0\equiv \left[
\begin{array}{cccc}
1 & 1 & 0 & 0 \\
1 & 1 & 0 & 0 \\
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 \\
\end{array}
\right]\ .$$ It is shown in Ref. [@DLM] that, any matrix $M$ as defined above can be written as $$\label{m}
M=pD+qG\ ,$$ where $D\in \mathcal{D}$, $G\in \mathcal{G}$, and $p,q\geq 0$. Then, any solution of (\[io\]) can be specified by giving $(\theta_x,p_x,q_x,D_x,G_x)$. Using the fact that $G \cdot N_\theta
\succeq \nv$ for all $\theta$ and $G\in \mathcal{G}$, all solutions of (\[io\]) must satisfy $$\label{D}
N_{\pi/4} \succeq \int\!\! dx
\ p_x\, D_x \cdot N_{\theta\!_x}\ .$$ Recall that this component-wise inequality entails four inequalities. Adding them together we obtain the condition $\int dx \, p_x \leq 1$. Now, denote by $\N$ the set of all vectors obtained by permuting the components of $N_\theta$ (\[N\]) when $\theta$ runs through $[0,\pi/4]$. The convex hull of $\N$ ($\co~\N$) is precisely the set of vectors that can be written as the right-hand side of (\[D\]) under the constraint $\int dx \, p_x \leq 1$. The first inequality of (\[D\]) is $$\label{N1}
1-\sqrt{2} \geq N^1\ ,$$ where $N^1$ is the first component of $N\in \co~\N$. All vectors $N\in
\co~\N$ satisfy the inequality $1-\sqrt{2} \leq N^1$, and only $N_{\pi/4}$ saturates it. Hence, the only possible value for the right-hand side of (\[D\]) is $N_{\pi/4}$. Substituting this into , and again using , we obtain $-\int dx\ q_x\, G_x
\cdot N_{\theta\!_x} \succeq \nv$. But as said above, $G \cdot
N_{\theta} \succeq \nv$ for all $\theta$ and $G\in\mathcal{G}$, which implies that for any solution $\int dx\ q_x\, G_x \cdot N_{\theta\!_x}
= \nv$. Therefore, inequality (\[io\]) becomes $N_{\pi/4} - M_0
\cdot N_{\pi/4} \succeq \nv$, where $M_0$ is a doubly-stochastic matrix such that $M_0\cdot N_{\pi/4}=N_{\pi/4}$. Due to the form of $N_{\pi/4}$, doubly-stochastic matrices that satisfy this equality must have the form of $$\label{M_0} M_0 = \left[
\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & 1- \eta & \eta & 0 \\
0 & \eta & 1- \eta & 0 \\
0 & 0 & 0 & 1 \\
\end{array}
\right]\ ,$$ where $\eta\in [0,1]$.
The fact that each of the four components of the left-hand side of (\[io\]) is zero, implies that the left-hand side of (\[uk\]) is traceless for all $i$. The only positive matrix with zero trace is the null matrix, therefore $$\label{fin}
\mu\t N_{\pi/4}^i =
\sum_{j=1}^4\, \omega_0^{ij} N_{\pi/4}^j,\quad i=1,2,3,4,$$ where $\omega_0$ is any $\omega_x$ that gives rise to $M_0$. Using the same argument, the pairs $(i,j)$ for which $M_0^{ij}=0$ are such that $\omega_0^{ij}=0$. Therefore, by adding the equalities in (\[fin\]) corresponding to $i=2,3$, and using the definition of $\omega_0^{ij}$ in , we obtain $$\label{fin2}
2\, \mu\t = \tr_{\A'' \B''}\!\left[\left(\id\o\Psi\right)
\Omega_0(\Psi)\right]\ ,$$ where $\Psi=\Pi_2+ \Pi_3$, and $\Omega_0$ is any $\Omega_x$ that gives rise to $\omega_0$. Using the PPT criterion [@Hppt] one can check that the (unnormalized) two-qubit state $\Psi$ is a separable state. Equation (\[fin2\]) implies that $\mu\t$ is the output of a separable map applied to a separable input state, and hence is a separable state as we wanted to prove. $\Box$
, for each entangled state $\sigma$ there exists a protocol (involving the state $\rho$ associated with the theorem) in which $\sigma$ cannot be substituted by an arbitrarily large amount of classical correlations, without changing the result:
This provides us with a new interpretation of entanglement in terms of the behavior of the states, in contrast with the usual definition in terms of the preparation of the states.
Differently, one can be interested in the set of bipartite states $\sigma$ which do not admit a LHVM description in scenarios where no other kind entanglement is present. That is, $\sigma$ may be processed with more copies of itself $\sigma^{\otimes n}$ but never with different entangled states $\rho$. Following [@Masymp] we say that a state $\sigma$ violates a Bell inequality asymptotically, if after jointly processing by LOCC a sufficiently large number of copies of a state, the result violates the Bell inequality. In Ref. [@Masymp] it is proven that the states violating CHSH asymptotically are the distillable ones. This, together with the results of the present paper, establishes an appealing picture: $$\begin{aligned}
\mbox{entangled} &\Longleftrightarrow& \mbox{non-simulable} \\
\mbox{distillable} &\Longleftrightarrow&
\mbox{asymptotic violation of CHSH}\end{aligned}$$\
Entangled states are the ones that cannot be generated from classical correlations plus local quantum operations. We show that in the bipartite case, one can equivalently define entangled states as the ones that cannot be simulated by classical correlations.\
The authors are thankful to Fernando Brandão and Nick Jones for useful comments. This work is supported by the EU Project QAP (IST-3-015848) and the Australian Research Council.
[99]{}
J. S. Bell, Physics [**1**]{}, 195 (1964).
R. F. Werner, Phys. Rev. A [**40**]{}, 4277 (1989).
J. Barrett, Phis. Rev. A [**65**]{}, 042302 (2002).
N. Gisin, Phys. Lett. A, [**162**]{}, 15 (1992).
S. Popescu and D. Rohrlich, Phys. Lett. A [**166**]{}, 293 (1992).
S. Popescu, , 2619 (1995)
N. Gisin, Phys. Lett. A [**210**]{}, 151 (1996).
A. Peres, Phys. Rev. A [**54**]{}, 2685 (1996).
Ll. Masanes, Phys. Rev. Lett. 97, 050503 (2006).
J. F. Clauser, M. A. Horne, A. Shimony, and R. A. Holt, Phys. Rev. Lett. [**23**]{}, 880 (1969).
Ll. Masanes, Phys. Rev. Lett. 96, 150501 (2006); Ll. Masanes, quant-ph/0510188.
A. Fine, , 291 (1982).
A. C. Doherty, Y.-C. Liang and Ll. Masanes [*In preparation.*]{}
F. Verstraete and M. M. Wolf, Phys. Rev. Lett. [**89**]{}, 170401 (2002).
R. Horodecki, P. Horodecki, and M. Horodecki, A, [**200**]{}, 340 (1995).
R. Bhatia, [*Matrix Analysis*]{}, Springer-Verlag, N.Y.
B. D. Craven, J. J. Koliha, SIAM J. Math. Anal. [ **8**]{}, pp. 983 (1977).
A. Peres, , 1413 (1995); M. Horodecki, P. Horodecki, R. Horodecki, Phys. Lett. A [**223**]{}, 1 (1996).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Canards are special solutions to ordinary differential equations that follow invariant repelling slow manifolds for long time intervals. In realistic biophysical single cell models, canards are responsible for several complex neural rhythms observed experimentally, but their existence and role in spatially-extended systems is largely unexplored. We describe a novel type of coherent structure in which a spatial pattern displays temporal canard behavior. Using interfacial dynamics and geometric singular perturbation theory, we classify spatio-temporal canards and give conditions for the existence of folded-saddle and folded-node canards. We find that spatio-temporal canards are robust to changes in the synaptic connectivity and firing rate. The theory correctly predicts the existence of spatio-temporal canards with octahedral symmetry in a neural field model posed on the unit sphere.'
author:
- 'D. Avitabile'
- 'M. Desroches'
- 'E. Knobloch'
bibliography:
- 'referencesFinal.bib'
title: 'Spatio-temporal canards in neural field equations'
---
\[sec:intro\]Introduction
=========================
Spatially extended, continuum, deterministic neural field models take the form [@Bressloff2012o; @Bressloff:2014cm; @Ermentrout:2010cga] $$\label{eq:neuralFieldModel}
\partial_t u(x,t) = - u(x,t) + \int_{{\mathbb{R}}} W(x,y) f(u(y,t) - h) \, d y,$$ where $u$ denotes the coarse-grained activity of a neural population at position $x\in {\mathbb{R}}$ and time $t \in {\mathbb{R}}^+$, $W$ is a synaptic kernel modelling the strength of connections from neurons at positions $y$ to those at position $x$, $f$ is a firing rate function converting neural activity into synaptic inputs and $h$ is a firing rate threshold. Nonlocal equations of this type, originally proposed by Wilson and Cowan [@Wilson1972aa] and Amari [@Amari1975aa], provide a coarse-grained model of macroscopic brain activity [@Folias2005aa], and have been used to explain experimental observations of cortical waves in vitro [@Richardson:2005cs] and in vivo [@Huang:2004kw; @GonzalezRamirez:2015gk], as well as electroencephalogram recordings [@SteynRoss:2003ep] and feature selectivity in the primary visual cortex [@Camperi:1998ji].
In this article we demonstrate that neural fields described by Eq. (\[eq:neuralFieldModel\]) support generically a novel type of coherent structure, in which a spatial pattern displays temporal canard behavior [@Benoit1981]. We refer to these solutions as *spatio-temporal canards*. Canards are considered to be a footprint of time scale separation in ordinary differential equations (ODEs): these special solutions follow (locally) invariant repelling slow manifolds for long time intervals, and manifest themselves via $O(1)$ amplitude changes that take place within an exponentially small range of parameter values. In planar systems, this brutal growth of solutions is referred to as a *canard explosion* [@Benoit1981; @Krupa2001].
It is widely accepted that canards have a *functional* role in biophysical single-neuron models of Hodgkin–Huxley-type, where they approximate excitability thresholds [@Desroches2013b; @Mitry2013] and organise abrupt transitions from resting to spiking states [@Moehlis2006], or from spiking to bursting regimes [@Kramer2008; @Rinzel1986]. In addition, canards underpin complex neural rhythms such as mixed-mode oscillations [@Desroches2012] or spike-adding phenomena [@Desroches2013] [in bursters]{}.
[An intriguing open question concerns the existence and role of temporal canards in spatially extended dynamical systems with time scale separation. Numerical simulations indicate that canards do indeed exist in such systems [@Gandhi16], but the absence of a rigorous geometric singular perturbation theory near non-hyperbolic slow manifolds for infinite-dimensional dynamical systems requires that the interpretation of such computations be treated with caution. The reduction procedure described in the first part of this paper overcomes this difficulty in a key example.]{}
In this article we identify canards in neural field models of the type shown in Eq. . When the threshold $h$ is constant, the neural field admits an $h$-dependent family of coexisting stationary localized solutions organised along a branch with one or more folds [@Laing2002aa; @Coombes2003aa]. When $h$ varies slowly with respect to the macroscopic characteristic time $t$ of Eq. the system may drift along this branch of equilibria but abrupt transitions, [excitable]{} dynamics on the faster time scale $t$, may occur in the vicinity of the folds where the state of the system ‘jumps’ to a different state. We remark that the time scales of interest in our study are different from those used in previous work on canards: these structures have thus far only been found when there exists a time scale separation at the level of a single cell, between the membrane potential (fast) and gating variables (slow); in the present study we find canards in neural fields, which are coarse-grained models of neural networks, and the time scale separation is between the threshold crossing dynamics (slow) and the activity variable $u$ (fast). Our findings can be summarized as follows: (i) If the firing threshold $h$ varies slowly, complex spatio-temporal patterns containing canard segments exist for steep firing rates $f$ and for generic choices of the synaptic kernel $W$; (ii) A theory for the classification of such spatio-temporal canards can be derived using interfacial dynamics; (iii) Spatio-temporal canards of *folded-node* or *folded-saddle* types are present, depending on the coupling between $h$ and $u$; (iv) The behavior described above is robust to changes in the synaptic kernel $W$ and to perturbations in the firing rate function $f$.
\[sec:interface\]Interface dynamics
===================================
The interfacial description [@Coombes2012aa] applies in the case $f(u)=\Theta(u)$, where $\Theta(u)$ is the Heaviside step function. As customary, we consider localized regions of activity $-\xi(t)\le x\le\xi(t)$, where the interfaces (or threshold crossings) $x=\pm \xi(t)$ are defined by the level set conditions $u(\pm \xi(t),t) = h(t)$ with $\partial_x u(\pm \xi(t),t) \lessgtr 0$, for all $t \in {\mathbb{R}}^+$, and take their width $2 \xi(t)$ as a measure of the spatial extent of the solution (see for instance Fig. \[fig:figs1-2\](b)). Integrating , we find that solutions $u(x,t)$ can be expressed in terms of the interfacial functions $\xi(t)$ and the initial datum $u(x,0)$, $$\label{eq:uSolution}
u(x,t) = e^{-t} u(x,0) + \int_0^t \int_{-\xi(s)}^{\xi(s)} e^{s-t} W(x,y) \, dy \, ds.$$
[The approach of Refs. [@Coombes2011aa; @Coombes2012aa] can be extended to the case of time-dependent $h$. In this case differentiation of the level set condition for $\xi$ with respect to time leads to a closed scalar evolution equation for the half-width of the pattern. Using we obtain]{} $$\label{eq:xiEvolution}
\varphi(\xi,t) \dot \xi = h + \dot h - \psi(\xi),$$ where $\varphi(\xi,t) = \partial_x u(\xi,t)$ and $\psi(\xi) = \int_{-\xi}^\xi W(\xi,y) \, dy$. By hypothesis $\varphi$ is strictly negative at all times. The function $\psi$ encodes the neural connectivity of the model, as it depends solely on the synaptic kernel $W$, which models arbitrary heterogeneous synaptic circuits. Figure \[fig:figs1-2\](d) shows $\psi(\xi)$ for several commonly used kernels $W(\xi,y)$ and [highlights that $\psi$ generically possesses folds. These are marked by circles in Fig. \[fig:figs1-2\](d) and correspond to locations where $\psi'=0$.]{} Equation represents an exact reduction of the field equation for $u$ with a time-dependent threshold and Heaviside firing rate, and constitutes a key tool for the study of spatio-temporal canards.
If $h$ is a *constant* control parameter, Eq. admits equilibria for all $h$ and $\xi$ such that $h = \psi(\xi)$. In other words, the curves in Fig. \[fig:figs1-2\](d) can be interpreted as branches of steady (patterned) states of the full system with the parameter $h$ identified with $\psi(\xi)$ in Fig. \[fig:figs1-2\](d). This strategy for constructing patterns, contained in the original work of Amari [@Amari1977aa], can be extended also to study stability: to each fold of $\psi$ corresponds a saddle-node bifurcation of the full system. In Ref. [@Avitabile2015aa] it was shown that [sinusoidal modulation of the kernel in space generates an infinite number of saddle-nodes organized]{} in a *snakes-and-ladders* bifurcation structure [@Knobloch15].
The firing rate threshold parameter, $h$, is therefore a common continuation parameter in neural field studies: as $h$ is varied, we obtain branches of patterned stationary states and, depending on the choice of the kernel, secondary symmetry-breaking bifurcations may occur. It is therefore natural to search for canards in cases where $h$ is slowly varying. Variations of $h$ have been considered before in the literature: in Ref. [@Brackley2007aa; @Thul:2016gr], the firing threshold was subject to fluctuations induced by noise, decoupled from the network activity; in Refs. [@Coombes:2005hp; @coombes2007exotic] the threshold $h$ was coupled directly to the local value of $u$, in order to mimic spike-frequency adaptation, observed experimentally in *in vitro* experiments of rat pyramidal neurons [@Madison:1984di].
In the following we study spatio-temporal canards by combining a spatially-extended neural field with a slowly-varying oscillatory threshold $h(t)$ which may arise, for instance, from the competition between *adaptation* and *facilitation* processes, coupled to the neural field via the macroscopic width of the pattern, and describe a simple example of the dynamics that result when $h$ evolves on a slow time scale. Depending on the choice of control parameters, we consider limits where $h$ influences $u$ (but not vice-versa), as well as cases where the dynamics of $h$ and $u$ are fully coupled, as previously done in Refs. [@Brackley2007aa; @Thul:2016gr] and [@Coombes:2005hp; @coombes2007exotic], respectively. Specifically, we study the extended neural field model $$\label{eq:extendedNeuralField}
\begin{aligned}
& \partial_t u(x,t) = - u(x,t) +\!\! \int_{{\mathbb{R}}} W(x,y)\Theta
\big[u(y,t)\!-\!h(t)\! \big]dy,
\\
& \ddot h(t) + {\varepsilon}^2 h(t) = {\varepsilon}^2 ( \alpha + \beta \xi(t)) + {\varepsilon}\gamma \dot \xi(t), \\
& \xi(t) = \frac{1}{2} \int_{\mathbb{R}}\Theta \big[ u(y,t) - h(t) \big]
\, dy.
\end{aligned}$$ Thus $h$ obeys a weakly forced oscillator equation with a low natural frequency ${\varepsilon}$ that is coupled to the neural field via both $\xi$ and $\dot \xi$. In Figs. \[fig:figs1-2\](a)–(c) we show direct simulations of the model , displaying strong sensitivity to changes in the parameter $\alpha$. We will show below that spatio-temporal canards organise abrupt transitions between branches patterned states, and are therefore responsible for the behavior shown in Figs. \[fig:figs1-2\](a)–(c).
Interactions between excitable systems and slow oscillations are known to produce canard-type dynamics in ODEs with folded saddles [@Mitry2013; @desroches2016spike]. This type of interaction motivated our choice of the coupling in model , which indeed produces canards in a spatially-extended system. In terms of the slow time $\tau = {\varepsilon}t$, system is equivalent to $$\label{eq:rescaledPhi}
\begin{aligned}
{\varepsilon}| \varphi_{\varepsilon}( \xi, \tau) | \dot \xi & = \psi(\xi) - h - {\varepsilon}(q + \gamma \xi), \\
\dot h & = q + \gamma \xi, \\
\dot q & = \alpha + \beta \xi - h,
\end{aligned}$$ where $\varphi_{\varepsilon}$ is a rescaled version of $\varphi$ and we used the fact that $\varphi$ and $\varphi_{\varepsilon}$ are both strictly negative at all times. Crucially, we passed from model , which involves an evolution equation for the scalar field $u(x,t)$, to model , whose state variables are the scalars $(\xi(t),h(t))$. Since $\lim_{{\varepsilon}\to 0^+} {\varepsilon}| \varphi_{\varepsilon}( \xi, \tau) | = 0$ for all $\tau\in {\mathbb{R}}^+$, Eqs. take the form of a singularly perturbed system, with one fast variable $\xi$ and two slow variables $h$ and $q$. An important object for understanding the dynamics of such systems is the *critical manifold* $S^0$, defined as the ${\varepsilon}=0$ limit of the fast nullsurface [@Krupa2001]. In the present case, this manifold is the folded surface $\{(h,q,\xi) \in {\mathbb{R}}^3 \colon h = \psi(\xi)\}$. The limit yields the differential-algebraic system $$\label{eq:DAE}
\begin{aligned}
0 & = \psi(\xi) - h,\\
\dot h & = q + \gamma \xi, \\
\dot q & = \alpha + \beta \xi - h,
\end{aligned}$$ or equivalently the reduced system (or slow subsystem) $$\label{eq:slowFlow}
\begin{aligned}
-\psi'(\xi) \dot \xi & = - q - \gamma \xi,\\
\dot q & = \alpha + \beta \xi - \psi(\xi).
\end{aligned}$$ This system is singular when $\psi'(\xi) = 0$, that is, at the folds of the critical manifold separating attracting sheets from repelling ones. For the problem under consideration, the singularity occurs at fold lines $\{ (h, q, \xi_*) \in {\mathbb{R}}^3
\colon h = \psi(\xi_*),\; \psi'(\xi_*) = 0\}$; in passing we note that $\xi_*$ can be any of the folds marked by circles in Fig. \[fig:figs1-2\](d). It is possible to remove this singularity by rescaling time by the factor $-\psi'(\xi)$, leading to the desingularised reduced system (DRS) $$\label{eq:desingSlowFlow}
\begin{aligned}
\dot \xi & = -q - \gamma \xi,\\
\dot q & = \psi'(\xi) \big[ \psi(\xi) - \alpha -\beta \xi\big].
\end{aligned}$$ We carry out this rescaling because it is helpful in deciphering the flow of system near the fold lines. Indeed, the rescaling has two major consequences: (i) Orbits of system are extended in system to the fold lines, where is undefined; (ii) System may possess equilibria on the fold lines. As we shall see below, these equilibria are related to canards in system .
Folded singularities and canards in the extended system {#sec:canards}
=======================================================
System has an equilibrium at $(\xi_*,-\gamma \xi_*)$, where $\xi_*$ satisfies $\psi'(\xi_*) = 0$, i.e., on a fold line of the surface [$S^0$]{}. This is not an equilibrium of the reduced system because of the time rescaling by $-\psi'(\xi)$, which reverses the orientation of trajectories on the repelling sheets of $S^0$. Therefore solutions to the reduced system approach the point $(\xi_*,-\gamma \xi_*)$ along an attracting sheet of $S^0$, cross it in finite time, and continue to flow along a repelling sheet of $S^0$: these solutions of system are called *singular canards* and persist for small ${\varepsilon}>0$ as *canard solutions* of system , and hence as *spatio-temporal canards* of system .
Equilibria $(\xi_*,-\gamma \xi_*)$ of the DRS are called *folded singularities* (of node, saddle or focus type) and are therefore important [in the classification of canards]{}. Other equilibria of the DRS may exist as true equilibria of the reduced system : these states are not generically related to canards and are not considered here. The Jacobian at $(\xi_*,-\gamma \xi_*)$ is given by $$\begin{pmatrix}
- \gamma & -1 \\
\pi(\xi_*) & 0
\end{pmatrix},
\qquad
\pi(\xi_*) = \psi''(\xi_*)\big[ \psi(\xi_*) - \alpha -\beta \xi_* \big],$$ and hence [$(\xi_*,-\gamma \xi_*)$ is either (i) a folded saddle (if $\pi(\xi_*) < 0$) or (ii) a folded node (if $0 < \pi(\xi_*) < \gamma^2/4$), corresponding in to (i) excitable dynamics and (ii) mixed-mode dynamics.]{}
[Classical theory [@Desroches2012] now guarantees the presence of canards in , and these correspond]{} to spatio-temporal canards in for sufficiently small ${\varepsilon}>0$, close to the above-mentioned folded singularities.
We have confirmed these predictions using the full model with the heterogeneous synaptic kernel $$\label{eq:heterokernel}
W(x,y) = \frac{1}{2} e^{-|x-y|} \Big(a + b \cos \frac{y}{\Lambda} \Big),$$ where $a,b \geq 0$, $\Lambda > 0$ and the firing rate function $$f(u) = (1 + e^{-\mu u})^{-1}.\label{eq:firingRateSupp}$$ For $\mu \gg 1$ this sigmoidal function approximates a Heaviside firing rate employed in the theory. We use the spectral algorithm developed in Ref. [@Rankin:2014bz] to solve the resulting equations. System , where $h$ is a fixed parameter, admits branches of localized steady states arranged in a characteristic *snakes-and-ladders* bifurcation structure exhibiting countably many folds at which $\psi'(\xi_*) = 0$ (Fig. \[fig:figs1-2\](d)).
Spatio-temporal folded-saddle canards
-------------------------------------
We first consider the uncoupled case with $\alpha=0.5$, $\beta=\gamma=0$, which leads to spatio-temporal folded-saddle canards. Figure \[fig:foldedSingCanards\](a) shows the solution of the full spatial system in the form of a space-time plot while Fig. \[fig:foldedSingCanards\](b) shows the same results but projected onto the [$(h,q,\xi)$]{} space (blue curve), compared with the singular canards of (red curves). For reference we plot $S^0$ in grey. In Fig. \[fig:foldedSingCanards\](c) we show a projection onto the $(\xi,q)$ plane, where we indicate folded saddles (open circles) and the attracting (A) and repelling (R) sheets of $S^0$. For these parameters the theory predicts the presence of a folded-saddle spatio-temporal canard in system and the projection indeed displays behavior typical of folded-saddle singularities in ODEs: the orbit follows the upper attracting sheet, passes the folded singularity from right to left (Fig. \[fig:foldedSingCanards\](c)) and then continues near a repelling sheet of $S^0$ for an $O(1)$ time, before a fast (anterior) jump leads to the lower attracting sheet; the orbit returns to the upper attracting sheet with a second (posterior) fast jump.
Since the latter jump occurs near a folded saddle, this opens the possibility of a *jump-on canard* segment, in which the orbit jumps and lands on the upper repelling sheet of $S^0$ before returning to the upper attracting one. We observe this behavior in canard cycles obtained with slightly different parameter values, as reported in Figs. \[fig:foldedSaddleSuppl\](a)–(b): the solution is periodic and passes near four different folded-saddle singularities, marked with red circles in Fig. \[fig:foldedSaddleSuppl\](b); this trajectory contains one clear canard segment near the topmost folded saddle; this segment is a jump-on canard, as the orbit makes a fast upward jump and then follows directly a repelling segment along the maximal canard of the folded saddle. The trajectory then passes near the other folded saddles, without displaying a clear canard segment.
In Figs. \[fig:foldedSaddleSuppl\](c)–(d) we present a family of solutions of Eqs. for different initial conditions near an attracting sheet of $S^0$. This experiment explains the sensitivity documented in Figs. \[fig:figs1-2\](a)–(c), and highlights the transition through the canard in the folded-saddle case. This corresponds—modulo a change of direction near the repelling sheets of $S^0$—to a perturbation of the stable manifold of the folded saddle (as a saddle equilibrium of the DRS). Indeed, trajectories approach the canard and follow it past the folded saddle; the trajectories are then repelled and jump to a lower or upper attracting sheet of $S^0$, depending on their initial condition. In Fig. \[fig:foldedSaddleSuppl\](d) we plot the singular canards associated with this folded saddle: the true canard (from A to R) and the so-called “false” (or *faux*) canard (from R to A), both shown in red. In this scenario, the true canard plays the role of a separatrix between trajectories that jump upwards, following the faux canard, and downwards, towards a different attracting sheet of $S^0$.
Spatio-temporal folded-node canards
-----------------------------------
{width="\textwidth"}
We next repeat our numerical analysis for the fully coupled system when $\alpha=1$, $\beta=0$, $\gamma=1$ for which the theory predicts spatio-temporal canards of folded-node type (Fig. \[fig:multiOscillationsFoldedNode\](a)–(c)). The folded-node scenario is richer than the folded-saddle one: first, solutions containing canard segments exist for $O(1)$ ranges of initial conditions and parameter values; second, there are many more possible waveforms due to the existence of a funnel region around the folded-node singularity that induces a rotation of the trajectories as they pass through it; this effect is clearly visible in Fig. \[fig:multiOscillationsFoldedNode\](b)–(c) as small amplitude spiraling motion in the vicinity of the fold, and rather less clearly as the minute oscillations for $t \in [0, 1.5]$ in Fig. \[fig:multiOscillationsFoldedNode\](a). As initial conditions change, the number of these small (subthreshold) oscillations in the funnel region varies and this phenomenon defines rotation sectors near $S^0$. The boundaries between different rotation sectors correspond to canard solutions generating mixed-mode dynamics in the system. For fixed parameter values, the maximum number of subthreshold oscillations is given by the eigenvalue ratio of the folded node [@Desroches2012], seen as an equilibrium of Eqs. . Thus trajectories with different initial conditions will be trapped in the funnel and pass near the folded node while making different numbers of subthreshold oscillations, thereby encoding the possible waveforms in this regime.
We exemplify this behavior in Figs. \[fig:multiOscillationsFoldedNode\](d)–(i) by time-stepping with slightly different initial conditions, close to a folded node, when $\alpha=1$, $\beta=0$, $\gamma=0.7$. In the experiment under consideration we pre-computed a stationary pattern $u_0(x)$ for the neural field equation with constant firing rate threshold, $h = 0.57$, $\dot h=0$, that is, we select a stationary state on $S^0$. We then perturb this state and compute two trajectories, with initial conditions close to the folded node by setting $u(x,0) =
u_0(x)$, $h(0) = 0.58$, $q(0) = -18.40$ (label 1) and $q(0) = -18.35$ (label 2). [Figures \[fig:multiOscillationsFoldedNode\](d,g)]{} show the corresponding space-time evolution $u(x,t)$ while Figs. \[fig:multiOscillationsFoldedNode\](e,h) show the corresponding trajectories in $(h,q,\xi)$ space. Panel (c) and the enlargement in (f) show the projections of these trajectories on the $(q,\xi)$ plane. The trajectories are initially close and exhibit the drifting and spiralling motion predicted by the theory [@Desroches2012], with respectively three and five subthreshold oscillations near the folded node. After an initial transient, in which trajectory 1 visits the upper attracting sheet of $S^0$, both trajectories wrap clockwise around the middle and bottom attracting sheets of $S^0$ (Figs. \[fig:multiOscillationsFoldedNode\](e,h)). Both also display jump-on canard segments at every turn, in the vicinity of the left boundary of the upper repelling sheet of $S^0$, although these become less pronounced as time increases.
Neural fields posed on a sphere
===============================
We have also studied neural field models posed on a more realistic spherical domain and identified spatio-temporal canards with octahedral symmetry where interfaces are no longer points but curves in 3D. The above theory does not readily generalize to this setting but we nevertheless successfully tested its predictions in the folded-saddle case, when $u$ and $h$ are decoupled. As shown in Fig. \[fig:spheresBifDiag\], the model displays orbits with canard segments (and canard cycles). In this case the system with constant $h$ admits an intricate bifurcation diagram (not shown), where coexisting stable states with octahedral, icosahedral and rotational symmetry are interconnected via symmetry-breaking bifurcations and saddle-node bifurcations.
The calculations for neural fields posed on the unit sphere were performed using a neural field model with a constant threshold crossing $h$, $$\begin{aligned}
\label{eq:sphereModelSupp}
\partial_t u({\ensuremath{\bm x}},t) &= -u({\ensuremath{\bm x}},t)\\
\nonumber &+ \kappa \int_{{\mathbb{S}^2}} W\big( \langle {\ensuremath{\bm x}},{\ensuremath{\bm y}} \rangle \big) f\big(
u({\ensuremath{\bm y}},t) - h\big) \, d\sigma({\ensuremath{\bm y}}), \end{aligned}$$ where ${\ensuremath{\bm x}} \in {\mathbb{S}^2}= \{ {\ensuremath{\bm z}} \in {\mathbb{R}}^3 \colon |
{\ensuremath{\bm z}} | = 1 \}$ and the integral is over ${\mathbb{S}^2}$. In this integro-differential equation the kernel $W$ models the synaptic wiring between two points ${\ensuremath{\bm x}}, {\ensuremath{\bm y}}$ on the surface of a sphere; we assume that this wiring depends solely on the great-circle distance (geodesic) between ${\ensuremath{\bm x}}$ and ${\ensuremath{\bm y}}$, hence the dependence on the scalar product $\langle
{\ensuremath{\bm x}},{\ensuremath{\bm y}} \rangle$. We use an excitatory-inhibitory Gaussian synaptic kernel $$\label{eq:sphereKernel}
W(\xi) = A_1 \exp(-\xi^2 /B_1) - A_2 \exp(-\xi^2 /B_2).$$ Stationary patterned states of were continued in the parameter $h$ using a Nyström scheme, combined with standard path-following techniques [as well as]{} high-order, highly efficient, icosahedral- or tetrahedral-invariant quadrature schemes. A comprehensive study of branches of patterned states supported by this model, their symmetries and stability, as well as the properties of the numerical scheme will be described in a separate publication [@Avitabile:2016]. A sample result showing a branch of states with octahedral symmetry is reported in Fig. \[fig:spheresBifDiag\](b). Solutions with this symmetry bifurcate transcritically from the homogeneous steady state, and then undergo a sequence of saddle-nodes and symmetry-breaking bifurcations shown in the figure.
We are interested in testing the predictions of the theory developed for 1D domains for more realistic cortical surfaces. For physical domains in higher dimensions, it is possible to reduce the equations as for 1D domains, but the reduction is still a spatially-extended dynamical system. In 1D, the activity set is given by $\mathcal{A}(t) = [-\xi(t), \xi(t)] \in {\mathbb{R}}$ and, differentiating one of the threshold conditions, say $u(\xi(t),t) = h(t)$, we obtain $$\label{eq:evolEq1DSupp}
\partial_x u(\xi(t),t) \dot \xi(t) = h(t) + \dot h(t) - \int_{-\xi(t)}^{\xi(t)}
W(x,y)\, dy$$ which is an evolution equation for the scalar variable $\xi$. To extend this procedure to the sphere, we assume that the activity set $
\mathcal{A}(t) = \{ {\ensuremath{\bm x}} \in {\mathbb{S}^2}\colon u({\ensuremath{\bm x}},t) \geq h(t) \}
$ has a boundary which can be parameterized as follows, $$\begin{aligned}
\partial \mathcal{A}(t) & = \bigcup_{k = 1}^K \mathcal{C}_k(t),\\
\mathcal{C}_k(t) & = \{ {\ensuremath{\bm x}} \in {\mathbb{S}^2}\colon {{\ensuremath{\bm x}}}= {{\ensuremath{\bm \xi}}}_k(s,t), \; s \in [0,2\pi) \},\end{aligned}$$ where the functions $\{ {{\ensuremath{\bm \xi}}}_k \}$ are $2\pi$-periodic and smooth in the variable $s$. In other words, we assume that the boundary of the activity set on the spherical domain is the union of $K$ disjoint curves on the spherical surface ${\mathbb{S}^2}$. We seek evolution equations for the functions $\{ {{\ensuremath{\bm \xi}}}_k \}$. Since the solution $u({{\ensuremath{\bm x}}},t)$ crosses the threshold $h(t)$ on each of the curves $\mathcal{C}_k$, we differentiate the threshold condition $u({{\ensuremath{\bm \xi}}}_k(s,t),t) = h(t)$ with respect to $t$ to obtain $$\begin{aligned}
\hspace*{-1cm}\label{eq:evolEqSupp}\big\langle {{\ensuremath{\bm \nabla}}}u \big({{\ensuremath{\bm \xi}}}_k(s,t),t\big),& \; \partial_t {{\ensuremath{\bm \xi}}}_k(s,t) \big\rangle = h(t) + \dot h(t)\\
\nonumber& - \int_{\mathcal{A}(t)} W\big( \langle {{\ensuremath{\bm \xi}}}_k(s,t), {{\ensuremath{\bm y}}}\rangle \big) \, d\sigma({{\ensuremath{\bm y}}}),\\
\nonumber& s \in [0,2\pi),\quad k = 1,\ldots,K,\\
{{\ensuremath{\bm \xi}}}_k(0,t) & = {{\ensuremath{\bm \xi}}}_k(2\pi,t), \quad k = 1,\ldots,K,
\label{eq:BCsSupp}\end{aligned}$$ where the gradient is in spherical coordinates. It can be shown that, under suitable assumptions on the kernel, the inner product on the left hand side and the surface integral on the right hand side of can be written [@Coombes2012aa; @Coombes2013aa] in terms of line integrals over the closed curves $\mathcal{C}_k(t)$. The system – is therefore closed and represents a generalization of . In this case, however, the state variables are the functions $\{ {{\ensuremath{\bm \xi}}}_k(s) \}$, as opposed to the scalar $\xi$, and a canard theory for this system is currently unavailable.
We can, however, simulate the system or the system – numerically and search for evidence of spatio-temporal canards. More precisely, we have performed numerical experiments to test the robustness of the 1D theory to
1. Changes in the geometry of the problem: the spherical model includes curvature effects via the great-circle distance $\langle {\ensuremath{\bm x}}, {\ensuremath{\bm y}} \rangle$ between points ${\ensuremath{\bm x}}$, ${\ensuremath{\bm y}}$ on the spherical cortex (see Eq. ).
2. Changes in the synaptic connectivity function: the kernel is different from that used in the 1D computations; in particular, the kernel is excitatory-inhibitory and homogeneous while kernel is purely excitatory and heterogeneous.
3. Changes in the firing rate function: the theory is valid for a Heaviside firing rate which is approximated in the 1D simulations by a steep sigmoid (Eq. with $\mu=50$); in the spherical simulations we employ a shallow firing rate ($\mu = 8$).
4. Changes in the evolution equation of the firing threshold $h$: in the spherical simulations, $h$ evolves slowly and independently from $u$, but not harmonically: $$\label{eq:hEvolution}
h(t) =
\begin{cases}
{\varepsilon}t + h_0 & 0 \leq t < (h_1 - h_0)/{\varepsilon}\\
-{\varepsilon}t + 2h_1 - h_0 & t \geq (h_1 - h_0)/{\varepsilon},
\end{cases}$$ where $h_1$ is a fold point in the bifurcation diagram (located using standard bifurcation analysis techniques) and $h_0<h_1$. Consequently, $h$ undergoes a slow linear increase up to the fold, followed by a slow linear decrease.
In each case we found that the qualitative predictions of the 1D theory carried over to this much more complicated situation.
Conclusions and perspectives
============================
To the best of our knowledge, this article presents the first theory for folded-singularity temporal canards in a spatially-extended system. This result paves the way towards a systematic study of spatio-temporal mixed-mode oscillations (MMOs) in spatially-extended systems, [with the view of explaining the origin of MMOs observed in spatio-temporal signals modelling spike-frequency adaptation and synaptic depression [@Folias2005aaa].]{} [The spatio-temporal structures discussed here are also directly relevant to neural mass and connectomic models, in which a discrete connectomic matrix replaces the heterogeneous kernel $W$ [@Haimovici2013]: canard structures in these models would offer a rigorous explanation of the brutal transitions observed, for instance, in models of partial epilepsy [@proix2014]]{}. There is a general consensus that spike (and more generally burst) timings, durations and rates are involved in information coding in the brain [@borst1999]. Being able to identify boundaries (represented by spatio-temporal canards) between different activity regimes (e.g. spiking/bursting or mixed-mode oscillations with different signatures) may shed further light on the transmission of information in the brain. **Acknowledgement**: This work was supported in part by the Engineering and Physical Sciences Research Council under grant EP/P510993/1 (DA) and by the National Science Foundation under grant DMS-1613132 (EK). DA thanks Luke Wood and Oliver Smith [for their work on neural field models during their final-year undergraduate dissertations.]{}
**Author contributions**: DA and MD contributed equally to this work.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the results of a near-infrared (NIR) spectroscopic follow-up survey of 182 M4–L7 low-mass stars and brown dwarfs (BDs) from the *BANYAN All-Sky Survey* (*BASS*) for candidate members of nearby, young moving groups (YMGs). We confirm signs of low-gravity for 42 new BD discoveries with estimated masses between 8–75[$M_{\mathrm{Jup}}$]{} and identify previously unrecognized signs of low gravity for 24 known BDs. This allows us to refine the fraction of low-gravity dwarfs in the high-probability *BASS* sample to $\sim$82%. We use this unique sample of 66 young BDs, supplemented with 22 young BDs from the literature, to construct new empirical NIR absolute magnitude and color sequences for low-gravity BDs. We show that low-resolution NIR spectroscopy alone cannot differentiate between the ages of YMGs younger than $\sim$120Myr, and that the BT-Settl atmosphere models do not reproduce well the dust clouds in field or low-gravity L-type dwarfs. We obtain a spectroscopic confirmation of low-gravity for 2MASS J14252798–3650229, which is a new $\sim$27[$M_{\mathrm{Jup}}$]{}, L4$\gamma$ bona fide member of AB Doradus. We identify in this work a total of 19 new low-gravity candidate members of YMGs with estimated masses below 13[$M_{\mathrm{Jup}}$]{}, seven of which have kinematically estimated distances within 40pc. These objects will be valuable benchmarks for a detailed atmospheric characterization of planetary-mass objects with the next generation of instruments such as the James Webb Space Telescope. We find 16 strong candidate members of the Tucana-Horologium association with estimated masses between 12.5–14[$M_{\mathrm{Jup}}$]{}, a regime where our study was particularly sensitive. This would indicate that for this association there is at least one isolated object in this mass range for every $17.5_{-5.0}^{+6.6}$ main-sequence stellar member, a number significantly higher than expected based on standard log-normal initial mass function, however in the absence of radial velocity and parallax measurements for all of them, it is likely that this over-density is caused by a number of young interlopers from other moving groups. Finally, as a byproduct of this project, we identify 12 new L0–L5 field BDs, seven of which display peculiar spectroscopic properties.'
author:
- 'Jonathan Gagné, Jacqueline K. Faherty, Kelle L. Cruz, David Lafrenière, René Doyon, Lison Malo, Adam J. Burgasser, Marie-Eve Naud, Étienne Artigau, Sandie Bouchard, John E. Gizis, and Loïc Albert'
title: 'BANYAN. VII. A NEW POPULATION OF YOUNG SUBSTELLAR CANDIDATE MEMBERS OF NEARBY MOVING GROUPS FROM THE BASS SURVEY'
---
INTRODUCTION
============
Young moving groups (YMGs) consist of stars that formed recently ($\lesssim$120Myr) from a molecular cloud and that are too young to have experienced significant gravitational perturbations from their environment. The members of YMGs share similar galactic velocities within a few . The closest and youngest moving groups include the TW Hydrae association (TWA; 5–15Myr; ), $\beta$ Pictoris ($\beta$PMG; 20–26Myr; [@2001ApJ...562L..87Z; @2014ApJ...792...37M; @2014MNRAS.438L..11B]), Tucana-Horologium (THA; 20–40Myr; [@2000AJ....120.1410T; @2000ApJ...535..959Z; @2001ApJ...559..388Z; @2014AJ....147..146K]), Carina (CAR; 20–40Myr; [@2008hsf2.book..757T]), Columba (COL; 20–40Myr; [@2008hsf2.book..757T]), Argus (ARG; 30–50Myr; [@2000MNRAS.317..289M]) and AB Doradus (ABDMG; 110–130Myr; [@2004ApJ...613L..65Z; @2005ApJ...628L..69L; @2013ApJ...766....6B]). The YMGs are ideal laboratories to measure fundamental properties of star formation such as the initial mass function (IMF) because their members are coeval. This is of particular interest in the case of very low-mass stars and substellar-mass objects (spectral types $\geq$M5) since these populations are still poorly characterized. The massive, bright population of YMGs has already been explored, thanks to the *Hipparcos* survey . However, fainter members are hard to identify mainly because of the lack of radial velocity (RV) and trigonometric distance measurements that are necessary to obtain their spacial velocities and galactic positions. Several efforts have been made to identify the very low-mass members of YMGs [@2011MNRAS.411..117K; @2012AJ....144..109S; @2012ApJ...758...56S; @2012ApJ...752...56F; @2013AJ....145....2F; @2013ApJ...774..101R; @2013ApJ...777L..20L; @2014ApJ...788...81M; @2014AJ....147..146K; @2014AJ....147...85R; @2015MNRAS.447.1267M]; however, as of today it is likely that most of them still remain to be identified.
The Bayesian Analysis for Nearby Young AssociatioNs tool[^1] (BANYAN; [@2013ApJ...762...88M]), which is based on naive Bayesian inference, identified promising candidate members of YMGs among a sample of low-mass stars that do not have prior RV or parallax measurements. The BANYAN II tool[^2] (@2014ApJ...783..121G; Paper II hereafter) was subsequently developed to identify substellar candidate members with a similar but improved algorithm. BANYAN II is an expansion on BANYAN I that is focused on very-low mass stars and brown dwarfs (BDs) with spectral types $\geq$M5. The *BANYAN All-Sky Survey* (*BASS*; @2015ApJ...798...73G; Paper V hereafter) was initiated by our team to search for the elusive late-type ($\geq$M5) members of YMGs, using the BANYAN II tool on an all-sky cross-match of the *Two Micron All-Sky Survey* (*2MASS*; [@2006AJ....131.1163S]) with the *AllWISE* survey [@2014ApJ...783..122K]. The *AllWISE* survey is based on a combination of the cryogenic phase of the *Wide-Field Survey Explorer mission* (*WISE*; [@2010AJ....140.1868W]) and the *Near-Earth Object WISE* (*NEOWISE*; [@2011ApJ...743..156M]) post-cryogenic phase.
We present here the results of a near-infrared (NIR) spectroscopic follow-up survey of substellar candidate members of YMGs identified in *BASS*. In Section \[sec:cand\], we summarize *BASS* and the method that we used to build the sample of candidate members from a cross-match of *2MASS* and *AllWISE*. We detail our NIR spectroscopic follow-up and its motivation in Section \[sec:obs\]. In Section \[sec:sptclass\], we present our method to assign a spectral and gravity classification. We present the resulting spectral types and updated YMG membership probability for our sample in Section \[sec:results\]. In Section \[sec:discussion\], we use new discoveries presented here and other known low-gravity BDs and low-mass stars to build empirical photometric sequences, and we then investigate the physical properties of young BDs. We summarize and conclude in Section \[sec:conclusion\].
THE *BASS* SURVEY {#sec:cand}
=================
*BASS* is a systematic all-sky search for later-than-M5 candidate members to nearby YMGs that was the focus of an earlier publication [@2015ApJ...798...73G]. In this work, we undertake a spectroscopic follow-up of the *BASS* sample, which we briefly summarize in this section. We refer the reader to [@2015ApJ...798...73G] for an extensive description of the *BASS* survey.
We cross-matched the *2MASS* and *AllWISE* catalogs outside of the galactic plane and crowded regions ($\geq 2.5$ objects per square arcminute) using a cross-match radius of 25 and applied color, confusion and photometric quality cuts to produce a starting sample of 98970 targets with NIR colors consistent with $\geq$ M5 spectral types and proper motion measurements larger than 30[$\mathrm{mas}\,\mathrm{yr}^{-1}$]{} at $\geq$5$\sigma$ (see @2015ApJ...798...73G for the detailed cross-match and selection algorithm). Astrometry provided in the *2MASS* and *AllWISE* catalogs as well as the mean epochs of observation for both surveys (*JD* keyword in *2MASS*; *W1MJDMEAN* keyword in *AllWISE*) were used to calculate proper motions. We used *W1MJDMAX*-*W1MJDMIN* as a conservative measurement error on the *AllWISE* astrometric epoch, which typically corresponds to $\sim$6 months to one year, compared to a $\sim$11yr baseline between *2MASS* and *AllWISE*. This uncertainty as well as those on astrometric measurements themselves were propagated to the proper motion measurement errors. We obtain a typical proper motion precision of $\sim$15[$\mathrm{mas}\,\mathrm{yr}^{-1}$]{}.
We used the BANYAN II tool to select only objects that have a Bayesian probability $>$10% of belonging to any YMG considered here (this threshold ensures that known bona fide members are recovered; see Paper V). The BANYAN II tool takes the sky position, proper motion and $J$, $H$, $K_S$, *W1* and *W2* photometry as input quantities. It then uses a naive Bayesian classifier to compare those measurements with spatial and kinematic models (SKMs) of YMGs, as well as with old and young color-magnitude diagram (CMD) sequences in both $M_{W1}(J - K_S)$ and $M_{W1}(H - W2)$ spaces. Those CMD sequences were chosen because they were found as the most efficient independent sequences to distinguish between young and field M6–L4 dwarfs. Probabilities generated from a naive Bayesian classifier can be biased when the input parameters are not independent (which is the case here); however, the relative ranking of hypotheses for a given object overcomes this bias [@Hand:2001tr].
It is known that there is a large scatter in the NIR colors of young BDs even though they are redder than field dwarfs on average (e.g. @2012ApJ...752...56F). The inclusion of the CMD sequences described above in BANYAN II will systematically bias our sample towards red NIR colors, and decrease our sample completeness for YMG members that are not especially red. However, this effect is likely less important than the color criteria that were applied in selecting the 98970 objects that were input to BANYAN II. Furthermore, a total of only two independent photometric observables (corresponding to the color-magnitude diagrams) are used in BANYAN II, compared to four kinematics observables when no RV or parallax is available; the relative weight of kinematics is thus twice that of photometry in the calculation of probabilities. Parallax motion was not accounted for in our proper motion measurements or in the BANYAN II tool; the maximal relative importance of this effect will become as large as our typical *2MASS*–*AllWISE* proper motion precision only for objects closer than $\sim$10pc (considering the 11yr baseline between *2MASS* and *AllWISE*). This correction will properly be accounted for in a future version of the BANYAN II tool.
We performed a Monte Carlo simulation based on the [Besançon]{} galactic model (A. C. Robin et al. in preparation, ) and the SKMs of YMGs to obtain a field contamination probability for each individual target in our sample, which allows for a more absolute interpretation in terms of the expected contamination fraction. We used the results of this simulation to reject any candidate member with a $> 50$% probability of being a field contaminant. Note that the contamination probability from this Monte Carlo analysis is not necessarily complementary with the YMG Bayesian probability (see Paper V for more detail). We refer the reader to Paper V for an extensive description of all filters that were used to build the *BASS* sample (e.g., minimal proper motion, color and quality filters, etc.).
There are three samples that are referred to in this Paper: (1) *PRE-BASS* consists of targets that were initially selected as potential members and followed up with spectroscopy, but that were later rejected as we modified our selection criteria to reject contaminants; (2) *Low-Priority BASS* (*LP-BASS*) consists of targets that have NIR colors only slightly redder than field dwarfs; and (3) *BASS* is the final sample presented in Paper V that contains targets at least 1$\sigma$ redder than field dwarfs and that has a lower fraction of contaminants. As discussed in Paper V, the statistical distance associated with the most probable YMG of a candidate member can be used to place it in two CMDs ($M_{W1}$ versus $J - K_S$ and $H - W2$) and compare its position to known field and young BDs and low-mass stars.
\[fig:spectra1b\]
All candidate members that were placed blueward of the field sequence in any of the two CMDs were rejected from *BASS* and *LP-BASS*. Those that were not at least 1$\sigma$ redder than both field sequences were grouped into the *LP-BASS* sample, which is expected to be more contaminated by field objects and young M dwarfs with spectral types earlier than M5. We note that the *PRE-BASS* sample does not necessarily consist of erroneous YMG candidate members; however, it likely suffers from a higher contamination rate from field interlopers or members of moving groups not considered in BANYAN II.
OBSERVATIONS {#sec:obs}
============
Because of their recent formation, young, low-mass objects have inflated radii compared to their field counterparts and are warmer for a given mass. As a consequence, they have a lower surface gravity at a given temperature (and spectral type). It is well known that these low-gravity dwarfs display weaker alkali and molecular absorption lines ( at 7665 & 7669Å in the optical and 1.17 & 1.25$\mu$m in the NIR; at 8183 & 8195Å in the optical and 1.14 & 2.21$\mu$m in the NIR; at 7800 & 7948Å; at 8521 & 8943Å; FeH at 8692Å in the optical and 0.99, 1.20 & 1.55$\mu$m in the NIR; TiO at 8432Å; and CrH at 8611Å). This is due to a lower-pressure in their photosphere, which is a direct consequence of their lower surface gravity. Collision-induced absorption (CIA) of the $H_2$ molecule is also decreased in this lower pressure environment, causing a flatter $K$-band plateau at 2.18–2.28$\mu$m (see the H$_2$($K$) index of @2013MNRAS.435.2650C), leaving the effect of water vapor to become apparent from the triangular-shaped continuum of the $H$ band ([@2001MNRAS.326..695L; @2006ApJ...639.1120K; @2007ApJ...657..511A; @2010ApJ...715L.165R]; see the $H$-cont index of @2013ApJ...772...79A). Furthermore, VO condensate clouds get thicker in the external layers of low-pressure atmospheres, causing deeper absorption bands at 7300-7550 and 7850–8000Å in the optical and 1.06$\mu$m in the NIR (These effects are discussed in more detail by [@2003ApJ...593.1074G; @2004ApJ...600.1020M; @2006ApJ...639.1120K; @2008ApJ...689.1295K; @2009AJ....137.3345C; @2013ApJ...772...79A] and @2013MNRAS.435.2650C). Gravity-sensitive features were initially identified by comparing the optical spectra of M-type giants and M-type dwarfs [@1986ApJS...62..501K; @1998AJ....116.2520J], and it was later demonstrated that the same features could be used to identify young, inflated M-type dwarfs by observing members of star-forming regions [@1996ApJ...469..706M; @1997ApJ...489L.165L; @2004ApJ...610.1045S; @2001MNRAS.326..695L; @2007ApJ...657..511A; @2008MNRAS.383.1385L].
A number of low-gravity features (CIA effects of $H_2$ on the continuum, weaker FeH absorption and stronger VO absorption) can be measured in low-mass stars and BDs with spectral types later than M6 using low-resolution ($R \sim 75$) NIR spectroscopy, providing an efficient way of identifying field interlopers in a set of YMG candidates. A higher spectral resolution ($R \sim 1000$) allows for a more robust determination of low gravity features through the measurement of the pseudo-equivalent width (EW) of the atomic lines listed above. We thus obtained low-resolution NIR spectra of 241 candidate YMG members from the *BASS*, *LP-BASS* and *PRE-BASS* samples. We describe in this section all observations and the individual instrumental configurations that were used. A description of individual observations is included in Table \[tab:obslog\].\
FIRE at Magellan {#sec:fire}
----------------
We obtained NIR spectroscopy for 17 targets with the Folded-port InfraRed Echellette (FIRE; [@2008SPIE.7014E..0US; @2013PASP..125..270S]) at the Magellan Telescopes in April and December 2013, as well as May, June, August and September 2014 and February 2015. We used both the cross-dispersed and high-throughput prism modes to obtain respective resolving powers $R \sim 450$ (prism mode) and $R \sim 6000$ (echelle mode) across the 0.8–2.45$\mu$m range. Total exposure times ranged from 200s to 1800s, depending on source brightness, instrument configuration and weather conditions. This allowed us to obtain a typical S/N $>$100 per resolution element. Science targets were observed in an ABBA pattern along the slit, and a standard A0-type star was observed immediately before or after each of them at a similar airmass to ensure a proper telluric correction. We obtained ThAr (prism mode) or NeNeAr (echellette mode) lamp exposures between every science target to perform wavelength calibration, as well as high- and low-illumination flat fields that were combined to obtain a flat-field image with a large S/N across all orders while avoiding saturation. We reduced all data using the Interactive Data Language (IDL) pipeline FIREHOSE, which is based on the MASE [@2009PASP..121.1409B] and SpeXTool [@2003PASP..115..389V; @2004PASP..116..362C] packages. We supplemented the list of Ar atomic lines with those listed in [@1973PhyS....8..249N] to allow a more robust wavelength solution in the $K$ band in the case of prism data.
The six echellette spectra that we obtained here have a sufficient resolution to measure radial velocities down to a precision down to a few . These measurements will be presented in a future publication along with a significant number of additional FIRE echellette spectra.
SpeX at IRTF {#sec:spex}
------------
We obtained NIR spectroscopy with SpeX [@2003PASP..115..362R] at the IRTF telescope for 118 targets from 2007 to 2015. We used the cross-dispersed and prism modes with slits of 06, 08 and 10 depending on the seeing to obtain resolving powers ranging from $R \sim 75$ to $R \sim 750$ over the 0.8–2.45$\mu$m range. We used ABBA nodding patterns along the slit with typical exposure times of 60s to 250s which yielded typical S/N$>$100 per resolution element. A standard early A-type star was observed immediately before or after every science target at a similar airmass to ensure a proper telluric correction. Several high-S/N quartz lamp and Ar lamp exposures were obtained immediately after every target to ensure a proper wavelength calibration and flat field correction. The data were reduced with the IDL SpeXTool package [@2003PASP..115..389V; @2004PASP..116..362C].
Flamingos-2 at Gemini-South {#sec:f2}
---------------------------
We used Flamingos-2 [@2004SPIE.5492.1196E] at Gemini-South to obtain NIR spectroscopy for 101 targets from 2013 to 2015. We observed each target with both the *JH* and *HK* low resolution grisms and the 072 slit to obtain a resolving power of $R \sim 500$ over 0.9–2.4$\mu$m. Targets were observed in an ABBA pattern along the slit, with total exposure times ranging from 120s to 3400s, to obtain S/N$>$80 per resolution element. Standard A0 to A6-type stars were observed immediately before or after every science target at a similar airmass to ensure a proper telluric correction. Several high-S/N quartz lamp and Ar lamp exposures were obtained immediately after every telluric standard star to ensure a proper wavelength calibration and flat field correction. Dark exposures were obtained at the end of each night, using similar exposure times than all of the science and calibration data to ensure a proper correction of the dark current. A numbers of observations were split between a few nights when observing conditions changed before the required S/N could be obtained.
We used a custom IDL pipeline to apply dark current subtraction and flat field calibrations, correct the trace curvature, optimally extract the spectrum [@1986PASP...98..609H] and perform a wavelength calibration using the Ar lamp observations. A dark current subtraction is usually not needed when data are reduced in $A-B$ pairs, like is the case here; however, we found that applying this correction improved the quality of the data. This is likely due to the large exposure times that were used for some targets, which resulted in a large contribution from the dark current that must be corrected both in the data and flat field exposures before applying the flat field correction. A low-pass filter was applied to the flat field exposures to avoid contaminating data with scattered light. We observed that the spectral dispersion (and thus wavelength solution) generally varied from one exposure to another; the wavelength solutions obtained from the Ar calibrations are hence only approximate.
To address this problem, we used several telluric absorption features in the raw spectra of the science and telluric observations to refine individual wavelength solutions. The $JH$ and $HK$ blocking filters also caused significant fringing in the data (up to $\sim$7%). We corrected this by adjusting a sinusoid fringing solution to the low frequencies of the raw spectra. We found that a complete fringing solution (which includes finesse as an additional parameter) did not improve the results; we thus chose the simpler sinusoid approach to have a more robust algorithm.
The extracted science and telluric spectra were combined and telluric-corrected using a modified version of the SpeXtool package adapted for Flamingos-2. We observed that the slope of the continuum in the overlapping region of both observing modes (in the $H$ band) varied in a systematic way at the edge of the detector. Hence, we removed these regions before combining the spectra. A few objects for which we obtained Flamingos-2 data (e.g. 2MASS J07083261–4701475, 2MASS 20414283–3506442 and 2MASS J12042529–2806364) turned out to be field dwarfs that closely match literature SpeX-prism spectra of other known objects of the same spectral type: this is an indication that the systematics mentioned above were accurately corrected.
GNIRS at Gemini-North {#sec:gnirs}
---------------------
We used GNIRS at Gemini-North to obtain NIR spectroscopy for three targets in 2013. We used the 32lmm$^{-1}$ grating centered at 1.65$\mu$m in the cross-dispersed mode with the 0675 slit to achieve a resolving power of $R \sim 750$ over 0.9–2.45$\mu$m. We nodded exposures along the slit in ABBA patterns with total exposure times ranging from 120s to 360s to reach S/N$>$100 per resolution element. A0-type telluric standard stars were observed immediately before or after science targets at a similar airmass to ensure a proper telluric correction. Several high-S/N quartz lamp and Ar lamp exposures were obtained immediately after every target to ensure a proper wavelength calibration and flat field correction. The data were reduced with the XDGNIRS IRAF package provided by Gemini.
TripleSpec at Hale {#sec:triplespec}
------------------
We used TripleSpec [@2008SPIE.7014E..0XH] at the Palomar Observatory 5m Hale Telescope to obtain NIR spectroscopy for one target in the cross-dispersed mode with the 10 slit, yielding a resolving power $R \sim 3\,800$ over 1.0–2.45$\mu$m. We observed the science target in 4-position ABBA nodding pattern along the slit with a total exposure time of 1200s to reach a S/N$>$100 per resolution element. High-S/N quartz lamp and NeAr lamp exposures were obtained to ensure a proper wavelength calibration and flat field correction. We reduced the data using an adapted version of SpeXtool (see Section \[sec:spex\]).
SPECTRAL TYPES AND LOW-GRAVITY CLASSIFICATION {#sec:sptclass}
=============================================
We describe in this section the method that we used to assign spectral types to our new observations. Our typing scheme consists of two distinct dimensions : the first dimension consists of the usual spectral subtypes and is mostly sensitive to [$T_{\mathrm{eff}}$]{}. The second dimension, introduced by and [@2006ApJ...639.1120K], aims at characterizing the surface gravity with the use of a greek-letter suffix. Field-gravity dwarfs are designated with the $\alpha$ suffix or no suffix, intermediate-gravity dwarfs with the $\beta$ suffix, and very low-gravity dwarfs with the $\gamma$ suffix. The $\delta$ suffix was also introduced by [@2006ApJ...639.1120K] to designate objects with an even younger age (typically less than a few Myrs) and lower surface gravity than those associated to the $\gamma$ suffix.
Optical spectral standards were used to classify NIR spectra of field K7–M9 spectral types. We used the NIR data of GJ 820 B (K7), Gl 229 A (M1), Gl 411 (M2), Gl 213 (M4), Gl 51 (M5), Gl 406 (M6), GJ 644 C (M7), GJ 752 B (M8) and LHS 2924 (M9) as field-gravity spectral standards for these respective spectral types. These standards were identified from the list maintained by Eric Mamajek[^3] [@1976PhDT........14B; @1991ApJS...77..417K; @2013ApJS..208....9P] and their spectra were downloaded from the IRTF spectral library[^4]. We did not use any of the suggested K8, K9, M0 and M3-type standards, since none of them were available in the IRTF spectral library.
While NIR L dwarfs spectral standards have been identified by [@2010ApJS..190..100K], we have opted to use optically-anchored NIR spectral average templates for classifying field L0–L9 dwarfs. Templates are constructed by median-combining all spectra of a given optical spectral type and gravity class. These templates were provided by K. Cruz and their creation will be discussed in detail and be made public as part of a forthcoming paper (Cruz et al. in preparation). The spectral morphology of these templates is consistent with the [@2010ApJS..190..100K] spectral standards but since they are an average of many objects, they also reflect the diversity of spectral morphologies present in each spectral type. Spectral standards have been determined for low-gravity M and L dwarfs by [@2013ApJ...772...79A], but we opted to use spectral average templates in this case too, for the reasons mentioned above. We generated M6–M9$\gamma$ templates with data published in [@2013ApJ...772...79A] and sent to us directly by the authors. These templates are available at the Montreal Spectral Library[^5]. The optically-anchored L0$\beta$, L1$\beta$ and L0–L4$\gamma$ templates were provided by K. Cruz. They will be discussed in detail and be made public as part of a forthcoming paper et al. (Cruz et al. in preparation).
All template, standard, and target spectra were re-sampled to the same resolution and wavelength grid as SpeX prism observations with the 06 slit ($R \sim 120$). Following the method of [@DisentanglingLDwar:db], the spectra were normalized in three sections in order to minimize the effect of large NIR color variations within a given spectral type. The spectra were broken into three sections: 0.80–1.35$\mu$m, 1.40–1.80$\mu$m and 1.95–2.40$\mu$m, roughly corresponding to the $zJ$, $H$, and $K$ bands.
[ll]{} USco J160603.75–221930.0 & L0$\delta$\
USco J160727.82–223904.0 & L0$\delta$\
USco J160737.99–224247.0 & L0$\delta$\
USco J160818.43–223225.0 & L0$\delta$\
USco J160828.47–231510.4 & L0$\delta$\
USco J160843.44–224516.0 & L0:$\delta$\
USco J160918.69–222923.7 & L0$\delta$\
USco J161228.95–215936.1 & L0$\delta$\
USco J161441.68–235105.9 & L0$\delta$\
USco J163919.15–253409.9 & L0$\delta$\
CD–35 2722 B & L3$\beta$\
2MASS J01531463–6744181 & L3$\beta$\
2MASS J17260007+1538190 & L3$\beta$\
2MASS J00011217+1535355 & L4$\beta$\
2MASS J05120636–2949540 & L5$\beta$\
2MASS J23174712–4838501 & L5$\beta$\
2MASS J03264225-2102057 & L5$\beta$/$\gamma$\
SIMP J21543454–1055308 & L5$\beta$/$\gamma$\
2MASS J03552337+1133437 & *J0355-type* (L3–L6$\gamma$)\
2MASS J16154255+4953211 & *J0355-type* (L3–L6$\gamma$)\
2MASS J23433470–3646021 & *J0355-type* (L3–L6$\gamma$)\
WISEP J004701.06+680352.1 & *J2244-type* (L6–L8$\gamma$)\
2MASS J22443167+2043433 & *J2244-type* (L6–L8$\gamma$)\
PSO J318.5338–22.8603 & *J2244-type* (L6–L8$\gamma$)\
In a first step to estimating a spectral type, we categorized our 245 new spectra with the spectral template and standard grid described above. There were 11 objects, however, that did not have a good visual match to any standard or template in the grid; this number excludes the early-type contaminants which are discussed later in this work. We collected additional low-gravity brown dwarf spectra in the literature to identify 19 more objects that do not match our standards.
We performed a visual analysis of all of the unclassifiable spectra and identified enough objects with similar spectral morphologies to create tentative new spectral types and templates for L0$\delta$, L3$\beta$, L4$\beta$ and L5$\beta$. The objects that were used in the creation of these templates are listed in Table \[tab:sptextlit\]. We list the revised spectral types that we obtain for other spectra from the literature in Table \[tab:retypes\]. We note that our L3$\beta$ template includes 2MASS J17260007+1538190, which was suggested by [@2013ApJ...772...79A] as a tentative template for the L3$\beta$ spectral type.
We could not build a template for the L2$\beta$ spectral type, as the only objects that were confirmed as L2$\beta$ from optical data have either very low signal-to-noise (S/N) ratios in the NIR or no NIR data. As we gather more high-S/N spectra of low-gravity L dwarfs, we expect to fill this gap.
The L0$\delta$ template was built from eight candidate members of Upper Scorpius [@2008MNRAS.383.1385L] and one candidate member of $\beta$PMG (2MASS 00464841+0715177) that are similar to the L0$\gamma$ template except that their $H$ band is even more triangular and their $K$ band has a redder continuum. It is also notable that the H$_2$O-dependent slope of the L0$\delta$ at 1.7–1.8$\mu$m is slightly steeper than what is seen in any other L-type template.
![Distribution of the differences between our visual and index-based spectral classifications, for young dwarfs and field objects. Spectral types generally agree within one subtype, with a standard deviation of 0.7 subtypes and a reduced $\chi^2$ value of 0.8. This is indicative that our measurement errors are representative of the observed differences between the two methods. It can also be seen that most of the outliers in the distribution correspond to objects with uncertain spectral types (i.e., measurement errors of one subtype or more.)[]{data-label="fig:hist_diff"}](Figure2.pdf){width="45.00000%"}
[lllllll]{} 2MASS J21324036+1029494 & & $\cdots$ & & L4.5: & 1 & L4:$\beta$/$\gamma$\
2MASS J14482563+1031590 & & L4: & 2 & L3.5 & 3 & L5:$\beta$\
WISE J174102.78–464225.5 & & $\cdots$ & & L7:: & 4 & L5:–L7:$\gamma$\
G 196–3 B & & L3$\beta$ & 5 & L3$\gamma$ & 6 & L2–L4$\gamma$\
2MASS J00303013–1450333 & & L7 & 7 & L4.5:: & 8 & L4–L6$\beta$\
2MASS J20025073–0521524 & & L6 & 9 & L7:: & 10 & L5–L7$\gamma$\
2MASS J08354256-0819237 & & L5 & 11 & L5 & 12 & L4pec\
2MASS J18212815+1414010 & & L4.5pec & 13 & L5pec & 14 & L4pec\
2MASS J21512543–2441000 & & L3 & 9 & $\cdots$ & & L4pec\
2MASS J01033203+1935361 & & L6$\beta$ & 7,15 & L6$\beta$ & 6 & L6pec\
2MASS J01075242+0041563 & & L8 & 16 & L8pec & 17 & L7pec\
2MASS J08251968+2115521 & & L7.5 & 7 & L6 & 18 & L7pec\
2MASS J08575849+5708514 & & L8 & 5 & L8$\pm$1 & 19 & L8–L9pec\
There are two sets of objects with similar spectra, each with three targets, that we identified via our visual analysis; however, we are unable to confidently assign them a spectral type that fits into our grid of templates. For the purposes of this paper, we label these objects as *J0355-type* and *J2244-type*. One set is composed of 2MASS J03552337+1133437, 2MASS J16154255+4953211 and 2MASS J23433470–3646021. Their spectra are similar to the L4$\gamma$ template except that they have a shallower CO band at 2.3$\mu$m. The other set is composed of 2MASS J00470038+6803543, PSO J318.5338–22.8603 and 2MASS J22443167+2043433. Their spectra display a significantly redder continuum than our templates, which might be indicative of a later spectral type. We note that two objects have previous classifications based on the index-based scheme of [@2013ApJ...772...79A]: 2MASS J00470038+6803543 was classified as an intermediate-gravity L7 dwarf by [@2015ApJ...799..203G] and PSO J318.5338–22.8603 was classified as a very low-gravity L7 dwarf by [@2013ApJ...777L..20L]. We listed these two sets of objects as well in Table \[tab:sptextlit\].
We adopt a conservative estimate of L3–L6$\gamma$ for the spectral type range of the *J0355-type*. The spectral features of the *J2244-type* are indicative of a spectral type in the range L6–L8$\gamma$ range. For both of these new spectral types, we refrain from assigning them a more precise location in the spectral sequence until more data are available at these late low-gravity types. It is unclear at this stage whether *J0355-type* and *J2244-type* objects are peculiar or a simple extension of low-gravity brown dwarfs at spectral types later than L5. A larger number of late-type, low-gravity L dwarfs will need to be identified before we can assess this. Our set of low-gravity templates is displayed in Figure \[fig:spectra1\].
We used the index-based classification method of [@2013ApJ...772...79A] to corroborate our visual classification. This method consists of measuring the slope of H$_2$O continuum features to assign a spectral type, and a combination of several gravity-sensitive spectroscopic indices to assign a gravity class. We found that spectral types obtained from the template grid system described above generally agree with index-based spectral types within one subtype (Figure \[fig:hist\_diff\]). The standard deviation between the two methods for the 163 non-peculiar objects that we categorized is of 0.7 subtypes, with a reduced $\chi^2$ value of 0.8. A reduced $\chi^2 \approx 1$ indicates that measurement errors are representative of the discrepancies. The reduced $\chi^2$ is given by $1/(N-1)\cdot\sum y/\sigma_y$, where $N$ is the number of objects, $y$ is the spectral type discrepancy and $\sigma_y$ is the quadrature sum of the index-based and visual-based spectral type measurement errors. All cases discrepant by more than 1.5 spectral types correspond to low-S/N data, except for 2MASS J21420580–3101162 that gets L1.5$ \pm 0.3$ from the index-based method and L3 from the visual-based method. This object does not display signs of youth or significantly peculiar features, but it has a slightly redder slope at 1.7–1.8$\mu$m. It unclear what is the cause of this discrepancy.
We used optical data to assign an adopted spectral type using a template-based visual classification method [@2009AJ....137.3345C] only for the 4 objects for which no NIR data were available. In all other cases, our adopted spectral types are based on NIR data only. Our NIR spectral types based on a visual comparison with templates show a standard deviation of 0.9 subtype with respect to optical spectral types in the literature, and the reduced $\chi^2$ of the differences is 1.5, hence slightly larger than what would be expected given the uncertainties. If we compare optical spectral types to the index-based spectral types of [@2013ApJ...772...79A], we obtain a slightly larger standard deviation (1.1 subtype) and reduced $\chi^2$ (2.4). This is indicative that our visual-based classification method is more consistent with spectral types based on optical data that were reported in the literature. This should be expected, as our templates are anchored on optical data. In both cases, we observe no systematic bias (the mean of the differences is smaller than 0.1 subtype). Several objects that deserve further discussion are presented in detail in the Appendix.
We note that the index-based field-gravity, intermediate-gravity and very low-gravity classes defined by [@2013ApJ...772...79A] were built to correspond to the optical $\alpha$, $\beta$ and $\gamma$ classes, which is what we observe in 143/176 (81%) of the cases. Some of the discrepancies arise for objects near the spectral type thresholds where the method of [@2013ApJ...772...79A] stops being applicable ($\lesssim$M6 or $\gtrsim$L6) or for data with a lower S/N. The $\delta$ gravity class does not have an equivalent in the index-based classification of [@2013ApJ...772...79A], but we note that all three of the young dwarfs that we categorized as $\delta$ are assigned with the maximal index-based gravity score (2222). It does not seem that this maximal index-based gravity score always translates as a $\delta$ visual classification though, as there are four additional objects in our sample that obtained the score 2222 but that we visually categorized as $\gamma$ (2MASS J00182834–6703130; L0$\gamma$, 2MASS J01205114–5200349; L1$\gamma$; 2MASS J20113196–5048112; L3$\gamma$; 2MASS J22351658–3844154; L1.5$\gamma$; all are THA candidate members). For consistency within this work, we have adopted the visual spectral types in the remaining sections, but we list all visual and index-based spectral types in Table \[tab:sptclass\]. We note that this choice does not affect the conclusions presented in this work.
RESULTS {#sec:results}
=======
In Figure \[fig:allspt\], we present the NIR spectra of several new intermediate ($\beta$) and very-low ($\gamma$) gravity dwarfs discovered in this work, as well as known dwarfs for which we have obtained new data. Several objects that were uncovered as candidate members of YMGs in *BASS* had NIR or optical spectroscopy readily available in the SpeX Prism Spectral Libraries [^6] [@2014ASInC..11....7B] or the RIZzo Ultracool Spectral Library[^7] with no discussion of low gravity in the literature; we included them in our present analysis and the resulting spectral classification is listed in Table \[tab:sptclass\] along with the new discoveries.
There are some cases where the BANYAN II tool yields ambiguous candidate membership to more than one association (i.e., at least a second moving group shares 10% of the total YMG Bayesian probabilities). In all such cases, we list in Table \[tab:sptclass\] all plausible YMGs with their relative share of the total YMG probability (i.e., excluding the field probability). An extensive RV and parallax follow-up will be required before more can be said on their YMG membership.
We have identified seven objects (Table \[tab:sptclass\]) that display signs of low gravity, but for which additional information was inconsistent with membership to any of the YMGs presented here (e.g., RV and distance measurements or the effect of interstellar extinction affecting the NIR spectrum which is not consistent with ages older than $\sim$5Myr). It is possible that these objects belong to YMGs or star-forming regions that are not considered here, that their RV or parallax measurements are affected by an unresolved binary companion (see the Appendix for a detailed discussion), or that other physical properties such as enhanced dust mimics a lower gravity.
However, [@2013ApJ...772...79A] noted that their index-based classification scheme should not be significantly contaminated by old, dusty brown dwarfs, which makes this last hypothesis less likely.
In Figure \[fig:spthist\], we show a histogram of all previously known low-gravity dwarfs along with new discoveries or confirmations of low gravity that are presented here. It might seem surprising that we did not identify any new low-gravity L2 dwarfs, however this is likely the effect of small number statistics and the fact that we still lack a template for the L2$\beta$ spectral type, e.g., some low-S/N low-gravity objects presently typed as L1: and L3: might turn out to be L2 dwarfs when more data becomes available. We anticipate our visual-based low-gravity classification scheme to improve as more data is obtained. If we account for the measurement errors on our spectral types using gaussian probability density functions (which softens the gap at L2) and use Poisson statistics to assess the significance of this lack of L2 dwarfs, we find that the differences between the number of known low-gravity L1, L2 and L3 dwarfs is insignificant (at the level of 0.2$\sigma$).
In Figures \[fig:I1\] and \[fig:I2\], we compare all new low-gravity confirmations with the field and low-gravity sequences defined by [@2013ApJ...772...79A]. The individual values for these gravity-sensitive spectroscopic indices are listed in Table \[tab:sptindices\]. There are 7 objects in our sample that did not have a discussion of low gravity in the literature and for which optical spectra were available in the Ultracool RIZzo Spectral Library. We used them to revise their spectral types and measure gravity-sensitive optical indices defined by [@1999ApJ...519..802K] and [@2009AJ....137.3345C]. These results, based on optical data only, are presented in Table \[tab:sptoptindices\]. The new spectroscopic observations presented here (95 from *BASS*, 26 from *LP-BASS* and 120 from *PRE-BASS*) allowed us to uncover a total of 108 new M6–L5 low-gravity dwarfs, doubling the number of such known objects (98 before this work).
In addition to several new candidate members of YMGs, we report here that 2MASS J14252798–3650229 (DENIS-P J142527.97-365023.4) is a new low-mass BD bona fide member of ABDMG. This object was identified by as an L5 dwarf with an estimated spectro-photometric distance of $\sim$10pc. [@2010ApJ...723..684B] measured an RV of $5.37 \pm 0.25$ and [@2014AJ....147...94D] measured a trigonometric distance of $11.57 \pm 0.11$pc. [@2015ApJ...798...73G] reported that the galactic position and space velocities of this object are a very good match to ABDMG (Figure \[fig:L3XYZUVW\]), suggesting that it would be a new bona fide member if low gravity would be confirmed. They also indicated that its NIR colors are redder than those of field dwarfs of the same spectral type, which hints at low gravity. The low gravity is indeed readily apparent in the new SpeX prism spectrum that we obtained for this object (Figure \[fig:L3ador\]): both a visual comparison and the index-based classification of [@2013ApJ...772...79A] indicate that this object is an L4$\gamma$ dwarf. We conclude that 2MASS J14252798–3650229 is a new bona fide member of ABDMG, making it the second latest-type confirmed member of this moving group after the L7$\beta$ member WISEP J004701.06+680352.1. At the age of ABDMG, 2MASS J14252798–3650229 has an estimated mass of $26.6^{+0.3}_{-1.0}$[$M_{\mathrm{Jup}}$]{}.
![NIR Spectral type histogram of all known low-gravity dwarfs and those presented in this work. Green bars delimited by dashed lines represent the known population prior to *BASS*, purple bars delimited by dash-dotted lines represent known dwarfs for which low-gravity features were identified here for the first time, and orange bars delimited by solid lines represent new discoveries from *BASS*. The *BASS* survey has contributed significantly in increasing the number of known low-gravity M6–L5 dwarfs.[]{data-label="fig:spthist"}](Figure4.pdf){width="49.00000%"}
Updated YMG Membership {#sec:mg}
----------------------
![NIR spectrum of the new L4$\gamma$ ABDMG bona fide member 2MASS J14252798–3650229 (thick black line), compared with various field and low-gravity L4 BDs. All spectra were degraded to a resolution of $R \sim 120$ and normalized at their median value in the $\sim$1.27–1.33$\mu$m range. The $H$-band continuum of 2MASS J14252798–3650229 has a typical triangular shape and its global slope is particularly red, which are both telltale signs of low gravity.[]{data-label="fig:L3ador"}](Figure8.pdf){width="49.00000%"}
![NIR spectra of typical contaminants in the *PRE-BASS* sample. Resolution was degraded in the same way as described in Figure \[fig:spectra1\]. All spectra were normalized to their median across the full wavelength range and shifted vertically for comparison purposes. The contaminants presented in this figure likely correspond to background K- and M-type stars reddened by interstellar dust. We used alternating colors for visibility.[]{data-label="fig:spectrac"}](Figure9.pdf){width="49.00000%"}
It is possible to use the spectral type information as well as the youth of candidate members determined from the spectroscopic follow-up presented here as additional inputs in BANYAN II to refine estimates of distance, RV and YMG membership and contamination probabilities. Spectral types are used to assess if the absolute $W1$ magnitude of a target is consistent with its spectral type at the statistical distance that corresponds to a given YMG membership (using distinct sequences for field and low-gravity dwarfs; see Paper V), whereas prior knowledge of youth reduces the number of potential contaminants from the field and thus improves the probability that the object belongs to a YMG. We reject all objects with spectral types $\geq$M5 that display no signs of low gravity (17 in *BASS*, 7 in *LP-BASS* and 41 in *PRE-BASS*), since this implies an age older than the Pleiades ($\sim$120Myr; [@2009AJ....137.3345C; @2013ApJ...772...79A]) and is not consistent with membership to any YMG considered here. These updated results are listed in Table \[tab:sptclass\], and individual objects of interest are discussed in the Appendix.
X-Ray Luminosity {#sec:xray}
----------------
We followed up several objects that turned out to have spectral types earlier than expected, some of them ($\leq$ M5) to the point where current NIR and optical index-based methods are unable to determine whether they are likely young or field objects. In this section, we take advantage of the ROSAT bright and faint source catalogs (; VizieR catalogs *IX/10A* and *IX/29*) to assess whether these objects are young candidate members of YMGs or field interlopers.
[@2014ApJ...788...81M] demonstrated that the distribution of absolute X-ray luminosity for M0–M5 dwarf members of ABDMG and $\beta$PMG is significantly distinct from that of field M0–M5 dwarfs. In particular, they showed that $\beta$PMG members are $\sim$4 times more X-ray luminous than ABDMG members, a factor that goes up to $\gtrsim$40 when instead compared with field dwarfs. We investigated whether any of our M0–M5 candidate members listed in Table \[tab:sptclass\] display X-ray emission by cross-matching their *2MASS* position with the ROSAT catalogs with a 15$"$ search radius. We computed the absolute X-ray luminosity for all X-ray sources recovered this way, using trigonometric distances when possible or kinematic distances otherwise.
We have identified ROSAT entries for only three objects: 2MASS J08540240–3051366 (M4 candidate member of $\beta$PMG; $\log L_X = 28.4 \pm 0.3$) has a low X-ray luminosity compared with M3–M5 members of ABDMG or $\beta$PMG (both have $\log L_X \approx 28.5-29.5$) and could thus be a field interloper ($\log L_X \approx 27-28.5$; see Figures 7 and 8 of @2014ApJ...788...81M). 2MASS J08194309–7401232 (M4.5 candidate member of COL; $\log L_X = 29.3 \pm 0.4$) and 2MASS J21490499–6413039 (M4.5 candidate member of THA; $\log L_X = 29.3 \pm 0.3$) both have X-ray luminosities consistent with an age similar or younger than that of ABDMG, making them likely members of their respective moving groups. We note that 2MASS J21490499–6413039 has already been reported as a candidate member of THA by [@2014AJ....147..146K], who measured its RV and found it to be consistent with other THA members. Objects that do not have a ROSAT counterpart do not necessarily have a low absolute X-ray luminosity, but might be too distant or located outside of the regions covered by the ROSAT survey.
Using the ROSAT bright catalog detection limit of $0.1$ct/s in the 0.1–2.4keV energy band and assuming a hardness ratio *HR1*$\approx 0$, we can only put an upper limit of $\log L_X = 28-29.8$ on the remaining targets, which is generally not sufficient to reject any more candidate members. Only 3/41 of these targets (2MASS J05484454–2942551, 2MASS J06494706–3823284 and 2MASS J07583098+1530146 AB) have $\log L_X < 28.5$, potentially making them less interesting candidate members. It should be noted however that one of these three objects (2MASS J06494706–3823284) has weak absorption consistent with a very low surface gravity. This demonstrates how the absence from the ROSAT catalog is not a strong enough constraint to reject any of our M0–M5 candidate members. [@2014AJ....147..146K] has demonstrated that surveys for M-type moving group members based on either X-ray or UV-bright samples are incomplete because of the sky coverage and detection limits of current X-ray and UV catalogs.
Sources of contamination {#sec:contam}
------------------------
In Paper II, we demonstrated that a fraction of candidate members identified by the BANYAN II tool are expected to be field interlopers, especially if no prior knowledge is available on age. This fraction of contaminants is dependent on the YMG considered: ARG, ABDMG and $\beta$PMG are expected to be the most contaminated, mostly due to their proximity and their overlap with the galactic plane. Counting the fraction of low-gravity dwarfs in the spectroscopic follow-up presented here allows us to estimate minimal contamination rates of 18% and 33% in the *BASS* and *LP-BASS* samples, respectively. These values are slightly larger than the estimates that we derived in Paper V (12.6% for *BASS* and 26% for *LP-BASS*). The most likely explanation is that the kinematic distribution of field BDs is not perfectly reproduced by the [Besançon]{} galactic model, on which our previous estimates were based. The reason why these updated estimates correspond to a minimal contamination fraction is that some low-gravity dwarfs in our candidate sample could still be contaminants from associations not considered in BANYAN II, e.g., the Ursa Majoris moving group (UMA; $\sim$300Myr; ), the Hercules-Lyrae moving group (250Myr; ), the $\epsilon$ Chamaeleontis association (also called Cha-Near; $\sim$10Myr; ), the Octans association (30–40Myr; [@2008hsf2.book..757T; @2015MNRAS.447.1267M]) and the Carina-Near moving group (200Myr; @2006ApJ...649L.115Z). Measurements of RV and trigonometric distance will be helpful to identify such contaminants. Besides field-gravity $\geq$M5 dwarfs, we identified other kinds of contaminants in our sample of candidates, based on our new NIR spectroscopy. We uncovered a number of objects with spectral types earlier than M5 (4 in *BASS*, 4 in *LP-BASS* and 28 in *PRE-BASS*), for which there is no known reliable low-gravity indicators in the NIR. In addition to those, we uncovered 27 contaminants mostly in the *PRE-BASS* sample (only one was found in *BASS*) that correspond to K- and M-type low-mass stars reddened by interstellar dust in the line of sight (Figure \[fig:spectrac\]). A number of these are likely located in star-forming regions, such as $\rho$ Ophiucus ($\rho$OPH), the Scorpius-Centaurus Complex (SCC) and Taurus-Aurigae (TAU; @1978ApJ...224..857E). These objects were all rejected from the *BASS* sample, mainly because (1) we avoided star-forming regions in the final survey; and (2) the extragalactic WISE color filter defined by [@2011ApJS..197...19K] and the *2MASS* crowding filter defined in Paper V efficiently rejected them.
DISCUSSION {#sec:discussion}
==========
Updated Color-Magnitude Sequences for Young Low-Mass Stars and Brown Dwarfs {#sec:cmd}
---------------------------------------------------------------------------
![Spectral type at which the young and field absolute magnitude polynomial sequences cross (see Figure \[fig:CMDs\]), as a function of the effective wavelength in which each sequence is defined. Young dwarfs are systematically brighter than their field counterparts because of their inflated radii; however, dust clouds are thicker in the high atmosphere of young L dwarfs, which counter-balances this effect and causes the young sequence to cross the field sequence. The fraction of Monte Carlo steps where the sequences crossed is indicated next to a given data point; see text for more detail. Dust clouds are more opaque in the $J$ band ($\sim$1.2$\mu$m), hence the crossing point for this sequence happens at earlier spectral types. At longer wavelengths ($\sim$4.5$\mu$m), dust clouds do not have as much effect. This causes the sequences to cross less often and when they do, they cross at later spectral types.[]{data-label="fig:crossing"}](Figure12.pdf){width="48.00000%"}
{width="99.50000%"}
We complemented the list of all spectroscopically confirmed $\geq$ L0 dwarfs as of February 2014 [@2014PhDT........56M] and the DwarfArchives online library[^8] with more recent discoveries, measurements of photometry in the literature and additional NIR photometry from a cross-match with *2MASS* and *WISE*, in order to build an up-to-date sequence of field dwarfs. This list currently contains $>$1800 published $\geq$ L0 low-mass stars and BDs[^9]. We compiled a similar list of $>$8700 M6–M9 low-mass stars and BDs[^10]. These two lists of dwarfs contain photometric data from articles referenced throughout the present work[^11]. In Figures \[fig:CMDs\] and \[fig:CC2\], we compare our updated population of known young low-mass stars and BDs to the field sequence in various spectral type-color and spectral type-absolute magnitude diagrams. We used data from the two aforementioned lists to build the photometric sequences. In the case of YMG candidate members that do not have a trigonometric distance measurement, we used the statistical distance from BANYAN II, associated with the most probable YMG hypothesis. In each case, we calculated the error-weighted median sequence in bins of 1 subtype and adjusted a polynomial relation by minimizing the $\chi^2$ value. We list in Table \[tab:coeff\] the coefficients of these polynomial fits as well as the respective standard deviation of the data with respect to the best fit. We note that our field sequences are slightly redder than those derived from samples based on the Sloan Digital Sky Survey (*SDSS*; [@2000AJ....120.1579Y]) such as those presented by [@2008AJ....135..785W] and [@2015AJ....149..158S]. This is true because *SDSS*-based surveys rely directly on spectra and are thus un-biased, whereas other surveys based on *2MASS* and/or *WISE* (e.g., [@2003AJ....126.2421C; @2008AJ....136.1290R; @2011ApJS..197...19K]) perform a spectroscopic follow-up only on targets that were pre-selected from color cuts, which makes them biased towards detecting red objects more easily. Since *2MASS*- and *WISE*-based surveys dominating the population of L dwarfs identified in the literature, our field sequences are consequently redder than those based on *SDSS* samples. This effect is also demonstrated in Figure 3 of [@2010AJ....139.1808S].
The radii of young low-mass stars and BDs are inflated compared with old objects of the same spectral type. For this reason, it could be expected that young absolute magnitude sequences fall above the field sequences across all spectral types. However, starting at spectral type $\sim$L0, dust clouds form in the photosphere of BDs. Young BDs have a lower atmospheric pressure, which allows the formation of thick clouds higher in their atmosphere [@2006ApJ...639.1120K; @2008ApJ...686..528L]. As a result, a fraction of the NIR light at $\sim$0.5–3$\mu$m gets redirected to longer wavelengths, causing young BDs to display similar absolute $J$ magnitudes to those of field BDs around spectral type $\sim$L0, as well as absolute $J$ magnitudes even fainter than those of field dwarfs at later spectral types . We could expect that this effect will eventually cease around spectral type T, where dust clouds fall below the photosphere. This has yet to be demonstrated, because there is only a very small number of young T dwarfs currently known (e.g., ). In Figure \[fig:crossing\], we show the spectral type at which the young and field sequences cross as a function of spectral band. Horizontal error bars represent the effective width of the photometric filters and vertical error bars are drawn from a 10000-step Monte Carlo simulation, introducing noise in the data that is representative of photometric uncertainties and repeating the polynomial fit every time. Cases where the sequences do not cross are not included in the calculation of the median and standard deviation of the crossing points. We note that the fraction of Monte Carlo steps where the sequences cross significantly decreases at increasing wavelengths. This is explained by the fact that the photometric sequences become gradually disjointed in the spectral range considered; it is thus possible that in reality the sequences generally cross at spectral types $\geq$L7 (or not at all) in the $W1$ and $W2$ bands. This figure shows a clear correlation which indicates that flux is redistributed out to longer wavelengths in low-gravity dwarfs, a likely effect of the dust clouds (J. K. Faherty et al., in preparation). This is a known effect which is in part due to the larger opacity from the H$_2$O, CO and H$_2$ molecules at wavelengths larger than $\sim$1$\mu$m that are masking the effects of clouds [@2001ApJ...556..872A]. The BT-Settl isochrones do not reproduce this effect, as the young ($\leq$100Myr) and old ($\geq$1Gyr) isochrones do not cross in neither of the $J$, $H$ or $K_S$ bands over the range of effective temperatures that correspond to the M and L spectral types ($\sim$1300–3000K; @2009ApJ...702..154S).
In Figure \[fig:JJK\], we show a $M_J$ versus $J - K$ CMD in the Mauna Kea Observatories NIR filter system (MKO; [@2002PASP..114..169S]) for low-gravity and field dwarfs. When MKO photometry was not available, we used *2MASS* photometry with the conversion relations of @2004PASP..116....9S ([[email protected]]; L and T dwarfs) and @2006MNRAS.373..781L ([[email protected]]; M dwarfs). The combined effects of redder colors due to thicker/higher clouds [@2002ApJ...568..335M] and brighter absolute $J$ magnitude due to inflated radii cause a systematic shift of the low-gravity sequence to the right compared to the field sequence. This Figure brings into evidence the fact that the currently known population of young BDs does not reach a color reversal similar to the L/T transition of field dwarfs (at $J-K \sim 1.8$ and $M_J \sim 14.5$), corresponding to the temperature at which dust clouds fall below the photosphere . We chose this parameter space because a significant amount of data are available in these filters and it is very efficient in displaying this color reversal. It can be expected that a color reversal would eventually be reached for young dwarfs around the T spectral type, corresponding to cooler temperatures than the currently known population. The coolest known directly imaged young exoplanets and low-mass BDs (blue stars in Figure \[fig:JJK\]) tentatively hint at such a color reversal.
Since $J - K$ and $M_J$ are generally correlated for a given spectral type, the $(J - K)$–spectral type and $M_J$–spectral type relations listed in Table \[tab:coeff\] are not the best representation for the low-gravity and field sequences in this CMD diagram. In the case of the young sequence, the absence of a color reversal allowed us to simply fit a polynomial sequence to the young dwarfs directly in the $M_J$–$(J-K)$ space; however, the field sequence cannot be represented by a simple polynomial relation across the M6–T9 range. We used a Markov Chain Monte Carlo algorithm to construct a parametrized polynomial sequence that fits the field sequence across its complete spectral range. We started from a parametrized equation obtained from the combination of the $J - K$ and $M_J$ polynomial relations described in Table \[tab:coeff\], and allowed the eight coefficients of each dimension to vary such that the sequence minimizes the quadrature sum of the bi-dimensional distance of all individual field dwarf positions in the CMD diagram relative to their error bars. This results in a parametrized sequence that describes $J - K$ and $M_J$ as a function of the parametric variable $\lambda$. Larger values of $\lambda$ correspond to later spectral types on average, but no relation between $\lambda$ and spectral types can be provided as the field sequence is a parametric equation that does not assign a $\lambda$ value to individual data points. We obtain:
$$\begin{aligned}
(M_J)_{\mathrm{Young}} =\ &1.61 \times 10^{3} - 8.92 \times 10^{2}\ (J - K)\notag\\ + &2.04 \times 10^{2}\ (J - K)^2 - 2.46 \times 10^{1}\ (J - K)^3\notag\\ + &1.65\ (J - K)^4 - 5.87 \times 10^{-2}\ (J - K)^5\notag\\ + &8.59 \times 10^{-4}\ (J - K)^6\end{aligned}$$
$$\begin{aligned}
(J - K)_{\mathrm{Field}} =\ &1.98 \times 10^{1} - 1.55 \times 10^{1}\ \lambda\notag\\ + &5.09\ \lambda^2 - 8.76 \times 10^{-1}\ \lambda^3\notag\\ + &8.68 \times 10^{-2}\ \lambda^4 - 5.12 \times 10^{-3}\ \lambda^5\notag\\ + &1.76 \times 10^{-4}\ \lambda^6 - 3.27 \times 10^{-6}\ \lambda^7\notag\\ + &2.53 \times 10^{-8}\ \lambda^8\end{aligned}$$
$$\begin{aligned}
(M_J)_{\mathrm{Field}} =\ &-5.61 \times 10^{1} + 3.73 \times 10^{1}\ \lambda\notag\\ - &8.45\ \lambda^2 + 1.04\ \lambda^3\notag\\ - &7.64 \times 10^{-2}\ \lambda^4 + 3.56 \times 10^{-3}\ \lambda^5\notag\\ - &1.06 \times 10^{-4}\ \lambda^6 + 1.84 \times 10^{-6}\ \lambda^7\notag\\ - &1.43 \times 10^{-8}\ \lambda^8\end{aligned}$$
\
where the young sequence is valid in the range ${8.8 \leq M_J \leq 14.7}$ and the field sequence is valid in the range ${4.5 \leq \lambda \leq 28.5}$ (i.e., ${J \geq -1.1}$ and ${9.3 \leq M_J \leq 19.9}$).
We note that the NIR colors of young BDs discovered in the BASS and LP-BASS surveys are likely affected by a form of the confirmation bias, in the sense that we specifically looked for red objects in our survey (see Paper V). Hence, this new photometric data should not be taken as additional evidence that young BDs are redder than field BDs. Reinforcing this result would require looking for signs of low gravity in a sample of BDs that were selected independently of their photometric colors. Directly imaged young planets and brown dwarf companions do not suffer from this potential bias however, and their colors seem consistent with those of isolated young BDs . This might be an indication that our confirmation bias is not significant.
An Updated Investigation on the Age Dependence of Spectroscopic Indices {#sec:ageinv}
-----------------------------------------------------------------------
Since they are the only BDs with a well calibrated age, members of YMGs provide the exciting opportunity of creating a spectroscopic age calibration applicable to all young BDs. Using Pleiades members and the fact that known low-gravity BDs were located away from star-forming regions, [@2008ApJ...689.1295K] and [@2009AJ....137.3345C] estimated that the very low-gravity ($\gamma$) and intermediate gravity ($\beta$) classifications likely correspond to $\sim$10Myr and $\sim$100Myr, respectively. [@2013ApJ...772...79A] extended this investigation by using a restrained sample of 25 M6–L5 dwarf members of young associations. They found that very low-gravity ($\gamma$) and intermediate-gravity ($\beta$) dwarfs likely correspond to ages of $\sim$10–30Myr and $\lesssim$200Myr; however, they note that BDs with ages older than $\sim$30Myr, such as the $\sim$120Myr ABDMG member 2MASS J03552337+1133437, can display very strong signs of low-gravity that correspond to the very low gravity ($\gamma$) classification.
We used our updated sample of low-gravity candidate members of YMGs to investigate this further. We inspected various spectroscopic index–spectral type relations of candidate members of different YMGs to identify any systematic correlation with age. We assigned the age of the most probable YMG to our candidates, while rejecting any candidate with ambiguous membership (Table \[tab:sptclass\]). We found that the strongest correlations with age in the $\sim$10–130Myr range resulted from: (1) the mean value of the EW of the three doublets at $1.169$$\mu$m, $1.177$$\mu$m and $1.253$$\mu$m; and (2) the mean gravity score defined by [@2013ApJ...772...79A]. The resulting sequences are presented in Figure \[fig:agecalib\]. Even though they do correlate with age on average, the scatter is too large to allow a precise determination of the age of an individual system from spectroscopic indices alone. We find that the H$_2$($K$) index defined by [@2013MNRAS.435.2650C] does not seem to correlate significantly with age in the 1–130Myr range. Our results seem to be in contradiction with the findings of [@2013MNRAS.435.2650C] that the H$_2$($K$) index is sufficient to differentiate between objects from populations of $\sim$1–2Myr, $\sim$3–10Myr and field dwarfs in the M8–L0 range: we observe an overlap of the typical values for H$_2$($K$) in populations of $\sim$1–2Myr and $\sim$5–15Myr. However, our results are consistent with H$_2$($K$) being a good gravity-sensitive index, as it discriminates between the field population and $\lesssim$100Myr dwarfs for spectral types in the M6–L1 range, or $\lesssim$130Myr for L2–L6. It is possible that interlopers from other young associations not considered in BANYAN II contaminate our sample, which would introduce noise in these relations. A full RV and parallax follow-up of the candidates presented here will be needed to assess this.
Model Comparison {#sec:models}
----------------
We used our sample of 86 new low-gravity M6–L5 dwarfs supplemented with 39 low-gravity and 131 field M6–L9 dwarfs from [@2013ApJ...772...79A] and the SpeX Prism Spectral Libraries to investigate the physical properties of our sample of young dwarfs, using BT-Settl atmosphere models . In Section \[sec:btsettl\], we focus on effective temperatures and surface gravities obtained from a comparison of our NIR spectra with atmosphere models. In Section \[sec:evmod\], we focus on the mass and radii that are obtained from a comparison of our photometry with evolution models.
### BT-Settl Atmosphere Models {#sec:btsettl}
used BT-Settl atmosphere models to determine the physical parameters of seven young L dwarfs and found that (1) low-gravity L0–L3 dwarfs fit models with similar temperatures of $\sim$1800K; (2) the continuum shape of the $H$ band is not well reproduced by solar-metallicity models; (3) the 1.1–2.5$\mu$m range in the $zJ$ bands is not well reproduced by models; and (4) the global continuum slope is not well reproduced by atmosphere models for L dwarfs.
We used a method similar to that of [@2008ApJ...678.1372C] and [@2014ApJ...787....5N] to identify the best fitting solar-metallicity CIFIST2011 BT-Settl atmosphere model for our observed spectra, on a grid of effective temperature and surface gravity ranging from [$T_{\mathrm{eff}}$]{} = 500–5000K and $\log g =$3.0–5.5dex with a grid spacing of 100K and 0.5dex, respectively. We computed the goodness-of-fit ($G_{k,j}$; @2008ApJ...678.1372C) in each case.
[@2008ApJ...678.1372C] demonstrated that [$T_{\mathrm{eff}}$]{} can only be recovered efficiently by performing such a model fitting on a very large spectral range in the case of field L1–L8 dwarfs; however, while fitting a single model spectrum in this way allows recovering a good [$T_{\mathrm{eff}}$]{} estimate, it does not reproduce well the general slope and the features in individual spectral bands. Since gravity-sensitive spectral features are generally narrow, this method will not yield good estimates of $\log g$. We have thus performed our model fitting in two different steps : (1) by fitting one single BT-Settl spectrum to the full 0.8–5$\mu$m range (*WISE* $W1$ and $W2$ magnitudes were added as additional data to our spectra in order to do this); and (2) by fitting one BT-Settl spectrum to each one of the $zJ$, $H$ and $K$ spectral bands. The first method allowed us to obtain an estimate of [$T_{\mathrm{eff}}$]{}, while the second one allowed us to obtain an estimate of $\log g$ for each object in our sample.
In order to append the *WISE* photometric data to an observed NIR spectrum, we compute the synthetic $J$, $H$ and $K_S$ *2MASS* magnitudes of the spectrum and determine the three corresponding normalization factors. We then use the median of these factors to bring back the two *WISE* photometric data points to the same scale as the observed spectrum. The dilution factor is treated as a free parameter in our analysis so that no estimate nor measurement of distance is needed in the model fitting. We thus choose the dilution factor that minimizes $G_{k,j}$ for each fitted model. We do so in an analytical way to decrease computing time. We thus define the goodness-of-fit as :
$$\begin{aligned}
G_{k,j} &= \sum_{i=1}^N W_{i,j} \left(\frac{F_{obs,i,j}-D_{k,j}F_{k,i,j}}{\sigma_{obs,i,j}}\right)^2\label{gof}, \\
\mbox{where } D_{k,j} &= \frac{\sum_{i=1}^N W_{i,j} \frac{F_{obs,i,j}F_{k,i,j}}{\sigma_{obs,i,j}^2}}{\sum_{i=1}^N W_{i,j} \frac{F_{k,i,j}^2}{\sigma_{obs,i,j}^2}}\label{dilute}, \\
W_{i,j} &= \frac{\left.\frac{\mathrm{d}\ln\lambda}{\mathrm{dx}}\right|_{x=x_{i,j}}}{\sum_{i^\prime=1}^N \left.\frac{\mathrm{d}\ln\lambda}{\mathrm{dx}}\right|_{x=x_{i^\prime,j}}},\label{eqweights}\end{aligned}$$
where $x_{i,j}$ is the pixel number (i.e., the spectral position), $j$ is the index of the spectral band (i.e., $zJ$, $H$ or $K$, applicable only when we fit by individual bands), $k$ is the atmosphere model index (each value of $k$ corresponds to a given combination of [$T_{\mathrm{eff}}$]{} and $\log g$), $N$ is the total number of pixels in the fitting range, $\lambda$ is the wavelength ($\mu$m), $W_{i,j}$ are the normalized weight factors, $D_{k,j}$ is the dilution factor that minimizes $G_{k,j}$, $F_{obs,i,j}$ and $\sigma_{obs,i,j}$ are the observed spectrum and its measurement error, and $F_{k,i,j}$ is an atmosphere model. The weights are chosen to ensure that equal wavelength ranges in log space equally contribute to the goodness-of-fit. For example, a broadband photometric measurement or one pixel of a low-dispersion spectroscopic order would be given a larger weight than one pixel of a high-dispersion spectroscopic order as it covers a larger wavelength range. [@2008ApJ...678.1372C] introduced this weighting method except that it was not done in log space; [@2014ApJ...787....5N] noted that the log space provides a more physically meaningful scale (i.e. using the log space prevents a bias that would be caused by working in wavelength space rather than frequency space).
We calculated errors on the adjusted parameters $a_l$ (i.e., [$T_{\mathrm{eff}}$]{} and $\log g$) from @Wolberg:2006ve ([-@Wolberg:2006ve]; p.50) :
$$\begin{aligned}
\sigma_{a_l,k,j} &= \sqrt{\frac{N}{N-2} G_{k,j}\ C_{l,l,k,j}^{-1}}\ \mbox{with}\label{erreq} \\
C_{l,m,k,j} &= \sum_{i=1}^N W_{i,j} \frac{\partial}{\partial a_l}\left(D_{k,j}F_{k,i,j}\right) \frac{\partial}{\partial a_m}\left(D_{k,j}F_{k,i,j}\right), \end{aligned}$$
where $C_{l,m,k,j}$ are elements of the correlation matrix. Equation and the equivalent expression of [@Wolberg:2006ve] differ by a factor $\sqrt{N}$ to compensate for our use of normalized weights in Equation . These error estimates do not take into account any systematic error in either our observations or the BT-Settl atmosphere models, and are thus only based on the variation of the goodness-of-fit with respect to each parameter. We show a few typical examples of per-band model fitting in Figure \[fig:spectra\_fitting1\].
As noted by , we find that the BT-Settl models generally fail to accurately reproduce the $zJ$-bands spectra of L dwarfs, especially at wavelengths smaller than $\sim$1$\mu$m; the general slope seems to be in agreement, but a high-gravity solution is almost always preferred for all L dwarfs. Moreover, the FeH absorption features at $\sim$1.6$\mu$m are not present at all in the atmosphere models, which could be explained by missing opacity sources in the synthetic models. For this reason, we have only kept results from the $H$ and $K$ bands to determine $\log g$. The adopted $\log g$ value is thus determined from the weighted mean of the values obtained from the $H$-band and $K$-band fitting, where the weights are set to the total values of $W_{i,j}$ (see Equation \[eqweights\]) within the fitting range divided by the inverse square of the individual measurement errors. This corresponds to the optimal weights that account both for the measurement error and the wavelength range used in the fitting process. Both the measurements and errors were rounded to the nearest half-integer and to the nearest factor of 100K in the case of $\log g$ and [$T_{\mathrm{eff}}$]{}, respectively. We imposed a floor on measurement errors that correspond to the grid size of our BT-Settl models, i.e. 0.5 and 100K for $\log g$ and [$T_{\mathrm{eff}}$]{}.
Our adopted [$T_{\mathrm{eff}}$]{} and $\log g$ values are listed in Table \[tab:physpar\] for our complete sample of low-gravity and field dwarfs. In Figures \[fig:spt\_teffk\], we show the spectral type–[$T_{\mathrm{eff}}$]{} sequence that we obtain, compared to various sequences from the literature [@2009ApJ...702..154S; @2004AJ....127.3516G; @2013AJ....146..161M]. We find [$T_{\mathrm{eff}}$]{} values that are consistent with the literature across the full range of spectral types, except for low-gravity objects which seem to be systematically cooler. This might be an additional indication that low-gravity brown dwarfs have cooler effective temperatures compared with field brown dwarfs of the same spectral types, an effect that was previously hypothesized and then demonstrated for the young, directly-imaged BD and exoplanet companions HD 203030 B, TWA 27 b, HR 8799 b and $\beta$ Pictoris b [@2006ApJ...651.1166M; @2011ApJ...733...65B; @2011ApJ...735L..39B; @2014ApJ...786...32M], as well as for young brown dwarfs ([@2012ApJ...752...56F; @2013ApJ...777L..20L]; Joseph C. Filippazzo et al., submitted to ApJ).
In Figure \[fig:spt\_logg\], we show the spectral type–$\log g$ sequence that we obtain for low-gravity and field dwarfs. The $\log g$ values that we derive for our low-gravity sample are systematically lower than those of our field sample, as expected. However, we observe a large scatter in the $\log g$ values of low-gravity dwarfs, although they are lower on average. This indicates that the model fitting method that we described above might not be very efficient in recovering low-gravity dwarfs in an ensemble of NIR spectra. Additionally, we derive slightly lower $\log g$ values for field dwarfs with spectral types M7 and L3–L6, indicating that the false positive rate might be larger when identifying low-gravity dwarfs based solely on model fitting in this range of spectral types. Our results also tentatively indicate that M7 dwarfs are systematically better fit by low-gravity atmosphere models, however this is based on only three objects and is thus possibly an effect of small number statistics.
We tried to reproduce the results of [@2008ApJ...678.1372C] showing that fitting individual bands yield systematically offset [$T_{\mathrm{eff}}$]{} values, and to extend this result to our full M6–L9 range as well as to low-gravity dwarfs. In Figure \[fig:spt\_teff\], we show the spectral type–[$T_{\mathrm{eff}}$]{} sequence that we obtain if we combine the $zJ$- and $H$-band measurements in a weighted mean (using similar weights than described above for $\log g$). We show that the systematic offsets in [$T_{\mathrm{eff}}$]{} values derived with this method are significant in the M9–L5 range, and independent of surface gravity. In Figure \[fig:Kbias\], we compare the difference of [$T_{\mathrm{eff}}$]{} values obtained from $zJ$- and $H$-band fitting to those obtained from the $K$-band fitting only. We show that [$T_{\mathrm{eff}}$]{} values derived from the $K$-band only are systematically warmer in the M9–L5 range. The values of [$T_{\mathrm{eff}}$]{} obtained from $K$-band fitting only are thus closer to those presented in Figure \[fig:spt\_teffk\], except that the scatter is much larger. These results confirm the findings of [@2008ApJ...678.1372C], while extending them to earlier spectral types (down to M9) and seem to indicate that the $zJ$ and $H$ bands are the most likely cause of the systematic offset in [$T_{\mathrm{eff}}$]{}.
It will be interesting to investigate whether fixing the [$T_{\mathrm{eff}}$]{} value using a large spectral coverage, and subsequently determining the best $\log g$ value using wavelength regions significantly smaller than a spectral band that are known to be gravity-sensitive, might provide a better way to determine accurate $\log g$ values for L dwarfs. This will be the subject of a future work, along with repeating this analysis with future generations of BT-Settl atmosphere models that include a more realistic treatment of dust clouds (see for a discussion on this topic).
### Evolution Models {#sec:evmod}
We estimated the physical parameters (mass, radius, [$T_{\mathrm{eff}}$]{}, $\log g$) of all low-gravity candidate members presented here from a comparison of their absolute *2MASS* and *WISE* photometry with isochrones from CIFIST2011 BT-Settl models using a likelihood analysis. The age range of the most probable host YMG was used in each case, and statistical distances from BANYAN II are used when a trigonometric distance is not available. These models do not account for magnetic fields and assume a hot-start formation (large initial entropy). Both effects could cause a systematic underestimation of mass [@2010ApJ...711.1087K; @2012ApJ...756...47S; @2014MNRAS.437.1378M; @2014ApJ...792...37M]. However, it has been demonstrated that BD masses derived from evolution models are systematically too large when compared to dynamical mass measurements . This seems in contradiction with what would be expected from the model limitations described above; instead, it is likely that the cooling rate of BDs is slowed down by atmospheric clouds, an effect that is not taken into account in current evolution models [@2015arXiv150306212D].
The resulting physical parameters are presented in Table \[tab:physpar\]. This allowed us to compile a total of 25 objects with an estimated mass in the planetary regime ($<$13[$M_{\mathrm{Jup}}$]{}); they are individually discussed in the Appendix. These objects are all likely located within 10–60pc and will constitute a sample of choice for a detailed study of the connection between the physical properties of BDs and giant, gaseous exoplanets, e.g. using the James Webb Space Telescope [@2006SSRv..123..485G].
In Figure \[fig:mr\_all\], we compare the masses and radii estimated for the objects in our sample with those of other known exoplanets and young BDs, as well as with BD radii measured by [@2014AJ....147...94D]. We show that our sample overlaps with the regime of giant, gaseous exoplanets. Our sample displays inflated radii and lower masses than field dwarfs, for given spectral types, which is expected for young, low-gravity low-mass stars and BDs.
Space Density at the Deuterium-Burning Limit {#sec:imf}
--------------------------------------------
Late-type members of YMGs provide the opportunity of measuring the low-mass end of the IMF which is still poorly constrained. The *BASS* survey is still not complete enough to construct individual IMFs for the YMGs under study, but we can already put constraints on the population of objects near the planetary-mass boundary where our survey is particularly sensitive.
We display in Figure \[fig:histmass\] a histogram of the estimated masses of all objects in our sample. We also display in this Figure a probability density function (PDF) that represents a continuous analog of the histogram which is independent on the binning and that includes individual measurement errors. This PDF is obtained by normalizing the integral of each individual mass estimation PDF to unity and summing them over the full sample. In the case of absolute $W1$ magnitudes, the PDFs that correspond to individual measurements were taken as normalized gaussian distributions with a characteristic width that corresponds to the measurement error.
There are 15 objects in our sample of THA candidates that have estimated masses in the 12.5–14[$M_{\mathrm{Jup}}$]{} range, which corresponds to the planetary-mass limit. This peak-shaped distribution of estimated masses for the THA candidate members uncovered here is the combined effect of a selection bias (we observed the latest-type objects first) and the distance distribution of THA members ($\sim$30–70pc; Paper II), as 12/15 of these objects are likely located within 50pc. Furthermore, we have identified a larger number of THA candidates compared to other YMGs, because its members are more easily identified in an all-sky search–the slightly larger distance of THA ensures that its members have a narrower distribution in space position and proper motion. The relatively large number of 12.5–14[$M_{\mathrm{Jup}}$]{} objects compared to objects in the 5–10[$M_{\mathrm{Jup}}$]{} or 15–75[$M_{\mathrm{Jup}}$]{} ranges is thus a selection effect.
Since our sample is biased on recovering objects more efficiently in the 12.5–14[$M_{\mathrm{Jup}}$]{} range, it remains useful to assess the space density of such objects. We will concentrate on the THA candidate members for this as they provide a larger sample. Assuming that we have uncovered all of the 12.5–14[$M_{\mathrm{Jup}}$]{} candidate members of THA in *BASS* within 50pc (accounting for 65.6% of the expected population according to our SKM model for THA) and correcting for the expected completeness of BASS for this association (90%; @2015ApJ...798...73G), we can expect that there are a total number of $20.3_{-5.1}^{+6.8}$ objects in THA that lie within this range of masses. The error was estimated assuming that the objects were drawn from a Poisson distribution, and they thus account for small number statistics.
Assuming that the population of 1.00–1.26[$M_{\odot}$]{} stars is complete in THA ($N=14_{-3.3}^{+4.3}$ using Poisson statistics, see Figure 8 of @2014AJ....147..146K) and adjusting a fiducial log-normal IMF peaking at 0.25[$M_{\odot}$]{}with a width $\sigma = 0.5$dex [@2012EAS....57...45J], we can expect a total of $356_{-47}^{+61}$ main-sequence stars in THA ($>$75[$M_{\mathrm{Jup}}$]{}) and only $0.56_{-0.13}^{+0.17}$ objects in the 12.5–14[$M_{\mathrm{Jup}}$]{} range (the ratio of 12.5–14[$M_{\mathrm{Jup}}$]{} to 1.00–1.26[$M_{\odot}$]{} objects derived from that IMF is $0.04$). We thus seem to be uncovering at least $36.4_{-12.5}^{+16.6}$ times too many objects in this mass range, compared to the predictions of a typical log-normal IMF anchored on the 1.00–1.26[$M_{\odot}$]{} population of THA.
It is possible that this is a consequence of a fault in the evolution models rather than a true over-population. For example, one could argue that the models fail to reproduce the effects of clouds which have a larger impact on the spectra of less massive, cooler objects. This could lead us to misinterpret the masses of our 15 low-gravity THA candidates, assigning them 12.5–14[$M_{\mathrm{Jup}}$]{} while their true masses span a larger range. If this effect alone is to explain the over-population, the true range of masses for our 15 objects would have to be extended by 190% in log space, which would mean that their true masses would span 4.5–39[$M_{\mathrm{Jup}}$]{}. This effect is thus unlikely to be the lone explanation of this over-population. It is also possible that the current age estimate of THA is wrong–e.g., $\beta$PMG, Upper Scorpius, AB Doradus and the Pleiades have recently been found to be slightly older than previously thought [@2005ApJ...628L..69L; @2012ApJ...746..154P; @2014ApJ...792...37M; @2014MNRAS.438L..11B; @2014MNRAS.445.2169M]. If it turns out that this is also the case for THA, our estimated masses would need to be shifted to larger values. As an example, doubling the age of THA would shift the estimated mass of a member from $\sim$13[$M_{\mathrm{Jup}}$]{} to $\sim$20[$M_{\mathrm{Jup}}$]{}. This effect alone would thus be insufficient to explain the large number of 12.5–14MJup THA candidates that we found. A similar shift of our estimated masses could be caused by systematics in evolution models (see our discussion in Section \[sec:evmod\]), although it is difficult to estimate the magnitude of this effect at this time. [@2015arXiv150306212D] has shown that masses from evolution model are likely under-estimated for dusty BDs at the L/T transition; it could be expected that the same effect is important in young L dwarfs. This would further accentuate the discrepancy between our observations and the predictions from a typical IMF.
[@2013ApJ...774...55B] noted that the age–absolute luminosity model sequences of $\sim$ 13[$M_{\mathrm{Jup}}$]{} and $\sim$ 25[$M_{\mathrm{Jup}}$]{} objects at different young ages overlap; such a pile-up in the isochrones could cause a degeneracy in our estimated masses and cause our method to mis-interpret true $\sim$ 25[$M_{\mathrm{Jup}}$]{} objects as planetary-mass objects. However, there are several observations that make this explanation unlikely: (1) The likelihood method with which we estimate masses not only generates a measurement and error bars, but it also provides a continuous PDF for each individual mass estimate. If this effect is important, we would thus be able to observe double-peaked individual measurement PDFs, as well as a peak at $\sim$ 25[$M_{\mathrm{Jup}}$]{} in the PDF displayed in Figure \[fig:histmass\]. Note that even if present, this effect would not introduce a second peak in the histogram, since it was constructed from the most probable values of the estimated masses only. (2) While the young age–absolute luminosity isochrones overlap at different masses, this effect is much more subtle in the individual $J$, $H$, $K_S$, $W1$ and $W2$ age–absolute magnitude isochrones. Furthermore, the slight overlap happens at slightly different ages and masses in the different filters, and allows to lift the degeneracy between $\sim$ 13[$M_{\mathrm{Jup}}$]{} and $\sim$ 25[$M_{\mathrm{Jup}}$]{} objects. This likely explains why we do not observe dual-valued mass estimate PDFs. (3) Performing a Monte Carlo analysis in which 20 and 40Myr isochrones are used to estimate the masses of a population of 20000 synthetic objects with true masses uniformly distributed between 4 and 80[$M_{\mathrm{Jup}}$]{} produces no over-density of estimated masses in the 12–14.5[$M_{\mathrm{Jup}}$]{} range. The absolute $J$, $H$, $K_S$, $W1$ and $W2$ magnitudes of these synthetic objects are obtained from the model isochrones themselves, hence this Monte Carlo analysis cannot be used to investigate systematics in the model cooling tracks. Instead, it only addresses the potential problem of overlapping isochrones that could produce degenerate mass estimates.
As a consequence of these observations, it does not appear that overlapping isochrones are the cause of the large population of 12–14.5[$M_{\mathrm{Jup}}$]{} THA candidates in our sample. We note that it is however possible that a fraction of these THA candidate members are contaminants in our analysis (i.e., young interlopers from other moving groups or associations, considered in BANYAN II or not) despite their high Bayesian probability and the low expected contamination rate in this particular YMG. It will be necessary to measure the RVs and parallaxes for all 12 objects discussed here to assess this, but at this stage it seems that this effect would be the most likely explanation for this over-density. For example, only 5/12 of these objects would need to be interlopers in order for the over-density to become a 1$\sigma$ result.
If we assume that the over-density is real, it would mean that there is at least one isolated dwarf in the 12.5–14[$M_{\mathrm{Jup}}$]{} range for every $17.5_{-5.0}^{+6.6}$ main-sequence star in THA. Comparing with the space density of main-sequence stars in the solar neighborhood ($9.3 \times 10^{-2}$ starspc$^{-3}$; @2005ASSL..327...41C) and assuming that the ratio we observed in THA is valid in the field, this would amount to a field density of $5.3_{-2.9}^{+3.8} \times 10^{-3}$ dwarfspc$^{-3}$ in the 12.5–14[$M_{\mathrm{Jup}}$]{} range in THA. At ages older than 2.5Gyr, they will all have temperatures below 450K that correspond to spectral types later than Y0, and will thus be hard to locate due to their extreme faintness . This is significantly larger than the lower limit measured by [@2011ApJS..197...19K] that corresponds to at least one $\geq$Y0 dwarf for every 78 main-sequence star (or $1.2 \times 10^{-3}$ dwarfspc$^{-3}$), especially when considering that the population of field $\geq$Y0 dwarfs is also probably composed of objects that span a large range of ages and thus masses. [@2011ApJS..197...19K] noted that their measurement is only a gross underestimation on the space density of Y-type dwarfs due to several biases. We note however that the IMF of YMGs might be different than that of the field, which could be yet another cause for this difference.
A less likely scenario is that our results could be an indication that we are approaching an up-turn in the IMF of isolated objects in THA with masses below the deuterium-burning limit: such an up-turn has already been hinted at by micro-lensing surveys in the galactic plane that measure $1.8^{+1.7}_{-0.8}$ Jupiter-mass object for every main-sequence star (corresponding to space density of $1.7^{+1.6}_{-0.7} \times 10^{-1}$ objectspc$^{-3}$; @2011Natur.473..349S). Measurements of RV and distance for the complete set of YMG candidates in *BASS* will be crucial to assess whether the observed over-density holds, and discovering YMG candidate members at even lower masses will provide a strong constraint on whether there is an up-turn in the IMF of YMGs.
Summary and Conclusions {#sec:conclusion}
=======================
We presented a NIR spectroscopic follow-up of 241 candidate members of YMGs identified through the *BASS*, *LP-BASS* and *PRE-BASS* samples. This allowed us to identify 108 new low-gravity M5–L5 candidate members of YMGs with estimated masses spanning the range of 7–189[$M_{\mathrm{Jup}}$]{}. Thirty-seven of these objects were previously known in the literature, but no signs of low gravity had been reported for them before this work. We complemented this unique sample with 22 low-gravity dwarfs from the literature to (1) build color–spectral type and absolute magnitude–spectral type sequences for field and young dwarfs; (2) show that some gravity-sensitive indices correlate with age in the 10–200Myr regime, albeit with a large scatter, such that low-resolution NIR spectroscopy does not allow a strong constraint on the age of an individual object; (3) we discuss some limitations of the current BT-Settl models, mainly their improper treatment of dust clouds in L-type dwarfs of all ages; and (4) show that we find an unexpectedly large number of isolated objects with estimated planetary masses in the Tucana-Horologium association, which might be caused by young interlopers from other moving groups. This study represents one of the first steps towards bridging the gap in our knowledge of the the space density of the lowest-mass BDs ($\sim$13[$M_{\mathrm{Jup}}$]{}; @2011ApJS..197...19K) and potential isolated giant planets that were ejected from their stellar system ($\sim$1[$M_{\mathrm{Jup}}$]{}; @2011Natur.473..349S). Additional figures, data and information on this work can be found on the website [www.astro.umontreal.ca/\\textasciitilde gagne](www.astro.umontreal.ca/\textasciitilde gagne) and in the Montreal Spectral Library, which is located at [www.astro.umontreal.ca/\\textasciitilde gagne/MSL.php](www.astro.umontreal.ca/\textasciitilde gagne/MSL.php).
The authors would like to thank the anonymous referee who suggested to improve BANYAN II with the inclusion of parallax motion and significantly helped to improve the quality of this paper, as well as make it more concise and clear. We would like to thank Robert Simcoe, Philippe Delorme, Michael C. Cushing, Rebecca Oppenheimer, Amélie Simon, Gilles Fontaine, Sergio B. Dieterich, Benjamin M. Zuckerman, André-Nicolas Chené, Sarah Jane Schmidt, Simon Coudé, Daniella C. Bardalez Gagliuffi and Jonathan B. Foster for useful comments and discussions. We thank Katelyn N. Allers, Michael C. Liu, Federico Marocco and Brendan P. Bowler for sharing data. We also thank all observatory staff and observers who helped us in this quest - Bernard Malenfant, Ghislain Turcotte, Pierre-Luc Lévesque, Alberto Pastén, Rachel Mason, Stuart Ryder, Rubén Díaz, Stéphanie Côté, John P. Blakeslee, Mischa Schirmer, Andrew McNichols, Dave Griep, Brian Cabreira, Tony Matulonis, German Gimeno, Steve Margheim, Percy L. Gomez, René Rutten, Bernadette Rodgers, Tim J. Davidge, Jaehyon Rhee Jay, Inger Jørgensen, Thomas L. Hayward, Andrew Cardwell, Blair C. Conn, Eleazar Rodrigo Carrasco Damele, David A. Krogsrud, Eduardo Marin, Erich Wenderoth, Fredrik T. Rantakyro, Joanna E. Thomas-Osip, Pablo Patricio Candia, Pascale Hibon, Cláudia Winge, Benoit Neichel, Peter Pessev, Matthew B. Bayliss and Anne Sweet. This work was supported in part through grants from the Fond de Recherche Québécois - Nature et Technologie and the Natural Science and Engineering Research Council of Canada. This research has benefited from the SpeX Prism Spectral Libraries, maintained by Adam Burgasser at [http://pono.ucsd.edu/\\textasciitilde adam/browndwarfs/spexprism](http://pono.ucsd.edu/\textasciitilde adam/browndwarfs/spexprism), as well as the M, L, T and Y dwarf compendium housed at <http://DwarfArchives.org> and maintained by Chris Gelino, Davy Kirkpatrick, and Adam Burgasser, whose server was funded by a NASA Small Research Grant, administered by the American Astronomical Society. This research made use of: the SIMBAD database and VizieR catalog access tool, operated at the Centre de Données astronomiques de Strasbourg, France ; data products from the Two Micron All Sky Survey (*2MASS*; [@2006AJ....131.1163S; @2003yCat.2246....0C]), which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center (IPAC)/California Institute of Technology (Caltech), funded by the National Aeronautics and Space Administration (NASA) and the National Science Foundation [@2006AJ....131.1163S]; the Extrasolar Planets Encyclopaedia ([exoplanet.eu](exoplanet.eu)), which was developed and is maintained by the exoplanet TEAM; data products from the *Wide-field Infrared Survey Explorer* (*WISE*; [@2010AJ....140.1868W]), which is a joint project of the University of California, Los Angeles, and the Jet Propulsion Laboratory (JPL)/Caltech, funded by NASA; the NASA/IPAC Infrared Science Archive (IRSA), which is operated by JPL, Caltech, under contract with NASA; the Infrared Telescope Facility (IRTF), which is operated by the University of Hawaii under Cooperative Agreement NNX-08AE38A with NASA, Science Mission Directorate, Planetary Astronomy Program; the Database of Ultracool Parallaxes maintained by Trent Dupuy [@2012ApJS..201...19D]; the Hale 5 m telescope at Palomar Observatory, which received funding from the Rockefeller Foundation; and of tools provided by Astrometry.net. This paper includes data gathered with the 6.5 meter Magellan Telescopes located at Las Campanas Observatory, Chile (CNTAC program CN2013A-135). Based on observations obtained at the Gemini Observatory through programs number GN-2013A-Q-118, GS-2013B-Q-79, GS-2014A-Q-55, GS-2014B-Q-72, GS-2014B-Q-47 and GS-2015A-Q-60. The Gemini Observatory is operated by the Association of Universities for Research in Astronomy, Inc., under a cooperative agreement with the National Science Foundation (NSF) on behalf of the Gemini partnership: the NSF (United States), the National Research Council (Canada), CONICYT (Chile), the Australian Research Council (Australia), Ministério da Ciência, Tecnologia e Inovação (Brazil) and Ministerio de Ciencia, Tecnología e Innovación Productiva (Argentina). All data were acquired through the Canadian Astronomy Data Center and part of it was processed using the Gemini IRAF package. This material is based upon work supported by AURA through the National Science Foundation under AURA Cooperative Agreement AST 0132798 as amended. This publication uses observations obtained at IRTF through programs number 2007B023, 2007B070, 2008A050, 2008B054, 2009A055, 2010A045, 2011B071, 2012A097, 2012B015, 2013A040, 2013A055, 2013B025, 2014B026 and 2015A026. The authors 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.
*Facilities:* IRTF (SpeX), Magellan:Baade (FIRE), Gemini:South (Flamingos-2), Gemini:North (GNIRS), Hale (TripleSpec).
DISCUSSIONS ON INDIVIDUAL OBJECTS {#App:discussion}
=================================
Several objects presented here deserve a detailed discussion, either because they display peculiar features, or were reported in the literature as candidate members of other YMGs. Additionally, optical spectra were available in the literature for some objects discussed here, and can serve as an independent assessment of low surface gravity.
Potential Planetary-Mass Low-Gravity Candidate Members of YMGs {#sec:potplan}
--------------------------------------------------------------
We list in Table \[tab:jwst\] twenty potential isolated planetary-mass objects in our sample, ten of which were discovered as part of this work. A few of these objects deserving further discussion are listed below.
**2MASS J05012406–0010452** was discovered by [@2008AJ....136.1290R] as an L4 dwarf in the optical, and was categorized as a low-gravity L4$\gamma$ by [@2009AJ....137.3345C], using its optical spectrum. [@2013ApJ...772...79A] categorized it as a very-low gravity L3 dwarf in the NIR, whereas we categorize it as an L4$\gamma$ dwarf. [@2012ApJ...752...56F] measured a trigonometric distance of $13.1 \pm 0.8$pc. We recovered this object in *BASS* as an ambiguous candidate member of Columba or Carina with respective Bayesian probabilities of 49% and 17%, taking the trigonometric distance measurement of [@2012ApJ...752...56F] into account. If this object is a member of either COL or CAR (both YMGs are coeval at 20–40Myr), it has an estimated mass of $10.2^{+0.8}_{-1.0}$[$M_{\mathrm{Jup}}$]{}. independently measured a trigonometric distance of $19.6 \pm 1.4$pc, which is discrepant with that of [@2012ApJ...752...56F] at the 5$\sigma$ level. The reason for this large discrepancy is unclear; the measurement of [@2012ApJ...752...56F] used a smaller number of epochs (11 versus 21); however, they were spread across a larger temporal coverage (3yr versus 2yr). If we adopt the distance measurement of , the CAR membership probability becomes negligible and that of COL becomes considerably smaller (7.2%), although we also calculate a low field contamination probability (1.3%). It will be necessary to better constrain the distance of this object to assess whether it is a viable candidate member of COL or CAR. Obtaining an RV measurement would also be useful for this.
**2MASS J05120636–2949540** has been identified as an L4.5 dwarf in the optical by [@2003AJ....126.2421C; @2008ApJ...689.1295K], and [@2014ApJ...794..143B] obtained a NIR spectrum to categorize it as an L4.5$ \pm 2$ dwarf. In Paper II, we determined that this object is a low-probability candidate member of $\beta$PMG. We used the NIR spectrum of [@2014ApJ...794..143B] to revisit its spectral classification: we find that this object is a very good match to our L5$\beta$ template; however, the method of [@2013ApJ...772...79A] assigns it an intermediate gravity. We note that the VO$_Z$ index is significantly larger than that of field L5 dwarfs, but [@2013ApJ...772...79A] only use this index within the L0–L4 spectral types, as later-type low-gravity dwarfs in their sample displayed similar VO absorption than that of field dwarfs of the same spectral types. However, only one low-gravity L5 dwarf was available at the time, hence it is possible that the VO$_Z$ index remains useful to discriminate low-gravity L5 dwarfs. For this reason, we adopt the L5$\beta$ spectral type. Due to its low-gravity features, this object is preserved as a candidate member of $\beta$PMG. This object has one of the lowest estimated masses among the YMG candidates presented here, with $6.7^{+1.0}_{-0.9}$[$M_{\mathrm{Jup}}$]{}. Its statistical distance associated with membership to $\beta$PMG is $10.9^{+4.4}_{-4.0}$pc, which makes it a valuable benchmark to study the atmosphere of planetary-mass objects.
**2MASS J12074836–3900043** (2MASS J1207–3900) was discovered as a candidate member of TWA in *BASS*. Its discovery and NIR spectroscopic follow-up have been presented in [@2014ApJ...785L..14G]. They reported an optical spectral type L0$\gamma$ and a NIR spectral type L1$\gamma$. Here we used the spectra of several low-gravity candidate members of Upper Scorpius obtained by [@2008MNRAS.383.1385L] to define tentative templates for the spectral type L0$\delta$, which likely correspond to objects younger than $\sim$15–20Myr and have an even more triangular $H$-band continuum than the L0$\gamma$ type. Given that both the optical and NIR spectra of 2MASS J1207–3900 are peculiar even in comparison to the best template matches (L0$\gamma$ and L1$\gamma$ respectively) and that its $H$ band continuum is more triangular than any $\beta$ or $\gamma$ template, we revised its spectral classification by comparing it to Upper Scorpius candidate members. We find that the best match is the L0$\delta$ template; however, 2MASS J1207–3900 displays features that are attributable to a later spectral type (redder slopes at 1.2–1.35$\mu$m and 1.5–1.6$\mu$m). We thus suggest a tentative spectral type of L1$\delta$ for this object, but identifying other similar objects will be necessary to confirm this. If it is a member of TWA (5–15Myr), this object has an estimated mass of $12.1^{+1.4}_{-2.0}$[$M_{\mathrm{Jup}}$]{} and a statistical distance of $58.2^{+6.8}_{-6.4}$pc.
**2MASS J12271545–0636458** was identified as an M9 dwarf by [@2003AJ....126.2421C] using optical spectroscopy. We identified it as a candidate member of TWA in *PRE-BASS*, and NIR spectroscopy allowed us to categorize it as a low-gravity M8.5$\beta$ dwarf. It was initially rejected from the *BASS* sample because of its low Bayesian probability, which is in part due to the fact that its kinematic distance of $32.5 \pm 3.2$pc if it is a member of TWA does not match its spectrophotometric distance ($63.2 \pm 11.4$pc). The latter estimate would place 2MASS J12271545–0636458 at the far-end of the TWA members ($\sim$40–62pc; Paper II; ). This is reminiscent of 2MASS J12474428–3816464, TWA 29 and TWA 31, which are young and seem to be located between TWA and SCC in terms of distance ([@2003ApJ...599..342S; @2012ApJ...754...39S; @2014ApJ...785L..14G]; Paper V). Measurements of distance and RV will be useful to assess whether this is a true member of TWA despite its small Bayesian probability. If it is a true member of TWA (5–15Myr) located at its statistical distance, this object has an estimated mass of $11.6^{+1.4}_{-1.9}$[$M_{\mathrm{Jup}}$]{}.
**2MASS J12563961–2718455** was identified in *PRE-BASS* as a low-probability candidate member of TWA. NIR spectroscopy revealed that this object is a low-gravity L3$ \pm 1$$\beta$ dwarf. The probability that this object belongs to TWA is lower than 20%, but the field contamination probability is also very low at $<$0.1%. This usually points out to either an incomplete SKM for the YMG or to contamination from a source not taken into account in BANYAN II. The most likely a priori explanation would be that this object is a contaminant from SCC (located at $\sim$100–150pc; ); however, the spectrophotometric distance of 2MASS J12563961–2718455 ($43.1 \pm 3$pc) is not consistent with this hypothesis, even when its low gravity is taken into account. Using its *2MASS* and *WISE* photometry and comparing it with other known low-gravity L4 dwarfs, we can rule out a distance larger than 48.5pc at a 95% confidence level, assuming this object is not an unresolved multiple system. We can hence conclude that as long as this object is not extremely peculiar for a low-gravity L4 dwarf or a multiple system composed of four equal-luminosity components, it cannot be a member of SCC. The statistical distance from BANYAN II which is associated to the TWA hypothesis ($46.2^{+4.8}_{-4.4}$pc) is similar to those of bona fide members of TWA ($\sim$40–62pc; Paper II; ), hence this case is different from those of 2MASS J12271545-0636458, 2MASS J12474428–3816464, TWA 29 and TWA 31, which are young and seem to be located between TWA and SCC in terms of distance ([@2003ApJ...599..342S; @2012ApJ...754...39S; @2014ApJ...785L..14G]; Paper V). Obtaining a distance measurement for this object will be helpful to assess whether it is a member of TWA. Assuming an age of 5–15Myr and comparing its statistical distance from BANYAN II with BT-Settl *2MASS* and *WISE* isochrones, the estimated mass of this object is $7.7^{+1.4}_{-1.5}$[$M_{\mathrm{Jup}}$]{}, amongst the lowest of all candidate YMG members reported here. Its statistical distance associated with membership to TWA is $44.6 \pm 5.2$pc.
**2MASS J21324036+1029494** was discovered as an L4.5$\pm 1$ dwarf by [@2006AJ....131.2722C] using low-S/N NIR spectroscopy. We identified it as a candidate member of ARG from *PRE-BASS*. The NIR spectrum obtained by [@2006AJ....131.2722C] is available in the SpeX PRISM Spectral Libraries, we thus retrieved it to assess whether it is a low-gravity dwarf. We categorize this object as an L4:$\beta$ dwarf. Its $H$-cont index [@2013ApJ...772...79A] is consistent with low-gravity objects; however, the quality of the data is not sufficient to assess whether its FeH$_Z$ and KI$_J$ indices are consistent with this. Obtaining a better-quality and higher-resolution NIR spectrum will be useful to confirm the spectral type of this object. If it is a member of ARG (30–50Myr), this object has an estimated mass of $11.4 \pm 0.4$[$M_{\mathrm{Jup}}$]{} and a statistical distance of $34.2 \pm 4.8$pc.
Low-Gravity Candidate members of YMGs {#sec:lowgm}
-------------------------------------
**2MASS J00413538–5621127** (DENIS-P J00041353–562112) has been identified as a candidate nearby, red dwarf by , and spectroscopically confirmed by [@2007AJ....133.2258S] as an active M8 dwarf. Using high resolution optical spectroscopy, [@2009ApJ...702L.119R] revised its spectral type to M7.5 and showed evidence that it displays Li and signatures of active accretion, which indicates that it is a young, $\sim$10Myr BD. Based on its position, proper motion and RV, they suggest that it could be a member of THA, or an ejected member of $\beta$PMG, which would make it the first accreting BD discovered in either of these associations. [@2010ApJ...722..311L] reported that it is a binary with estimated spectral types of M6.5$\pm 1$ and M8 from photometry. In Paper II, we corroborated that it is a high-probability candidate member of THA, with estimated masses of 14–41[$M_{\mathrm{Jup}}$]{} and 18–41[$M_{\mathrm{Jup}}$]{} for the individual components. This object was retrieved in *BASS* as a high-probability candidate of THA. We obtained NIR spectroscopy for the unresolved system, and categorize it as a very low-gravity M7.5$\gamma$ BD system, which is consistent with its young age.
**2MASS J02590146–4232204** was identified by [@2013ApJ...774..101R] as a candidate member of COL with infrared excess indicative of the presence of a circumstellar disk host. We independently identified this object as a candidate member of COL in *PRE-BASS*; however, it was subsequently rejected from *BASS* because of its low membership probability and the fact that its *WISE* colors did not survive the extragalactic filter defined by [@2011ApJS..197...19K], which is likely a consequence of its infrared excess. NIR spectroscopy allowed us to categorize it as an M5$\gamma$ dwarf. This is consistent with the results of [@2013ApJ...774..101R], who reported that this object displays weak absorption that is indicative of a low surface gravity. Including the RV measurement of $15.3 \pm 1.5$ from [@2013ApJ...774..101R], we find that this object is a low-probability candidate member of COL: this conclusion differs from that of [@2013ApJ...774..101R], which found that 2MASS J02590146–4232204 is a candidate member of THA. Obtaining a trigonometric distance will be useful to assess whether this object is a member of COL or THA.
**2MASS J03264225–2102057** has been identified as an L4 dwarf with Li absorption by [@2007AJ....133..439C]. Using the DUSTY evolution models [@2000ApJ...542..464C], the presence of Li and the spectrophotometric absolute magnitude of this object, they determined that it should be younger than 500Myr and less massive than 50[$M_{\mathrm{Jup}}$]{}. We identified this object as a highly probable L5$\beta$/$\gamma$ candidate member of ABDMG in *PRE-BASS*. [@2002AJ....124.1170D] measured a trigonometric distance of $32.3 \pm 1.6$pc that is consistent with membership to ABDMG. The presence of low-gravity feature in its optical and NIR and optical spectra puts a slightly stronger constraint on the age of 2MASS J03264225–2102057, since it is expected that gravity-sensitive spectral indices remain useful only up to $\sim$200Myr [@2009AJ....137.3345C; @2013ApJ...772...79A]. We therefore categorize this object as a low-gravity L5$\beta$/$\gamma$ dwarf. An RV measurement is needed before it can be assessed whether this object is a bona fide member of ABDMG.
**2MASS J04493288+1607226** was identified in *PRE-BASS* as a candidate member of $\beta$PMG, but was rejected from the *BASS* sample because of its proximity with TAU. NIR spectroscopy revealed that it is a low-gravity M9$\gamma$ dwarf. We estimate a distance of $54.9 \pm 10.0$pc for this object by comparison with other low-gravity dwarfs. A distance larger than 82pc can be excluded at a 99% confidence level, which is incompatible with membership to TAU ($140 \pm 20$pc; @2007ApJ...671.1813T) unless it is an unresolved multiple with at least 3 individual equal-luminosity components. This scenario is unlikely, especially considering that the NIR spectrum of 2MASS J04493288+1607226 is not reddened. We thus preserve this object as a candidate member of $\beta$PMG.
**2MASS J11083081+6830169** has been discovered by [@2000AJ....120.1085G] as an L1 dwarf in the optical with H$\alpha$ emission. We recovered this object in *BASS* as a candidate member of ABDMG, and [@2002ApJ...575..484G] identified it as a candidate member of TWA. The RV of $-9.8 \pm 0.1$ measured by [@2010ApJ...723..684B] does not match the predicted RV of $-18.9 \pm 1.5$ for membership to ABDMG. It closely matches that of the CAR hypothesis ($-9.7 \pm 0.8$), but it still obtains a very low Bayesian probability of being a member of CAR. Its statistical distance ($15.3 \pm 0.8$pc) places it right into the locus of known young L dwarfs in both an $M_{W1}$ versus $J - K_S$ and $M_{W1}$ versus $H - W2$ CMDs. This distance places it at only 0.27 of the CAR bona fide member HIP 33737 in $UVW$ space, and at 17.2pc of the CAR bona fide member GJ 2079 in $XYZ$ space. We show in Figure \[fig:J1108XYZUVW\] its $XYZUVW$ position at its most probable distance: it seems that this object has a most probable position that is consistent with bona fide members of CAR, but our SKM fails to represent this. It can be expected that our SKM of CAR is not accurate because it was derived from a small number of bona fide members. Furthermore, both the NIR spectrum that we obtained and the optical spectrum from the RIZzo spectral library display clear signs of low-gravity and allowed us to categorize it as an L1$\gamma$ dwarf, which is consistent with membership to a YMG. A measurement of this object’s trigonometric distance will be useful to assess whether or not it is a member of CAR, but we note that it is likely a member despite its low Bayesian membership probability.
**2MASS J12265135–3316124** (TWA 32) has been identified by [@2011ApJ...727....6S] as an UV-bright M6.5 low-mass star. They measured strong H$\alpha$ emission and Li absorption, as well as an RV of $7.15 \pm 0.26$. They used this information as well as a photometric distance ($53 \pm 5$pc) to identify it as a new member of TWA, and they noted that it is a $656.1 \pm 0.4$mas visual binary with near-equal luminosity. [@2011ApJ...727...62R] independently discovered this object and measured strong H$\alpha$ and emission at 5876Å and 6678Å, as well as strong Li absorption. They argued that the H$\alpha$ full width at 10% of 270 is consistent with this object being a classical T Tauri star. They measured an RV of $14.8 \pm 3$ and note that its $UVW$ space velocity is consistent with TWA and SCC. We recovered this object in *PRE-BASS* as a candidate member of TWA, and obtained NIR spectroscopy that allowed us to assign it a spectral type of M5.5$\gamma$. At this spectral type, only the weaker absorption is a useful low-gravity indicator. We adopted the RV measurement of [@2011ApJ...727....6S] which is more precise, and, like them, found that this object is a strong candidate member of TWA. The BANYAN II statistical distance corresponding to the TWA hypothesis is $61.8^{+6.4}_{-6.0}$pc, which is consistent with the photometric estimate of [@2011ApJ...727....6S] that takes its binary nature into account. The SKMs of BANYAN II do not take SCC into account, which includes the Lower Centaurus Crux (LCC) and the Upper Centaurus Lupus (UCL) regions, hence our result does not preclude membership to SCC. The space velocity $UVW$ for this object is ($-8.6 \pm 1.4$, $-15.7 \pm 1.1$, $-3.4 \pm 1.1$) [@2011ApJ...727....6S], at 4.6 from the kinematic center of TWA (Paper II), 4.6 from that of UCL and 4.2 from that of LCC . Its kinematics are thus consistent with SCC and TWA; however, its photometric distance is not consistent with the distance of this complex ($\sim$100–150pc; ), whereas it is consistent with that of TWA members ($\sim$40–62pc; Paper II; [@2013ApJ...762..118W]; ). We conclude that TWA 32 is a likely member of TWA, unless it is a multiple system composed of at least three equal-luminosity components. A trigonometric distance measurement will be useful to assess this.
**2MASS J20391314–1126531** was discovered as an M9 dwarf by [@2003AJ....126.2421C] using optical spectroscopy. [@2010MNRAS.409..552G] reported that it is a candidate member of the Pleiades stream. demonstrated that the Pleiades stream is not a moving group but rather a dynamical stream of stars without a common origin. We identified 2MASS J20391314–1126531 as a candidate member of ABDMG as part of *PRE-BASS* and obtained NIR spectroscopy which revealed that this is a low-gravity M7$\beta$ dwarf. The RV of $-18.0 \pm 2$ that was measured by [@2010MNRAS.409..552G] is consistent with a membership to ABDMG, and the fact that it has a low gravity indicates that it might not be a contaminant from the Pleiades stream. A measurement of its distance will be needed to assess this.
Candidate members of YMGs with no age constraint {#sec:noconstraintmg}
------------------------------------------------
**2MASS J03582255–4116060** has been discovered by [@2007AJ....133..439C] as an L5 BD in the optical. We identified it as a low-probability candidate member of $\beta$PMG as part of *BASS*. $R \sim 75$ NIR spectroscopy allowed us to categorize it as a peculiar L6 dwarf. Its continuum is redder and its $H$ band is slightly more triangular than our field L6 template, however it is unclear at this time if these effects are due to a low gravity or not. Obtaining a higher-resolution spectrum would be useful to assess this. If we assume an age of 20–26Myr and the BANYAN II statistical distance associated with the $\beta$PMG hypothesis ($18.1 \pm 3.2$pc) and compare its *2MASS* and *WISE* photometry with the BT-Settl isochrones, we find that this object has one of the lowest estimated mass of all candidate YMG members reported here, with $8.2 \pm 0.6$[$M_{\mathrm{Jup}}$]{}.
**2MASS J08095903+4434216** was identified by [@2004AJ....127.3553K] and confirmed by [@2006AJ....131.2722C] as an L6 dwarf. In Paper V, we identified it as a candidate member of ARG as part of *BASS*. We used its NIR spectrum to revise its spectral type to L6pec(red) from a visual comparison with field and low-gravity templates. This object has a red continuum and red NIR colors for its spectral type, with $J - K_S = 2.02$ and $J - W2 = 3.63$, compared with median values of $J - K_S = 1.7 \pm 0.3$ and $J - W2 = 2.9 \pm 0.4$ for field L6 dwarfs (Figure \[fig:CC2\]). The low-resolution gravity classification scheme of [@2013ApJ...772...79A] categorizes it as an intermediate-gravity L5.4 dwarf due to its $H$–cont, KI$_J$ and FeH$_Z$ indices. However, it is visually a better match to the field L6 template than the field L5 template, albeit it displays a slightly redder continuum. Adopting a spectral type of L6, only the $H$–cont index remains useful and categorizes it as an intermediate-gravity dwarf, but this index alone does not reject the possibility that this object is a dusty dwarf in the field. [@2014AJ....147...34S] demonstrated that the H$_2(K)$ index defined by [@2013MNRAS.435.2650C] seems to be gravity-sensitive up to at least L8; we obtain a value of H$_2(K) = 1.056 \pm 0.008$ for 2MASS J08095903+4434216, which is slightly lower than the typical values for field L6 dwarfs ($1.06 \pm 0.01$; see Figure \[fig:agecalib2k\] of this work and Figure 10 of @2014AJ....147...34S). It is unclear at this time if this object is a low-gravity L6 dwarf; a higher resolution (R$\gtrsim$750) NIR spectrum will be useful to confirm if this object is a very low-mass, very late-type candidate member of ARG, or more massive and dusty field interloper. At the age of ARG (30–50Myr), this object would have one of the lowest estimated masses amongst all YMG candidates presented here, with $8.1 \pm 0.8$[$M_{\mathrm{Jup}}$]{}. Its statistical distance associated with membership to ARG is $15.3 \pm 2.0$pc.
**2MASS J23512200+3010540** was discovered by [@2010ApJS..190..100K] as L5.5 dwarf in the optical, and as an unusually red L5.5 dwarf in the NIR. In Paper II, we identified this object as a candidate member of ARG, and it was recovered as such in *PRE-BASS*. We used its NIR spectrum to categorize it as a peculiar L5 dwarf. Only the $H$–cont index is indicative of a possible young gravity; it is thus a likely scenario that this object is a dusty field interloper. Obtaining a higher-resolution spectrum would be useful to assess this. If it is a member of ARG (30–50Myr), this object has an estimated mass of $10.0^{+0.6}_{-0.7}$[$M_{\mathrm{Jup}}$]{} and a statistical distance of $20.1 \pm 1.8$pc.
Interlopers from the Field or Other Regions {#sec:fieldinterlop}
-------------------------------------------
**2MASS J00174858-0316334** was identified as a candidate member of ABDMG as part of *PRE-BASS*. NIR spectroscopy revealed that this is a reddened low-gravity M7$\beta$ dwarf. We de-reddened its spectrum using the *fm\_unred.pro* IDL routine based on the extinction law of [@1999PASP..111...63F] and visually compared it with our M7$\beta$ template to determine that its total extinction is $A(V) = 2.5$. We used the parametrization of [@1999PASP..111...63F] with a total-to-selective extinction of $R(V) = 3.1$. This reddening is unlikely caused by interstellar dust, since this object is far from the galactic plane ($b = -64.8$) and has a spectro-photometric distance of only $65.1 \pm 11.5$pc (low gravity was considered in this estimate). It can be expected that this object is still embedded in its formation material, which indicates that it is not a member of ABDMG, but rather a member of another young star-forming region that might not be known. It would be interesting to investigate whether other very young objects can be found in its vicinity.
**2MASS J00461551+0252004** was identified as a candidate member of ABDMG in *PRE-BASS*. NIR spectroscopy revealed that it is a peculiar L0 dwarf with no indication of low gravity. The $H$-band bump at 1.57$\mu$m is significantly stronger than that of field dwarfs, which could hint at an unresolved T-type component (Figure \[fig:J0046\_pec\]); however, the spectral indices constructed by [@2014ApJ...794..143B] do not categorize it as a likely L-type + T-type binary. The cause of its peculiar properties is thus unclear.
![NIR spectrum of the peculiar L0 dwarf 2MASS J00461551+0252004 that was recovered as a candidate member of ABDMG in *PRE-BASS*. The $H$-band bump at 1.57$\mu$m could be a hint of an unresolved T-type companion.[]{data-label="fig:J0046_pec"}](Figure20.pdf){width="48.00000%"}
**2MASS J02441019–3548036** was discovered as candidate member of THA in *BASS*. NIR spectroscopy ($R\sim 750$) reveals that it is an L2 dwarf that lacks the weaker alkali lines or stronger VO absorption that are typical of low-gravity dwarfs. However, its continuum is unusually red for an L2 dwarf and the shape of its $H$ band is unusual (Figure \[fig:J0244\_pec\]). This could be explained by an unusually dusty atmosphere or an unresolved later-type companion; however, the classification of [@2014ApJ...794..143B] based on various spectral indices does not categorize it as a candidate binary. We reject it as a candidate member of THA, as the weaker-than-usual alkali lines are not consistent with a young age even if this object is dusty or multiple.
![NIR spectrum of the peculiar L1 dwarf 2MASS J02441019–3548036 that was recovered as a candidate member of THA in *BASS*. Its NIR continuum is redder than usual and the shape of its $H$ band is peculiar. This could be explained by a dusty atmosphere; however, this object is likely older than the YMGs considered here.[]{data-label="fig:J0244_pec"}](Figure21.pdf){width="48.00000%"}
**2MASS J02530084+1652532** (Teegarden’s star) was discovered by [@2003ApJ...589L..51T] as a nearby ($2.43 \pm 0.54$pc) M6.5 dwarf; [@2006AJ....132.2360H] refined its distance measurement to $3.85 \pm 0.01$pc. identified this object as a potential low-gravity dwarf from atmosphere model fitting, and we identified it as a candidate member of ARG in Paper II, but its distance is not consistent with this possibility. We obtained the NIR spectrum of [@2012ApJ...745...26B] from the SpeX PRISM spectral library and categorized it as an M7.5$\beta$ dwarf. The classification scheme of [@2013ApJ...772...79A] assigns it an intermediate gravity, which is consistent with our visual comparison. This is due to a low FeH$_Z$ index ($1.0699 \pm 0.0040$), a low KI$_J$ index ($1.0511 \pm 0.0065$) and a high $H$-cont index ($0.9974 \pm 0.0049$) compared with field M7.5 dwarfs. This object has a relatively blue *2MASS* $J - K_S$ color ($0.809 \pm 0.053$) for an M7.5 dwarf (Figure \[fig:JK\]), which is not expected for a low-gravity dwarf. We used the NIR spectrum to measure its synthetic NIR colors and obtained $J - K_S = 0.869 \pm 0.077$ (assuming a 5% uncertainty in the *2MASS* photometric zero points), which places it closer to the locus of low-gravity and field M7 dwarfs, albeit still on the blue end. Its *WISE* $W1 - W2$ color ($0.265 \pm 0.034$) is consistent with field and low-gravity M7.5 dwarfs (Figure \[fig:W1W2\]). Obtaining a higher-resolution NIR spectrum will be useful to assess whether alkali lines are weaker than usual, which would confirm if this object has a low gravity or not. Another explanation could be that this object is an unresolved binary. If Teegarden’s star is young, it could be a member of a YMG that is not considered here, and it would thus be interesting to measure its RV. It is worthwhile mentioning that this object would be the nearest low-gravity dwarf if this is confirmed, a record currently held by LP 944–20 (an L0$\beta$ at $6.41 \pm 0.04$pc that is a candidate member of the Castor stream; ).
**2MASS J03140344+1603056** was identified by [@2007AJ....133.2258S] as an L0 dwarf with H$\alpha$ emission. measured its RV and used its kinematics to assign it as a candidate member of UMA. We initially identified this object as a low-probability candidate member of $\beta$PMG in *PRE-BASS*, but was later rejected because of its large contamination probability, as well as its position on a $M_{W1}$ versus $H-W2$ diagram that is not consistent with young BDs at the most probable statistical distances obtained from BANYAN II. A NIR follow-up allowed us to categorize it as a peculiar M9 dwarf with no apparent sign of low gravity from the classification scheme of [@2013ApJ...772...79A] or a visual comparison with spectroscopic standards. However, it is unclear at what exact age signs of low-gravity stop being apparent in moderate-resolution NIR spectra, and [@2013ApJ...772...79A] suggest that this might take place around $\sim$200Myr. We thus reject any possible membership with the younger moving groups considered here, but our data is insufficient to corroborate its possible membership to UMA.
**2MASS J04070752+1546457** has been identified as an L3.5 dwarf by [@2008AJ....136.1290R] from optical spectroscopy. We identified this BD as an ambiguous candidate member of $\beta$PMG and COL as part of *PRE-BASS*, but we subsequently rejected it because of its alignment with TAU. NIR spectroscopy allowed us to categorize it as field L3 BD. It displays marginal signs of low-gravity (weaker FeH and slightly weaker alkali line widths); however, all other features as well as a visual comparison with spectroscopic standards are consistent with a field L3 BD. It could be interesting to investigate whether this object has a peculiar metallicity or a slightly young age ($\sim$200 to a few hundred Myrs), but it is most probably not a member of any YMG considered here.
**2MASS J05243009+0640349** has been identified as a potential member of $\beta$PMG in *PRE-BASS*. It has been subsequently excluded from *BASS* because of its low Bayesian probability, but its NIR spectrum allowed us to categorize it as a low-gravity M5.5$\beta$. The only useful sign of low gravity for this spectral type is the weaker absorption. This object has a low galactic latitude ($b = -15.92$) and is located within the Orion II super bubble [@1974ApJ...191L.121G] at only 106 of the Ori C 11 core (see Figure 12b of [@1994ApJS...95..457W]; the B1950 coordinates of this object are 05h21m48.67s, +0637$'$545). These clouds are located at significantly larger distances ($\sim$400–500pc; [@2014ApJ...786...29S]) compared to the YMGs considered here. Using the absolute *2MASS* and *WISE* photometry of known young M5.5 dwarfs, we estimate a spectrophotometric distance of $42.0 \pm 7.7$pc for 2MASS J05243009+0640349 and exclude a distance larger than 63pc at a 99% confidence level, assuming it is not a multiple system. This discrepancy, supplemented with the fact that its spectrum does not seem reddened by interstellar dust, makes it unlikely that this object is a member of the Orion Molecular Complex (OMC) even though it is clearly young. Even when its youth is taken into account, this object has a very low probability of being a member of $\beta$PMG. It will be useful to obtain a distance and RV measurement to investigate whether this object is a member of another YMG that is not considered here.
**2MASS J05271676+0007526** has been identified as a potential member of $\beta$PMG in *PRE-BASS*. Its low Bayesian probability as well as color filters ($H - K_S > 0.269$ and $VR - J \geq 2.63$; Paper II) excluded it from the *BASS* sample. This object has a low galactic latitude ($b = -18.57$) and is located in the vicinity of the OMC [@1974ApJ...191L.121G; @2014ApJ...786...29S]. Acquisition images obtained with SpeX revealed that this is a 24 visual binary. We obtained resolved NIR spectroscopy and determined that both components are reddened early M dwarfs. We de-reddened both spectra using the *fm\_unred.pro* IDL routine based on the extinction law of [@1999PASP..111...63F] and visually compared the results with NIR spectroscopic standards to determine the best matching spectral types and total extinction $A(V)$, using the parametrization of [@1999PASP..111...63F] with a total-to-selective extinction of $R(V) = 3.1$. We find a best match of $A(V) = 0.93$ with spectral types M0 + M3. We note that the $H$-band continuum of both objects has a rounded triangular shape, which is only seen at those spectral types for very young ($\lesssim$5Myr) objects. This system is thus likely very young and still embedded in its formation material. Using the *2MASS* $J$ magnitude of the unresolved system with the young absolute magnitude-spectral type sequences of [@2013ApJ...762...88M], we estimate a distance of $\sim$450pc. We conclude that this system is a probable very young low-mass star member of OMC.
**2MASS J08503593+1057156** (2MASS J0850+1057) was first identified from the *2MASS* survey as an L6 BD by [@1999ApJ...519..802K]. Subsequently, [@2001AJ....121..489R] and identified and confirmed that this object is a 016 binary system, and [@2011AJ....141...70B] used a template fitting method constrained by the flux ratio of its individual components to assign them spectral types of L7 and L6. They noted the surprising fact that the brighter primary component is the one that gets assigned a later spectral type. They argue that this could be explained either by youth or the latest-type component being an unresolved binary. [@2011AJ....141...71F] subsequently identified the NLTT 20346 M5+M6 binary system as a very wide ($\sim$7700AU) co-moving companion to 2MASS J0850+1057. They assign an age estimate of 250–450Myr for NLTT 20346 based on X-ray luminosity, but they note that this estimate is discrepant with that based on H$\alpha$ emission ($6.3 \pm 1.0$Gyr and $6.5\pm 1.0$Gyr for its respective components). They measure a systemic RV of $26 \pm 9$ for NLTT 20346, and a trigonometric distance of $29 \pm 7$pc for 2MASS J0850+1057. They note that this latter measurement is not precise enough to discriminate between two previous inconsistent measurements in the literature ($38 \pm 6$pc from [@2004AJ....127.2948V] and $25.6 \pm 2.3$ pc from [@2002AJ....124.1170D]). Using their proper motion measurement of $\mu_\alpha = -144 \pm 6$[$\mathrm{mas}\,\mathrm{yr}^{-1}$]{}and $\mu_\delta = -38 \pm 6$[$\mathrm{mas}\,\mathrm{yr}^{-1}$]{}, they argued that a faint background contaminant was blended at the epochs used for previous distance measurements, which could explain the discrepancy.
[@2012ApJS..201...19D] independently measured a proper motion of $-144.2 \pm 0.6$[$\mathrm{mas}\,\mathrm{yr}^{-1}$]{}, $\mu_\delta = -12.6 \pm 0.6$[$\mathrm{mas}\,\mathrm{yr}^{-1}$]{} and a distance of $33.2 \pm 0.9$pc for 2MASS J0850+1057. They also refined the photometry of its resolved components, and used these new measurements to draw different conclusions than those outlined above. First, they used a similar analysis to that of [@2011AJ....141...70B] with their updated photometry to argue that the spectral types of the components are rather L6.5$\pm 1$ and L8.5$\pm 1$, with the fainter component now associated with the later spectral type. This conclusion does away with the need to invoke youth or any additional component, which was previously based on flux reversal [@2011AJ....141...70B]. They thus argued that 2MASS J0850+1057 is a BD system displaying no notable peculiarity. Furthermore, they use their new proper motion measurement at 6.7$\sigma$ from that of NLTT 20346 with the criterion of [@2007AJ....133..889L] to argue that the two systems are likely random alignments, and thus not gravitationally linked.
We measure a proper motion of $-141.1 \pm 7.7$[$\mathrm{mas}\,\mathrm{yr}^{-1}$]{} and $\mu_\delta = -13.1 \pm 9.5$[$\mathrm{mas}\,\mathrm{yr}^{-1}$]{}, based on *2MASS* and *ALLWISE*. The $\mu_\alpha$ component is consistent with both measurements, whereas the $\mu_\delta$ component is at 2.2$\sigma$ and 0.05$\sigma$ respectively from the measurements of [@2011AJ....141...71F] and [@2012ApJS..201...19D]. Our measurement thus favors the later one, but our precision is $\sim\,16$ times lower. We find that even if both components seem to display no peculiarity in their relative fluxes, the unresolved system seems to be unusually red for its absolute magnitude (Figure \[fig:J0850\_pec\]). This could be explained either by additional unresolved later-type components, or the presence of thicker/higher clouds in their atmosphere compared to field BDs. We used the $R \sim 120$ NIR spectrum for the unresolved BD system to assign a spectral type of L7. For such a late spectral type, the only features known to be gravity-sensitive in a low-resolution NIR spectrum are the $H$-cont index of [@2013ApJ...772...79A] and the H$_2(K)$ index of [@2013MNRAS.435.2650C]. We find values of $H$-cont$= 0.872 \pm 0.025$ and H$_2(K) = 1.055 \pm 0.014$, both being only marginally consistent with a low surface gravity. The parallax and proper motion measurements of [@2012ApJS..201...19D] preclude a possible membership to ARG; however, obtaining a higher-resolution NIR spectrum for this system would be interesting to assess whether it displays signs of low surface gravity. If this system is younger than $\sim$200Myr, the individual mass of each component would be well below the deuterium burning limit, which would make it a remarkable benchmark system to understand the properties of planetary-mass objects.
![NIR spectrum of the peculiar L7 dwarf 2MASS J0850+1057 that was recovered as a candidate member of ARG in *BASS*. This system is a binary with resolved spectral types of L6.5$\pm 1$ and L8.5$\pm 1$ estimated from photometry. The NIR continuum of 2MASS J0850+1057 is redder than field L7 dwarfs which is likely an effect of its binary nature, and its kinematics are not consistent with those of ARG.[]{data-label="fig:J0850_pec"}](Figure22.pdf){width="48.00000%"}
**2MASS J08575849+5708514** has been discovered by [@2002ApJ...564..466G] as an L8$\pm 1$ BD. [@2009ApJ...702..154S] used atmosphere model fitting to determine that this object is unusually cloudy and seems to have a low surface gravity ($\log g = 4.5$). Using our visual comparison with spectral templates, we categorize this object as a peculiar L8 dwarf (Figure \[fig:J0857\_pec\]). No indices from [@2013ApJ...772...79A] are gravity-sensitive for such a late spectral type, but the H$_2(K)$ index defined by [@2013MNRAS.435.2650C] seem to remain useful [@2014AJ....147...34S]. We find a weaker H$_2(K)$ value ($1.102 \pm 0.006$) compared with typical field L8 dwarfs ($1.12 \pm 0.02$), which could be an indication of a lower surface gravity. We identified this object as a highly probable candidate member of ARG in *BASS*, with an estimated mass of $8.5 \pm 0.8$[$M_{\mathrm{Jup}}$]{} and an estimated distance of $8.9 \pm 0.8$pc. However, [@2010AJ....139.1808S] measured an RV of $-123.5 \pm 20.0$ , which is not consistent with membership to ARG or even with the kinematics of any young BD in the solar neighborhood. This large RV thus seems contradictory with its unusually red colors and tentative indications of a lower surface gravity, but it was measured from a low-signal optical spectrum. It is likely that this object is not a young member of ARG but rather an interloping cloudy object from the field. However, obtaining an RV measurement from higher-S/N data will be useful to assess this.
![NIR spectrum of the peculiar L8 dwarf 2MASS J08575849+5708514 that was recovered as a candidate member of ARG in *BASS*. Its red NIR continuum could be an indication of a low surface gravity, but an RV measurement that was obtained from low-S/N data is not consistent with membership to ARG.[]{data-label="fig:J0857_pec"}](Figure23.pdf){width="48.00000%"}
**2MASS J11335700–7807240** was identified by [@2007ApJS..173..104L] in the optical as an M8 dwarf in a search for new members of the Chamaeleon I (CHA) star-forming region. They rejected it because it lacks low-gravity indications in its optical spectrum. We independently recovered this object in *PRE-BASS* as a candidate member of CAR and obtained a NIR spectrum. We categorize this object as a peculiar M6$ \pm 1$ dwarf; its $H$ band is more triangular and the slope of its $K$ band is bluer. This could be indicative of a low surface gravity, but it lacks all the other usual signatures: only the 1.253$\mu$m line is slightly weaker than that of field M6 dwarfs. All other lines, the doublet and FeH absorption are all consistent with a field M6 dwarf : the classification scheme of [@2013ApJ...772...79A] thus categorizes this object as a field-gravity M6 dwarf. It is unclear what is the source of the peculiar features in this object’s NIR spectrum. It is possible that the triangular-shaped $H$ band could be caused by dust in its photosphere [@2013ApJ...772...79A], but this would be unusual at such an early spectral type, and neither its $J - K_S$ color ($1.01 \pm 0.04$) or its $J - W2$ color ($1.53 \pm 0.04$) are redder than those of field M6 dwarfs, which would be unexpected for a dusty object. We thus categorize this object as a peculiar M6 dwarf and reject it as a candidate member of CAR.
**2MASS J11555389+0559577** was discovered by [@2004AJ....127.3553K] as an L7.5 dwarf using NIR spectroscopy. We recovered this object in *PRE-BASS* as a candidate member of ARG. [@2012ApJ...752...56F] measured a trigonometric distance of $17.27 \pm 3.04$pc and [@2010AJ....139.1808S] used a low-quality optical spectrum from SDSS to categorize it as an L0 dwarf and measure an RV of $136.8 \pm 20.0$. If we include only the trigonometric measurement, it remains a modest candidate of ARG; however, the RV measurement is not consistent with this, nor with the kinematics of nearby, young dwarfs [@2009AJ....137....1F], much like the case of 2MASS J08575849+5708514. We retrieved the NIR spectrum of this object from the SpeX PRISM Libraries and categorize it as a peculiar L6–L8 dwarf. It lacks the triangular-shaped $H$-band continuum that would be expected for a young object ($H$-cont $= 0.8230 \pm 0.0087$), and its NIR colors are consistent with those of field dwarfs. We measured the gravity-sensitive H$_2$($K$) index defined by [@2013MNRAS.435.2650C] and find a value of $1.0862 \pm 0.0085$, which is consistent with field L6–L7 dwarfs [@2014AJ....147...34S]. Higher-resolution NIR spectroscopy as well as an RV measurement derived from a high signal-to-noise spectrum would be needed to completely rule out low gravity, but it is very likely that this object is a regular BD; we thus reject it as a candidate member of ARG.
![NIR spectrum of the peculiar L2 dwarf 2MASS J20484222–5127435 that was recovered as a candidate member of THA in *BASS*. It lacks indications of a low surface gravity and is thus not a likely member of THA. Its $K$ band is peculiar, as it resembles those of later-type L dwarfs.[]{data-label="fig:J2048_pec"}](Figure24.pdf){width="48.00000%"}
![NIR spectrum of the peculiar L2 dwarf 2MASS J22062157–6116284 that was recovered as a candidate member of THA in *PRE-BASS*. It lacks indications of a low surface gravity and is thus not a likely member of THA. Furthermore, its $H$-band bump at $\sim$1.57$\mu$m is stronger than usual.[]{data-label="fig:J2206_pec"}](Figure25.pdf){width="48.00000%"}
![NIR spectrum of the peculiar L3 dwarf 2MASS J23155665–4747315 that was recovered as a candidate member of THA in *BASS*. It lacks indications of a low surface gravity and is thus not a likely member of THA. Its $J$ band is similar to a field L3 dwarf with stronger FeH absorption, and its $H$ and $K$ bands are consistent with a later spectral type.[]{data-label="fig:J2315_pec"}](Figure26.pdf){width="48.00000%"}
![NIR spectrum of the peculiar L4 dwarf 2MASS J23392527+3507165 that was recovered as a candidate member of $\beta$PMG in *BASS*. It lacks indications of a low surface gravity and is thus not a likely member of $\beta$PMG. Furthermore, its $H$-band bump at $\sim$1.57$\mu$m is stronger than usual.[]{data-label="fig:J2339_pec"}](Figure27.pdf){width="48.00000%"}
**2MASS J20484222–5127435** was identified as a candidate member of THA as part of *BASS*. NIR spectroscopy revealed that its $J$ and $H$ bands are similar to a field L2 dwarf, but its $K$ band is significantly different, and similar to the $K$ band of field L5 dwarfs (Figure \[fig:J2048\_pec\]). This object is thus unlikely young and we reject it as a candidate member of THA; however, it is unclear what is the cause of its peculiar $K$ band. It would be worthwhile investigating whether this is an early-L / mid-L binary from high-resolution imaging or an RV follow-up.
**2MASS J22062157–6116284** has been identified as a candidate member of THA in *PRE-BASS*. We obtained NIR spectroscopy and categorized it as a peculiar L0 $\pm 1$ dwarf because its $H$-band flux at $\sim$1.57$\mu$m is stronger than usual (Figure \[fig:J2206\_pec\]). This could indicate the presence of an unresolved T-type component; however, the index-based scheme of [@2014ApJ...794..143B] indicates that this scenario is unlikely.
**2MASS J23155665–4747315** has been identified as a candidate member of THA in *BASS*. We obtained NIR spectroscopy and categorized it as a peculiar L3 dwarf; its $J$ band is similar to an L3 dwarf albeit with stronger FeH absorption, and its $H$ and $K$ bands are similar to our field L5 template (Figure \[fig:J2315\_pec\]).
**2MASS J23310161–0406193** (Koenigstuhl 3 BC) was discovered by [@2000AJ....120.1085G] as an M9 dwarf in the optical, and [@2003AJ....125.3302G] demonstrated that it is an M8 + L3, 058 binary system. [@2007ApJ...667..520C] discovered that this system is a very wide 4518 comoving system to the F8 star HR 8931 (HD 221356). Koenigstuhl 3 BC was identified as a candidate member of ABDMG in *LP-BASS*, but measurements of RV and distance for the co-moving star HR 8931 ($-12.86 \pm 0.09$ and $26.12 \pm 0.37$pc; ) preclude a possible membership to all YMGs considered here. We categorize its unresolved spectrum as a peculiar M8 dwarf; our best-matching NIR template is M8$\gamma$, however it presents several differences with it and lacks several low-gravity indications such as weaker alkali lines. We thus conclude that the peculiar nature of this spectrum is likely related to its binary nature, which further rules out a possible membership to ABDMG.
**2MASS J23392527+3507165** has been discovered as an L3.5 BD in the optical by [@2008AJ....136.1290R], and [@2010ApJ...710.1142B] categorized it as an L4.5 BD in the NIR. We recovered this object as a candidate member of $\beta$PMG in *BASS*. We used its NIR spectral type to categorize it as a peculiar L4 BD that has a stronger $H$-band peak at $\sim$1.57$\mu$m (Figure \[fig:J2339\_pec\]). However, it is unlikely that this object is young as it lacks the usual low-gravity indications. We thus reject it as a candidate member of $\beta$PMG. The peculiar $H$-band feature described above can be an effect of an unresolved T-type companion; however, the index-based scheme of [@2014ApJ...794..143B] indicates that this scenario is unlikely.
[^1]: Publicly available at [http://www.astro.umontreal.ca/\\textasciitilde malo/banyan.php](http://www.astro.umontreal.ca/\textasciitilde malo/banyan.php)
[^2]: Publicly available at [http://www.astro.umontreal.ca/\\textasciitilde gagne/banyanII.php](http://www.astro.umontreal.ca/\textasciitilde gagne/banyanII.php)
[^3]: [http://www.pas.rochester.edu/\\textasciitilde emamajek/spt/](http://www.pas.rochester.edu/\textasciitilde emamajek/spt/)
[^4]: Maintained by Michael C. Cushing and available at [http://irtfweb.ifa.hawaii.edu/\\textasciitilde spex/IRTF\_Spectral\_Library/](http://irtfweb.ifa.hawaii.edu/\textasciitilde spex/IRTF_Spectral_Library/).
[^5]: [www.astro.umontreal.ca/\\textasciitilde gagne/MSL.php](www.astro.umontreal.ca/\textasciitilde gagne/MSL.php)
[^6]: [http://pono.ucsd.edu/\\textasciitilde adam/browndwarfs/spexprism](http://pono.ucsd.edu/\textasciitilde adam/browndwarfs/spexprism)
[^7]: [@2000AJ....120..447K; @2003AJ....126.2421C; @2007AJ....133..439C; @2008ApJ...689.1295K; @2008AJ....136.1290R; @2009AJ....137.3345C; @2010ApJS..190..100K]; see [http://www.astro.umontreal.ca/\\textasciitilde gagne/rizzo](http://www.astro.umontreal.ca/\textasciitilde gagne/rizzo)
[^8]: <http://dwarfarchives.org>
[^9]: Publicly available at [www.astro.umontreal.ca/\\textasciitilde gagne/listLTYs.php](www.astro.umontreal.ca/\textasciitilde gagne/listLTYs.php)
[^10]: Publicly available at [www.astro.umontreal.ca/\\textasciitilde gagne/listMs.php](www.astro.umontreal.ca/\textasciitilde gagne/listMs.php)
[^11]: In addition to the following references: ; and [@1995gcts.book.....V]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present spatially- and spectrally-resolved photoluminescence measurements of indirect excitons in high magnetic fields. Long indirect exciton lifetimes give the opportunity to measure magnetoexciton transport by optical imaging. Indirect excitons formed from electrons and holes at zeroth Landau levels ($0_{\rm e} - 0_{\rm h}$ indirect magnetoexcitons) travel over large distances and form a ring emission pattern around the excitation spot. In contrast, the spatial profiles of $1_{\rm e} - 1_{\rm h}$ and $2_{\rm e} - 2_{\rm h}$ indirect magnetoexciton emission closely follow the laser excitation profile. The $0_{\rm e} - 0_{\rm h}$ indirect magnetoexciton transport distance reduces with increasing magnetic field. These effects are explained in terms of magnetoexciton energy relaxation and effective mass enhancement.'
author:
- 'Y. Y. Kuznetsova'
- 'C. J. Dorow'
- 'E. V. Calman'
- 'L. V. Butov'
- 'J. Wilkes'
- 'K. L. Campman'
- 'A. C. Gossard'
title: Transport of Indirect Excitons in High Magnetic Fields
---
Studies of cold fermions (electrons) in high magnetic fields have led to exciting findings, including the integer and fractional quantum Hall effects [@QH]. Excitons are composite neutral bosons. The high magnetic field regime for excitons is realized when the cyclotron splitting becomes comparable to the exciton binding energy. The fulfillment of this condition for atoms requires magnetic fields $B \sim 10^6$ Tesla, and studies of cold atoms therefore use synthetic magnetic fields in rotating systems [@Madison00; @Abo-Shaeer01; @Schweikhard04] and optically synthesized magnetic fields [@Lin09]. Due to the small exciton mass and binding energy, the high magnetic field regime for excitons is realized with magnetic fields of a few Tesla, achievable in lab.
There are exciting theoretical predictions for cold two-dimensional (2D) neutral exciton and electron-hole (e–h) systems in high magnetic fields. Predicted collective states include a paired Laughlin liquid [@Yoshioka90], an excitonic charge-density-wave state [@Chen91], and a condensate of magnetoexcitons [@Kuramoto78; @Lerner81]. Predicted transport phenomena include the exciton Hall effect [@Dzyubenko84; @Paquet85], superfluidity [@Paquet85; @Imamoglu96], and localization [@Dzyubenko95a]. Excitons also play an important role in the description of a many-body state in bilayer electron systems in high magnetic fields [@Eisenstein04]. 2D neutral exciton and e–h systems in high magnetic fields were studied experimentally in single quantum wells (QWs), and excitons and deexcitons were observed in dense e–h magnetoplasmas [@Butov91; @Butov92]. However, low exciton temperatures were not achieved in the studied single QWs due to short exciton lifetimes.
{width="\linewidth"}
{width="\textwidth"}
An indirect exciton (IX) in a coupled QW structure (CQW) is composed of an electron and a hole in separate QWs [@Lozovik76; @Fukuzawa90] (Fig. 1a). Lifetimes of IXs are orders of magnitude longer than lifetimes of regular direct excitons and long enough for the IXs to cool below the temperature of quantum degeneracy $T_0 = 2 \pi \hbar^2 n/(gk_{\rm B}M)$ [@Butov01] (for a GaAs CQW with the exciton spin degeneracy $g = 4$ and mass $M = 0.22m_0$, $T_0 \sim 3$ K for the exciton density per spin state $n/g = 10^{10}$ cm$^{-2}$). Furthermore, due to their long lifetimes, IXs can travel over large distances before recombination, allowing the study of exciton transport by optical imaging. Finally, the density of photoexcited e–h systems can be controlled by the laser excitation, which allows the realization of virtually any Landau level filling factor, ranging from fractional $\nu < 1$ to high $\nu$, even at fixed magnetic field. The opportunity to implement low temperatures, the high magnetic field regime, long transport distances, and controllable densities make IXs a model system for studying cold bosons in high magnetic fields.
Earlier studies of IXs in magnetic fields addressed IX energies [@Butov95; @Dzyubenko96; @Butov99; @Butov01a; @Kowalik-Seidl11; @Schinner13], dispersion relations [@Lozovik97; @Butov01b; @Lozovik02; @Wilkes16], and spin states [@Gorbunov13; @High13]. Here, we present spatially- and spectrally-resolved measurements as well as theoretical simulations of IX emission, which probe IX transport and relaxation in high magnetic fields.
The theory of Mott excitons in high magnetic fields, magnetoexcitons (MXs), was developed in Refs. [@Elliott; @Hasegawa; @Gor'kov68; @Lerner80; @Kallin84]. 2D MXs are shown schematically in Fig. 1. Optically active MXs are formed from electrons and holes at Landau levels (LLs) with $N_{\rm e} = N_{\rm h}$ (Fig. 1b,c). The MX dispersion is determined by the coupling between the MX center-of-mass motion and internal structure: MX is composed of an electron and a hole forced to travel with the same velocity, producing on each other a Coulomb force that is canceled by the Lorentz force. At small momenta $k$, the magnetoexciton dispersion is parabolic and can be described by an effective MX mass. At high $k$, the MX energy tends toward the sum of the electron and hole LL energies (Fig. 1c). The MX mass and MX binding energy increase with $B$ [@Gor'kov68; @Lerner80; @Kallin84; @Paquet85]. With reducing $B$, $N_{\rm e} - N_{\rm h}$ MX states transform to $(N+1)$s exciton states (Fig. 1d). These properties are characteristic of both direct MXs (DMXs) and indirect MXs (IMXs). Due to the separation ($d$) between the e and h layers, IMX energies are lower by $\sim edF_z$ and grow faster with $B$ [@Butov95; @Dzyubenko96; @Butov99; @Butov01a; @Kowalik-Seidl11; @Schinner13; @Lozovik97] ($F_z$ is the electric field in the $z$ direction), IMX binding energies are smaller [@Lozovik97; @Dzyubenko96; @Butov01b; @Lozovik02; @Wilkes16], and IMX effective masses grow faster with $B$ [@Lozovik97; @Butov01b; @Lozovik02; @Wilkes16].
Free 2D MXs can recombine radiatively when their momentum $k$ is inside the intersection between the dispersion surface $E_{\rm MX}(k)$ and the photon cone $E = \hbar k c / \sqrt{\varepsilon}$, called the radiative zone [@Feldman87; @Hanamura88; @Andreani91] (Fig. 1c). In GaAs QW structures, the radiative zone corresponds to $k \lesssim k_0 \approx E_{\rm g} \sqrt{\varepsilon} / (\hbar c) \approx 2.7 \cdot 10^5$ cm$^{-1}$ ($\varepsilon$ is the dielectric constant, $E_{\rm g}$ the semiconductor gap). In GaAs QW structures, excitons may have four spin projections on the $z$ direction $J_z = \pm 2, \pm 1$; the $J_z = \pm 1$ states are optically active [@Maialle93]. Free MXs with $k \lesssim k_0$, $N_{\rm e} = N_{\rm h}$, and $J_z = \pm 1$ recombine radiatively directly contributing to MX emission. Free MXs with $k > k_0$, $N_{\rm e} \ne N_{\rm h}$, or $J_z = \pm 2$ are dark.
Experiments were performed on a $n^+ - i - n^+$ GaAs CQW. The $i$-region consists of a single pair of 8-nm GaAs QWs separated by a 4-nm Al$_{0.33}$Ga$_{0.67}$As barrier, surrounded by 200-nm Al$_{0.33}$Ga$_{0.67}$As layers. The $n^+$ layers are Si-doped GaAs with Si concentration $5 \cdot 10^{17}$ cm$^{-3}$. The indirect regime, characterized by IX being the lowest energy state, was implemented by applying voltage $V = - 1.2$ V between the $n^+$ layers. The 633 nm laser excitation was focused to $\sim 6$ $\mu$m spot. The $x-$energy images were measured with a liquid-nitrogen-cooled charge-coupled device (CCD) placed after a spectrometer with resolution 0.18 meV. Spatial resolution was $\approx 2.5$ $\mu$m. The measurements were performed in an optical dilution refrigerator at temperature $T_{\rm bath} = 40$ mK and magnetic fields $B = 0 - 10$ T perpendicular to the CQW plane.
Figure 2 shows the evolution of measured $x-$energy emission patterns with increasing $B$. Horizontal cross sections of the $x-$energy emission pattern reveal the spatial profiles at different energies (Fig. 3a,b), while vertical cross sections present spectra at different distances $x$ from the excitation spot center (Fig. 3c,d). More spatial profiles and spectra of the IMX emission at different $B$ and $P_{\rm ex}$ are presented in [@suppl_mat]. Spatial profiles of the amplitudes of IMX emission lines are shown in Fig. 4a,b.
![(a,b) Spatial profiles of the IMX emission at different energies at $P_{\rm ex} = 1$ (a) and 260 $\mu$W (b). (c,d) MX spectra at different distances $x$ from the excitation spot center at $P_{\rm ex} = 1$ (c) and 260 $\mu$W (d). IMX and DMX lines correspond to indirect and direct MX emission, respectively. The low-energy bulk emission was subtracted from the spectra [@suppl_mat]. The laser excitation profile is shown in Fig. 4a. $B$ = 3 T.](Figure3.pdf){width="\linewidth"}
{width="\linewidth"}
The magnetic field dependence (Fig. 4d) identifies the $0_{\rm e} - 0_{\rm h}$, $1_{\rm e} - 1_{\rm h}$, and $2_{\rm e} - 2_{\rm h}$ IMX emission lines. The DMX emission is also observed at high energies (Figs. 3, 4d).
At low $P_{\rm ex}$ and, in turn, low IMX densities, the $0_{\rm e} - 0_{\rm h}$ IMX emission essentially follows the laser excitation profile (Figs. 3a,c and 4b). This indicates that at low densities, $0_{\rm e} - 0_{\rm h}$ IMXs are localized in the in-plane disorder potential and practically do not travel beyond the excitation spot. However, at high densities, transport of $0_{\rm e} - 0_{\rm h}$ IMXs is observed as the $0_{\rm e} - 0_{\rm h}$ IMX emission extends well beyond the excitation spot (Figs. 2, 3b,d, and 4a). Furthermore, the $0_{\rm e} - 0_{\rm h}$ IMX emission shows a ring structure around the excitation spot (Figs. 2, 3b,d, and 4a). This structure is similar to the inner ring in the IX emission pattern at $B = 0$ [@Ivanov06; @Hammack09; @Kuznetsova12]. The enhancement of $0_{\rm e} - 0_{\rm h}$ IMX emission intensity with increasing distance from the center originates from IMX transport and energy relaxation as follows. IMXs cool toward the lattice temperature when they travel away from the laser excitation spot, thus forming a ring of cold IMXs. The cooling increases the occupation of the low-energy optically active IMX states (Fig. 1c), producing the $0_{\rm e} - 0_{\rm h}$ IMX emission ring. The ring extension $R$ characterizing the $0_{\rm e} - 0_{\rm h}$ IMX transport distance is presented in Fig. 4e,f. The IMX transport distance increases with density (Fig. 4e,f). This effect is explained by the theory presented below in terms of the screening of the structure in-plane disorder by the repulsively interacting IMXs.
In contrast to the $0_{\rm e} - 0_{\rm h}$ IMX emission, the spatial profile of the $1_{\rm e} - 1_{\rm h}$ IMX emission closely follows the laser excitation profile (Fig. 4a). The data show that the high-energy $1_{\rm e} - 1_{\rm h}$ IMX states are occupied in the excitation spot region (where the IMX temperature and density is maximum). Long transport is not observed here because the $1_{\rm e} - 1_{\rm h}$ IMXs effectively relax in energy and transform to $0_{\rm e} - 0_{\rm h}$ IMXs beyond the laser excitation spot where the IMX temperature drops down. $1_{\rm e} - 1_{\rm h}$ IMX transport distance within this relaxation time $\lesssim 3$ $\mu$m (Fig. 4a,f).
Additionally, the $0_{\rm e} - 0_{\rm h}$ and $1_{\rm e} - 1_{\rm h}$ IMX energies are observed to reduce with $x$ (Fig. 4c). This energy reduction follows the IMX density reduction away from the excitation spot. The density reduction can lower IX energy due to interactions and localization in the disorder potential. IMXs have a built-in dipole moment $\sim ed$ and interact repulsively. The repulsive IMX interaction causes the reduction of the IMX energy with reducing density. In contrast, direct MXs in single QWs are essentially noninteracting particles and their energy practically does not depend on density [@Lerner81; @Butov92].
Figure 4e,f also shows that the $0_{\rm e} - 0_{\rm h}$ IMX transport distance reduces with $B$. This effect is explained below in terms of the enhancement of the MX mass. The reduction of the IMX transport distance causes the IMX accumulation in the excitation spot area. The IMX accumulation contributes to the observed enhancement of both the IMX emission intensity and energy in the excitation spot area with increasing $B$ (Fig. 2).
We note also that IX emission patterns may contain the inner ring, which forms due to IX transport and thermalization [@Ivanov06; @Hammack09; @Kuznetsova12], and external ring, which forms on the interface between the hole-rich and electron-rich regions [@Butov04; @Rapaport04; @Chen05; @Haque06; @Yang10]. The data presented in Fig. S1 [@suppl_mat] show that the external ring and the presence of the charge-rich regions associated with it play no major role in the IMX transport and relaxation phenomena described in this work.
![(a) Optical transition energies of $k=0$ single MX states measured from the band gap vs $B$ calculated using a multi-sub-level approach. (b) $0_{\rm e} - 0_{\rm h}$ IMX mass renormalization due to the magnetic field calculated via perturbation theory. (c) Simulated ring radius $R$ (defined as HWHM of $I_{\rm PL}$ shown in (f)) vs $B$ for different injection rates, $P_\Lambda=2\pi\int_0^\infty\Lambda(r)r\,dr$. Spatial profiles of the (d) density, (e) diffusion coefficient and (f) emission intensity of $0_{\rm e} - 0_{\rm h}$ IMX from solving Eqns.(\[transportEqn\])-(\[thermalizationEqn\]) for different $B$ with injection rate $150\,{\rm ns^{-1}}$ and $T_{\rm bath} = 0.5$ K.](Figure5.pdf){width="\linewidth"}
The following two-body Hamiltonian is used to describe MXs in CQWs under external bias, $$\hat{H}({\bf r}_e,{\bf r}_h) = \hat{H}_e({\bf r}_e) + \hat{H}_h({\bf r}_h) - \frac{e^2}{\varepsilon|{\bf r}_e - {\bf r}_h|} + E_g. \label{Hamiltonian}$$ Here, ${\bf r}_{e(h)}$ and $\hat{H}_{e(h)}$ are the electron (hole) coordinates and single-particle Hamiltonians, respectively. The latter are given by $$\hat{H}_{e(h)}({\bf r}) = \hat{p}_{e(h)}({\bf r})\frac{1}{2}\hat{m}_{e(h)}^{-1}(z)\hat{p}_{e(h)}({\bf r}) + U_{e(h)}(z).$$ The magnetic field $B$ contributes to the momentum operators $\hat{p}_{e(h)}({\bf r}) = -i\hbar\nabla_{\bf r} \pm (e/c) {\bf A}({\bf r})$ via the magnetic vector potential $\bf A$. The mass tensor $\hat{m}_{e(h)}(z)$ contains the electron (hole) effective masses which are step functions along $z$ due to the electron and hole confinement in the corresponding QW layers (Fig. 1a). $U_{e(h)}(z)$ contain the QW confinement and the potential due to the applied electric field. The third term in Eq. \[Hamiltonian\] is the e-h Coulomb interaction. After a factorization of the wave function to separate the center of mass and relative coordinates [@Lozovik97], eigen states of the Hamiltonian describing the relative motion of an exciton with $k = 0$ are found using a multi-sub-level approach [@Sivalertporn12; @Wilkes16]. This allows the extraction of the $B$-field dependence of the $k = 0$ IMX energy, $E_{\rm IMX}$, and radiative lifetime, $\tau_R$. Treating $k$ as a perturbation, we then use perturbation theory to second order to determine the exciton in-plane effective mass enhancement due to the magnetic field, $M(B)$. The calculated $E_{\rm IMX}$ and $M(B)$ are in agreement with the measured $E_{\rm IMX}$ (compare Fig. 5a and 4d) and $M(B)$ (compare Fig. 5b and $M(B)$ in Ref. [@Butov01b; @Lozovik02]).
The $B-$dependence of the ring in the IMX emission pattern is simulated by combining the microscopic description of a single IMX with a model of IMX transport and thermalization. The following set of coupled equations were solved for the $0_{\rm e} - 0_{\rm h}$ IMX density $n(r,t)$ and temperature $T(r,t)$ in the space-time ($r$,$t$) domain, $$\begin{aligned}
\frac{\partial n}{\partial t} = \nabla \left[D\nabla n + \mu_{\rm x} n \nabla (u_0 n)\right] + \Lambda - \frac{n}{\tau}, \label{transportEqn} \\
\frac{\partial T}{\partial t} = S_{\rm pump} - S_{\rm phonon}. \label{thermalizationEqn}\end{aligned}$$ The two terms in square brackets in the transport equation (\[transportEqn\]) describe IMX diffusion and drift currents. The latter originates from the repulsive dipolar interactions approximated by $u_0 = 4 \pi e^2 d/\varepsilon$ within the model [@Hammack09]. The diffusion coefficient $D$ and mobility $\mu_{\rm x}$ are related by a generalized Einstein relation, $\mu_{\rm x} = D(e^{T_0/T} - 1)/(k_BT_0)$. An expression for $D$ is derived using a thermionic model to account for the screening of the random QW disorder potential by dipolar excitons [@Ivanov06; @Hammack09; @Kuznetsova12]. $D$ is inversely proportional to the exciton mass $M$. The enhancement of $M$ with $B$ describes the magnetic field induced reduction in exciton transport. The last two terms on the RHS of (\[transportEqn\]) describe creation and decay of excitons. $\Lambda(r)$ has a Gaussian profile chosen to match the excitation beam. The optical lifetime $\tau$ is the product of $\tau_R$ and a factor that accounts for the fraction of excitons that are inside the radiative zone. The effects of IMX in higher levels are included via $\Lambda$, since they relax to the $0_{\rm e} - 0_{\rm h}$ level within the excitation region.
The thermalization equation (\[thermalizationEqn\]) describes the balance between heating of excitons by non-resonant photoexcitation and cooling via interaction with bulk longitudinal acoustic phonons. Both rates are modified by the magnetic field due to their dependence on $M(B)$. The emission intensity is extracted from $n/\tau$. In the simulations, $T_{\rm bath} = {\rm 0.5\,K}$ was used to avoid the excessive computation times incurred by the dense grids needed to handle the strongly non-linear terms in (\[transportEqn\])-(\[thermalizationEqn\]) that are most prominent at low $T$. Below $T_{\rm bath}$ $\approx {\rm 1\,K}$, the results of the model are qualitatively similar with the ring radius only slowly varying with $T$. Modifying the computations for lower $T_{\rm bath}$ forms the subject of future work. Details of the transport and thermalization model including parameters and expressions for $D$, $\tau$, $S_{\rm pump}$ and $S_{\rm phonon}$ can be found in [@suppl_mat].
The simulations show the ring in the IMX emission pattern (Fig. 5f) in agreement with the experiment (Fig. 2, 3b,d, 4a). The increase of the IMX mass causes the reduction of the IMX diffusion coefficient (Fig. 5e), contributing to the reduction of the IMX transport distance with magnetic field (Fig. 5c). The measured and simulated reduction of the IMX transport distance with magnetic field are in agreement (compare Fig. 4e and 5c).
This work was supported by NSF Grant No. 1407277. J.W. was supported by the EPSRC (grant EP/L022990/1). C.J.D. was supported by the NSF Graduate Research Fellowship Program under Grant No. DGE-1144086.
For reviews, see [*Perspectives in Quantum Hall Effects*]{}, Ed. S. Das Sarma, A. Pinczuk (Wiley, New York, 1997); H.L. St[ö]{}rmer, D.C. Tsui, A.C. Gossard, Rev. Mod. Phys. [**71**]{}, S298 (1999).
K.W. Madison, F. Chevy, W. Wohlleben, J. Dalibard, Phys. Rev. Lett. [**84**]{}, 806 (2000).
J.R. Abo-Shaeer, C. Raman, J.M. Vogels, W. Ketterle, Science [**292**]{}, 476 (2001).
V. Schweikhard, I. Coddington, P. Engels, V.P. Mogendorff, E.A. Cornell, Phys. Rev. Lett. [**92**]{}, 040404 (2004).
Y.-J. Lin, R.L. Compton, K. Jim[é]{}nez-Garc[í]{}a, J.V. Porto, I.B. Spielman, Nature [**462**]{}, 628 (2009).
D. Yoshioka, A.H. MacDonald, J. Phys. Soc. Jpn. [**59**]{}, 4211 (1990).
X.M. Chen, J.J. Quinn, Phys. Rev. Lett. [**67**]{}, 895 (1991).
Y. Kuramoto, C. Horie, Solid State Commun. [**25**]{}, 713 (1978).
I.V. Lerner, Yu.E. Lozovik, Sov. Phys. JETP [**53**]{}, 763 (1981).
A.B. Dzyubenko, Yu.E. Lozovik, Sov. Phys. Solid State [**26**]{}, 938 (1984).
D. Paquet, T.M. Rice, K. Ueda, Phys. Rev. B [**32**]{}, 5208 (1985).
A. Imamoglu, Phys. Rev. B [**54**]{}, R14285 (1996).
P.I. Arseyev, A.B. Dzyubenko, Phys. Rev. B [**52**]{}, R2261 (1995).
J.P. Eisenstein, A.H. MacDonald, Nature [**432**]{}, 691 (2004).
L.V. Butov, V.D. Kulakovskii, E.I. Rashba, Sov. Phys. JETP Lett. [**53**]{}, 109 (1991).
L.V. Butov, V.D. Kulakovskii, G.E.W. Bauer, A. Forchel, D. Gr[ü]{}tzmacher, Phys. Rev. B [**46**]{}, 12765 (1992).
Yu.E. Lozovik, V.I. Yudson, Sov. Phys. JETP [**44**]{}, 389 (1976).
T. Fukuzawa, S. Kano, T. Gustafson, T. Ogawa, Surf. Sci. [**228**]{}, 482 (1990).
L.V. Butov, A.L. Ivanov, A. Imamoglu, P.B. Littlewood, A.A. Shashkin, V.T. Dolgopolov, K.L. Campman, A.C. Gossard, Phys. Rev. Lett. [**86**]{}, 5608 (2001.)
L.V. Butov, A.Zrenner, G. Abstreiter, A.V. Petinova, K. Eberl, Phys. Rev. B [**52**]{}, 12153 (1995).
A.B. Dzyubenko, A.L. Yablonski, Phys. Rev. B [**53**]{}, 16355 (1996).
L.V. Butov, A.A. Shashkin, V.T. Dolgopolov, K.L. Campman, A.C. Gossard, Phys. Rev. B [**60**]{}, 8753 (1999).
L.V. Butov, A. Imamoglu, K.L. Campmana, A.C. Gossard, J. Exp. Theor. Phys. [**92**]{}, 260 (2001).
K. Kowalik-Seidl, X.P. V[ö]{}gele, F. Seilmeier, D. Schuh, W. Wegscheider, A.W. Holleitner, J.P. Kotthaus, Phys. Rev. B [**83**]{}, 081307(R) (2011).
G.J. Schinner, J. Repp, K. Kowalik-Seidl, E. Schubert, M.P. Stallhofer, A.K. Rai, D. Reuter, A.D. Wieck, A.O. Govorov, A.W. Holleitner, J.P. Kotthaus, Phys. Rev. B [**87**]{}, 041303(R) (2013).
Yu.E. Lozovik, A.M. Ruvinskii, Sov. Phys. JETP [**85**]{}, 979 (1997).
L.V. Butov, C.W. Lai, D.S. Chemla, Yu.E. Lozovik, K.L. Campman, A.C. Gossard, Phys. Rev. Lett. [**87**]{}, 216804 (2001).
Yu.E. Lozovik, I.V. Ovchinnikov, S.Yu. Volkov, L.V. Butov, D.S. Chemla, Phys. Rev. B [**65**]{}, 235304 (2002).
J Wilkes, E.A. Muljarov, New J. Phys. [**18**]{}, 023032 (2016).
A.V. Gorbunov, V.B. Timofeev, Solid State Commun. [**157**]{}, 6 (2013).
A.A. High, A.T. Hammack, J.R. Leonard, Sen Yang, L.V. Butov, T. Ostatnick[ý]{}, M. Vladimirova, A.V. Kavokin, T.C.H. Liew, K.L. Campman, A.C. Gossard, Phys. Rev. Lett. [**110**]{}, 246403 (2013).
R.J. Elliott, R. Loudon, J. Phys. Chem. Solids [**8**]{}, 382 (1959); [**15**]{}, 196 (1960).
H. Hasegawa, R.E. Howard, J. Phys. Chem. Solids [**21**]{}, 179 (1961).
L.P. Gor’kov, I.E. Dzyaloshinskii, Sov. Phys. JETP [**26**]{}, 449 (1968).
I.V. Lerner, Yu. E. Lozovik, Sov. Phys. JETP [**51**]{}, 588 (1980).
C. Kallin, B.I. Halperin, Phys. Rev. B [**30**]{}, 5655 (1984).
J. Feldmann, G. Peter, E.O. G[ö]{}bel, P. Dawson, K. Moore, C. Foxon, R.J. Elliott, Phys. Rev. Lett. [**59**]{}, 2337 (1987).
E. Hanamura, Phys. Rev. B [**38**]{}, 1228 (1988).
L.C. Andreani, F. Tassone, F. Bassani, Solid State Commun. [**77**]{}, 641 (1991).
M.Z. Maialle, E.A. de Andrada e Silva, L.J. Sham, Phys. Rev. B [**47**]{}, 15 776 (1993).
Supplementary materials present IMX spatial profiles and spectra at different $B$ and $P_{\rm ex}$, external ring role, bulk emission subtraction, and theory details.
A.L. Ivanov, L.E. Smallwood, A.T. Hammack, Sen Yang, L.V. Butov, A.C. Gossard, Europhys. Lett. 73, 920 (2006).
A.T. Hammack, L.V. Butov, J. Wilkes, L. Mouchliadis, E.A. Muljarov, A.L. Ivanov, A.C. Gossard, Phys. Rev. B [**80**]{}, 155331 (2009).
Y.Y. Kuznetsova, J.R. Leonard, L.V. Butov, J. Wilkes, E.A. Muljarov, K.L. Campman, A.C. Gossard, Phys. Rev. B [**85**]{}, 165452 (2012).
L.V. Butov, L.S. Levitov, A.V. Mintsev, B.D. Simons, A.C. Gossard, D.S. Chemla, Phys. Rev. Lett. [**92**]{}, 117404 (2004).
R. Rapaport, G. Chen, D. Snoke, S.H. Simon, L. Pfeiffer, K. West, Y. Liu, S. Denev, Phys. Rev. Lett. [**92**]{}, 117405 (2004).
G. Chen, R. Rapaport, S.H. Simon, L. Pfeiffer, K. West, Phys. Rev. B [**71**]{}, 041301(R) (2005).
M. Haque, Phys. Rev. E [**73**]{}, 066207 (2006).
Sen Yang, L.V. Butov, L.S. Levitov, B.D. Simons, A.C. Gossard, Phys. Rev. B [**81**]{}, 115320 (2010).
K. Sivalertporn, L. Mouchliadis, A.L. Ivanov, R. Philp, E.A. Muljarov, Phys. Rev. B [**85**]{}, 045207 (2012).
![image]("S1".pdf){width="116.00000%"}
![image]("S2".pdf){width="116.00000%"}
![image]("S3".pdf){width="116.00000%"}
![image]("S4".pdf){width="116.00000%"}
![image]("S5".pdf){width="116.00000%"}
![image]("S6".pdf){width="116.00000%"}
![image]("S7".pdf){width="116.00000%"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We classify, according to their computational complexity, integer optimization problems whose constraints and objective functions are polynomials with integer coefficients and the number of variables is fixed. For the optimization of an integer polynomial over the lattice points of a convex polytope, we show an algorithm to compute lower and upper bounds for the optimal value. For polynomials that are non-negative over the polytope, these sequences of bounds lead to a fully polynomial-time approximation scheme for the optimization problem.'
title: Integer Polynomial Optimization in Fixed Dimension
---
Introduction
============
Mixed integer non-linear programs combine the hardness of combinatorial explosion with the non-convexity of non-linear functions. For example, the well-known optimality conditions developed for differentiable objective functions have no meaning when the variables are discrete. Thus, it is perhaps not surprising that already linear integer programming with general quadratic constraints is undecidable [@jeroslow]. Nevertheless, when the number of variables is fixed discrete optimization problems often become tractable and efficient polynomial algorithms exist (e.g.[@bar; @khachiyanporkolab; @lenstra]). It is thus natural to ask [ *what is the complexity of integer non-linear optimization assuming that the number of variables is fixed?*]{} We study the problem $$\label{model} \hbox{maximize} \ f(x_1,\dots,x_d) \
\hbox{subject to} \ g_i(x_1,\dots,x_d) \geq 0, \ x \in {\mathbb Z}^d.$$ Here $f,g_i$ are polynomials with integral coefficients. Note that all throughout the paper we assume that the number of variables is fixed. Here are our two contributions to the theory:
*(1)* We give a classification of the computational complexity of Problem (\[model\]) according to special cases. Section \[bounds\] of this article presents the details, but the reader can see the classification in Table \[gogo\]. New results are marked with letters, known results are marked with asterisks, arrows indicate implications:
[cc@[ ]{}c@[ ]{}c@[ ]{}c@[ ]{}c]{} &\
& & & convex & & arbitrary\
Type of constraints & linear & & polynomial & & polynomial\
Linear constraints, integer variables & polytime ($*$) & $\Leftarrow$ & polytime ($**$) & & NP-hard (a)\
& $\Uparrow$ & & $\Uparrow$ & & $\Downarrow$\
Convex semialgebraic constraints, integer variables & polytime ($**$) & $\Leftarrow$ & polytime ($**$) & & NP-hard (c)\
\
Arbitrary polynomial constraints, integer variables & undecidable (b) & $\Rightarrow$ & undecidable (d) & $\Rightarrow$ & undecidable (e)\
[*(2)*]{} For problem $(a)$, that of optimizing an arbitrary integral polynomial over the lattice points of a convex rational polytope with fixed number of variables, we present an algorithm to compute a sequence of upper and lower bounds for its optimal value. Our bounds can be used, for instance, in a branch-and-bound search for the optimum. We use Barvinok’s algebraic encoding of the lattice points of polytopes via rational functions [@BarviPom]. In Section \[algo\] we prove:
\[main\] Let the number of variables $d$ be fixed. Let $f(x_1,\dots,x_d)$ be a polynomial of maximum total degree $D$ with integer coefficients, and let $P$ be a convex rational polytope defined by linear inequalities in $d$ variables. We obtain an increasing sequence of lower bounds $\{L_k\}$ and a decreasing sequence of upper bounds $\{U_k\}$ to the optimal value $$f^*=\hbox{maximize} \ f(x_1,x_2,\dots,x_d) \ \hbox{subject to} \ x
\in P \cap {\mathbb Z}^d.\label{eq:maximize}$$ The bounds $L_k$, $U_k$ can be computed in time polynomial in $k$, the input size of $P$ and $f$, and the maximum total degree $D$ and they satisfy the inequality $U_k-L_k \leq f^* \cdot (\sqrt[k]{|P
\cap {\mathbb Z}^d|}-1).$
More strongly, if $f$ is non-negative over the polytope (i.e. $f(x)\geq 0$ for all $x \in P$), there exists a fully polynomial-time approximation scheme (FPTAS) for the optimization problem .
We conclude with examples and a brief look at the mixed integer problem.
Computational Complexity Bounds {#bounds}
===============================
All the results we present refer to the complexity model where the number of operations is given in terms of the input size measured in the standard binary encoding. The results of H. W. Lenstra Jr. [@lenstra] imply the entry of Table \[gogo\] marked with $(*)$, i.e. solving linear integer programming problems with a fixed number of variables can be done in time polynomial in the size of the input. More recently, Khachiyan and Porkolab [@khachiyanporkolab] have proved that in fixed dimension, the problem of minimizing a convex polynomial objective function over the integers, subject to polynomial constraints that define a convex body, can be solved in polynomial time in the encoding length of the input. Thus, they settled all entries marked by $(**)$. By the natural containment exhibited by these complexity classes, to show the validity of the remaining entries of Table \[gogo\] is enough to prove the following lemma:
1. The problem of minimizing a degree four polynomial over the lattice points of a convex polygon is NP-hard (entry $(a)$ in Table \[gogo\]).
2. The problem of minimizing a linear form over polynomial constraints in at most 10 integer variables is not computable by a recursive function (entry $(b)$ in Table \[gogo\]).
\(1) We use the NP-complete problem AN1 on page 249 of [@GarJohn79]. This problem states it is NP-complete to decide whether, given three positive integers $a,b,c$, there exists a positive integer $x<c$ such that $x^2$ is congruent with $a$ modulo $b$. This problem is clearly equivalent to asking whether the minimum of the quartic polynomial function $(x^2-a-by)^2$ over the lattice points of the rectangle $\{(x,y) | \ 1 \leq x \leq c-1,\ \frac{1-a}{b} \leq y
\leq \frac{(c-1)^2-a}{b} \}$ is zero or not. This settles part (1).
\(2) In 1973 Jeroslow [@jeroslow] proved a similar result without fixing the number of variables. We follow his idea, but resorting to a stronger lemma. More precisely our proof relies on a 1982 result [@jones] which states that there is no recursive function that, given an integer polynomial $f$ with *nine* variables, can determine whether $f$ has a non-negative integer zero, in the sense that it finds an explicit zero or returns null otherwise. Jones paper is a strengthening of the original solution of Hilbert’s tenth problem [@mati]. Now to each polynomial $f$ in ${\mathbb Z}[x_1,x_2\dots,x_9]$ associate the ten-dimensional minimization problem $$\hbox{minimize} \ y \quad \hbox
{subject to} \quad (1-y)f(x_1,x_2,\dots,x_9)=0, \ (y,x_1,\dots,x_9)
\in {\mathbb Z}_{\geq 0}^{10}.$$ The minimum attained by $y$ is either zero or one depending on whether $f$ has an integer non-negative solution or not. Thus part (2) is settled.
FPTAS for Optimizing Non-Negative Polynomials over Integer Points of Polytopes {#algo}
==============================================================================
Consider now a polynomial function $f \in
{\mathbb Z}[x_1,x_2, \dots ,x_d]$ of maximum total degree $D$ and a convex polytope $P=\{x | Ax \leq b\}$ where $A$ is an $m \times
d$ integral matrix and $b$ is an integral $m$-vector. The purpose of this section is to present an algorithm to generate lower and upper bounds $L_k,U_k$ to the integer global optimum value of $$\hbox{maximize} \ f(x_1,\dots,x_d) \ \hbox
{subject to} \ (x_1,\dots,x_d) \in P \cap {\mathbb Z}^d.$$ We should also remark that in our algorithm the polynomial objective function $f$ can be arbitrary (e.g. non-convex). As we have seen, the optimization problem is NP-hard already for two integer variables and polynomials of degree four. Nevertheless we will see that, in fixed dimension and when $f(x) \geq 0$ for all $x \in P$, the algorithm gives a fully polynomial time approximation scheme or FPTAS. This means that, in polynomial time on the input and $(1/\epsilon)$, one can compute a $(1-\epsilon)$-approximation to the maximum. The algorithm we present is based on A. Barvinok’s theory for encoding all the lattice points of a polyhedron in terms of short rational functions. See [@bar; @BarviPom] for all details. Lattice points are thought of as exponent vectors of monomials. For example, $z_1^2z_2^{-11}$ encodes the lattice point $(2,-11)$. The set of lattice points is represented by a Laurent polynomial: $g_P(z)=\sum_{\alpha \in P\cap{\mathbb Z}^d}
z^\alpha.$ From Barvinok’s theory this exponentially-large sum of monomials $g_P(z)$ can instead be written as a polynomial-size sum of rational functions (assuming the dimension $d$ is fixed) of the form: $$\label{eq:aa}
g_P(z) = \sum_{i\in I} {E_i \frac{z^{u_i}} {\prod\limits_{j=1}^d
(1-z^{v_{ij}})}},$$ where $I$ is a polynomial-size indexing set, and where $E_i\in\{1,-1\}$ and $u_i, v_{ij} \in{\mathbb Z}^d$ for all $i$ and $j$. For details see [@BarviPom; @latte1; @latte2].
We need a way to encode via rational functions the values of the polynomial $f$ over all the lattice points in a polytope. The key idea, first introduced in Lemma 9 of [@latte2] and generalized in [@hugginsthesis], is that differential operators associated to $f$ can be used to compute a rational function representation of $\sum_{a\in P \cap {\mathbb Z}^d} f(a) z^a.$ The following Lemma recently appeared in [@newbar]:
\[operators\] Let $g_P(z)$ be the Barvinok representation of the generating function of the lattice points of $P$. Let $f$ be a polynomial in ${\mathbb Z}[x_1,\dots,x_d]$ of maximum total degree $D$. We can compute, in time polynomial on $D$ and the size of the input data, a Barvinok rational function representation $g_{P,f}(z)$ for the generating function $\sum_{a\in P \cap {\mathbb Z}^d} f(a) z^a.$
We give here the author’s original proof the lemma for $D$ fixed. The first proof without this assumption was recently given by A. Barvinok in [@newbar].
We begin assuming $f(z)=z_r$, the general case will follow from it: Consider the action of the differential operator $z_r\frac{\partial}{\partial z_r}$ in the generating function $g_P(z)$ and on its Barvinok representation. On one hand, for the generating function $$z_r\frac{\partial}{\partial z_r}\cdot g_P(z)=\sum\limits_{\alpha\in
P \cap {\mathbb Z}^d} z_r \frac{\partial}{\partial z_r} z^\alpha=
\sum\limits_{\alpha\in P \cap {\mathbb Z}^d}\alpha_r z^\alpha.$$ On the other hand, by linearity of the operator, we have that in terms of rational functions $$z_r\frac{\partial}{\partial z_r}\cdot g_P(z)=\sum\limits_{i \in I} E_i
z_r\frac{\partial}{\partial z_r}\cdot \left( \frac{z^{u_i}} {\prod
\limits_{j=1}^d (1-z^{v_{ij}})}\right).$$ Thus it is enough to prove that the summands of the expression above can be written in terms of rational functions computable in polynomial time. The standard quotient rule for derivatives says that $$\frac{\partial}{\partial z_r} \left(\frac{z^{u_i}} {\prod
\limits_{j=1}^d (1-z^{v_{ij}})}\right)=\frac{(\frac{\partial z^{u_i} }{\partial
z_r}) \prod^d_{j=1}(1-z^{v_{ij}})
-z^{u_i}(\frac{\partial}{\partial z_r}
\prod_{j=1}^d(1-z^{v_{ij}}))} {\prod_{j=1}^d(1-z^{v_{ij}})^2}.$$ We can expand the numerator as a sum of no more than $2^d$ monomials. This is a constant number because $d$, the number of variables, is assumed to be a constant. This argument completes the proof of our lemma when $f(z)=z_r$.
For the case when $f(z)$ is a general monomial, i.e. $f(z)=c\cdot
z_1^{\beta_1}\cdot\ldots\cdot z_d^{\beta_d}$, then we can compute again a rational function representation of $g_{P,f}(z)$ by repeated application of basic differential operators: $$c\left(z_1\frac{\partial}{\partial z_1}\right)^{\beta_1}
\cdot\ldots\cdot \left(z_d \frac{\partial}{\partial
z_d}\right)^{\beta_d}\cdot g_P(z)=\sum\limits_{\alpha\in P \cap {\mathbb Z}^d} c\cdot
\alpha^{\beta} z^\alpha.$$ Thus we require no more than $O(D^d)$ repetitions of the single-variable case.
Finally, if we deal with a polynomial $f$ of many monomial terms, we compute and add up all such expressions that we get for each term of $f(x)$ and obtain our desired short rational function representation for the generating function for $ \sum_{\alpha\in P \cap {\mathbb Z}^d} f(\alpha)
z^\alpha.$ Note that only polynomially many steps are needed because $d$ is fixed and the largest number of possible monomials in $f$ of degree $s$ is $\left(\begin{smallmatrix}d+s-1\\ d-1\\
\end{smallmatrix}\right)$, thus for fixed $d$ we will do no more than $O(D^{d})$ repetitions of the monomial case.
Now we are ready to present our algorithm to obtain bounds $U_k,L_k$ that reach the optimum. Step 1 of preprocessing is necessary because we rely on the elementary fact that, for a collection $S=\{s_1,\dots,s_r\}$ of non-negative real numbers, $\hbox{\rm
maximum} \{s_i | s_i \in S \}$ equals $\lim_{k \rightarrow \infty}
\sqrt[k]{\sum_{j=1}^r s_j^k}.$
[[**Algorithm**]{}]{} \[Algorithm\]
[*Input:*]{} A rational convex polytope $P \subset {\mathbb R}^d$, a polynomial objective $f \in {\mathbb Z}[x_1,\dots,x_d]$ of maximum total degree $D$.
[*Output:*]{} An increasing sequence of lower bounds $L_k$, and a decreasing sequence of upper bounds $U_k$ reaching the maximal function value $f^*$ of $f$ over all lattice points of $P$.
[**Step 1.**]{} If $f$ is known to be non-negative in all points of $P$, then go directly to Step 2. Else, solving $2d$ linear programs over $P$, we find lower and upper integer bounds for each of the variables $x_1,\ldots,x_d$. Let $M$ be the maximum of the absolute values of these $2d$ numbers. Thus $|x_i|\leq M$ for all $i$. Let $C$ be the maximum of the absolute values of all coefficients, and $r$ be the number of monomials of $f(x)$. Then $$L:=-rCM^D\leq f(x)\leq rCM^D=:U,$$ as we can bound the absolute value of each monomial of $f(x)$ by $CM^D$. Replace $f$ by $\bar{f}(x)=f(x)-L\leq U-L$, a non-negative polynomial over $P$. Go to Steps 2, 3, etc. and return the optimal value of $\bar{f}.$ Trivially, if we find the optimal value of $\bar{f}$ over $P$ we can extract the optimal value for $f$.
[**Step 2.**]{} Via Barvinok’s algorithm (see [@bar; @BarviPom; @newbar]), compute a short rational function expression for the generating function $g_P(z)=\sum_{\alpha\in
P\cap{\mathbb Z}^d} z^\alpha$. From $g_P(z)$ compute the number $|P \cap
{\mathbb Z}^d|=g_P(1)$ of lattice points in $P$ in polynomial time.
[**Step 3.**]{} From the rational function representation $ g_P(z)$ of the generating function $\sum\limits_{\alpha\in P \cap {\mathbb Z}^d}
z^\alpha$ compute the rational function representation of $g_{P,f^k}(z)$ of $\sum_{\alpha\in P \cap {\mathbb Z}^d} f^k(\alpha) z^\alpha$ in polynomial time by application of Lemma \[operators\]. We define
$$L_k:=\sqrt[k]{g_{P,f^k}(1)/g_{P,f^0}(1)}\;\;\;\text{and}\;\;\;
U_k:=\sqrt[k]{g_{P,f^k}(1)}.$$
When $\lfloor U_{k}\rfloor-\lceil L_{k}\rceil<1$ stop and return $\lceil L_{k} \rceil=\lfloor U_{k} \rfloor$ as the optimal value.
[**End of Algorithm**]{}.
The algorithm is correct.
Using the fact that the arithmetic mean of a finite set of nonnegative values is at most as big as the maximum value, which in turn is at most as big as the sum of all values, we obtain the sequences of lower and upper bounds, $L_{k}$ and $U_{k}$, for the maximum: $$L_{k}=
\sqrt[k]{\frac{\sum\limits_{\alpha\in P \cap {\mathbb Z}^d}
f(\alpha)^{k}}{|P\cap{\mathbb Z}^d|}} \leq
\max\{f(\alpha):\alpha\in P\cap{\mathbb Z}^d\}\leq
\sqrt[k]{\sum\limits_{\alpha\in P \cap {\mathbb Z}^d}
f(\alpha)^{k}}=U_{k}.$$ Note that as $s\rightarrow\infty$, $L_{k}$ and $U_{k}$ approach this maximum value monotonously (from below and above, respectively). Trivially, if the difference between (rounded) upper and lower bounds becomes strictly less than $1$, we have determined the value $\max\{f(x):x\in P\cap{\mathbb Z}^d\}=\lceil L_{k}\rceil$. Thus the algorithm terminates with the correct answer.
Theorem \[main\] will follow from the next lemma:
\[lemma:bounds\] Let $f$ be a polynomial with integer coefficients and maximum total degree $D$. When the dimension $d$ is fixed,
1. the bounds $L_k$, $U_k$ can be computed in time polynomial in $k$, the input size of $P$ and $f$, and the total degree $D$. The bounds satisfy the following inequality: $$U_k-L_k \leq f^* \cdot \left(\sqrt[k]{|P \cap {\mathbb Z}^d|}-1 \right).$$
2. In addition, when $f$ is non-negative over $P$ (i.e. $f(x)\geq 0$ for all $x \in P$), for $k=(1+1/\epsilon)\log({|P \cap {\mathbb Z}^d|})$, $L_k$ is a $(1-\epsilon)$-approximation to the optimal value $f^*$ and it can be computed in time polynomial in the input size, the total degree $D$, and $1/\epsilon$. Similarly, $U_k$ gives a $(1+\epsilon)$-approximation to $f^*$. Moreover, with the same complexity, one can also find a feasible lattice point that approximates an optimal solution with similar quality.
Part (i). From Lemma \[operators\] on fixed dimension $d$, we can compute $g_{P,f}=\sum_{\alpha\in P \cap {\mathbb Z}^d} f(\alpha) z^\alpha$ as a rational function in time polynomial in $D$, the total degree of $f$, and the input size of $P$. Thus, because $f^k$ has total degree of $Dk$ and the encoding length for the coefficients of $f^k$ is bounded by $k \log (kC)$ (with $C$ the largest coefficient in $f$), we can also compute $g_{P,f^k} = \sum_{\alpha\in P \cap {\mathbb Z}^d}
f^k(\alpha) z^\alpha$ in time polynomial in $k$, the total degree $D$, and the input size of $P$. Note that using residue techniques [@newbar], we can evaluate $g_{P,f^k}(1)$ in polynomial time. Finally observe
$$\begin{aligned}
U_k-L_k&= \sqrt[k]{\sum_{\alpha \in P \cap {\mathbb Z}^d}
f^k(\alpha)}-\sqrt[k]{\frac{\sum_{\alpha \in P \cap {\mathbb Z}^d}
f^k(\alpha)}{|P \cap {\mathbb Z}^d|}} = \sqrt[k]{\frac{\sum_{\alpha \in P
\cap {\mathbb Z}^d} f^k(\alpha)}{|P \cap {\mathbb Z}^d|}} \left( \sqrt[k]{|P \cap
{\mathbb Z}^d|}-1 \right) \\
& =L_k \left( \sqrt[k]{|P \cap {\mathbb Z}^d|}-1 \right) \leq f^*
\left(\sqrt[k]{|P \cap {\mathbb Z}^d|}-1 \right). \end{aligned}$$
Part (ii). Note that if $\left(\sqrt[k]{|P \cap {\mathbb Z}^d|}-1 \right) \leq
\epsilon$ then $L_k$ is indeed a $(1-\epsilon)$-approximation because $$f^* \leq U_k =L_k+(U_k-L_k) \leq L_k + f^*\left(\sqrt[k]{|P \cap
{\mathbb Z}^d|}-1 \right) \leq L_k +f^* \epsilon.$$
Observe that $\phi(\epsilon) := (1+1/\epsilon)/(1/\log(1+\epsilon))$ is an increasing function for $\epsilon <1$ and $\lim_{\epsilon \rightarrow
0}\phi(\epsilon)=1$, thus $\phi(\epsilon)\geq 1$ for $0<\epsilon\leq1$. Hence, for all $k \geq \log({|P \cap {\mathbb Z}^d|})+\log({|P \cap
{\mathbb Z}^d|})/\epsilon \geq \log({|P \cap {\mathbb Z}^d|})/\log(1+\epsilon),$ we have indeed $\left(\sqrt[k]{|P \cap {\mathbb Z}^d|}-1 \right) \leq \epsilon.$ Finally, from Lemma \[operators\], the calculation of $L_k$ for $k=\log({|P \cap {\mathbb Z}^d|})+\log({|P \cap {\mathbb Z}^d|})/\epsilon$ would require a number of steps polynomial in the input size and $1/\epsilon$. A very similar argument can be written for $U_k$ but we omit it here.
To complete the proof of part (ii) it remains to show that not only we approximate the optimal value $f^*$ but we can also efficiently find a lattice point $\alpha$ with $f(\alpha)$ giving that quality approximation of $f^*$. Let $k = (1+1/\epsilon)\log({|P \cap
{\mathbb Z}^d|})$, thus, by the above discussion, $L_k$ is an $(1-\epsilon)$-approximation to $f^*$. Let $Q_0 := [-M,M]^d$ denote the box computed in Step 1 of the algorithm such that $P\subseteq
Q_0$. By bisecting $Q_0$, we obtain two boxes $Q_1'$ and $Q_1''$. By applying the algorithm separately to the polyhedra $P\cap Q_1'$ and $P\cap Q_1''$, we compute lower bounds $L_k'$ and $L_k''$ for the optimization problems restricted to $Q_1'$ and $Q_1''$, respectively. Because $L_k^k$ is the arithmetic mean of $f^k(\alpha)$ for $\alpha\in
P\cap{\mathbb Z}^d$, clearly $$\min\{L_k', L_k''\} \leq L_k \leq \max\{L_k', L_k''\}.$$ Without loss of generality, let $L_k'\geq L_k''$. We now apply the bisection procedure iteratively on $Q_k'$. After $d\log M$ bisection steps, we obtain a box $Q_k'$ that contains a single lattice point $\alpha\in P\cap Q_k'\cap Z^d$, which has an objective value $f(\alpha) = L_k' \geq L_k \geq (1-\epsilon) f^*$.
We remark that if we need to apply the construction of Step 1 of the algorithm because $f$ takes negative values on $P$, then we can only obtain an $(1-\epsilon)$-approximation (and $(1+\epsilon)$-approximation, respectively) for the modified function $\bar f$ in polynomial time, but not the original function $f$. We also emphasize that, although our algorithm requires the computation of $\sum_{\alpha \in P} f^q(\alpha)$ for different powers of $f$, these numbers are obtained without explicitly listing all lattice points (a hard task), nor we assume any knowledge of the individual values $f(\alpha)$. We can access the power means $\sum_{\alpha \in P}
f^q(\alpha)$ indirectly via rational functions. Here are two small examples:
#### Example 1, monomial optimization over a quadrilateral:
The problem we consider is that of maximizing the value of the monomial $x^3y$ over the lattice points of the quadrilateral $$\{(x,y) | 3991 \leq 3996\,x-4\,y \leq 3993, \ 1/2\leq x \leq 5/2\}.$$ It contains only 2 lattice points. The sum of rational functions encoding the lattice points is $$\frac{{x}^{2}{y}^{1000}}{\left( 1-(xy^{999})^{-1} \right)
\left( 1-{y}^{-1} \right)}+\frac{xy}{\left( 1-x{y}^{999} \right)
\left( 1-{y}^{-1} \right)}+
{\frac {xy}{ \left( 1-x{y}^{999}
\right) \left( 1-y \right) }}+\frac{{x}^{2}{y}^{1000}}{\left( 1-(xy^{999})^{-1} \right)\left( 1-y \right)}.$$ In the first iteration $L_1=4000.50$ while $U_1=8001$. After thirty iterations, we see $L_{30}=7817.279750$ while $U_{30}=8000$, the true optimal value.
#### Example 2, nvs04 from [MINLPLIB]{}:
A somewhat more complicated example, from a well-known library of test examples (see <http://www.gamsworld.org/minlp/>), is the problem given by $$\begin{aligned}
\min\quad & 100 \left(\frac12 + i_2 - \left(\frac35 + i_1\right)^2 \right)^2 + \left(\frac25 -
i_1\right)^2\\
\text{s.\,t.}\quad & i_1, i_2 \in [0,200] \cap {\mathbb Z}.
\end{aligned}$$ Its optimal solution as given in [MINLPLIB]{} is $i_1=1$, $i_2=2$ with an objective value of $0.72$. Clearly, to apply our algorithm from page literally, the objective function needs to be multiplied by a factor of $100$ to obtain an integer valued polynomial.
Using the bounds on $i_1$ and $i_2$ we obtain an upper bound of $165
\cdot 10^9$ for the objective function, which allows us to convert the problem into an equivalent maximization problem, where all feasible points have a non-negative objective value. The new optimal objective value is $164999999999.28.$ Expanding the new objective function and translating it into a differential operator yields $$\begin{aligned}
&\frac{4124999999947}{25} \mathrm{Id} - 28 z_2 \frac{\partial}{\partial
z_2}+\frac{172}{5} z_1\frac{\partial}{\partial z_1}-117 \left(z_1
\frac{\partial}{\partial z_1} \right) ^{(2)} - 100
\left(z_2 \frac{\partial}{\partial z_2}\right)^{(2)} \\
&+ 240 \left(z_2
\frac{\partial}{\partial z_2}\right) \left(z_1 \frac{\partial}{\partial
z_1}\right)
+ 200 \left(z_2
\frac{\partial}{\partial z_2}\right) \left(z_1 \frac{\partial}{\partial
z_1}\right) ^{(2)} - 240 \left(z_1 \frac{\partial}{\partial
z_1}\right)^{(3)}
- 100 \left(z_1 \frac{\partial}{\partial
z_1}\right)^{(4)}.
\end{aligned}$$ The short generating function can be written as $ g(z_1,z_2) = \left( \frac1{1-z_1} - \frac{z_1^{201}}{1-z_1} \right)
\left( \frac1{1-z_2} - \frac{z_2^{201}}{1-z_2} \right).$
In this example, the number of lattice points is $|P\cap {\mathbb Z}^2|=40401.$ The first bounds are $L_1=139463892042.292155534$, $U_1=28032242300500.723262442$. After 30 iterations the bounds become $L_{30}=164999998845.993553019$ and $U_{30}=165000000475.892451381.$
An extension to the mixed integer case
======================================
Now, we wish to discuss extensions of Theorem \[main\] to the mixed integer scenario. If some of the variables are allowed to be continuous then we can describe the task as $$\hbox{maximize} \ f(x,y) \ \hbox{subject to} \ \{ (x,y) | Ax+By \leq
b, \ \hbox{with} \ x_i \in {\mathbb Z}\ \hbox{and} \ y_i \in {\mathbb R}\}.$$ Consider the sequence of integer polynomial optimization problems $\hbox{optimize} \ f(x,\frac{y}{n}) \ \hbox{subject to}\ (x,y) \in
\Gamma_n=\{ (x,y) | Ax+B(\frac{y}{n}) \leq b, \ \hbox{with} \ x_i \in
{\mathbb Z}\ \hbox{and} \ y_i \in {\mathbb Z}\},$ where each of the subproblems is equivalent to optimizing a polynomial over a “semi-dilated” polytope (in some coordinate directions but not others). As $n$ goes to infinity, the sequence of optimal solution values can have several limit points. Nevertheless, it is still possible to construct a subsequence of problems whose optimal values approximate the mixed integer optimum to arbitrary precision:
\[mip\] With the hypotheses of Theorem \[main\], we can construct a sequence of integer polynomial programming problems, over finer and finer grids, whose optimal values converge to the optimal value of the mixed integer program $$f^*=\hbox{maximize} \ f(x_1,x_2,\dots,x_d) \ \hbox{subject to} \ x \in P, \ \hbox{and} \ x_i \in {\mathbb Z}\
\hbox{for} \ i \in I \subseteq \{1,2,\dots,d\}.$$ By applying the algorithm of Theorem \[main\] to the subproblems, we can approximate the optimum to arbitrary precision.
When all variables are continuous, the original polytope $P$ is actually dilated uniformly in all directions by a parameter $n$, or equivalently, the integer grid is refined. Define $$r(q,n)=\sum_{\alpha \in (nP) \cap {\mathbb Z}^d} f^q(\alpha) z^\alpha,$$ for each power $q$ and dilation factor $n$. Note that for fixed $q$, then we can easily see that the sum $\sum_{\alpha \in (nP) \cap
{\mathbb Z}^d} \frac{1}{n^d} f^q(\alpha) z^\alpha$ is essentially an approximation to the Riemann integral of $f^q$; thus $$\lim_{n \rightarrow \infty} r(q,n)/n^d = \int_P f^q(\bar{x})
d\bar{x}.$$ As $n,q$ grow, the values $\sqrt[q]{r(q,n)}$ approximate the sequence $\sqrt[q]{\int_P f^q(\bar{x}) d\bar{x}}$ which converges to $(\hbox{max}_{x \in P} f(x))\cdot \hbox{volume} (P)$. This is related to recent work (see [@deklerketal; @lasserre; @parrilosturmfels] and references therein) where the global optimum of a polynomial over a compact domain is investigated as the result of a grid refinement and properties of sums of squares.
[99]{}
. Math of Operations Research 19, 1994, 769–779.
. In: [*New Perspectives in Algebraic Combinatorics*]{} (Berkeley, CA, 1996-1997), 91–147, Math. Sci. Res. Inst. Publ. 38, Cambridge Univ. Press, Cambridge, 1999.
Manuscript 2005. Available at [front.math.ucdavis.edu](front.math.ucdavis.edu) math.CO/0504444.
Manuscript 2004, available at <http://homepages.cwi.nl/~monique/>
Journal of Symbolic Computation, vol 38, 4, 2004, 1273–1302.
. Journal of Symbolic Computation, Vol. 38, 2, 2004, 959–973.
. Freeman, San Francisco, 1979.
Senior undergraduate thesis, Department of mathematics, University of California, Davis, 2004.
Discrete Comput. Geom. 23, 2000, 207–224.
. Operations Research 21:1, 1973, 221–224.
J. of Symbolic Logic, 47 (3), 1982, 403–410.
Mathematics of Operations Research, 28, 3, 2003, 470–496.
. SIAM J. Optimization 11, 2001, 796–817.
. [Mathematics of Operations Research]{}, 8, 538–548
The MIT Press, Cambridge, London, 1993.
In: [*Algorithmic and quantitative real algebraic geometry*]{}, DIMACS Series in Discrete Mathematics and Theoretical Computer Science, Vol. 60, pp. 83–99, AMS.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'To understand the rich electronic phase diagram in FeSe under pressure that vividly demonstrates the strong interplay between the nematicity, magnetism and superconductivity, we analyze the electronic states by including the higher-order many-body effects called the vertex correction (VC). We predict the pressure-induced emergence of $\dxy$-orbital hole-pocket based on the first-principles analysis. Due to this pressure-induced Lifshitz transition, the spin fluctuations on the $\dxy$ orbital are enhanced, whereas those on $\dxz,\dyz$ orbitals are gradually reduced. For this reason, nonmagnetic orbital order $O=n_{xz}-n_{yz}$, which is driven by the spin fluctuations on $\dxz,\dyz$ orbitals through the intra-orbital VCs, is suppressed, and it is replaced with the magnetism of $\dxy$-orbital $d$-electrons. The nodal $s$-wave state at ambient pressure ($O\ne0$) and the enhancement of $T_{\rm c}$ under pressure are driven by the cooperation between the spin and orbital fluctuations.'
address: ' Department of Physics, Nagoya University, Furo-cho, Nagoya 464-8602, Japan. '
author:
- 'Youichi Yamakawa, and Hiroshi Kontani'
title: |
Nematicity, magnetism and superconductivity in FeSe under pressure:\
Unified explanation based on the self-consistent vertex correction theory
---
Introduction
============
Close relationship between the nematicity, magnetism, and high-$T_{\rm c}$ superconductivity is an essential electronic properties in Fe-based superconductors. The nematic transition is the rotational symmetry breaking in the electronic states at $T_{\rm str}$, and the strong nematic fluctuations are observed above $T_{\rm str}$ [@Yoshizawa; @Egami; @Gallais]. Inside the nematic phase below $T_{\rm str}$, large orbital polarization ($E_{yz}-E_{xz}\sim60$meV) is observed by ARPES studies [@Shen]. As the nematic order parameter, the spin-nematic order [@Kivelson; @Fernandes] and orbital/charge order [@Kruger-OO; @Lee-OO; @Onari-SCVC] have been discussed actively. In both scenarios [@Kivelson; @Fernandes; @Onari-SCVC], the nematicity and magnetism are closely related, so the magnetic transition slightly below $T_{\rm str}$ observed in many Fe-based superconductors is naturally explained.
To uncover the origin of the nematicity in Fe-based superconductors, very rich phase diagram of FeSe attracts increasing attention. Above $T_{\rm str}=90$K in FeSe at ambient pressure, strong nematic fluctuations emerges whereas low-energy spin fluctuations are weak [@Baek; @Ishida-NMR; @Raman-Gallais; @Shibauchi-nematic]. In the nematic state below $T_{\rm str}$, large orbital polarization emerges [@FeSe-ARPES2; @FeSe-ARPES22; @FeSe-ARPES3; @FeSe-ARPES4; @FeSe-ARPES5; @FeSe-ARPES6], whereas no magnetism appears down to the superconducting temperature $T_{\rm c}=9$K. Both the spin nematic [@Valenti-FeSe; @Wang-nematic; @Yu-nematic; @Fanfarillo] and the orbital order [@Yamakawa-FeSe; @Onari-FeSe; @Jiang-FeSe; @Chubukov-nemaic2; @Fanfarillo-FeSe] scenarios have been applied for FeSe so far. In the latter scenario, the orbital nematicity is unable to be explained unless the higher-order electronic correlations beyond the mean-field, called the vertex corrections (VCs), is taken into consideration. Recently, it was revealed that the Aslamazov-Larkin (AL) type VC, which expresses the strong interference between the orbital and spin fluctuations, gives rise to the nematicity due to the orbital order $O=n_{xz}-n_{yz}$. The nontrivial “sign-reversal orbital polarization” observed by ARPES studies below $T_{\rm str}$ [@FeSe-ARPES6] is reproduced by taking both the AL-VC and the Maki-Thompson VC [@Onari-FeSe].
The pressure-induced change in the electronic states in FeSe is an important open problem [@Imai; @Takano-rev; @mSR1; @mSR2; @Terashima-pFeSe; @Bohmer-pFeSe; @NMR-pFeSe; @Shibauchi-pFeSe]. Under pressure, $T_{\rm str}$ decreases whereas $T_c$ increases, and the nematic phase is replaced with the magnetic order ($T_m\sim30$K) at $P=2$GPa. This pressure-induced magnetism in FeSe may remind us of the second magnetic phase in heavy H-doped region in LaFeAsO [@Onari-SCVCS]. Around 4GPa, $T_m$ reaches its maximum ($45$K), and $T_c$ in the magnetic phase exceeds $20$K. It is a significant challenge for theorists to explain such rich pressure-induced electronic states based on the realistic Hubbard model for FeSe.
In this paper, we predict the pressure-induced emergence of $\dxy$-orbital hole-pocket based on the first-principles analysis. Due to this pressure-induced Lifshitz transition, the spin fluctuations on the $\dxy$ orbital increase, whereas those on the $\dxz,\dyz$ orbitals gently decrease. [@Onari-SCVC]. For this reason, nematic orbital order $n_{xz}\ne n_{yz}$, which is driven by the spin fluctuations on $\dxz,\dyz$ orbitals through the intra-orbital VCs, is suppressed and replaced with the magnetism on the $\dxy$ orbital. The nodal $s$-wave state at 0GPa below $T_{\rm str}$ [@Song; @Hanaguri; @Feng-FeSe-gap] and the enhancement of $T_{\rm c}$ under high pressures are given by the orbital+spin fluctuations. The close interplay between the nematicity, magnetism, and superconductivity due to the VC in FeSe would be the universal feature of Fe-based superconductors.
In FeSe at ambient pressure, weak spin fluctuations are enough for realizing the orbital order due to the AL-VC, because of the smallness of the ratio $J/U$ and the “absence of the $\dxy$-orbital hole-pocket” as discussed in Ref. [@Yamakawa-FeSe]. Since the Fermi energy of each pocket is very small in FeSe, the nature of spin fluctuations is sensitively modified by the pressure-induced parameter change. Then, spin-fluctuation-driven orbital order is also sensitively modified. Therefore, rich $T$-$P$ phase diagram in FeSe is naturally explained.
We study the following two-dimensional eight-orbital Hubbard model: $$\begin{aligned}
H= H_0 +H_P+ r H_U
\label{eqn:Ham} ,\end{aligned}$$ where $H_0$ is the kinetic term for FeSe at ambient pressure introduced in Ref. [@Yamakawa-FeSe], and $H_U$ is the first-principles multiorbital interaction for FeSe [@Arita]. In FeSe, ${\bar U}=7.21$eV and ${\bar J}=0.68$eV on average. Here, $r$ is the reduction factor for the interaction term [@Yamakawa-FeSe]. We denote $d_{z^2},d_{xz},d_{yz},d_{xy},d_{x^2-y^2}$ orbitals as 1, 2, 3, 4, 5, and $p_x,p_y,p_z$ orbitals as 6, 7, 8. We introduce the renormalization factor $z_{4}=1/1.2$ and $z_{l}=1$ for $l\ne4$ [@Yamakawa-FeSe], consistently with the microscopic theory in Ref. [@DMFT1].
In Eq. (\[eqn:Ham\]), $H_P$ represents the change in the kinetic term under pressure. To derive $H_P$, we perform the band calculation using the crystal structure of FeSe under pressure [@Millican; @Bohmer]: The most dominant pressure-induced change is the lift of the $\dxy$-orbital level around $\q=(\pi,\pi)$ due to the increase of the Se-atom height [@Kuroki-z; @Haule]. To shift the $E_{xy}$-level by $\Delta E_{xy}$ at $\q=(\pi,\pi)$, we set $H_P = \sum_{i,j,\s} \delta t_{i,j}c^\dagger_{i,4\s}c_{j,4\s}$, where $\delta t_{i,j}=\Delta E_{xy}/4$ for the on-site ($i=j$), $\delta t_{i,j}=-\Delta E_{xy}/8$ and $\delta t_{i,j}=\Delta E_{xy}/16$ for the first- and second-nearest sites [@Yamakawa-FeSe]; see the Supplemental Material (SM):A [@SM]. Figure \[fig:fig1\] (a) shows the bandstructure for $\Delta E_{xy}=0$, 0.06, and 0.12 eV, given by the solutions of ${\rm det}\{ {\hat Z}\cdot\e-({\hat H}_0(\k)+{\hat H}_P(\k))\}=0$, or equivalently given by the eigenvalues of ${\hat Z}^{-1/2}({\hat H}_0(\k)+{\hat H}_P(\k)){\hat Z}^{-1/2}$, where ${Z}_{l,m}=(1/z_l)\delta_{l,m}$ [@Yamakawa-FeSe; @Onari-FeSe].
The Fermi surfaces (FSs) for $\Delta E_{xy}=0$ and 0.12 eV are shown in Figs. \[fig:fig1\] (b) and (c), respectively. In the present model, $\dxy$-orbital pocket appear for $\Delta E_{xy}\gtrsim 0.1$eV. According to the first principles study, $\Delta E_{xy}\sim0.1$eV is realized under $\sim4$ GPa [@SM]. In FeSe at ambient pressure, the top of the $\dxy$-hole pocket is about $40$meV below the Fermi level experimentally [@FeSe-ARPES6], which corresponds to $\Delta E_{xy}\sim 0.05$eV in the present model.
![ (color online) (a) Band structure of the present FeSe model for $\Delta E_{xy}=0$, 0.06 and and 0.12 eV. (b) FSs for $\Delta E_{xy}=0$ and (c) FSs for $\Delta E_{xy}=0.12$eV. The hFS3 is the $\dxy$-orbital hole-pocket. The colors green, red, and blue represents the $\dxz$, $\dyz$, and $\dxy$ orbitals, respectively. []{data-label="fig:fig1"}](fig1.eps){width=".7\linewidth"}
Hereafter, we analyze the model Hamiltonian (\[eqn:Ham\]) based on the self-consistent VC (SC-VC) theory. The spin- or charge-channel susceptibility is given as $$\begin{aligned}
{\hat \chi}^x(q)
={\hat \Phi}^x(q)[{\hat 1}-{\hat U}^{0x}{\hat \Phi}^x(q)]^{-1}
\ \ \ (x=s,c)
\label{eqn:chi} ,\end{aligned}$$ where $x=s$ or $c$, and $q=(\w_l=2\pi l T, \q)$. ${\hat U}^{0x}$ is the first-principles Coulomb interaction for FeSe derived in Ref. [@Arita], and it was introduced as ${\hat \Gamma}^x$ in Ref. [@Yamakawa-FeSe]. The irreducible susceptibility is ${\hat \Phi}^x(q)= {\hat \chi}^0(q)+{\hat X}^x(q)$, where ${\hat \chi}^0(q)$ is the bare susceptibility and ${\hat X}^x(q)$ is the VC that is dropped in the random-phase-approximation (RPA). The spin (charge) Stoner factor $\a_{S(C)}$ is defined as the largest eigenvalue of ${\hat U}^{0s(c)}{\hat \Phi}^{s(c)}(\q)$: The magnetic order (orbital order) is established when $\a_{S(C)}=1$.
In the SC-VC theory, we calculate both the Maki-Thompson-type VC and the AL-type VC, which are the first-order and the second-order diagrams with respect to $\chi^{x}$, respectively [@Onari-SCVC; @Yamakawa-FeSe]. Figure \[fig:fig2\] (a) shows the diagrammatic expression for the AL-type VC $X^c_{yz}({\bm 0})\sim
|\Lambda^0_{yz}|^2T\sum_{\q}\{\chi^s_{yz}(\q)\}^2$. Here, ${\hat \Lambda}^0_{yz}$ is the three-point vertex for $yz$-orbital, given by ${\hat \Lambda}^0_{3,3;3,3;3,3}({\bm 0};\Q)$ in Fig. \[fig:fig2\] (b) [@Onari-SCVC]. Thus, $X^c_{yz}({\bm 0})$ increases in proportion to $\sum_{\q}\{\chi^s_{yz}(\q)\}^2\sim \chi^s_{yz}(\Q)$. The increment of the AL-VC drives the strong nematic orbital susceptibility for $O=n_{xz}-n_{yz}$, $\chi^{\rm orb}({\bm 0}) \equiv \sum_{l,m}^{2,3}(-1)^{l+m}\chi^c_{l,l;m,m}({\bm 0})$. Here, we calculate only the charge-channel VC $X_{l,l';m,m'}^c(q)$ for $l,l',m,m'=2\sim4$ except for $l=l'=m=m'=4$, by following Ref. [@Onari-SCVC], which is justified for various Fe-based superconductors [@Onari-SCVC; @Yamakawa-FeSe]. In the SM:B [@SM], we explain that essentially similar results are obtained if we perform the self-consistent calculation for both ${\hat X}^{s}(q)$ and ${\hat X}^c(q)$ for all $d$-orbitals.
![ (color online) (a) Charge-channel AL-type VC for the $\dyz$-orbital and (b) three-point vertex. (c) $\chi^s_{xy}(\q)$, (d) $\chi^s_{yz}(\q)$, and (e) $\chi^{\rm orb}(\q)$ or $\Delta E_{xy}=0$ and $0.12$eV at $r=0.266$ and $T=30$meV. (f) Stoner factors as functions of $\Delta E_{xy}$ at $r=0.266$ and $T=30$meV. (g) $T$-dependences of $\a_{S,C}$ for $\Delta E_{xy}=0.08$ eV and (h) those for $\Delta E_{xy}=0.12$ eV. (i) $T_m$ and $T_{\rm str}$ defined by the relation $\a_{S,C}=1$ and $0.96$. []{data-label="fig:fig2"}](fig2.eps){width=".99\linewidth"}
In Fig. \[fig:fig2\] (c), we show the spin susceptibility on $xy$-orbital, $\chi^s_{xy}(\q) \equiv \chi^s_{4,4;4,4}(\q)$, at $r=0.266$ and $T=30$meV. For $\Delta E_{xy}=0.12$eV, the hFS3 at $(\pi,\pi)$ appears, and therefore $\chi^s_{xy}(\Q)$ $(\Q=(\pi,0))$ is drastically enlarged due to the good nesting between eFS1,2 and the hFS3 on $\dxy$-orbital. We stress that the emergence of the hFS3 could lead to the inhomogeneous phase separation as discussed in Ref. [@Nori]. In contrast, $\chi^s_{yz}(\q) \equiv \chi^s_{3,3;3,3}(\q)$ in Fig. \[fig:fig2\] (d) gently decreases with $\Delta E_{xy}$, since the nesting condition on $\dyz$-orbital becomes worse due to the shrinkage (expansion) of hFS1,2 (eFS1,2). This suppression of $\chi_{yz}(\Q)$ caused the reduction of the orbital susceptibility in Fig. \[fig:fig2\] (e), due the reduction of the AL-term $X_{yz}^c({\bm 0})\propto \chi^s_{yz}(\Q)$. Such drastic $\Delta E_{xy}$-dependences of the spin and orbital susceptibilities are contributed by the the smallness of the FSs in FeSe.
Figure \[fig:fig2\] (f) shows the Stoner factors $\a_{S,C}$ obtained by the SC-VC theory for $r=0.266$ at $T=30$meV. The obtained $\a_{S(C)}$ is qualitatively proportional to $T_{m({\rm str})}$. At $\Delta E_{xy}=0$, which corresponds to $P=0$GPa, $\a_C$ is close to the unity whereas $\a_S\sim0.9$. When $\Delta E_{xy}>0$, the chemical potential $\mu$ increases even for $\Delta E_{xy}\ll0.1$ eV due to the finite-$T$ effect. Then, the nesting between hFS1,2 and eFS1 becomes worse, and therefore $\chi^s_{yz}(\Q)$ gradually decreases. At the same time, $\a_C$ decreases since $X^c_{yz}({\bm 0})\sim
|\Lambda^0_{yz}|^2T\sum_{\q}\{\chi^s_{yz}(\q)\}^2$ becomes smaller, consistently with the experimental reduction of $T_{\rm str}$ under $P=0\sim2$GPa. In contrast, $\chi^s_{xy}(\Q)$ starts to increase for $\Delta E_{xy}\gtrsim 0.08$ eV, due to the nesting between the hFS3 and eFS1,2 on $xy$-orbital. For this reason, $\a_S$ rapidly increases in Fig. \[fig:fig2\] (f). When $\Delta E_{xy}\gtrsim0.1$ eV, $\a_C$ starts to increase since $\chi^s_{xy}(\q)$ also contribute to the “AL-term on $yz$-orbital through $|\Lambda_{yz\mbox{-}xy}^0|^2T\sum_{q}{\chi^s_{xy}(q)}^2$”, where $\Lambda_{yz\mbox{-}xy}^0\equiv \Lambda_{3,3;4,4;4,4}^0({\bm0};\Q)$. Although the off-diagonal three-point vertex $\Lambda_{yz\mbox{-}xy}^0$ is much smaller than the diagonal one $\Lambda_{yz}^0$, it is finite since the electron-FSs are composed of three orbitals $2 \sim 4$.
To confirm the validity of the simplified pressure term $H_P$ introduced in the present study, we derive the pressure term from the first principles study, $H_P^{\rm 1st}$, and perform the numerical study based on the SC-VC theory in the SM: D [@SM]. The obtained $P$-dependences of $\a_S$ and $\a_C$ are consistent with Fig. \[fig:fig2\] (f).
We also estimate the transition temperatures $T_m$ and $T_{\rm str}$. As shown in Figs. \[fig:fig2\] (g) and (h), the Stoner factors $\a_{S(C)}$ follow the $T$-linear relations for $T\gtrsim30$meV, at $\Delta E_{xy}=0.08$ eV and $0.12$ eV. (Then, $\chi^s(\Q)$ and $\chi^{\rm orb}({\bm 0})$ follow the Curie-Weiss relations.) By extrapolating the $T$-linear relations, we estimate $T_m$ and $T_{\rm str}$ under the condition $\a_{S,C}=1$ and $\a_{S,C}=0.96$ in Fig. \[fig:fig2\] (i). In both cases, $T_{\rm str}$ decreases with $\Delta E_{xy}$ for $\Delta E_{xy}<0.1$ eV, whereas $T_m$ drastically increases for $\Delta E_{xy}>0.1$ eV, consistently with experimental phase diagram. In the region $T_m>T>T_{\rm str}$ for $\a_{S,C}=0.96$ with $\Delta E_{xy}>0.11$ eV, the tetragonal ($C_4$) stripe magnetic phase appears, if the spin-lattice coupling is negligibly small.
![ (color online) (a) $V_{\rm (Migdal)}$, (b) $V^{\Lambda}$, and (c) $V^{\rm cross}$. (d) $U$-VC $\Lambda^x$. (e) Eigenvalues for various pairing interactions: $\lambda^{\rm tot}$ for $V^{\rm tot}$, $\lambda^{\Lambda}$ for $V^{\Lambda}$, and $\lambda_{(\rm Migdal)}$ for $V_{(\rm Migdal)}$. (f) $s_\pm$-wave gap given by $V^{\Lambda}$ for $\Delta E_{xy}=0.12$eV. (g) $s_{++}$-wave gap given by $V^{\rm tot}$. The FSs are shown in Fig. \[fig:fig1\] (c). Numerical study is performed in the nonmagnetic state. []{data-label="fig:fig3"}](fig3.eps){width=".85\linewidth"}
In the next stage, we study the superconducting state. Both the spin [@Kuroki-z; @Mazin; @Chubukov; @Graser; @Chubu-Rev] and/or orbital [@Kontani-RPA; @Onari-SCVCS; @Yamakawa-eFeSe] fluctuation mediated mechanisms have been studied. The linearized gap equation is $$\begin{aligned}
\lambda\Delta_\a(k)=-T\sum_{p,\b}V_{\a,\b}^{\rm SC}(k,p)|G_\b(p)|^2\Delta_\b(p)
\label{eqn:gapeq} ,\end{aligned}$$ where $k=((2n+1)\pi T,\k)$, $\Delta_\a(k)$ is the gap function on the $\a$-FS, and $\lambda$ is the eigenvalue that reaches unity at $T=T_c$. $V_{\a,\b}^{\rm SC}$ is the pairing interaction. In the conventional Migdal approximation, ${\hat V}^{\rm SC}$ in the orbital basis is given as ${\hat V}_{(\rm Migdal)}(k,p)=\frac32{\hat I}^s(k-p)
+\frac12{\hat I}^c(k-p)-{\hat U}^{0s}$, where ${\hat I}^{x}(q)= {\hat U}^{0x}+
{\hat U}^{0x}{\hat \chi}^{x}(q){\hat U}^{0x}$ $(x=s,c)$. It is schematically shown in Fig. \[fig:fig3\] (a). However, since this approximation is not justified for Fe-based superconductors, we introduce the following “beyond-Migdal-Eliashberg pairing interaction” introduced in Ref. [@Yamakawa-eFeSe]: $$\begin{aligned}
&&{\hat V}^{\rm tot}(k,p)={\hat V}^{\Lambda}(k,p)+
{\hat V}^{\rm cross}(k,p)
\label{eqn:V-total} ,\end{aligned}$$ which are depicted in Figs. \[fig:fig3\] (b) and (c). ${\hat \Lambda}^x$ in Fig. \[fig:fig3\] (b) is the VC for the coupling constant ${\hat U}^{0x}$, called the $U$-VC. Since $|\Lambda^c|^2\gg1$ due to the AL-type VC shown in Fig. \[fig:fig3\] (d) [@Yamakawa-eFeSe; @Tazai], moderate orbital fluctuations give sizable attractive interaction. Also, $V^{\rm cross}$ is the “AL-type crossing fluctuation interaction” [@Yamakawa-eFeSe]. The analytic expressions for Figs. \[fig:fig3\] (b)-(d) are explained in Refs. [@Tazai; @Yamakawa-eFeSe] and in the SM: E [@SM]. In Ref. [@Tazai], we verified the significance of the $U$-VC by applying the functional-renormalization-group (fRG) method to the two-orbital Hubbard model. The fRG method enables us to generate the higher-order VCs (including the higher-order processes other than Figs. \[fig:fig3\] (a)-(d)) in a systematic and unbiased way.
The eigenvalue $\lambda^{\rm tot}$ obtained for $V^{\rm tot}$ in the absence of the nematicity and magnetism is shown in Fig. \[fig:fig3\] (e), together with $\lambda_{(\rm Migdal)}$ for $V_{(\rm Migdal)}$ and $\lambda^{\Lambda}$ for $V^{\Lambda}$: The relation $\lambda_{(\rm Migdal)} \ll \lambda^\Lambda$ means that the charge $U$-VC strongly enlarges $T_{\rm c}$. For $V^{\rm SC}=V^{\Lambda}$ at $\Delta E_{xy}=0.12$eV, the $s_\pm$-wave state with sign reversal between electron- and hole-pockets is obtained as shown in Fig. \[fig:fig3\] (f). In contrast, the “$s_{++}$-wave state without sign reversal” is obtained for the total pairing interaction $V^{\rm tot}$, since $V^{\rm cross}$ gives large inter-pocket attractive interaction: The obtained nodal $s_{++}$-wave state at $\Delta E_{xy}=0.12$eV is shown in Fig. \[fig:fig3\] (g). The $s_{++}$-state will be stabilized by the pressure-induced strong $e$-ph interaction [@Haule].
![ (color online) (a) FSs and bandstructure in the presence of the sign-reversing orbital polarization $\delta E_{\rm nem}(\Gamma)= -30$meV and $\delta E_{\rm nem}({\rm X})= 60$meV. Here, $\Delta E_{xy}=0$. (b) $\chi^s_{xz}(\q)$, $\chi^s_{yz}(\q)$, $\chi^c_{xz}(\q)$, and $\chi^c_{yz}(\q)$ obtained by the SC-VC theory. (c) $s$-wave gap functions obtained for $V^{\rm tot}$: The gap functions on each FS is nodal. []{data-label="fig:fig4"}](fig4.eps){width=".85\linewidth"}
Finally, we study the superconductivity in FeSe in the nematic state at $P=0$ [@Mukherjee]. Figure \[fig:fig4\] (a) shows the FSs and bandstructure under the sign-reversing orbital polarization reported in Ref. [@FeSe-ARPES6]: We set $\delta E_{\rm nem}\equiv E_{yz}-E_{xz} = -30$meV at $\Gamma$ point, and $\delta E_{\rm nem}=60$meV at X and Y points by following Ref. [@Yamakawa-FeSe]. The susceptibilities obtained by the SC-VC theory for $z_4^{-1}=1.2$ and $r=0.266$ at $T=30$ meV \[$(\a_S,\a_C)=(0.914,0.934)$\] are shown in Fig. \[fig:fig4\] (b). By introducing $\delta E_{\rm nem}$, $\chi^s_{yz}(\pi,0)$ increases whereas $\chi^s_{xz}(0,\pi)$ decreases as explained in Ref. [@Kontani-RPA]. Then, the AL-term $X_{yz}^c(\q)$ is enlarged due to the large $\chi^s_{yz}(\pi,0)$. The obtained strong ferro- and antiferro-orbital fluctuations give significant pairing interaction on the $\dyz$-orbital.
Figure \[fig:fig4\] (c) shows the $s$-wave gap function obtained for $V^{\rm tot}$. Here, the gap is large on the FSs with large $\dyz$-orbital component, since strong intra-pocket force and inter-pocket attractive one are caused by large $\chi^c_{yz}(\q)$ at $\q\sim{\bm 0}$ and $\q\sim(\pi,0)$ in addition to $V^{\rm cross}$. The position of the nodal part on hole-FS is consistent with the experimental report [@Feng-FeSe-gap]. The nodal structure on $\Delta_{\rm hFS}$ is unchanged when two Dirac cones appear near X-point for larger $\delta E_{\rm nem}({\rm X})$, as we show in SM: C [@SM]. Thus, the nodal gap structure is robust against model parameters, whereas the ratio between $|\Delta_{\rm hFS}|$ and $|\Delta_{\rm eFS}|$ is sensitive to parameters, so it is our future problem to explain the experimental data quantitatively.
In summary, we studied the electronic states in FeSe at ambient pressure and under pressure. We predicted that the hFS3 appears under pressure. Due to this pressure-induced Lifshitz transition, the spin fluctuations on the $\dxy$ orbital are enlarged, whereas those on the $\dxz,\dyz$ orbitals are reduced gently. Since the orbital order is driven by the spin fluctuations on $\dxz,\dyz$ orbitals via the intra-orbital VCs, the nematicity is suppressed and replaced with the magnetism on the $\dxy$ orbital electrons under pressure. Also, both the nodal $s$-wave state in the nematic state at ambient pressure and the enlargement of $T_{\rm c}$ under pressure are satisfactorily explained, due to the novel cooperation between spin and orbital fluctuations.
We are grateful to A. Chubukov, P.J. Hirschfeld, R. Fernandes, J. Schmalian, D.L. Feng, S. Shin, Y. Matsuda, T. Shibauchi, K. Okazaki and T. Shimojima for useful discussions. This study has been supported by Grants-in-Aid for Scientific Research from MEXT of Japan.
[99]{}
M. Yoshizawa, D. Kimura, T. Chiba, S. Simayi, Y. Nakanishi, K. Kihou, C.-H. Lee, A. Iyo, H. Eisaki, M. Nakajima, and S. Uchida, J. Phys. Soc. Jpn. [**81**]{}, 024604 (2012).
J. L. Niedziela, D. Parshall, K. A. Lokshin, A. S. Sefat, A. Alatas, and T. Egami, Phys. Rev. B [**84**]{}, 224305 (2011).
Y. Gallais, R. M. Fernandes, I. Paul, L. Chauvière, Y.-X. Yang, M.-A. Méasson, M. Cazayous, A. Sacuto, D. Colson, and A. Forget, Phys. Rev. Lett. [**111**]{}, 267001 (2013).
M. Yi, D. Lu, J.-H. Chu, J. G. Analytis, A. P. Sorini, A. F. Kemper, B. Moritz, S.-K. Mo, R. G. Moore, M. Hashimoto, W.-S. Lee, Z. Hussain, T. P. Devereaux, I. R. Fisher, and Z.-X. Shen, Proc. Natl. Acad. Sci. USA [**108**]{}, 6878 (2011).
C. Fang, H. Yao, W.-F. Tsai, J. P. Hu, and S. A. Kivelson, Phys. Rev. B [**77**]{}, 224509 (2008).
R. M. Fernandes, L. H. VanBebber, S. Bhattacharya, P. Chandra, V. Keppens, D. Mandrus, M. A. McGuire, B. C. Sales, A. S. Sefat, and J. Schmalian, Phys. Rev. Lett. [**105**]{}, 157003 (2010).
R. M. Fernandes, A. V. Chubukov, J. Knolle, I. Eremin, and J. Schmalian, Phys. Rev. B [**85**]{}, 024534 (2012).
F. Krüger, S. Kumar, J. Zaanen, and J. van den Brink, Phys. Rev. B [**79**]{}, 054504 (2009).
C.-C. Lee, W.-G. Yin, and W. Ku, Phys. Rev. Lett. [**103**]{}, 267001 (2009).
S. Onari and H. Kontani, Phys. Rev. Lett. [**109**]{}, 137001 (2012).
S.-H. Baek, D. V. Efremov, J. M. Ok, J. S. Kim, J. van den Brink, and B. Büchner, Nat. Mater. [**14**]{}, 210 (2015).
A. E. Böhmer, T. Arai, F. Hardy, T. Hattori, T. Iye, T. Wolf, H. v. Löhneysen, K. Ishida, and C. Meingast, Phys. Rev. Lett. [**114**]{}, 027001 (2015).
P. Massat, D. Farina, I. Paul, S. Karlsson, P. Strobel, P. Toulemonde, M.-A. Méasson, M. Cazayous, A. Sacuto, S. Kasahara, T. Shibauchi, Y. Matsuda, and Y. Gallais, Proc. Natl. Acad. Sci. USA [**113**]{}, 9177 (2016).
S. Hosoi, K. Matsuura, K. Ishida, H. Wang, Y. Mizukami, T. Watashige, S. Kasahara, Y. Matsuda, and T. Shibauchi, Proc. Natl. Acad. Sci. USA [**113**]{}, 8139 (2016).
K. Nakayama, Y. Miyata, G. N. Phan, T. Sato, Y. Tanabe, T. Urata, K. Tanigaki, and T. Takahashi, Phys. Rev. Lett. [**113**]{}, 237001 (2014).
M. D. Watson, T. K. Kim, A. A. Haghighirad, N. R. Davies, A. McCollam, A. Narayanan, S. F. Blake, Y. L. Chen, S. Ghannadzadeh, A. J. Schofield, M. Hoesch, C. Meingast, T. Wolf, and A. I. Coldea, Phys. Rev. B [**91**]{}, 155106 (2015).
T. Shimojima, Y. Suzuki, T. Sonobe, A. Nakamura, M. Sakano, J. Omachi, K. Yoshioka, M. Kuwata-Gonokami, K. Ono, H. Kumigashira, A. E. Böhmer, F. Hardy, T. Wolf, C. Meingast, H. v. Löhneysen, H. Ikeda, and K. Ishizaka, Phys. Rev. B [**90**]{}, 121111(R) (2014).
P. Zhang, T. Qian, P. Richard, X. P. Wang, H. Miao, B. Q. Lv, B. B. Fu, T. Wolf, C. Meingast, X. X. Wu, Z. Q. Wang, J. P. Hu, and H. Ding, Phys. Rev. B [**91**]{}, 214503 (2015).
Y. Zhang, M. Yi, Z.-K. Liu, W. Li, J. J. Lee, R. G. Moore, M. Hashimoto, N. Masamichi, H. Eisaki, S.-K. Mo, Z. Hussain, T. P. Devereaux, Z.-X. Shen, and D. H. Lu, Phys. Rev. B [**94**]{}, 115153 (2016). Y. Suzuki, T. Shimojima, T. Sonobe, A. Nakamura, M. Sakano, H. Tsuji, J. Omachi, K. Yoshioka, M. Kuwata-Gonokami, T. Watashige, R. Kobayashi, S. Kasahara, T. Shibauchi, Y. Matsuda, Y. Yamakawa, H. Kontani, and K. Ishizaka, Phys. Rev. B [**92**]{}, 205117 (2015).
J. K. Glasbrenner, I. I. Mazin, H. O. Jeschke, P. J. Hirschfeld, R. M. Fernandes, and R. Valentí, Nat. Phys. [**11**]{}, 953 (2015).
F. Wang, S. A. Kivelson, and D.-H. Lee, Nat Phys [**11**]{}, 959 (2015).
R. Yu and Q. Si, Phys. Rev. Lett. [**115**]{}, 116401 (2015).
L. Fanfarillo, A. Cortijo, and B. Valenzuela, Phys. Rev. B [**91**]{}, 214515 (2015).
Y. Yamakawa, S. Onari, and H. Kontani, Phys. Rev. X [**6**]{}, 021032 (2016).
S. Onari, Y. Yamakawa, and H. Kontani, Phys. Rev. Lett. [**116**]{}, 227001 (2016).
A. V. Chubukov, M. Khodas, and R. M. Fernandes, arXiv:1602.05503.
K. Jiang, J. Hu, H. Ding, and Z. Wang, Phys. Rev. B [**93**]{}, 115138 (2016).
L. Fanfarillo, G. Giovannetti, M. Capone, and E. Bascones, Phys. Rev. B [**95**]{}, 144511 (2017). T. Imai, K. Ahilan, F. L. Ning, T. M. McQueen, and R. J. Cava, Phys. Rev. Lett. [**102**]{}, 177005 (2009).
Y. Mizuguchi and Y. Takano, J. Phys. Soc. Jpn. [**79**]{}, 102001 (2010).
M. Bendele, A. Amato, K. Conder, M. Elender, H. Keller, H.-H. Klauss, H. Luetkens, E. Pomjakushina, A. Raselli, and R. Khasanov, Phys. Rev. Lett. [**104**]{}, 087003 (2010).
M. Bendele, A. Ichsanow, Y. Pashkevich, L. Keller, T. Strässle, A. Gusev, E. Pomjakushina, K. Conder, R. Khasanov, and H. Keller, Phys. Rev. B [**85**]{}, 064517 (2012).
T. Terashima, N. Kikugawa, S. Kasahara, T. Watashige, T. Shibauchi, Y. Matsuda, T. Wolf, A. E. Böhmer, F. Hardy, C. Meingast, H. v. Löhneysen, and S. Uji, J. Phys. Soc. Jpn. [**84**]{}, 063701 (2015).
K. Kothapalli, A. E. Böhmer, W. T. Jayasekara, B. G. Ueland, P. Das, A. Sapkota, V. Taufour, Y. Xiao, E. Alp, S. L. Bud’ko, P. C. Canfield, A. Kreyssig, and A. I. Goldman, Nat. Commun. [**7**]{}, 12728 (2016). P. Wang, S. Sun, Y. Cui, W. Song, T. Li, R. Yu, H. Lei, and W. Yu, Phys. Rev. Lett. [**117**]{}, 237001 (2016). J. P. Sun, K. Matsuura, G. Z. Ye, Y. Mizukami, M. Shimozawa, K. Matsubayashi, M. Yamashita, T. Watashige, S. Kasahara, Y. Matsuda, J.-Q. Yan, B. C. Sales, Y. Uwatoko, J.-G. Cheng, and T. Shibauchi, Nat. Commun. [**7**]{}, 12146 (2016).
S. Onari, Y. Yamakawa, and H. Kontani, Phys. Rev. Lett. [**112**]{}, 187001 (2014).
C.-L. Song, Y.-L. Wang, P. Cheng, Y.-P. Jiang, W. Li, T. Zhang, Z. Li, K. He, L. Wang, J.-F. Jia, H.-H. Hung, C. Wu, X. Ma, X. Chen, and Q.-K. Xue, Science [**332**]{}, 1410 (2011).
S. Kasahara, T. Watashige, T. Hanaguri, Y. Kohsaka, T. Yamashita, Y. Shimoyama, Y. Mizukami, R. Endo, H. Ikeda, K. Aoyama, T. Terashima, S. Uji, T. Wolf, H. von Löhneysen, T. Shibauchi, and Y. Matsuda, Proc. Natl. Acad. Sci. USA [**111**]{}, 16309 (2014). H. C. Xu, X. H. Niu, D. F. Xu, J. Jiang, Q. Yao, M. Abdel-Hafiez, D. A. Chareev, A. N. Vasiliev, R. Peng, and D. L. Feng, Phys. Rev. Lett. [**117**]{}, 157003 (2016). T. Miyake, K. Nakamura, R. Arita, and M. Imada, J. Phys. Soc. Jpn. [**79**]{}, 044705 (2010). Z. P. Yin, K. Haule, and G. Kotliar, Nat. Mater. [**10**]{}, 932 (2011).
J. N. Millican, D. Phelan, E. L. Thomas, J. B. Leão, and E. Carpenter, Solid State Commun. [**149**]{}, 707 (2009).
A. E. Böhmer, F. Hardy, F. Eilers, D. Ernst, P. Adelmann, P. Schweiss, T. Wolf, and C. Meingast, Phys. Rev. B [**87**]{}, 180505(R) (2013).
K. Kuroki, H. Usui, S. Onari, R. Arita, and H. Aoki, Phys. Rev. B [**79**]{}, 224511 (2009).
S. Mandal, R. E. Cohen and K. Haule, Phys. Rev. B [**89**]{}, 220502(R) (2014).
Supplemental Material.
A. O. Sboychakov, A. V. Rozhkov, K. I. Kugel, A. L. Rakhmanov, and F. Nori, Phys. Rev. B [**88**]{}, 195142 (2013).
I. I. Mazin, D. J. Singh, M. D. Johannes, and M. H. Du, Phys. Rev. Lett. [**101**]{}, 057003 (2008). A. V. Chubukov, D. V. Efremov, and I. Eremin, Phys. Rev. B [**78**]{}, 134512 (2008). S. Graser, G. R. Boyd, C. Cao, H.-P. Cheng, P. J. Hirschfeld, and D. J. Scalapino, Phys. Rev. B [**77**]{}, 180514(R) (2008). A. Chubukov and P. J. Hirschfeld, Physics Today [**68**]{}, 46 (2015).
H. Kontani, T. Saito, and S. Onari, Phys. Rev. B [**84**]{}, 024528 (2011).
Y. Yamakawa, and H. Kontani, arXiv:1611.05375. R. Tazai, Y. Yamakawa, M. Tsuchiizu, and H. Kontani, Phys. Rev. B [**94**]{}, 115155 (2016). S. Mukherjee, A. Kreisel, P. J. Hirschfeld, and B. M. Andersen, Phys. Rev. Lett. [**115**]{}, 026402 (2015): The authors studied the gap equation for FeSe model with the constant orbital polarization using the RPA. The obtained gap on the hole-pocket is fully-gapped.
[ **Nematicity, magnetism and superconductivity in FeSe under pressure:\
Unified explanation based on the self-consistent vertex correction theory** ]{}
Youichi Yamakawa, Hiroshi Kontani
A: First principles model Hamiltonian
-------------------------------------
In the main text, we studied the electronic states in FeSe under pressure by using the SC-VC theory. To describe the change in the bandstructure under pressure, we introduced the additional term $H_P$ into the model Hamiltonian for FeSe at ambient pressure. We set $H_P = \sum_{i,j,\s} \delta t_{i,j}c^\dagger_{i,4\s}c_{j,4\s}$, where $\delta t_{i,j}=\Delta E_{xy}/4$ for the on-site ($i=j$), $\delta t_{i,j}=-\Delta E_{xy}/8$ and $\delta t_{i,j}=\Delta E_{xy}/16$ for the first- and second-nearest sites. By introducing $H_P$, the $d_{xy}$-orbital level around $(\pi,\pi)$ is lifted by $\Delta E_{xy}$.
To justify the validity of $H_P$, we perform the band calculation by using the WIEN2k code. We use the crystal structure of FeSe under pressure reported in Ref. [@MillicanS] for $0\sim0.6$GPa, which are shown in Fig. \[fig:AP1\] (a). We perform the first principles calculation using the WIEN2k code for $P=0$GPa and 0.6GPa, and find that $E_{xy}$-level at $\Gamma$-point is lifted by 13meV, whereas $E_{xz(yz)}$-level at $\Gamma$-point is lowered by 9meV by applying 0.6GPa. That is, the difference $E_{xy}-E_{xz(yz)}$ increases by 25meV by applying 0.6GPa. This result is consistent with the theoretical report in Ref. [@Kuroki-zS] that the top of the $d_{xy}$-orbital hole-pocket is lifted as increasing $z_{\rm Se}$.
![ (color online) (a) Crystal structure parameters $a$, $b$ and $z_{\rm Se}$ for $P=0\sim0.6$GPa reported by Millican [*et al*]{} [@MillicanS]. The Crystal structure parameters at ambient pressure reported in Ref. [@BohmerS] are also shown. (b) Dispersions on the $k_x$-$k_y$ plane in the three-dimensional WIEN2k bandstructure for $P=0$ and $P=4$GPa. The characteristic change in the bandstructure is reproduced by introducing $\Delta E_{xy}$, as shown in Fig. 1 (a) in the main text. []{data-label="fig:AP1"}](figs1.eps){width=".9\linewidth"}
According to the recent study for FeSe under high pressure [@MizukamiS], the lattice constants $a$ and $c$ linearly decrease in proportion to $P$ for $P<4$GPa. Here, we perform the band calculations for $P=0$GPa and $4$GPa, by assuming the linear extrapolations of $a$, $b$ and $z_{\rm Se}$ from Fig. \[fig:AP1\] (a). The disperdion on the $k_z=0$ plane in the three-dimensional WIEN2k bandstructure are shown in Fig. \[fig:AP1\] (b). It is predicted that $E_{xy}-E_{xz(yz)}$ increases by 170meV by applying 4GPa. Similar orbital-dependent change in the bandstructure is obtained even if we drop the inter-layer hoppings, as we show in Fig. \[fig:AP5\] (b). Therefore, it is natural to expect that the top of the $d_{xy}$-hole pocket, which is about $40$meV below the Fermi level at 0GPa [@FeSe-ARPES6S], is shifted to above the Fermi level under pressure. Thus, the simplified pressure term $H_P$ in the main text is justified.
In the main text, we analyze the total Hamiltonian $H=H_0+H_P+rH_U$, where $H_0$ is the kinetic term at $P=0$ introduced in Ref. [@Yamakawa-FeSeS], and $H_U$ is the screened Coulomb interaction due to the valence-bands obtained by the constraint-RPA (cRPA) method [@MiyakeS]. We neglect the pressure dependence of $H_U$ since the screening would be insensitive to the pressure-induced change in the valence-band structure. Here, we introduce the reduction factor $r$ and the orbital-dependent renormalization factor $z_l$ ($l=1\sim5$). In the case of $z_l=z$ for all $l$, the Stoner factors for the parameters $(r,T)$ at $z=1$ are equal to those for $(r/z,zT)$ at $z<1$. [@Yamakawa-FeSeS].
In this study, we put $z_4 < 1$ and $z_l=1$ ($l\ne4$), which is consistent with the experimental relation by ARPES. In the present numerical study at $T\le30$ meV, we put $1/z_4=1.2$ in order to reproduce the moderate spin fluctuations at $\q=(\pi,0)$ in FeSe at ambient pressure. (When $z_4=1$, $\chi^s(\q)$ has the maximum at $\q=(\pi,\pi)$ due to the nesting between eFS1 and eFS2 on $xy$-orbital.) At $T=50$ meV, $1/z_4$ should be larger than $1.5$ to satisfy the relation $\chi^s(\pi,0)>\chi^s(\pi,\pi)$ [@Onari-FeSeS]. The reason is that the hFS1,2 shrink at $T=50$ meV due to the temperature-induced shift in the chemical potential, so $\chi^s_{yz}(\pi,0)$ is relatively reduced. Note that the numerical results for $1/z_4=1.2$ at $T= 30$ meV are essentially similar to the results for $1/z_4=1.6$ at $T= 50$ meV.
B: Full self-consistent vertex correction analysis
--------------------------------------------------
In the main text, we analyzed the eight-orbital Hubbard model for FeSe by using the SC-VC theory. In this theory, the VC for the spin and charge irreducible susceptibilities, ${\hat X}^s(q)$ and ${\hat X}^c(q)$, are calculated self-consistently. For ${\hat X}^{s,c}(q)$, both the Maki-Thompson-type VC and the AL-type VC are analyzed [@Onari-SCVC-S; @Yamakawa-FeSeS; @Onari-FeSeS]. In the main text, we drop ${\hat X}^s(q)$ and calculate only the charge-channel VC $X_{l,l';m,m'}^c(q)$ for $l,l',m,m'=2\sim4$ except for $l=l'=m=m'=4$, by following the simplification introduced in Ref. [@Onari-SCVC-S]. This simplification is justified for Fe-based superconductors as verified in our previous studies [@Yamakawa-FeSeS; @Onari-FeSeS],
![ (color online) Stoner factors ($\a_S$, $\a_C$) and gap equation eigenvalue for $V^{\rm tot}$ ($\lambda^{\rm tot}$) as functions of $\Delta E_{xy}$ given by the full SC-VC analysis. []{data-label="fig:AP2"}](figs2.eps){width=".9\linewidth"}
Here, we perform the time-consuming full self-consistent calculation for both ${\hat X}^{s}(q)$ and ${\hat X}^c(q)$ for all $d$-orbitals, in order to verify the validity of the numerical study in the main text. To reflect the increase of the bandwidth under pressure, we put $r=r_0-a\cdot \Delta E_{xy}$, where $r_0=0.287$ and $a=0.19$. Then, $r=0.287 \ (0.268)$ for $\Delta E_{xy}=0 \ (0.1)$ eV. In Fig. \[fig:AP2\], we show the obtained Stoner factors $\a_S$ and $\a_C$ for $z_4^{-1}=1.4$ at $T=30$meV. For $\Delta E_{xy}=0$, the obtained relation $\a_C>\a_S$ means that the orbital order is realized. When $\Delta E_{xy}$ increases to 0.05eV, both $\a_C$ and $\a_S$ decreases and the relation $\a_S>\a_C$ holds. For $\Delta E_{xy}>0.05$eV, both $\a_C$ and $\a_S$ starts to increase, meaning that $T_m$ increases with the pressure, consistently with the phase diagram in FeSe under pressure. In addition, we show the eigenvalue of the $s$-wave gap equation $\lambda_{\rm tot}$ obtained for $V^{\rm tot}$. The increment of $\lambda_{\rm tot}$ for $\Delta E_{xy}>0.08$eV is consistent with the increment of $T_{\rm c}$ under pressure observed experimentally. In the present calculation, the intra-pocket attractive interaction is large whereas the interaction between hole- and electron-pockets is small because of the cancellation between attractive and repulsive interactions. For this reason, we find that both the $s_{++}$-wave and $s_\pm$-wave states can appear depending of the model parameters, and the impurity-induced $s_\pm \rightarrow s_{++}$ crossover is easily realized. Note that $\lambda^{\rm tot}$ in Fig. \[fig:AP2\] is smaller than unity, meaning that $\lambda^{\rm tot}$ given in the main text is overestimated due to the simplified analysis.
C: Gap functions for larger orbital polarization
------------------------------------------------
![ (color online) (a) FSs and bandstructure for $\delta E_{\rm nem}(\Gamma)= -30$meV and $\delta E_{\rm nem}({\rm X})= 120$meV. Here, $\Delta E_{xy}=0$. (b) $\chi^s_{xz}(\q)$, $\chi^s_{yz}(\q)$, $\chi^c_{xz}(\q)$, and $\chi^c_{yz}(\q)$ obtained by the SC-VC theory. (c) Gap functions obtained for $V^{\rm tot}$ ($\lambda^{\rm tot}=0.66$). []{data-label="fig:AP3"}](figs3.eps){width=".99\linewidth"}
![ (color online) (a) FSs and bandstructure for $\delta E_{\rm nem}(\Gamma)= -30$meV and $\delta E_{\rm nem}({\rm X})= 240$meV. (b) $\chi^s_{xz}(\q)$, $\chi^s_{yz}(\q)$, $\chi^c_{xz}(\q)$, and $\chi^c_{yz}(\q)$ obtained by the SC-VC theory. (c) Gap functions obtained for $V^{\rm tot}$ ($\lambda^{\rm tot}=0.79$). []{data-label="fig:AP4"}](figs4.eps){width=".99\linewidth"}
In the main text, we studied the gap function in FeSe for $\delta E_{\rm nem}({\rm X})= 60$meV, in which the horizontally-long electron-like FS is realized around the X-point. Here, we study the gap function for $\delta E_{\rm nem}({\rm X})>100$meV, in which the electron-like FS is divided into two Dirac cones. We note that the experimentally observed orbital splitting is $z\cdot \delta E_{\rm nem}$, where $z\ (\gtrsim3)$ is the renormalization factor for $xz,yz$-orbitals [@DMFT1S].
First, we study the case $\delta E_{\rm nem}(\Gamma)= -30$meV and $\delta E_{\rm nem}({\rm X})= 120$meV. The FSs and bandstructure are shown in Fig. \[fig:AP3\] (a): The Dirac cones around X-point are electron-like. The susceptibilities obtained by the SC-VC theory for $r=0.266$ at $T=30$ meV \[$(\a_S,\a_C)=(0.953,0.802)$\] are shown in Fig. \[fig:AP3\] (b). Here, we dropped the VC for the spin susceptibility as we did in the main text. Figure \[fig:AP3\] (c) shows the $s$-wave gap function for $V^{\rm SC}=V^{\rm tot}$. Here, the gap is large on the FSs with large $\dyz$-orbital component, since large intra-pocket and inter-pocket attractive force are caused by large $\chi^c_{yz}(\q)$ at $\q\sim{\bm 0}$ and $\q\sim(\pi,0)$. (We find that $V^{\rm cross}$ give weak repulsive interaction in this case.) The position of the nodal part on hole-FS is consistent with the experimental report [@Feng-FeSe-gap-S].
Next, we study the case $\delta E_{\rm nem}(\Gamma)= -30$meV and $\delta E_{\rm nem}({\rm X})= 240$meV. The Dirac cones around X-point in Fig. \[fig:AP4\] (a) are hole-like. The susceptibilities for $r=0.266$ at $T=30$ meV \[$(\a_S,\a_C)=(0.963,0.842)$\] are shown in Fig. \[fig:AP4\] (b). Figure \[fig:AP4\] (c) shows the $s$-wave gap function for $V^{\rm SC}=V^{\rm tot}$. In both cases, the gap function on the Dirac cone takes the largest value, which is consistent with the recent experimental report [@BEC-S].
D: Numerical study based on $H_P^{\rm 1st}$ given by the first-principles study
-------------------------------------------------------------------------------
Here, we derive the pressure effect Hamiltonian $H_P^{\rm 1st}\equiv H^{\rm 1st}(P)-H^{\rm 1st}(0)$, where $H^{\rm 1st}(P)$ is the first principles tight-binding model for FeSe under $P$ GPa in the tetragonal phase given by the WIEN2k software. The inter-layer hopping integrals are dropped to obtain the two-dimensional model. Hereafter, we study the total Hamiltonian $H=H_0+H^{\rm 1st}(P)+ rH_U$. Here, we introduce $H_0$ by shifting the orbital level $\delta E_l ( \Gamma ), \delta E_l ({\rm X}), \delta E_l ({\rm Y}),
\delta E_l ({\rm M})$ as $(-0.24, -0.38, 0.12, 0.5)$ for $l=2$, $(-0.24, 0.12, -0.38, 0.5)$ for $l=3$, and $(0.5, -0.24, -0.24, 0.2)$ for $l=4$ in unit eV. Here, ${\rm M}=(\pi,\pi)$. Figure \[fig:AP5\] (a) shows the bandstructure for $P=0$, $2$, and $4$ GPa given by the eigenvalues of $H_0+H^{\rm 1st}(P)$. The FSs are sensitively modified by the following orbital levels given by $H^{\rm 1st}(P)$: $E_2(\Gamma)=E_3(\Gamma)$, $E_2({\rm Y})=E_3({\rm X})$, $E_4({\rm X})$, and $E_4({\rm M})$ These $P$-dependences are shown in Fig. \[fig:AP5\] (b). Among them, $E_4({\rm M})$ shows the most prominent $P$-dependence. The obtained FSs for $P=0$ and $2$ GPa are shown in Figs. \[fig:AP5\] (c) and (d), respectively. Since the top of the $xy$-orbital hole-band at $P=0$ is $-40$ meV in the present model, the hFS3 appears at $P=2$GPa.
![ (color online) (a) Band structure of the two-dimensional FeSe model $H_0+H_P^{\rm 1st}$ at $P=0$, $2$, and $4$ GPa. (b) $E_3(\Gamma)$, $E_3({\rm X})$, $E_4({\rm X})$ and $E_4({\rm M})$ given by $H_P^{\rm 1st}(P)$. (c) FSs at $P=0$ GPa. (d) FSs at $P=2$ GPa: The hFS3 is the $\dxy$-orbital hole-pocket. []{data-label="fig:AP5"}](figs5.eps){width=".99\linewidth"}
Figure \[fig:AP6\] (a) shows the Stoner factors $\a_{S,C}$ obtained by the SC-VC theory for $r=0.263$ at $T=20$meV. At ambient pressure, $\a_C$ is much larger than $\a_S$. For $P\lesssim2$GPa, $\a_C$ decreases in proportion to $P$, reflecting the decrease of $X_{yz}^c({\bm 0})\sim \chi^s_{yz}(\Q)$. In this model, the minimum of $\a_C$ is 0.86 at $P=2.5$ GPa, which becomes much smaller than that in Fig. 2 (e) in the main text ($\a_C=0.925$ at $\Delta E_{xy}=0.1$eV). For $P\gtrsim2$GPa, the hFS3 appears, and then $\a_S$ starts to increase due to the nesting between hFS3 and eFS1,2. In addition, $\a_C$ also starts to increase for $P>2.5$ GPa since $\chi^s_{xy}(\q)$ also contribute to the “AL-term on $yz$-orbital due to $|\Lambda_{3,3;4,4;4,4}^0|^2T\sum_{\q}\{\chi^s_{xy}(\q)\}^2$”. These results are essentially equivalent to Fig. \[fig:fig2\] (f) in the main text.
Figure \[fig:AP6\] (b) shows the eigenvalue of the gap equation obtained for $V^{\rm tot}$ ($\lambda^{\rm tot}$) and that for $V^{\Lambda}$ ($\lambda^{\Lambda}$). Both $V^{\rm tot}$ and $V^{\Lambda}$ includes the beyond-Migdal-Eliashberg effects discussed in the main text. For comparison, we also show $\lambda_{\rm (Migdal)}$. For $P=4$GPa, the $s_\pm$-wave state with sign reversal between electron- and hole-pockets is obtained for $V^{\Lambda}$ as shown in Fig. \[fig:AP6\] (c). In contrast, the “$s_{++}$-wave state without sign reversal” is obtained for the total pairing interaction $V^{\rm tot}$ as shown in Fig. \[fig:AP6\] (d), since $V^{\rm cross}$ gives large inter-pocket attractive interaction: These results are essentially equivalent to Figs. \[fig:fig3\] (f) and (g) in the main text.
![ (color online) (a) Stoner factors $\a_{S,C}$ for $P=0\sim4$ GPa, at $T=20$meV and $r=0.263$. (b) Eigenvalues $\lambda^{\Lambda}$ and $\lambda^{\rm tot}$ as functions of $P$. For comparison, we also show $\lambda_{\rm (Migdal)}$. (c) $s_\pm$-wave gap given by $V^{\Lambda}$ at $P=4$ GPa. (d) $s_{++}$-wave gap given by $V^{\rm tot}$ at $P=4$ GPa. Numerical study is performed in the nonmagnetic state. []{data-label="fig:AP6"}](figs6.eps){width=".99\linewidth"}
The low-energy bandstructure of $H_0$ introduced in this section is very similar to that in the main text near the Fermi level ($|\mu-\e|<0.5$ eV). (Note that the bandstructure away from the Fermi level is unknown because of the broadening of the ARPES spectra.) Even if $H_0$ in the main text are used, the obtained $P$-dependences of the Stoner factors in Fig. \[fig:AP6\] (a) are qualitatively unchanged.
E: Pairing interaction beyond the Migdal-Eliashberg theory
----------------------------------------------------------
In conventional Migdal-Eliashberg (ME) gap equation, the pairing interaction is given by the single-fluctuation exchange process, and the VC for the coupling constant ($U$-VC) is dropped. In the main text, we studied the gap equation by including the pairing interaction beyond the ME formalism, such as the $U$-VC and the double-fluctuation exchange processes $V^{\rm cross}(\k,\p)$. Here, we explain the analytic expressions for these beyond ME processes [@Yamakawa-eFeSeS].
The analytic expression of $V^{\Lambda}$ in Fig. 3 (b) in the main text is $$\begin{aligned}
{\hat V}^{\Lambda}(k,p)
=\frac32{\hat I}^{\Lambda,s}(k,p)
-\frac12{\hat I}^{\Lambda,c}(k,p)-{\hat U}^{0s},\end{aligned}$$ where $$\begin{aligned}
{\hat I}^{\Lambda,x}(k,p)= {\hat \Lambda}^x(k,p)
{\hat I}^{x}(k-p){\hat {\bar \Lambda}}^x(-k,-p).\end{aligned}$$ Here, $\Lambda^x$ ($x=s,c$) is the $U$-VC shown in Fig. 3 (d). From now on, we explain the analytic expressions for the $U$-VC due to the Maki-Thompson (MT) and Aslamazov-Larkin (AL) processes which were already given in Ref. [@Yamakawa-eFeSeS]. First, we explain the charge- and spin-channel MT-terms: $$\begin{aligned}
\Lambda^{{\rm MT}, c}_{l,l';m,m'} (k,k')
&=& \frac{T}{2} \sum_{p} \sum_{a,b}
\left\{
I^{c}_{b,l';a,l} (p) + 3 I^{s}_{b,l';a,l} (p)
\right\}
\nonumber \\
&&\times G_{a,m} (k+p) G_{m',b} (k'+p)
\label{eqn:UMTc} ,\end{aligned}$$
$$\begin{aligned}
\Lambda^{{\rm MT}, s}_{l,l';m,m'} (k,k')
&=& \frac{T}{2} \sum_{p} \sum_{a,b}
\left\{
I^{c}_{b,l';a,l} (p) - I^{s}_{b,l';a,l} (p)
\right\}
\nonumber \\
&&\times G_{a,m} (k+p) G_{m',b} (k'+p)
\label{eqn:UMTs} ,\end{aligned}$$
where ${\hat I}^x(q)= {\hat U}^{0x}{\hat \chi}^x(q){\hat U}^{0x}+{\hat U}^{0x}$.
Next, we explain the charge- and spin-channel AL-terms: $$\begin{aligned}
&&\Lambda^{{\rm AL}, c}_{l,l';m,m'} (k,k')
\nonumber \\
&&\quad = \frac{T}{2} \sum_{p} \sum_{a,b,c,d,e,f}
G_{a,b} (k'-p) {\Lambda^0}'_{m,m';c,d;e,f} (k-k',p)
\nonumber \\
&&\quad \times
\left\{
I^{c}_{l,a;c,d} (k-k'+p) I^{c}_{b,l';e,f} (-p) \right.
\nonumber \\
&&\quad \left. + 3 I^{s}_{l,a;c,d} (k-k'+p) I^{s}_{b,l';e,f} (-p)
\right\}
\label{eqn:UALc} ,
$$ $$\begin{aligned}
&&\Lambda^{{\rm AL}, s}_{l,l';m,m'} (k,k')
\nonumber \\
&&\quad = \frac{T}{2} \sum_{p} \sum_{a,b,c,d,e,f} G_{a,b} (k'-p)
{\Lambda^0}'_{m,m';c,d;e,f} (k-k',p)
\nonumber \\
&& \quad \times
\left\{
I^{c}_{l,a;c,d} (k-k'+p) I^{s}_{b,l';e,f} (-p)
\right.
\nonumber \\
&& \quad \left.
+ I^{s}_{l,a;c,d} (k-k'+p) I^{c}_{b,l';e,f} (-p)
\right\}
\nonumber \\
&& \quad
+ \delta\Lambda^{{\rm AL}, s}_{l,l';m,m'} (k,k')
\label{eqn:UALs} ,\end{aligned}$$ where ${\Lambda^0}'_{m,m';c,d;g,h}(q,p)\equiv
\Lambda^0_{c,h;m,g;d,m'}(q,p)+\Lambda^0_{g,d;m,c;h,m'}(q,-p-q)$. ${\hat \Lambda}^0$ is the three-point vertex in Fig. 2 (b) in the main text. The last term in Eq. (\[eqn:UALs\]) is given as $$\begin{aligned}
&&\delta \Lambda^{{\rm AL}, s}_{l,l';m,m'} (k,k')
= T \sum_{p} \sum_{a,b,c,d,e,f} G_{a,b} (k'-p)
\nonumber \\
&&\quad \times I^{s}_{l,a;c,d} (k-k'+p) I^{s}_{b,l';e,f} (-p)
{\Lambda^0}''_{m,m';c,d;e,f} (k-k',p) ,
\nonumber \\\end{aligned}$$ which is found to be very small [@Onari-SCVC-S]. Here, ${\Lambda^0}''_{m,m';c,d;g,h}(q,p)\equiv
\Lambda^0_{c,h;m,g;d,m'}(q,p)-\Lambda^0_{g,d;m,c;h,m'}(q,-p-q)$.
The ($U^0$)-linear terms in Eqs. (\[eqn:UMTc\]) and (\[eqn:UMTs\]) should be dropped to avoid the double counting of the RPA-type diagrams. We also carefully drop the double counting $(U^0)^2$-terms included in both MT and AL terms. We verified numerically that the large charge-channel $U$-VC ($|\Lambda^c|^2\gg1$) originates from the $\chi^s$-square term in Eq. (\[eqn:UALc\]). We also verified that the relation $|\Lambda^s|^2\ll1$ is realized mainly by the $(U^0)^2$-term [@Yamakawa-eFeSeS].
Next, we explain the double-fluctuation exchange process $V^{\rm cross}(k,p)$, which gives sizable pairing interaction in Fe-based superconductors [@Yamakawa-eFeSeS]. It is schematically expressed as Fig. 3 (c) in the main text. Physically, $V^{\rm cross}(k,p)$ represents the pairing interaction due to the “multi-fluctuation exchange processes”. This term is expressed as [@Yamakawa-eFeSeS] $$\begin{aligned}
&&V^{\rm cross}_{l,l';m,m'} ( k, p )
= \frac{T}{4} \sum_{q} \sum_{a,b,c,d}
G_{a,b} (p-q)
G_{c,d} (-k-q)
\nonumber \\
&& \times
\left\{
3 I'^{s}_{l,a;m,d} (k-p+q) I'^{s}_{b,l';c,m'} (-q)
\right.
\nonumber \\
&&\quad
+ 3 I'^{s}_{l,a;m,d} (k-p+q) I'^{c}_{b,l';c,m'} (-q)
\nonumber \\
&&\quad
+ 3 I'^{c}_{l,a;m,d} (k-p+q) I'^{s}_{b,l';c,m'} (-q)
\nonumber \\
&&\quad \left.
- I'^{c}_{l,a;m,d} (k-p+q) I'^{c}_{b,l';c,m'} (-q)
\right\}
\label{eqn:Vcross} ,\end{aligned}$$ where we put ${\hat I}'^x= {\hat I}^x-{\hat U}^{0x}$ to avoid the double counting of diagrams included in other terms. According to Eq. (\[eqn:Vcross\]), $V^{\rm cross}_{m,m;m,m}$ can take large negative value when the spin fluctuations develop on the $m$-orbital [@Yamakawa-eFeSeS].
[99]{}
J. N. Millican, D. Phelan, E. L. Thomas, J. B. Leão, and E. Carpenter, Solid State Commun. [**149**]{}, 707 (2009).
A. E. Böhmer, F. Hardy, F. Eilers, D. Ernst, P. Adelmann, P. Schweiss, T. Wolf, and C. Meingast, Phys. Rev. B [**87**]{}, 180505(R) (2013).
K. Kuroki, H. Usui, S. Onari, R. Arita, and H. Aoki, Phys. Rev. B [**79**]{}, 224511 (2009).
Y. Mizukami and T. Shibauchi, private communication.
Y. Suzuki, T. Shimojima, T. Sonobe, A. Nakamura, M. Sakano, H. Tsuji, J. Omachi, K. Yoshioka, M. Kuwata-Gonokami, T. Watashige, R. Kobayashi, S. Kasahara, T. Shibauchi, Y. Matsuda, Y. Yamakawa, H. Kontani, and K. Ishizaka, Phys. Rev. B [**92**]{}, 205117 (2015). Y. Yamakawa, S. Onari, and H. Kontani, Phys. Rev. X [**6**]{}, 021032 (2016).
T. Miyake, K. Nakamura, R. Arita, and M. Imada, J. Phys. Soc. Jpn. [**79**]{}, 044705 (2010).
S. Onari, Y. Yamakawa, and H. Kontani, Phys. Rev. Lett. [**116**]{}, 227001 (2016).
S. Onari and H. Kontani, Phys. Rev. Lett. [**109**]{}, 137001 (2012). Z. P. Yin, K. Haule, and G. Kotliar, Nat. Mater. [**10**]{}, 932 (2011).
H. C. Xu, X. H. Niu, D. F. Xu, J. Jiang, Q. Yao, M. Abdel-Hafiez, D. A. Chareev, A. N. Vasiliev, R. Peng, and D. L. Feng, Phys. Rev. Lett. [**117**]{}, 157003 (2016). S. Kasahara, T. Yamashita, A. Shi, R. Kobayashi, Y. Shimoyama, T. Watashige, K. Ishida, T. Terashima, T. Wolf, F. Hardy, C. Meingast, H. v Löhneysen, A. Levchenko, T. Shibauchi, and Y. Matsuda, arXiv:1608.01829.
Y. Yamakawa, and H. Kontani, arXiv:1611.05375.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Understanding the causes and effects of spatial patterns of vegetation is a fundamental problem in ecology, especially because these can be used as early predictors of catastrophic shifts such as desertification processes. Empirical studies of the vegetation cover of drylands and semiarid regions have revealed the existence of vegetation patches of broadly diverse sizes, such that the probability distribution of their sizes can be fitted by a power law, i.e. vegetation patches are approximately scale free. Different explanatory mechanisms, such as plant-plant interactions and plant-water feedback loops have been proposed to justify such a finding, yet a full understanding has not been reached. Using a simple individual-based model for vegetation dynamics, we show that environmental variability –a well-recognized characteristic feature of semiarid environments– promotes the emergence of power-law vegetation patches in a robust way (i.e. for a broad range of parameter values). Furthermore, these are related to a percolation phenomenon that occurs in an intermittent or fluctuating way. The model also reveals that the emerging power-law exponents depend on the overall vegetation-cover density exhibiting the same trend as empirical data and, remarkably, the fitted exponents agree reasonably well with the ones measured in well-known empirical studies. We conclude that environmental variability plays a key role in the formation of such vegetation patterns. From a practical viewpoint, this may be of importance to predict the effects that changes in environmental conditions may have in real ecosystems. From a theoretical side our study shed new light on intermittent percolation phenomena occurring under fluctuating conditions.'
author:
- |
Paula Villa Mart[í]{}n\
Okinawa Institute of Science and Technology Graduate University,\
Biological Complexity Unit, Onna, Okinawa 904-0495, Japan\
Virginia Dom[í]{}nguez-Garc[í]{}a\
ISEM, CNRS, Univ Montpellier, EPHE, IRD, 34095 Montpellier, France\
Miguel A. Muñoz\
Departamento de Electromagnetismo y F[í]{}sica de la Materia\
and Instituto Carlos I de F[í]{}sica Te[ó]{}rica y Computacional,\
Universidad de Granada, Granada, Spain\
`[email protected]`\
title: 'Intermittent percolation and the scale-free distribution of vegetation clusters'
---
Introduction
============
Living organisms are typically distributed in space neither uniformly nor randomly, but forming structured patterns of aggregation. Understanding the causes and effects of such spatial aggregation is one of the most fundamental problems in ecology [@Levin; @Durrett-Levin; @Legendre; @Sole; @Sole+Bascompte]. An archetypical example is found in arid and semiarid environments, which are characterized by the presence of vegetation patches forming either regular or irregular patterns even in seemingly spatially homogeneous environments. [@Regular; @semiarid1; @semiarid2; @semiarid3; @Lejeune]. A large number of hypotheses have been proposed to explain the origin of such patterns. These include, among others, the interplay between competition and facilitation mechanisms in plant-plant interactions and plant-water feedbacks [@Scanlon2007; @Kefi2007; @Manor-Shnerb-2008a; @Manor-Shnerb-2008b; @Manor-Shnerb-2009], but, still, we are far from reaching a full understanding. Importantly, variations in the shape of vegetation patterns as well as in the density cover have been proposed as early-warning indicators of catastrophic shifts, such as desertification [@Scheffer2001; @Rietkerk2004; @Kefi2007; @Kefi2011; @Kefi-Nature2017]. Thus, understanding, characterizing, and eventually controlling vegetation patterns is a problem of outmost theoretical and practical relevance.
While some of the patterns formed by vegetation, such as stripes, have a characteristic size scale, ground-breaking empirical studies showed a decade ago that vegetation aggregates in semi-arid environments lack a well-defined characteristic scale (i.e. vegetation patches of greatly diverse sizes can be found together) [@Scanlon2007; @Kefi2007]. More specifically, in this latter situation patch-size distributions can be fitted by power-laws, the hallmark of scale-free systems [@Newman-powerlaws; @Mitzenmacher; @Sornette-book; @RMP]. From a theoretical perspective, scale-free cluster-size distributions are often the fingerprint of critical points [@Binney; @Henkel-book; @Stauffer-Aharony; @Christensen-book; @GG]. A classical and archetypical example of a critical point appears in the problem of standard, also called “isotropic”, percolation.
In the simplest percolation setup there are two possible states for the sites of a given lattice (which represents the physical space): empty or occupied. In the percolating phase the largest aggregate of occupied sites spans the whole system; on the other hand, in the non-percolating phase only localized finite clusters exist. In between these two phases –right at the percolation threshold or critical point– clusters of all possible sizes –from tiny to system-size-spanning ones– emerge, generating a power-law (scale-free) distribution of cluster sizes with well-known universal exponent values and characteristic scaling features [@Stauffer-Aharony; @Christensen-book].
If the ecological processes giving rise to patterned vegetation patches in semiarid ecosystems would follow dynamical rules operating close to a percolation phase transition [@Roy2003; @Pascual2005], scale invariance should be universal; in other words, there should emerge robust (detail-independent) power-law exponents [@Binney; @Henkel-book]. However, contrary to this expectation, the empirically found distributions of patch sizes can be fitted by power laws whose exponents depend on the level of vegetation coverage in the considered area [@Scanlon2007; @Kefi2007; @Kefi-Nature2017]. This variability in scaling exponents suggests a lack of universality and sheds doubts on the –maybe exceedingly naive– alleged connection with standard/isotropic percolation and universality, calling for more elaborated theoretical approaches to understand the dynamics and patterns of these ecological communities.
Diverse relatively-simple theoretical models have been proposed to account for the empirical finding of scale-free vegetation patches [@Scanlon2007; @Kefi2007; @Manor-Shnerb-2008a; @Manor-Shnerb-2008b; @Manor-Shnerb-2009]. Some of these include three possible states to describe the system: occupied, empty, and disturbed/degraded (e.g. unavailable to be colonized by vegetation) sites [@Kefi2007; @Roy2003]. This is in contrast with the case of standard/isotropic percolation models as described above [@Stauffer-Aharony; @Christensen-book]. In such three-state models, an external control parameter (which can be interpreted as the grazing pressure, the yearly rainfall, etc.) regulates the level of vegetation cover under stationary conditions and as it is varied –e.g. to mimic harsher conditions– a phase transition from the vegetated to the deserted state appears [@Roy2003; @Kefi2007; @Kefi2011]. Remarkably, it was found –relying mostly on computational analyses– that cluster size distributions resembling power-laws appear in some of these models for a wide range of external conditions (control parameter values) rather than just around a fixed percolation or critical threshold [@Roy2003; @Pascual2005]. Thereby, the influential concept of “robust criticality” was introduced by Pascual and colleagues to describe this phenomenon [@Roy2003; @Pascual2005]. Some other models with only two states (occupied and empty) and some type of facilitation mechanism –as, for instance, a positive feedback between vegetation and water supplies– were claimed to support the idea of robust criticality [@Scanlon2007; @Manor-Shnerb-2008b]. From a theoretical side it is still not clear where such a generic or robust scaling comes from, nor if it is just a transient (not asymptotic) effect. In fact, van den Berg et al. shed some doubts on the theoretical foundations of the concept by demonstrating analytically that in some simple version of these models scaling occurs at a single point, i.e. there is an underlying sharp transition and not a broad range of criticality [@van2011; @van2015].
The aim of the present work is therefore to scrutinize whether some type of generic scaling –occurring in broad regions of parameter space–can actually emerge in very simple dynamical models (such as the contact process [@Henkel-book]) as well as to uncover the basic mechanisms promoting the emergence of scale invariance in the vegetation-patch sizes within broad parameter ranges.
To this end, we introduce a simple spatially-explicit individual-based model for vegetation dynamics and analyze its behaviour. In particular, we examine if the emergence of scale-free vegetation patches requires a fine tuning of the control parameter (i.e. if it behaves as a standard percolation-like transition) or if, on the contrary, scale-free size distributions emerge in a robust way for a broad range of parameter values. We show that implementing temporal environmental variability in the model –an ingredient inherent to actual ecological systems and of special relevance in semi-arid environments– induces the emergence of scale-free patch-size distributions in a wide region in parameter space , while such a region shrinks to a single point in the absence of temporal variability. Thus, generic scaling in this case emerges as a sort of intermittent or fluctuating percolation phenomenon, in which clusters nearby a percolation threshold fluctuate in size owing to external variability. Remarkably, we show that our model –even if with some limitations induced by its simplicity– is able to reproduce with reasonable accuracy the values and overall dependence on vegetation density of some empirically observed non-universal scaling exponents.
Model building
==============
The parsimonious individual-based model that we analyze is defined on a two-dimensional square lattice, with each lattice site being in one of two possible states: empty ($0$) or occupied ($1$). At each discrete time step an occupied site is randomly selected and, with probability $p$, a new offspring is created at a random nearest-neighbor provided it was empty [@Marro; @Henkel-book]. Alternatively, with probability $1-p$ the selected occupied site is emptied; i.e. an individual dies. The time counter is increased as $t \rightarrow t + \Delta t$ with $\Delta t =1/N(t)$, where $N(t)$ is the total number of occupied sites at time $t$. This way, we ensure that in a (Monte Carlo) step all are updated once on average. This stochastic rules are nothing but the “contact process”, a prototypical model that captures the essence of a large variety of spreading systems in physics, biology, epidemiology, [@Marro; @Henkel-book; @GG] as well as in spatial ecology [@Durrett-Levin; @Neuhauser; @Durrett2009; @Paula; @Paula-Q; @Bonachela-Munoz-Levin]. It can be summarized as the following set of [processes]{}: $$\arraycolsep=1.6pt\def\arraystretch{1}
\begin{array}{cccc}
A &\xrightarrow{p} & 2A & (birth), \\
A &\xrightarrow{1-p} & 0 & (death)
\end{array}
\label{test}$$ where $A$ indicates “particles” or “plants” in our interpretation.
In the special case of a well-mixed population in which all sites are taken to be neighbors of all others and assuming an infinitely large system (i.e. in the mean-field case) the dynamics of the overall density of occupied sites $\rho(t)=1/N \sum_i^N s_i $ obeys [@Binney; @Henkel-book; @GG]: $$\dot{\rho}(t)=p \rho(t)(1- \rho(t))- (1-p) \rho(t)= (2p-1) \rho(t)- p\rho^2(t)
\label{MF}$$ in some unspecified time units[^1]. This equation exhibits a transcritical bifurcation at the reproduction probability $p_c=1/2$. This is a *directed-percolation* critical point [@Marro; @Henkel-book; @Odor; @GG], above which the population survives for arbitrarily long time in sufficiently large lattices reaching a stationary density $\rho_s=(2p-1)/p$. Below such a value, the only stable solution is the empty state $\rho_s=0$. Thus the critical point separates an empty (or “absorbing”) phase from an occupied (or “active”) phase.
This type of exact solution cannot be straightforwardly extended to analytically solve the case of a two-dimensional lattice in simple terms; however, it is well-known that the overall phenomenology is qualitatively (though not quantitatively) preserved [@Marro; @Henkel-book]. This is illustrated in Figure \[phase\_diagram\] which shows the phase diagram for the two-dimensional contact process resulting from computer simulations. Observe, in particular, that the directed-percolation critical point –separating the empty from the active phase– is shifted to $p_c \approx 0.6225$ owing to fluctuations, i.e. to the finite size of local neighborhoods [@Marro; @Henkel-book; @GG].
![[**Phase diagram of the contact process**]{} in its “pure" (gray line, $\sigma=0$) and temporally heterogeneous (light orange line, $\sigma=0.25$) versions. For the pure contact process, at the directed-percolation critical point $p_c \approx 0.6225$ (separating empty from populated states), the stationary density is $\rho=0$, and it is not possible to measure any non-trivial steady-state spatial cluster distribution. Observe that within the active phase –in which activity propagates, reverberating indefinitely across time– clusters of adjacent active sites may [**(i)**]{} not percolate in space (as occurs, e.g. for $p=0.65$, which corresponds to $\rho\simeq0.26$; see lower inset) or [**(ii)**]{} do so, deeper into the active phase (as occurs for e.g. $p=0.725$, corresponding to $\rho\simeq0.56$ for large system sizes; see upper inset). In the insets we have highlighted the largest cluster (dark orange color) in snapshots for the above two cases, illustrating the absence of percolation for $\rho\simeq0.26$, and the existence of a percolating cluster for $\rho\simeq 0.56$. For the temporally heterogeneous case, the displayed density is the average over time for an average reproduction probability $\bar{p}\in[\sigma,1-\sigma]$, such that $p(t)=\bar{p}+ \sigma$ lay in $[0,1]$. In this case, the absorbing-to-active critical point is shifted to a value ($p_c \approx 0.6350$) larger than that of the pure model.[]{data-label="phase_diagram"}](FIGURE1.pdf){width="75.00000%"}
Most of the literature on the contact process is focused in analyzing the non-trivial features of this absorbing-active phase transition [@Marro; @Henkel-book; @Odor; @GG]. However, here we are interested in another “hidden” critical point whose very existence is usually neglected. This –as illustrated schematically in Figure 1– is a percolation-like transition at which a cluster of occupied sites percolates in space [@van2011; @van2015]. More specifically, this second transition is –as we shall show– a standard/isotropic *percolation* phase transition [@Stauffer-Aharony; @Christensen-book] above which, at any given time within the steady state, the probability to find a percolating cluster of adjacent occupied sites spanning the whole two-dimensional space is larger than zero. In two-dimensions, one needs to reach stationary activity densities of the order of $\rho \approx 0.56$ (deep inside the active phase, i.e. $p\approx 0.725 >p_c$) to observe the emergence of a spatially percolating cluster (Fig. \[phase\_diagram\]). Let us stress, even at the risk of being redundant, that this transition at which clusters *percolate in space* is conceptually different from the absorbing-to-active directed-percolation transition, at which clusters of occupied sites can *percolate in time*.
We first study the nature of the percolation transition within the active phase of the “pure” contact process. Although our main focus will be on a variant of the contact process aimed at modelling environmental variability, which we proceed now to describe.
In the *temporally-disordered contact process*, at each Monte Carlo timestep the reproduction probability, $p(t)$, takes a uniformy-distributed random value in the interval $[\bar{p}+\sigma,\bar{p}-\sigma]$. The values of $\bar{p}$ and $\sigma$ are kept constant in time in each simulation, but $\bar{p}$ can change for different realizations to explore diverse regions of parameter space, and in particular to sample different densities. $\bar{p}$ is taken within the interval $[\sigma,1-\sigma]$ so that $p(t) \in [0,1]$. The width $\sigma$ is kept as a control parameter characterizing the environment variability, e.g. the year-to-year demographic (i.e. birth and death) probabilities[^2]. In this temporally-disordered version of the model (fixing $\sigma=0.25$), the absorbing-to-active critical point is located at $p_c\approx 0.6350$, a value slightly larger than the corresponding value for the homogeneous (or “pure”) case $p_c\approx 0.6225$ (see Fig.\[phase\_diagram\]) [@TGP].
Let us stress that in order to properly compare our computational results for either of the considered models with field observations for vegetation cover at a given vegetation density $\rho$ [@Scanlon2007; @Kefi2007], we need to consider the “$constant-\rho$ ensemble”. This means that statistics should be collected not for a constant birth probability $p$ (which is difficult to measure empirically) but for constant density $\rho$ (a more accesible quantity). Although in the infinitely-large system-size limit, “constant-$p$ ensemble” and “$constant-\rho$ ensemble” are expected to be equivalent [@equivalence; @Henkel-book], in finite systems –as there is not a one-to-one map between $p$ and $\rho$– results can be (and actually are) different in both ensembles.
In practical terms, to obtain results for a given density $\rho$, we run computer simulations for different values of $p$ and for long-enough times as to reach the stationary state (typically $t>10^3$). Unless specifically mentioned, we consider a system size $N=1024\times1024$ which is enough to collect good statistics at a reasonable computational cost. Then, we take a “snapshot” of the final configuration and collect the statistics of patches sizes conditioned to the value of its overall density $\rho$ (using a binning of size $\Delta \rho=0.001$). As the output, we determine the path-size distribution for each possible density value.
Results
=======
The (isotropic) percolation transition in the “pure” contact process {#the-isotropic-percolation-transition-in-the-pure-contact-process .unnumbered}
--------------------------------------------------------------------
First, we analyze the percolation transition of the homogeneous or “pure” contact process –without additional heterogeneity– and use it as a benchmark for further analyses. In two-dimensional percolation on a square lattice the percolation threshold is $p^{occ}_c\approx0.5927...$ and at such a critical point the cluster-size distribution decays as $$P(s)=C s^{\tau}
\label{tau}$$ with an exponent value $\tau=187/91 \approx 2.05$ [@Stauffer-Aharony; @Christensen-book]. Importantly, the value of $p_c$ can change if other types of latices (e.g. hexagonal) are considered, but the exponent remains unchanged, i.e. it is universal.
Our dynamical models are (obviously) not identical to the standard/isotropic percolation model. Among other aspects, the models here are dynamical and the states of nearby sites are clearly correlated. The question we address now is whether such differences are relevant, i.e. whether the universal aspects of the percolation transition are preserved or not. To provide an answer to this question we resort to computational analyses.
Fig.\[CP\_collapse-Ps\]A and B show the cluster-size distribution $P(s)$ for a wide range of occupation-densities $\rho$ of the pure contact process (low (A) and high density (B) curves are separated for the sake of visual clarity). Observe that at a critical occupation-density value $\rho_c\approx 0.56$ (and only at this value) the cluster-size distribution decays as a power-law, i.e. cluster vegetation patches of broadly different sizes emerge. Furthermore, the scaling behavior is fully compatible with the expected universal behavior of standard/isotropic percolation, fitted value for the cluster-size distribution exponent $\tau$ is in agreement with the above-mentioned value $\tau \approx 2.05$ [@Stauffer-Aharony; @Christensen-book].
![[**Standard isotropic percolation scaling within the active phase of the pure contact process.**]{} The scaling in standard (isotropic/static/uncorrelated) percolation at the critical occupation probability threshold $p^{occ}_c$ is characterized by a cluster-size distribution $P(s)$ such that $P(s) \propto s^{-\tau}$ with $\tau=187/91\approx 2.05$ and the scaling relationship $\chi=|p^{occ}-p^{occ}_c|^{-\gamma}\tilde{\cal{F}}(|p^{occ}-p^{occ}_c| L^{1/\nu}$with $\nu=4/3$ and $\gamma=43/18$ [@Stauffer-Aharony; @Christensen-book]. We show $P(s)$ for low ([**A**]{}) and high ([**B**]{}) densities (plotted separately for the sake of clarity) as obtained in computer simulations of the contact process on a square lattice, for different occupation densities $\rho$ in the range $\rho \in [0.02,0.70]$ (color coded). Panel [**C**]{} shows the curve collapse for four different system sizes, using the scaling exponents of standard/isotropic percolation. We conclude that the contact process displays clean scaling behavior for a critical density $\rho_c \approx 0.56$, fully compatible with that of the standard/isotropic percolation class. []{data-label="CP_collapse-Ps"}](FIGURE2.pdf){width="\textwidth"}
Other quantities usually employed in studies of standard/isotropic percolation can also be determined at the percolation transition of the pure contact process. In particular, right at the percolation threshold, the mean cluster size $\chi$, defined as $$\chi=\sum_{i}s_{i}^2/n_{occ},$$ where $i$ runs over all clusters, $s_i$ is the corresponding cluster size and $n_{occ}$ is the total number of occupied sites; is governed by the general scaling law $$\chi=\Delta^{-\gamma}\tilde{\cal{F}}(\Delta L^{1/\nu}),$$ where $\Delta=|p^{occ}-p^{occ}_c|$ is the distance to the critical occupation probability $p^{occ}_c$, $L$ the system linear size, and $\nu=4/3$ and $\gamma=43/18$ are exactly-known critical exponents [@Stauffer-Aharony; @Christensen-book]. This implies that there exists a master curve $\tilde{\cal{F}}(\Delta L^{1/\nu})$ in which data should collapse for different system sizes $L$ and $\Delta$ values (if the right exponent values are used for $\nu$ and $\gamma$). Figure \[CP\_collapse-Ps\]C shows such a curve collapse for the contact process at fixed density values –at the critical point $\rho_c \approx 0.56$– providing us with strong evidence that it exhibits the same universal scaling behavior as standard/isotropic percolation.
In order to rationalize this conclusion from a theoretical viewpoint one can argue as follows. At the percolation critical point of the contact process –which lies well inside the active phase– occupied sites are not randomly placed, as dynamically-generated spatial correlations exist. However, as the dynamics at such a density is far away from the absorbing-active phase transition (where long-range correlations emerge), correlations are only short-ranged. As a consequence and from a renormalization-group perspective, at a sufficiently “coarse-grained” scale, the effective density becomes uncorrelated. Thus, standard/isotropic percolation scaling emerges at sufficiently large scales. In other words, the presence of weak short-range correlations changes the location of the critical density but does not affect the large-scale behavior of percolation, in perfect agreement with what computationally observed.
Summing up, our simple pure model exhibits scale invariance in its cluster-size distribution if, and only if, it is fine tuned to have a density close to the percolation threshold. Around such a density it exhibits scaling features of standard percolation, and no trace of generic scale invariance for arbitrary density values, for which exponential behavior (with a characteristic scale) is observed.
### Constant-$\rho$ ensemble properties with temporal heterogeneity. {#constant-rho-ensemble-properties-with-temporal-heterogeneity. .unnumbered}
When temporal heterogeneity, $\sigma > 0$, is introduced, the scenario described in the previous section for the case of the pure contact process ($\sigma=0$) changes dramatically, as we show in what follows.
Unless explicitly said, all simulations are ran for $\sigma=0.25$, without loss of generality. Given the subtleties usually encountered when determining whether a given distribution follows a power law or not, let us stress that in the forthcoming analyses we employ the standard most-stringent statistical tests –following the methods introduced in [@Clauset] (see also Appendix)– to fit power laws and to decide whether they provide better fits than e.g. exponentials. In particular, for each distribution $P(s)$ at a given density $\rho$ we obtain the parameters $C, \tau$ and $\Delta s$ in Eq.(\[tau\]) that maximize the normalized log-likelihood of the fit (see Appendix). Our analyses reveal that cluster-size distributions conform to a power-law curve for at least two (and up to four) decades (see Fig. \[TCP\_collapse-Ps\]A and B, and Table \[table:fit\_averaged\]). In this case, contrary to the pure contact process, the scale-free behavior of the cluster-sizes holds within a wide range of densities, with continuously varying exponents. Simple eye inspection of Fig. \[TCP\_collapse-Ps\]A and B already reveals that the exponent $\tau$, characterizing the observed power-law decay, displays a non-monotonic variation with population density $\rho$ (see Fig. \[TCP\_collapse-Ps\]C). In particular, the fitted value decreases from $\tau\simeq2.59$ (for $\rho= 0.02$) up to $\tau\simeq1.63$ (for $\rho \approx 0.5$), that value $\tau$ increases again up to $\tau\simeq3.12$ (for $\rho \approx 0.7$) (see Fig. \[TCP\_collapse-Ps\]C and Table \[table:fit\_averaged\]).
Observe that –similarly to the standard/isotropic percolation case– the average cluster size of finite clusters (i.e. the mean of these distributions) increases monotonously as the control parameter is increased in the subcritical phase and it starts decreasing once the percolation threshold is reached and finite clusters start merging together or with the percolating giant cluster (if it exists) [@Christensen-book]. However, it is remarkable that opposed to the standard/isotropic percolation case, power laws provide a good fit for the cluster-size distribution not only at the critical point but also in a wide parameter range above and below such a value, i.e. we observe scale-free cluster sizes through a wide range of densities.
Owing to this peculiarity, the percolation threshold $\rho_c$ cannot be straightforwardly obtained –as usually done– just by looking for scale-free cluster-size distributions, since there is not a unique density where scaling is observed. Thus, we resort to obtain an estimate of the critical point given by the density for which the fit to a power law extends for the largest possible range of cluster sizes with a fit exhibiting a maximal normalized log-likelihood. This is equivalent to looking for the point with largest averaged finite-cluster size, as usually done in standard/isotropic percolation. Such an optimal value turns out to be around $\rho \approx 0.56$, roughly speaking, the same occupation density obtained for the isotropic percolation transition of the pure contact process without temporal disorder (see Table I where details of the fits are given).
\[h!\] ![[**Isotropic percolation scaling of the temporally heterogeneous contact process.**]{} In analogy with Fig.\[CP\_collapse-Ps\], here we plot cluster-size distributions (two plots for low density values $[0.02,0.5]$ ([**A**]{}) and high-density values $[0.52,0.7]$ ([**B**]{}) in increments of $0.02$ (color coded)). Observe that contrarily to the pure contact process case, power law behaviour can be fitted for a wider range of densities (at least two or more decades for almost all densities, and up to four decades around the critical point). For each density $\rho$, we obtain a power-law exponent $\tau$ such that it maximizes the normalized log-likelihood (see Appendix, [@Clauset], panel **C**) giving raise to a non-monotonic dependence. The best fit is obtained for $\rho=0.56$ with an exponent $\tau \simeq 1.81$, slightly smaller than the expected for the standard percolation $\tau=187/91 \simeq 2.05$. Taking this value as the critical percolation point $\rho^{occ}_c=0.56$ we obtain a rather good scaling collapse (panel **D**) for $\chi=\Delta^{-\gamma}\tilde{\chi}(\Delta L^{1/\nu})$ for values $\nu\simeq3.5$ and $\gamma\simeq4$, much larger than expected for standard/isotropic percolation $\nu=4/3$ and $\gamma=43/18$.[]{data-label="TCP_collapse-Ps"}](FIGURE3.pdf "fig:"){width="\textwidth"}
The exponent fitted at $\rho=0.56$ is $\tau\simeq1.81$, slightly smaller than the expected value in two-dimensional percolation $\tau\approx 2.05 $ (see Table \[table:fit\_averaged\]). Moreover, considering $\rho=0.56$ as the percolation critical point, we obtain a scaling collapse (see Fig. \[TCP\_collapse-Ps\]D) similar to that of the homogeneous case, represented in Fig.\[CP\_collapse-Ps\]C but with exponents $\nu\simeq3.5$ and $\gamma\simeq4$ that are considerably larger than the percolation values ($\nu=4/3$ and $\gamma=43/18$), suggesting a different type of critical behavior (or universality class). A more precise determination of the critical point and the associated critical exponents is beyond the scope of the present manuscript and will be reported elsewhere.
### Individual snapshot properties with temporal heterogeneity. {#individual-snapshot-properties-with-temporal-heterogeneity. .unnumbered}
We have just shown that our model equipped with temporal disorder is able to reproduce scale-free cluster-size distributions without the need of careful fine tuning to a threshold point, i.e. clusters of largely different scales emerge in a rather robust way for a broad range of parameter values. Nevertheless, in order to compare with empirical observations, an important caveat remains. The results obtained so far correspond to ensemble averages where many different realizations of the system are considered and averages are performed over many different times in the steady state, with the only restriction that the value of the control parameter $\bar{p}$ is kept constant. On the other hand, empirically measured exponents of vegetation-patch-size distributions correspond to single snapshots of the (unknown) ecological dynamics.
To bridge this gap, we also computed the statistics of individual (or “instantaneous”) snapshots within our model. In other words, we performed a set of simulations of the model for diverse values of $\bar{p}$ and took instantaneous pictures of the system configuration in the steady state. For each of such snapshots we compute its overall density and measure the cluster-size distribution (histogram). Obviously, these distributions –not being ensemble averaged– are much noisier than those reported above. Eye inspection of Figure \[Snapshot1\] already reveals that the cluster-size distribution of an individual snapshot –even for a density far away from the percolation threshold, $\rho=0.3$ in this case– is broadly distributed and can be fitted as a power-law (straight line in the double-logarithmic plot).
\[h!\] ![[**Clusters and cluster-size distribution in the temporally-varying system for density $\rho=0.3$.**]{} ([**A**]{}) We show a snapshot of the system (with size $256\times256$) at some given time at which the density happens to take the value $\rho=0.3$; the colors are correlated with the cluster size: from red for the smaller ones to blue for the largest ones. ([**B**]{}) This panel shows the cluster-size distribution as obtained from panel ([**A**]{}); the distribution is obviously noisy given the relatively small system size, but it is clearly a broad one, and can be fitted as a power law (black dashed line).[]{data-label="Snapshot1"}](FIGURE4.pdf "fig:"){width="\textwidth"}
A collection of individual-snapshot cluster-size distributions together with the averaged distribution are shown in Fig.\[Snapshots2\] for different densities. Observe that in some cases the averaged distribution extends to sizes larger than those of individual snapshots. This is just because, for the sake of clarity, not all individual snapshots in the average are plotted. This figure illustrates that even for a fixed density $\rho$ there is a large variability in the histograms of cluster-sizes and that single snapshots may differ notably from the ensemble-averaged (constant-$\rho$) distribution. In particular, fast exponential-like and slow power-law-like decays can be identified for individual snapshots. Although exponential decays are seen, especially for low densities, power-laws are present for snapshots of all density values; however, they are more prominent and broad in the neighborhood of the percolation transition (see Table II where details of the fits are reported).
\[h!\] ![[**Cluster size distributions for individual snapshots of the model with temporal variability.**]{} Cluster-size distributions are shown –together with their corresponding averaged curves (dark gray curve)– for snapshots with diverse density values not far for the percolation threshold: ([**A**]{}) $\rho=0.54$, ([**B**]{}) $\rho=0.56$, and ([**C**]{}) $\rho=0.58$, and ([**D**]{}) $\rho=0.66$ respectively. Due to the high variability of cluster size distributions, individual realizations –rather than ensemble averages– are more adequate to characterize snapshots of real systems. Observe that approximate power-law distributions appear for snapshots of very different densities.[]{data-label="Snapshots2"}](FIGURE5.pdf "fig:"){width="\textwidth"}
In this case of histograms for individual snapshots, we also compared the goodness of power-law fits with that of exponential-distribution fits $P(s)=Ce^{-\alpha s}$. For most of the snapshots power-law distributions fit better the data than exponential ones, i.e. power-laws give a higher normalized log-likelihood in the majority of the cases, but not always, especially for low densities $\rho\in[0.04,0.24]$ (see fourth column of Table \[table:fit\_snapshots\]). However, even for densities in this interval, power-law fits are better in at least $60\%$ of the snapshots in all cases (see Table \[table:fit\_snapshots\]). On the other hand, for intermediate densities, $\rho\in[0.26,0.68]$, power laws outperform exponential fits in almost all snapshots. Observe that in spite of the limitations imposed, the finite system size and the lack of an averaging procedure, power-laws are observed for at least $1.5$ decades in most of the cases (fifth column of Table \[table:fit\_snapshots\]).
\[h\] ![[**Fitted exponents agree with field observations.**]{} We fitted power-law regimes of cluster-size distributions $P(s)$ of snapshots for different densities $\rho$ and $\sigma=0.25$ (see Appendix and Table \[table:fit\_snapshots\] for details). We obtained the mean (gray lines) and standard deviation (shaded area) of the exponents $\tau$ obtained for up to $200$ snapshots (see sixth columns of Table \[table:fit\_snapshots\]) for density values $[0.04,0.7]$ equally spaced in increments of size $0.01$. Observe that empirical data [@Scanlon2007; @Kefi2007] (black dots) are in good agreement with the model predictions (at least, for enough environmental variability $\sigma=0.25$).[]{data-label="Empirical"}](FIGURE6.pdf "fig:"){width="70.00000%"}
The average of the exponents obtained by fitting power-law distributions for individual snapshots (obtained excluding samples that are best fitted by exponentials) exhibits a non-monotonic behavior as a function of the vegetation density (see gray line of Fig. \[Empirical\]). The mean exponent $\tau(\rho)$ reaches a minimum value $\tau\simeq1.54<\tau_c$ at a density value below the percolation threshold $\rho\simeq0.44<\rho_c$, and increases up to $\tau\simeq4$ for larger and smaller densities.
We also performed simulations for smaller values of the noise amplitude $\sigma$, for which the shape of the curves remain similar and the widths of the distributions of $\tau$-values (i.e. the errorbars in Fig.6) decrease. Actually, in the limit in which $\sigma$ goes to $0$ one should recover power-law scaling only in a vanishing region around the critical point. However, a detailed comparison between the resulting curves for diverse values of $\sigma$ is not straightforward given that –in the way we chose to parametrize– the possible values of $\bar{p}$ are limited by the choice of $\sigma$ and thus the sampling of values of $\rho$ is also affected by $\sigma$. Disentangling these effects from the final results is therefore not straightforward in the present setting; thus, we do not pursue further such a goal here.
### Comparison with empirical data. {#comparison-with-empirical-data. .unnumbered}
We now compare the obtained exponents $\tau$ of the snapshots distributions $P(s)$ with available empirical data [@Scanlon2007; @Kefi2007]. In particular, the black dots of Fig. \[Empirical\] correspond to empirical data for different vegetation-cover densities. Remarkably, field data are clearly compatible with our model results (for large enough environmental variability $\sigma=0.25$) and within one standard deviation (p-value $p=0.04<0.05$ of Pearson’s $\chi^2$ statistics).
In the light of these results, we can conclude that a very simple model of vegetation dynamics –when equipped with temporal environmental variability as a key ingredient– is able to produce highly structured (vegetation) configurations which can be well fitted by power laws in a rather robust though non-universal way. Furthermore, and rather remarkably, the model also shows exponents that change with the cover density, having values compatible with those found in empirical communities.
Discussion
==========
The study of vegetation patterns have a long tradition in ecology and environmental sciences. This interest has been reinforced in recent years owing to the growing awareness that the statistics of such patterns can be used as a predictor of early warning of ecological transitions or regime shifts such as desertification in semiarid environments. From the empirical side, two very influential papers reported a decade ago that vegetation patterns in some arid and semiarid ecosystems are scale free, thus lacking characteristic scales [@Scanlon2007; @Kefi2007]. Puzzled by these field observations, theoreticians were challenged to understand the origin of such scale-free patterns and alternative explanatory mechanisms have been proposed to account for them. For instance, some models put the emphasis on (i) existing plant-plan interactions (competition and facilitation), (ii) feedback loops between vegetation dynamics and abiotic resources such as water [@Scanlon2007; @Manor-Shnerb-2008a], or (iii) the presence of heterogeneity in the environment [@Kefi2007; @Shnerb2003-glass; @Manor-Shnerb-2008b; @Manor-Shnerb-2009].
For the sake of generality, let us stress here that the theoretical challenge of modelling approaches is to explain how scale-free patterns can possibly emerge in a robust way, i.e. without the need of invoking a high degree of parameter fine tuning. In particular, experience in statistical physics tells us that scale-free patterns typically emerge at very special points of the parameter space, such as critical points. To circumvent this theoretical conumdrum and in order to explain generic power-law decays, two different types of approaches are considered. The first one consists in considering mechanisms of self-organization, inspired in self-organized criticality [@BTW; @Pruessner-book; @BJP; @JABO1], which would justify why critical-like scale-free features appear without the apparent need of parameter fine tuning (although they usually require an unrealistic separation of time scales). This type of approach leads, however, to universal –as opposed to continuously varying– exponents. The second strategy consists in devising models in which scale-free patterns –with their concomitant power law distributions– emerge for broad ranges of parameter values [@GG-generic]. Such type of models already exist in statistical mechanics (for an in-depth discussion of these issues we refer to [@GG-generic]), but, to the best of our knowledge, they are not well-suited to be translated to the ecological problem under study here.
A particularly relevant idea in this context is that of “robust criticality”, stating that scale-free clusters of vegetation are akin to percolation clusters at a percolation phase transition, but that –for reasons to be fully clarified– it could be observed in simple models of vegetation dynamics without the need of fine tuning exactly to the percolation transition [@Pascual2002simple; @Pascual2002cluster; @Roy2003; @Pascual2005; @Pascual2006]. In other words, it is claimed that these models exhibit a broad region of power-laws around a transition point. However, the origin of such a broad region of criticality is still controversial in the sense that there is no actual theoretical understanding of why generic scale invariance could possibly emerge.
In this paper we have proposed a very simple model that is able to justify in a clean way the emergence of scale-free vegetation patterns in a broad range of parameter values.
Our model is individual-based and spatially explicit, based on the two-dimensional contact process. This model is the simplest possible to mimic vegetation dynamics (as a birth-and-death process). It has been very deeply studied in the literature and it is well-known to exhibit a quiescent-to-active phase transition, separating a quiescent phase on which all activity/vegetation ceases/disappear, from another in which activity/vegetation survives indefinitely with a given averaged density. In other words, the model exhibits a “directed-percolation” type of phase transition separating phases where activity may, or may not, percolate in time.
On the other hand, it is often neglected that there is another phase transition hidden in this prototypical model. It occurs within the active phase (as illustrated in Figure 1) and separates a phase in which the largest cluster of active sites percolates in space from another one in which it does not. We have confirmed computationally that, indeed, as theoretically predicted by van den Berg *et al.* [@van2011; @van2015] the transition is sharp, i.e. there is not such a thing as a broad range of scaling for such a simple model. Moreover, we have shown for the first time that such a transition is perfectly described by the standard scaling theory of (isotropic) percolation, with its well-known set of scaling exponents. Let us emphasize that such a result is not trivial as in standard percolation sites are occupied in a completely random way, while in the active phase of the contact process activity occupied sites are spatially correlated. Readers familiar with the renormalization-group theory could have correctly anticipated though –employing renormalization-group arguments– that such short-ranged correlations should not be able to affect the scaling exponents. Thus, summing up, the standard contact process only exhibit scale-free clusters at a very precise point in parameter space and is therefore *not* an adequate model to explain the broad/generic emergence of scale-free vegetation patterns in a robust way.
Inspired by previous work of our group and others showing that temporal variability can lead to generic power laws in some simple models [@TGP; @TGP-Shnerb], we moved on to study a variant of the contact process implementing temporal variability in external conditions, i.e. a randomly changing control parameter. With this simple –and ecologically well justified– novel ingredient the model changes dramatically.
[**[ i) ]{}**]{} First of all, the percolation phase transition within the active phase is not in the standard/isotropic percolation universality class, i.e. its scaling exponents are significantly different form the standard ones (see below for a further discussion of this fact).
[**[ ii) ]{}**]{} Second, generic scale-free distributed clusters emerge for a broad range of parameter values and not just at the critical point.
[**[iii) ]{}**]{} Third, individual snaphots of the system state –i.e. single configurations in which no averaging is performed– also exhibit clusters of highly variable sizes that can be fit as power laws for a broad range of possible vegetation densities. Even more remarkably, the obtained power-law exponents are density dependent –as also occur in empirical observations– and are in reasonably good agreement with those observed in the field for each single density for which empirical results were accessible to us.
Let us now briefly discuss each of these three results.
[**[i) ]{}**]{} To understand in a qualitative way why the scaling at the percolation point does not coincide with that of standard/isotropic percolation, one could argue as follows. When unfavorable external conditions appear, small clusters are very likely to be removed, while large ones, although also affected by poor conditions, can remain connected or divided in smaller-sized clusters. When favorable periods return, the latter clusters become more compact and/or are re-merged with others. Contrarily, the removed clusters are not able to re-appear from the empty sites, so that empty regions persist. As a result of these effects, the active sites in the active phase are strongly correlated and not equivalent to a standard/isotropic percolation process. Of course, in order to turn these hand-waving arguments into a more formal proof, one would need to perform some type of renormalization-group calculation, which is far from our scope here. However, let us remark, that usually the presence of temporal disorder dramatically affects the universal behavior of well known universality classes [@Vojta2015; @Alonso] (including the quiescent-active critical point of the contact process, which is dramatically affected by temporal noise [@Moreira; @Cafiero; @TGP]). Thus, it does not come as a surprise that our model equipped with temporal disorder differs essentially from standard percolation. In any case, a systematic account of this (novel?) critical universality class, establishing all critical exponents and scaling relationships with good accuracy and precision, is left for a future publication.
[**[ii) ]{}**]{} The emergence of a broad region of densities for which scale-free behaviour emerges is in line with the idea of generic scaling emerging in the presence of temporal variability. Actually, it was recently shown that in the presence of temporal heterogeneity a novel type of phase may emerge in systems with an absorbing-active phase transition such as the contact process. Such a novel phase was named *temporal Griffiths phase* [@TGP] in analogy with the well-known *Griffiths phases* that emerge in the presence of spatial (quenched) heterogeneity [@Vojta]. In particular, it was shown in [@TGP] (see also [@Savanna; @TGP-Shnerb]) that for models similar to the contact process studied here, different quantities scale as power laws, not just at a critical transition point, but in a whole broad region in parameter space in the presence of temporal disorder.
[**[iii) ]{}**]{} Directly related to our findings is the important work of Manor and Shnerb who discussed –in a mean-field set up different from our spatially explicit one– how multiplicative (environmental) noise can lead to scale-free distributed cluster sizes [@Manor-Shnerb-2008a; @Manor-Shnerb-2009]. In a nutshell, each existing cluster experiences a multiplicative random-walk process in which the change in cluster size either increases or decreases at each time step, but, crucially, it does so in an amount which is proportional to its present size, i.e. in a multiplicative way. It is very well-known that multiplicative processes lead generically to power-law distributions (of sizes in this case) [@Mitzenmacher; @Solomon; @Sornette-multiplicative]. We believe that a multiplicative process similar to this one is at the basis of the power-laws found in our model, but it can not be analyzed in simple mean-field terms.
Thus, the reason why our model produces generic power-law distributions is twofold: (i) on the one hand, nearby the standard percolation phase transition clusters of very different sizes are generated and (ii) owing to the multiplicative process induced by changing environmental conditions such cluster fluctuate wildly. The combination of these two features leads to –at least approximate– scale-free cluster distributions. In other words, the model can exhibit a sort of *intermittent percolation* in which the overall density fluctuates around the critical (isotropic) percolation threshold. This sweeping through the transition (or even near the transition) may suffice to generate power-laws in a rather robust fashion. In fact, D. Sornette proposed years back a generic mechanism like this –alternative to self-organized criticality– to explain the robust emergence of power-laws [@Sornette1994].
Finally, in order to test the robustness of our conclusions, and following some proposals in the literature, we have constructed versions of our temporally-varying model implementing spatial heterogeneity (spatial-dependent $p$) and a facilitation mechanism ($p$ depends on the number of occupied nearest neighbors). Preliminary studies reveal that results do not change significantly with respect to what is reported here for the pure case. More detailed and careful analyses of such extensions will be presented elsewhere.
Conclusion
==========
The conclusion of the present study is that a very simple individual-based model with fluctuating external conditions is able to explain in a simple but robust way the emergence of scale-free vegetation patterns. The exponents associated with the concomitant power-law distributions are non-universal and density dependent, and agree remarkably well with those fitted to empirical data. This conclusion does not imply that other mechanisms, such as plant-plant interactions or plant-water feedbacks are not important to explain vegetation patterns in general, but reveals that temporal variability is a key ingredient to explain the emergence of scale-free vegetation patterns in semiarid environments.
Acknowledgements
================
We acknowledge the Spanish Ministry and Agencia Estatal de investigaci[ó]{}n (AEI) through grant FIS2017-84256-P (European Regional Development Fund), as well as the Consejería de Conocimiento, Investigación y Universidad, Junta de Andalucía and European Regional Development Fund (ERDF), ref. SOMM17/6105/UGR for financial support. We thank P. Moretti, V. Buend[í]{}a, G. Barrios, and P. Villegas for a critical reading of the manuscript.
Appendix
========
Fitting power laws {#fitting-power-laws .unnumbered}
------------------
We fit cluster-size distributions $P(s)$ –both averaged over different realizations and times and for individual snapshots– as power-laws $P(s)=C s^{\tau}$ within an optimal interval $[s_{\min},s_{\max}]$. The exponent $\tau$, the normalization constant $C$ as well as the values of $s_{\min}$ and $s_{\max}$ are fitted by maximizing the normalized log-likelihood assuming Poissonian counts $\ln L=(1/N)\sum_i[s_i \ln P(s_i)-P(s_i)+\ln(s_i!)]$, where $N$ is the number of non-zero cluster sizes in a given range of $s$ [@Clauset; @Klaus2011statistical; @Alstott2014powerlaw]. To avoid overfitting, we discard snapshots whose optimal range $[s_{\min},s_{\max}]$ includes less than $10$ points. In all cases, we compare the power-law fit with an exponential one $P(s)=Ce^{-c s}$ within the obtained interval $[s_{\min},s_{\max}]$; on the case of snapshots, we report on the number of trials for which the power-law constitutes a better fit than the exponential.
Density Fitted exponent $s_{min}$ $s_{max}$
------------------ ------------------ ---------------- -------------------- -------------------
0.02 2.59 25 5012 -4.25
0.04 2.59 25 5623 -5.74
0.06 2.63 25 17783 -6.80
0.08 2.62 25 14125 -11.16
0.10 2.6 25 22387 -22.81
0.12 2.56 22 25119 -47.11
0.14 2.57 25 25119 -43.07
0.16 2.56 25 28184 -58.00
0.18 2.5 22 31623 -103.96
0.20 2.49 25 39811 -88.51
0.22 2.46 25 44668 -116.06
0.24 2.42 25 50119 -141.10
0.26 2.37 25 70795 -170.19
0.28 2.33 25 89125 -177.45
0.30 2.28 25 79433 -187.49
0.32 2.23 25 89125 -199.60
0.34 2.16 25 112202 -240.01
0.36 2.09 25 141254 -264.38
0.38 2.01 25 141254 -318.41
0.40 1.93 25 177828 -324.45
0.42 1.85 25 177828 -405.49
0.44 1.77 25 199526 -367.81
0.46 1.7 25 177828 -332.26
0.48 1.65 25 223872 -281.94
0.50 1.63 25 177828 -137.16
0.52 1.64 25 177828 -44.50
0.54 1.7 25 177828 -36.13
[**[0.56]{}**]{} [**[1.81]{}**]{} [**[25]{}**]{} [**[112202]{}**]{} [**[-5.91]{}**]{}
0.58 1.99 25 158489 -16.94
0.60 2.19 25 25119 -22.78
0.62 2.41 25 10000 -14.14
0.64 2.63 25 8913 -9.58
0.66 2.81 25 2818 -4.31
0.68 2.93 22 3162 -2.18
0.70 3.12 25 3162 -0.71
: [**[Fits for averaged cluster-size distributions $P(s)$ with different densities.]{}**]{} Power-laws have been fitted (see Appendix) to averaged $P(s)$ for different densities (first column). We show the fitted exponents (second column), the optimal interval of clusters sizes –$s_{min}$ (third column) to $s_{max}$ (fifth column)– and the corresponding normalized log-likelihood for each density. The best fit is obtained for density $\rho=0.56$ while for $\rho=0.48$ the power-law fit extends for the broadest range. []{data-label="table:fit_averaged"}
Density
--------- ------ ------ -------- -------
0.04 2.14 1.10 71.74 71.74
0.06 2.38 1.26 82.54 61.90
0.08 2.46 1.21 81.48 59.26
0.10 2.63 1.33 78.95 40.00
0.12 2.67 1.42 70.16 32.26
0.14 2.75 1.32 64.29 26.62
0.16 2.79 1.30 62.23 31.38
0.18 2.70 1.34 62.00 28.50
0.20 2.57 1.24 73.50 32.50
0.22 2.49 1.24 81.00 43.00
0.24 2.35 1.13 93.00 37.00
0.26 2.43 1.36 100.00 45.50
0.28 2.33 1.24 100.00 55.00
0.30 2.16 1.05 100.00 54.00
0.32 2.05 0.89 100.00 60.00
0.34 1.93 0.63 100.00 66.50
0.36 1.80 0.37 100.00 70.00
0.38 1.71 0.23 100.00 75.50
0.40 1.62 0.11 100.00 84.00
0.42 1.58 0.07 100.00 87.50
0.44 1.60 0.07 100.00 92.00
0.46 1.62 0.06 100.00 94.97
0.48 1.66 0.07 100.00 94.95
0.50 1.70 0.06 100.00 96.97
0.52 1.77 0.07 100.00 96.48
0.54 1.86 0.11 99.49 95.43
0.56 1.96 0.18 100.00 93.91
0.58 2.01 0.15 100.00 92.89
0.60 2.10 0.18 100.00 89.23
0.62 2.15 0.23 99.48 88.02
0.64 2.24 0.32 98.97 63.92
0.66 2.39 0.57 99.47 61.38
0.68 2.50 0.71 97.30 80.54
: [**[Fits for individual snapshots with different densities.]{}**]{} We fit power-law curves (see Appendix) for snapshots of the model with temporal variability for different densities (as listed in the first column). The second column shows the mean fitted exponents; the third column shows the variance of exponent values (only snapshots whose optimal range considers more than $10$ points are considered); the fourth one shows the percentage of snapshots for which power-law fits are better than exponential ones. The fifth column shows the fraction for which the power-law fit extends for at least $1.5$ decades.[]{data-label="table:fit_snapshots"}
\#1
[64]{} \[1\][\#1]{} \[1\][`#1`]{} urlstyle \[1\][doi: \#1]{}
Juan Jos[é]{} Alonso and Miguel A Muñoz. Temporally disordered ising models. *EPL (Europhysics Letters)*, 560 (4):0 485, 2001.
Jeff Alstott, Ed Bullmore, and Dietmar Plenz. powerlaw: a python package for analysis of heavy-tailed distributions. *PloS one*, 90 (1):0 e85777, 2014.
Per Bak, Chao Tang, and Kurt Wiesenfeld. Self-organized criticality: An explanation of the 1/f noise. *Phys. Rev. Lett.*, 590 (4):0 381, 1987.
Miguel Berdugo, Sonia K[é]{}fi, Santiago Soliveres, and Fernando T Maestre. Plant spatial patterns identify alternative ecosystem multifunctionality states in global drylands. *Nature ecology & evolution*, 10 (2):0 0003, 2017.
J.J. Binney, N.J. Dowrick, A.J. Fisher, and M.E.J. Newman. *The Theory of Critical Phenomena*. Oxford University Press, Oxford, 1993.
J. A. Bonachela and M. A. Mu[ñ]{}oz. Self-organization without conservation: true or just apparent scale-invariance? *J. Stat. Mech.*, page P09009, Sep 2009.
Juan A Bonachela, Miguel A Mu[ñ]{}oz, and Simon A Levin. Patchiness and demographic noise in three ecological examples. *Journal of Statistical Physics*, 1480 (4):0 724–740, 2012.
Raffaele Cafiero, Andrea Gabrielli, and Miguel A Mu[ñ]{}oz. Disordered one-dimensional contact process. *Physical Review E*, 570 (5):0 5060, 1998.
Kim Christensen and Nicholas R Moloney. *Complexity and criticality*, volume 1. Imperial College Press, London, 2005.
Aaron Clauset, Cosma Rohilla Shalizi, and M. E. J. Newman. Power-law distributions in empirical data. *SIAM Rev.*, 510 (4):0 661–703, November 2009. ISSN 0036-1445.
Ronald Dickman, Miguel A Mu[ñ]{}oz, Alessandro Vespignani, and Stefano Zapperi. Paths to self-organized criticality. *Braz. J. Phys.*, 300 (1):0 27–41, 2000.
R. Durrett and S. Levin. The importance of being discrete (and spatial). *Theoretical Population Biology*, 460 (3):0 363 – 394, 1994. ISSN 0040-5809. [doi: ]{}[https://doi.org/10.1006/tpbi.1994.1032]{}. URL <http://www.sciencedirect.com/science/article/pii/S004058098471032X>.
Rick Durrett. Coexistence in stochastic spatial models. *Annals of Applied Probability*, 19, 06 2009. [doi: ]{}[10.1214/08-AAP590]{}.
G Grinstein. Generic scale invariance in classical nonequilibrium systems. *Journal of Applied Physics*, 690 (8):0 5441–5446, 1991.
G Grinstein and Miguel A Mu[ñ]{}oz. The statistical mechanics of absorbing states. In *Fourth Granada Lectures in Computational Physics*, pages 223–270. Springer, Heidelberg, Berlin, 1997.
Malte Henkel and Michel Pleimling. *Non-Equilibrium Phase Transitions: Volume 2: Ageing and Dynamical Scaling Far from Equilibrium*. Springer Science & Business Media, Heidelberg, Berlin, 2011.
H. J. Hilhorst and F. van Wijland. Equivalence of stationary state ensembles. *Phys. Rev. E*, 65:0 035103, Feb 2002. [doi: ]{}[10.1103/PhysRevE.65.035103]{}. URL <https://link.aps.org/doi/10.1103/PhysRevE.65.035103>.
Reinier HilleRisLambers, Max Rietkerk, Frank van den Bosch, Herbert H. T. Prins, and Hans de Kroon. Vegetation pattern formation in semi-arid grazing systems. *Ecology*, 820 (1):0 50–61, 2001. [doi: ]{}[10.1890/0012-9658(2001)082\[0050:VPFISA\]2.0.CO;2]{}.
Sonia K[é]{}fi, Max Rietkerk, Concepci[ó]{}n L Alados, Yolanda Pueyo, Vasilios P Papanastasis, Ahmed ElAich, and Peter C De Ruiter. Spatial vegetation patterns and imminent desertification in mediterranean arid ecosystems. *Nature*, 4490 (7159):0 213–217, 2007.
Sonia K[é]{}fi, Max Rietkerk, Manojit Roy, Alain Franc, Peter C De Ruiter, and Mercedes Pascual. Robust scaling in ecosystems and the meltdown of patch size distributions before extinction. *Ecology letters*, 140 (1):0 29–35, 2011.
Andreas Klaus, Shan Yu, and Dietmar Plenz. Statistical analyses support power law distributions found in neuronal avalanches. *PloS one*, 60 (5):0 e19779, 2011.
Christopher A Klausmeier. Regular and irregular patterns in semiarid vegetation. *Science*, 2840 (5421):0 1826–1828, 1999.
Pierre Legendre and Marie Jos[é]{} Fortin. Spatial pattern and ecological analysis. *Vegetation*, 800 (2):0 107–138, 1989.
Olivier Lejeune, Mustapha Tlidi, and Pierre Couteron. Localized vegetation patches: a self-organized response to resource scarcity. *Physical Review E*, 660 (1):0 010901, 2002.
Simon A Levin. Pattern formation in ecological communities. In *Spatial pattern in plankton communities*, pages 433–465. Springer, Heidelberg, Berlin, 1978.
Alon Manor and Nadav M Shnerb. Origin of pareto-like spatial distributions in ecosystems. *Physical Review Letters*, 1010 (26):0 268104, 2008.
Alon Manor and Nadav M Shnerb. Facilitation, competition, and vegetation patchiness: from scale free distribution to patterns. *Journal of Theoretical Biology*, 2530 (4):0 838–842, 2008.
Alon Manor and Nadav M Shnerb. Multiplicative noise and second order phase transitions. *Physical Review Letters*, 1030 (3):0 030601, 2009.
Joaqu[í]{}n Marro and Ronald Dickman. *Nonequilibrium phase transitions in lattice models*. Cambridge University Press, Cambridge, 2005.
M Mitzenmacher. A brief history of generative models for power law and lognormal distributions. *Internet Mathematics*, 1:0 226–251, 2002.
Adriana G Moreira and Ronald Dickman. Critical dynamics of the contact process with quenched disorder. *Physical Review E*, 540 (4):0 R3090, 1996.
Miguel A Muñoz. Colloquium: Criticality and dynamical scaling in living systems. *Reviews of Modern Physics*, 900 (3):0 031001, 2018.
Claudia Neuhauser. Mathematical challenges in spatial ecology. *Notices of the AMS*, 480 (11):0 1304–1314, 2001.
M. E. J. Newman. . *Contemp Phys*, 460 (5):0 323–351, May 2005. ISSN 0010-7514. [doi: ]{}[10.1080/00107510500052444]{}. URL <http://dx.doi.org/10.1080/00107510500052444>.
G[é]{}za [Ó]{}dor. *Universality in Nonequilibrium Lattice Systems: Theoretical Foundations*. World Scientific, Singapore, 2008.
Mercedes Pascual and Fr[é]{}d[é]{}ric Guichard. Criticality and disturbance in spatial ecological systems. *Trends in ecology & evolution*, 200 (2):0 88–95, 2005.
Mercedes Pascual, Manojit Roy, and Alain Franc. Simple temporal models for ecological systems with complex spatial patterns. *Ecology Letters*, 50 (3):0 412–419, 2002.
Mercedes Pascual, Manojit Roy, Fr[é]{}d[é]{}ric Guichard, and Glenn Flierl. Cluster size distributions: signatures of self–organization in spatial ecologies. *Philosophical Transactions of the Royal Society of London B: Biological Sciences*, 3570 (1421):0 657–666, 2002.
Mercedes Pascual, Jennifer A Dunne, and Simon A Levin. Challenges for the future: integrating ecological structure and dynamics. *Ecological networks: linking structure to dynamics in food webs*, pages 351–373, 2006.
Gunnar Pruessner. *Self-organised criticality: theory, models and characterisation*. Cambridge University Press, Cambridge, 2012.
Peter Richmond and Sorin Solomon. Power laws are disguised boltzmann laws. *Int. J. Mod. Phys. C*, 120 (03):0 333–343, 2001.
Max Rietkerk and Johan Van de Koppel. Regular pattern formation in real ecosystems. *Trends in ecology & evolution*, 230 (3):0 169–175, 2008.
Max Rietkerk, Stefan C Dekker, Peter C De Ruiter, and Johan van de Koppel. Self-organized patchiness and catastrophic shifts in ecosystems. *Science*, 3050 (5692):0 1926–1929, 2004.
Manojit Roy, Mercedes Pascual, and Alain Franc. Broad scaling region in a spatial ecological system. *Complexity*, 80 (5):0 19–27, 2003.
Todd M Scanlon, Kelly K Caylor, Simon A Levin, and Ignacio Rodriguez-Iturbe. Positive feedbacks promote power-law clustering of kalahari vegetation. *Nature*, 4490 (7159):0 209–212, 2007.
Marten Scheffer, Steve Carpenter, Jonathan A Foley, Carl Folke, and Brian Walker. Catastrophic shifts in ecosystems. *Nature*, 4130 (6856):0 591, 2001.
M Shachak, Y Zarmi, J von Hardenberg, and E Meron. Diversity of vegetation patterns and desertification. *Physical Review Letters*, 87, 2001.
NM Shnerb, P Sarah, H Lavee, and S Solomon. Reactive glass and vegetation patterns. *Physical Review Letters*, 900 (3):0 038101, 2003.
Ricard Sol[é]{}. Scaling laws in the drier. *Nature*, 4490 (7159):0 151, 2007.
Ricard Sol[é]{} and Jordi Bascompte. *Self-Organization in Complex Ecosystems.(MPB-42)*, volume 58. Princeton University Press, Princeton, 2012.
D. Sornette. *Critical Phenomena in Natural Sciences*. Springer, Heidelberg, Berlin, 2006.
Didier Sornette. Sweeping of an instability: an alternative to self-organized criticality to get powerlaws without parameter tuning. *Journal de Physique I*, 40 (2):0 209–221, 1994.
Didier Sornette. Multiplicative processes and power laws. *Phys. Rev. E*, 570 (4):0 4811, 1998.
Dietrich Stauffer and Ammon Aharony. *Introduction to percolation theory*. CRC press, Boca Raton, Florida, 1994.
J Van den Berg. Sharpness of the percolation transition in the two-dimensional contact process. *The Annals of Applied Probability*, pages 374–395, 2011.
J van den Berg, Jakob E Bj[ö]{}rnberg, and M Heydenreich. Sharpness versus robustness of the percolation transition in 2d contact processes. *Stochastic Processes and their Applications*, 1250 (2):0 513–537, 2015.
Nicolaas Godfried Van Kampen. *Stochastic processes in physics and chemistry*, volume 1. Elsevier, Amsterdam, 1992.
Federico Vazquez, Crist[ó]{}bal L[ó]{}pez, Justin M Calabrese, and Miguel A Mu[ñ]{}oz. Dynamical phase coexistence: a simple solution to the “savanna problem”. *Journal of Theoretical Biology*, 2640 (2):0 360–366, 2010.
Federico Vazquez, Juan A Bonachela, Crist[ó]{}bal L[ó]{}pez, and Miguel A Mu[ñ]{}oz. Temporal griffiths phases. *Physical Review Letters*, 1060 (23):0 235702, 2011.
Paula Villa Mart[í]{}n, Juan A Bonachela, and Miguel A Mu[ñ]{}oz. Quenched disorder forbids discontinuous transitions in nonequilibrium low-dimensional systems. *Phys. Rev. E*, 890 (1):0 012145, 2014.
Paula Villa Mart[í]{}n, Juan A Bonachela, Simon A Levin, and Miguel A Mu[ñ]{}oz. Eluding catastrophic shifts. *Proceedings of the National Academy of Sciences*, 1120 (15):0 E1828–E1836, 2015.
T. Vojta. . *J. Phys. A*, 390 (22):0 R143–R205, May 2006. ISSN 0305-4470. [doi: ]{}[10.1088/0305-4470/39/22/R01]{}. URL <http://dx.doi.org/10.1088/0305-4470/39/22/R01>.
Thomas Vojta and Jos[é]{} A Hoyos. Infinite-noise criticality: Nonequilibrium phase transitions in fluctuating environments. *EPL (Europhysics Letters)*, 1120 (3):0 30002, 2015.
Yitzhak Yahalom and Nadav M Shnerb. Phase diagram for logistic systems under bounded stochasticity. *Physical Review Letters*, 1220 (10):0 108102, 2019.
[^1]: This mean-field equation is straightforwardly derived by performing a size-expansion of the Master equation defining the model while keeping just the leading/deterministic term [@vanKampen].
[^2]: Let us note that one could have also defined the model by introducing temporal disorder in the “microscopic” transition rates in a continuous-time implementation of the contact process rather than in the probabilities [@Henkel-book]. We do not expect such a change to qualitatively affect the main results presented here.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Several exact cosmological solutions of a metric-affine theory of gravity with two torsion functions are presented. These solutions give a essentially different explanation from the one in most of previous works to the cause of the accelerating cosmological expansion and the origin of the torsion of the spacetime. These solutions can be divided into two classes. The solutions in the first class define the critical points of a dynamical system representing an asymptotically stable de Sitter spacetime. The solutions in the second class have exact analytic expressions which have never been found in the literature. The acceleration equation of the universe in general relativity is only a special case of them. These solutions indicate that even in vacuum the spacetime can be endowed with torsion, which means that the torsion of the spacetime has an intrinsic nature and a geometric origin. In these solutions the acceleration of the cosmological expansion is due to either the scalar torsion or the pseudoscalar torsion function. Neither a cosmological constant nor dark energy is needed. It is the torsion of the spacetime that causes the accelerating expansion of the universe in vacuum. All the effects of the inflation, the acceleration and the phase transformation from deceleration to acceleration can be explained by these solutions. Furthermore, the energy and pressure of the matter without spin can produce the torsion of the spacetime and make the expansion of the universe decelerate as well as accelerate.'
author:
- Guoying Chee
- Yongxin Guo
title: Torsion and accelerating expansion of the universe in quadratic gravitation
---
$\text{Introduction}$
=====================
In the last few years the realization that the universe is currently undergoing an accelerated expansion phase and the quest for the nature of dark energy has renewed interest in so-called modified gravity theories (for a review see \[1\]). In these theories one modifies the laws of gravity so that a late-time accelerated expansion is produced without recourse to a dark energy component, a fact which renders these models very attractive. The simplest family of modified gravity theories is obtained by replacing the Ricci scalar $R$ in the usual Hilbert- Einstein Lagrangian with some function $f(R)$ (for reviews, see , \[2-5\]).
There are actually three versions of $f(R)$ gravity: Metric $f(R)$ gravity, Palatini $f(R)$ gravity, and metric-affine $f(R)$ gravity. In fact, these are physically different theories rather than manifestations of the same theory in different guises, as the different variational principles yield inequivalent equations of motion (except when the action is the Einstein-Hilbert and matter is minimally coupled to geometry). In metric $f(R)$ gravity, the action is varied with respect to the metric as usual (for an introduction see \[6\]). Palatini $f(R)$ gravity comes about from the same action if we decide to treat the connection as an independent quantity. The connection, however, does not enter the matter action. Such a approach were introduced and initially studied by Buchdahl \[7\] and has attracted a lot of interest as possible infrared modifications of general relativity (for a shorter review of metric and Palatini $f(R)$ gravity see \[8\]). It has recently been generalized to $f(R)$ theories with non-symmetric connections, i.e. theories that allow for torsion \[9\] and $f(R,R_{\mu \nu }R^{\mu \nu })$ theories \[10\]. In metric-affine $f(R)$ gravity the matter action is allowed to depend also on the connection. In addition, the connection[* *]{}can include both torsion and non-metricity \[11\].
It has been shown that even in the most general case of Palatini $f(R)$ gravity where both torsion and non-metricity are allowed, the connection can still be algebraically eliminated in favor of the metric and the matter fields \[12\]. Clearly, $f(R)$ actions do not carry enough dynamics to support an independent connection which carries dynamical degrees of freedom. However, this is not a generic property of generalized Palatini gravity. [* *]{}The addition of the $R_{\mu \nu }R^{\mu \nu }$ term to the Lagrangian radically changes the situation and excites new degrees of freedom in the connection. The connection (or parts of it) becomes dynamical and so, it cannot be eliminated algebraically. If the connection is torsion free, the dynamical degrees of freedom reside in the symmetric part of the connection \[13\].
In generic metric-affine theories the addition of the $R_{\mu \nu }R^{\mu
\nu }$ term to the Lagrangian makes the propagating degrees of freedom reside in both the antisymmetric and symmetric parts of the connection. In other words, the dynamical degrees of freedom can be both torsion and non-metricity. In these theories torsion field plays a fundamental role: it contributes, together with curvature degrees of freedom, to the dynamics. Propagating torsion is the key feature of these theories \[14, 15\].
Torsion proves to be essential for total angular momentum conservation when intrinsic spin angular momentum is relevant (for reviews on torsion, see \[16, 17, 18\]). It has been argued that torsion must be present in a fundamental theory of gravity \[19, 20\]. In the teleparallel gravity, for example, torsion plays a central role (for a shorter review see \[21\]). Recently, models based on modified teleparallel gravity, namely $f(T)$, were presented. In these models the torsion proves to be the responsible of the observed acceleration of the universe \[22\].
The rediscovery of the metric-affine (Palatini) formulation was mainly driven by the interest in finding cosmological scenarios able to explain the current observations. Using the respective dynamically equivalent scalar-tensor representation of Palatini $f(R)$ gravity some cosmological models with asymptotically de Sitter behavior have been presented \[23\]. It was shown that adopting the metric-affine formulation together with an action that includes a term[* *]{}inversely proportional to the scalar curvature, such as the one in \[24\], can address the problem of the current accelerated expansion equally well as when using the purely metric formalism \[25\]. Additionally, it was found that $f(R)$ theories of gravity in the metric-affine formulation do not suffer from the problems for the metric formulation. On the other hand, although cosmology in the theories with the Lagrangian including $R^2$ and $R_{\mu \nu }R^{\mu \nu }$ terms have been studied in the purely metric formulation (for example see \[26\]) and the Palatini formulation \[27\], the similar cosmological models in the metric-affine formulation have not been discussed thoroughly in the literature. Especially, the cosmological effect of torsion in metric-affine theories of gravity has not been explored extensively. We have not known whether the dynamical torsion could lead to a de Sitter solution and then be used to explain the observed acceleration of the universe.[* *]{}An answer will be given in this paper.
The metric-affine approach has been widely used in order to interpret gravity as a gauge theory many times over the years (see, for example, \[28\] for a study on $f(R)$ actions and \[29\] for a thorough review). In recent years it has bee used in cosmology to interpret the accelerating expansion of the universe \[30, 31\]. In this approach the structure of the gravitational equations and physical consequences of cosmology, in particular, the situation concerning the accelerating expansion depend essentially on the form of the Lagrangian. The metric-affine gravity can be divided into different sectors in dependence on the number of nonvanishing components of the torsion tensor and the order of the differential equations. One sector of the metric-affine gravity is so-called dynamical scalar torsion sector considered in \[30\]. Starting from a Lagrangian consisting of $R^2$ and the quadratic torsion terms a cosmological model has been constructed. This model can contribute an oscillating aspect to the expansion rate of the universe. A different model of acceleration with torsion but without dark matter and dark energy has been presented in \[31\]. The Lagrangian of it is the most general form including the linear in the scalar curvature term as well as 9 quadratic terms (6 invariants of the curvature tensor and 3 invariants of the torsion tensor with indefinite parameters). Its Lagrangian involves too many terms and indefinite parameters, which make the field equations complicated and difficult to solve and the role of each term obscure. In order to simplify the field equations some restrictions on indefinite parameters have to be imposed. Under these restrictions, especially, all the higher derivatives of the scale factor are excluded from the cosmological equations. The question is whether such a complicated Lagrangian is necessary. Can we use a simpler Lagrangian to construct a model of cosmic acceleration? In fact all the indefinite parameters in the Lagrangian in \[31\] have been combined into four new ones, which implies that some terms are not necessary and the Lagrangian can be simplified. In this paper we will show that a rather simpler Lagrangian, $R+\alpha R^2+\beta R_{\mu \nu }R^{\mu \nu }+\gamma T{}^\mu
{}_{\nu \rho }T{}_\mu {}^{\nu \rho }$, is sufficient and necessary to construct a model of cosmic acceleration. The terms $\beta R_{\mu \nu
}R^{\mu \nu }$ and $\gamma T{}^\mu {}_{\nu \rho }T{}_\mu {}^{\nu \rho }$ play different roles in the theory: the former determines the structure of the field equations while the latter determines the behavior and the stability of the solutions. The $\beta R_{\mu \nu }R^{\mu \nu }$ term leads to different structure of the cosmological equations from the one in \[30\]. In addition to the simplicity the main advantage of this Lagrangian is to permit exact or analytic solutions which have not been found in previous works. For any physical theories, to find exact or analytic solutions is an important topic. Next comes the physical interpretation of the solutions thus obtained. Mathematically de Sitter spacetime as the maximally space is undoubtedly important for any gravity theories. From the observational side, recent studies illuminate that both the early universe (inflation) and the late-time universe (cosmic acceleration) can be regarded as fluctuations on a de Sitter background. So de Sitter solutions take a pivotal status in gravitational theories, especially in modern cosmology.
We will follow the approach of \[26, 30,31\] rather than the one in \[27\] to avoid getting involved in debate on the transformation from one frame to another\[32\]. We choose the tetrad$ e_I^{\mu}$ and the spin connection $\Gamma {}^{IJ}{}_\mu $ instead of the metric $g_{\mu \nu }$ and the affine connection $\Gamma {}^\lambda {}_{\mu \nu }$ as the dynamical variables following the gauge theory approach \[30\]. The descriptions in terms of the variables ($e^I{}_\mu $, $\Gamma {}^{JK}{}_\nu $) and ($g_{\mu
\nu }$, $T^\lambda {}_{\rho \sigma }$) are equivalent in our approach (the argument in detail see \[16\]). We will concentrate on the role of torsion subject to the metricity. In this case only the torsion part of the connection is independent of the metric (or tetrad).
Because the field equations can result of order higher than second and very difficult to handle, the theory of dynamical systems provides a powerful scheme for investigating the physical behavior of such theories \[33\] for a wide class of cosmological models. The dynamical system approach has acquired great importance in the investigation on various theories of gravity. Some works have been done in the case of scalar fields in cosmology and for scalar-tensor theories of gravity \[34\]. This approach has the advantage of offering a relatively simple method to obtain exact solutions (even if these only represent the asymptotic behavior) and to obtain a (qualitative) description of the global dynamics of the models. Such results are very difficult to obtain by other methods. The application of this method has allowed new insights on higher order cosmological models and has shown a deep connection between these theories and the cosmic acceleration phenomenon. It makes possible not only to develop experimental test for alternative gravity but also to allow a better understanding of the reasons underlying the success of the theory. The dynamical systems approach has been used to investigate universes in theories of gravity \[26,30, 31\]. In contrast with \[31\] we allow the field equations to contain higher derivatives. We will see that for the Lagrangian of the form $R+\alpha
R^2+\beta R_{\mu \nu }R^{\mu \nu }+\gamma T{}^\mu {}_{\nu \rho }T{}_\mu
{}^{\nu \rho }$ the field equations can be simplified and solved exactly for some choices of $\alpha $, $\beta $ and $\gamma $. Some meaningful consequences can be inferred from the solutions obtained. The accelerating expansion of the universe can be explained without a cosmological constant or dark energy. A vacuum spacetime can possess torsion which causes the acceleration of the cosmological expansion. The conception of vacuum as physical notion is changed essentially. Instead of it as passive receptacle of physical objects and processes, the vacuum assumes a dynamical properties as a gravitating object. The torsion of the spacetime can be produced by the energy and pressure besides the spin of matter.
The paper is organized as follows. In section II the gravitational field equations are derived following the approach of \[29, 30, 31\]. Using them to the spatially flat Friedmann-Robertson-Walker metric a system of cosmological equations is obtained in section III. Since the spin orientation of particles in ordinary matter is random, the macroscopic spacetime average of the spin vanishes. In this case, the solutions of the cosmological equations are divided into two classes. Each of them is related with only one torsion function, the scalar or the pseudoscalar torsion function. They are obtained in section IV and V, separately, using different methods. For the scalar torsion function the equations take the form of a dynamical system, of which asymptotically stable critical points represent the exact de Sitter solutions. For the pseudoscalar torsion function an exact analytic solution of the cosmological equations is presented in section V. In terms of this solution the acceleration and the phase transformation from decelerating to accelerating expansion of the universe can be explained. All of these solutions indicate that in vacuum the spacetime possesses an intrinsic torsion which does not originate from the spin of matter. It is the torsion that causes the acceleration of the cosmological expansion in vacuum. The torsion of the spacetime can be produced by the energy and pressure besides the spin of matter. In section VI we obtain some exact analytic solutions of the cosmological equations in the case $\gamma =0$. These solutions can only describe the inflation (in the early epoch) or the decelerating expansion (in the later epoch) of the universe. This means that the term $\gamma T{}^\mu {}_{\nu \rho }T{}_\mu
{}^{\nu \rho }$ is necessary to construct a model of cosmic acceleration. The section VII is devoted to conclusions.
$\text{Gravitational field equations}$
======================================
We start from the action $$S\left[ g_{\mu \nu },\Gamma ^\alpha {}_{\beta \gamma },\psi \right] =\frac \hbar {8\pi l^2}\int d^4x\sqrt{-g}\left[ \frac 12R+\alpha R^2+\beta R_{\mu
\nu }R^{\mu \nu }+\gamma T{}^\mu {}_{\nu \rho }T{}_\mu {}^{\nu \rho }\right]
+S_m\left[ g_{\mu \nu },\Gamma ^\alpha {}_{\beta \gamma },\psi \right] ,$$ where $l=\sqrt{\hbar G/c^3}$ is the Planck length, $\alpha $, and $\beta $ are two parameters with the dimension of $l^2$, $\gamma $ is a parameter of dimensionless, $\psi $ denotes matter fields. In contrast with \[27\], here the connection is not symmetric, i.e. $\Gamma ^\alpha {}_{\beta \gamma }\neq
\Gamma ^\alpha {}_{\gamma \beta }$, and appears in the action of matter $S_m$. In other words, we are dealing with a metric-affine theory rather than a Palatini one. By the same way used in \[29, 30, 31\], the variational principle yields the field equations for the tetrad[* *]{} $e_I{}^\mu $ and the spin connection $\Gamma {}^{IJ}{}_\mu $:
$$\begin{aligned}
&&e^{I\nu }R{}_{\nu \mu }-\frac 12e{}^I{}_\mu R \nonumber \\
&=&E^I{}_\mu -\alpha \left( 4e^{I\nu }R{}_{\nu \mu }-e{}^I{}_\mu R\right)
R-\beta \left( 2e{}^{I\sigma }R{}^\rho {}_\sigma R{}_{\rho \mu }+2e^J{}_\rho
R{}^{\rho \sigma }{}R{}^I{}_{J\mu \sigma }-e{}^I{}_\mu R_{\rho \sigma
}R^{\rho \sigma }\right) \nonumber \\
&&+\gamma \left( 4\partial _\nu \left( e{}^{I\lambda }T{}_{\mu \lambda
}{}^\nu \right) -4e{}^K{}_\tau e{}^{I\lambda }T{}_{\mu \lambda }{}^\nu
\partial _\nu e_K{}^\tau +e{}^I{}_\mu T{}^\lambda {}_{\rho \sigma
}T{}_\lambda {}^{\rho \sigma }-4e{}^{I\nu }T{}^\lambda {}_{\nu \tau
}T{}{}_{\lambda \mu }{}^\tau \right) ,\end{aligned}$$
$$\begin{aligned}
&&e_{[I}{}^\nu e_{J]}{}^\mu e{}^K{}_\tau \partial _\nu e_K{}^\tau
+e_{[I}{}^\nu e_{J]}{}^\tau \Gamma ^\mu {}_{\nu \tau }+e_{[I}{}^\nu
e_{J]}{}^\mu \Gamma ^\lambda {}_{\lambda \nu } \nonumber \\
&=&s_{IJ}{}^\mu -4\alpha \left( e_{[I}{}^\nu e_{J]}{}^\tau \Gamma ^\mu
{}_{\nu \tau }R+e_{[J}{}^\mu e_{I]}{}^\nu \left( \Gamma ^\lambda {}_{\lambda
\nu }R-\partial _\nu R\right) +e_{[I}{}^\nu e_{J]}{}^\mu Re{}^K{}_\tau
\partial _\nu e_K{}^\tau \right) \nonumber \\
&&-4\beta e_J{}^\lambda \left( e_I{}^{[\mu }\partial _\nu R_\lambda {}^{\nu
]}+e_I{}^{[\nu }R_\lambda {}^{\mu ]}e{}^K{}_\tau \partial _\nu e_K{}^\tau
+e_I{}^\tau \Gamma ^{[\nu }{}_{\nu \tau }R_\lambda {}^{\mu ]}+e_I{}^{[\nu
}R_\tau {}^{\mu ]}\Gamma ^\tau {}_{\nu \lambda }\right) \nonumber \\
&&-4\gamma e_{I\nu }e{}_J{}^\tau T{}^{\nu \mu }{}_\tau .\end{aligned}$$
where $E^I{}_\mu $ and $s_{IJ}{}^\mu $ are energy- momentum and spin tensors of the matter source, respectively. We use the Greek alphabet ($\mu $, $\nu $, $\rho $, $...=0,1,2,3$) to denote (holonomic) indices related to spacetime, and the Latin alphabet ($I,J,K,...=0,1,2,3$) to denote algebraic (anholonomic) indices, which are raised and lowered with the Minkowski metric $\eta _{IJ}$ $=$ diag ($-1,+1,+1,+1$). If $\alpha =\beta =\gamma =0$, these equations become the field equations of Einstein-Cartan-Sciama-Kibble theory.. Especially, (20) becomes the Einstein equation. To understand these equations, we will do a translation of (2, 3) into a certain effective Riemannian form–transcribing from quantities expressed in terms of the tetrad $e_I{}^\mu $ and spin connection $\Gamma {}^{IJ}{}_\mu $ into the ones expressed in terms of the metric $g_{\mu \nu }$ and torsion $T^\lambda
{}_{\mu \nu }$ (or contortion $K^\lambda {}_{\mu \nu }$), as was done in \[30\]. It should be noted \[16\] that the set ($e^I{}_\mu $, $\Gamma
{}^{JK}{}_\nu $) corresponds to the first order formalism, while the set ($g_{\mu \nu }$, $T^\lambda {}_{\rho \sigma }$) to the second order formalism. The origin of this is that in the last case the non-torsional part of the affine connection is a function of the metric, while, within the gauge approach, the variables ($e^I{}_\mu $, $\Gamma {}^{JK}{}_\nu $) are mutually independent completely. The descriptions in terms of the variables ($e^I{}_\mu $, $\Gamma {}^{JK}{}_\nu $) and ($g_{\mu \nu }$, $T^\lambda
{}_{\rho \sigma }$) are equivalent in our approach (the argument in detail see \[16\]).
Subject to the metricity,[* *]{} the affine connection $\Gamma ^\lambda
{}_{\mu \nu }$ is related to the tetrad[* *]{} $e_I{}^\mu $ and the spin connection $\Gamma {}^{IJ}{}_\mu $ by $$\begin{aligned}
\Gamma ^\lambda {}_{\mu \nu } &=&e_I{}^\lambda \partial _\mu e^I{}_\nu
+e_J{}^\lambda e^I{}_\nu \Gamma {}^J{}_{I\mu } \nonumber \\
&=&\left\{ _\mu {}^\lambda {}_\nu \right\} +K^\lambda {}_{\mu \nu },\end{aligned}$$ where $\left\{ _\mu {}^\lambda {}_\nu \right\} $, $K^\lambda {}_{\mu \nu }$ are the Levi-Civita connection and the contortion, separately, with $$\begin{aligned}
K^\lambda {}_{\mu \nu } &=&-\frac 12\left( T^\lambda {}_{\mu \nu }+T_{\mu
\nu }{}^\lambda +T_{\nu \mu }{}^\lambda \right) , \nonumber \\
T^\lambda {}_{\mu \nu } &=&e_I{}^\rho T^I{}_{\mu \nu }=\Gamma ^\lambda
{}_{\mu \nu }-\Gamma ^\lambda {}_{\nu \mu }, \nonumber \\
T{}^I{}_{\mu \nu } &=&\partial _\mu e{}^I{}_\nu -\partial _\nu e{}^I{}_\mu
+\Gamma {}^I{}_{J\mu }e{}^J{}_\nu -\Gamma {}^I{}_{J\nu }e{}^J{}_\mu .\end{aligned}$$ Accordingly the curvature $R^\rho {}_{\sigma \mu \nu }$ can be represented as $$\begin{aligned}
R^\rho {}_{\sigma \mu \nu } &=&e_I{}^\rho e^J{}_\sigma R^I{}_{J\mu \nu
}=\partial _\mu \Gamma ^\rho {}_{\sigma \nu }-\partial _\nu \Gamma ^\rho
{}_{\sigma \mu }+\Gamma ^\rho {}_{\lambda \mu }\Gamma ^\lambda {}_{\sigma
\nu }-\Gamma ^\rho {}_{\lambda \nu }\Gamma ^\lambda {}_{\sigma \mu }
\nonumber \\
&=&R_{\left\{ {}\right\} }^\rho {}_{\sigma \mu \nu }+\partial _\mu K^\rho
{}_{\sigma \nu }-\partial _\nu K^\rho {}_{\sigma \mu }+K^\rho {}_{\lambda
\mu }K^\lambda {}_{\sigma \nu }-K^\rho {}_{\lambda \nu }K^\lambda {}_{\sigma
\mu } \nonumber \\
&&+\left\{ _\lambda {}^\rho {}_\mu \right\} K^\lambda {}_{\sigma \nu
}-\left\{ _\lambda {}^\rho {}_\nu \right\} K^\lambda {}_{\sigma \mu
}+\left\{ _\sigma {}^\lambda {}_\nu \right\} K^\rho {}_{\lambda \mu
}-\left\{ _\sigma {}^\lambda {}_\mu \right\} K^\rho {}_{\lambda \nu },\end{aligned}$$ where $R_{\left\{ {}\right\} }^\rho {}_{\sigma \mu \nu }=\partial _\mu
\left\{ _\sigma {}^\rho {}_\nu \right\} -\partial _\nu \left\{ _\sigma
{}^\rho {}_\mu \right\} +\left\{ _\lambda {}^\rho {}_\mu \right\} \left\{
_\sigma {}^\lambda {}_\nu \right\} -\left\{ _\lambda {}^\rho {}_\nu \right\}
\left\{ _\sigma {}^\lambda {}_\mu \right\} $ is the Riemann curvature of the Levi-Civita connection. In view of this, we can identify the actual degrees of freedom of the theory with the (independent) components of the metric $g_{\mu \nu }$ and the tensor $K^\lambda {}_{\mu \nu }$.
$\text{Cosmological equations}$
===============================
For the spatially flat Friedmann-Robertson-Walker metric $$g_{\mu \nu }=\text{diag}\left( -1,a\left( t\right) ^2,a\left( t\right)
^2,a\left( t\right) ^2\right) ,$$ the non-vanishing components of the Levi-Civita connection are $$\begin{aligned}
\left\{ _0{}^0{}_0\right\} &=&0,\left\{ _0{}^0{}_i\right\} =\left\{
_i{}^0{}_0\right\} =0,\left\{ _i{}^0{}_j\right\} =a\stackrel{\cdot }{a}\delta _{ij}, \nonumber \\
\left\{ _0{}^i{}_0\right\} &=&0,\left\{ _j{}^i{}_0\right\} =\left\{
_0{}^i{}_j\right\} =\frac{\stackrel{\cdot }{a}}a\delta _j^i,\left\{
_j{}^i{}_k\right\} =0,i,j,k,...=1,2,3.\end{aligned}$$ The non-vanishing torsion components with holonomic indices are given by two functions, the scalar torsion $h$ and the pseudoscalar torsion $f$ \[35\]: $$\begin{aligned}
T_{110} &=&T_{220}=T_{330}=a^2h, \nonumber \\
T_{123} &=&T_{231}=T_{312}=2a^3f,\end{aligned}$$ and then the contortion components are $$\begin{aligned}
K^1{}_{10} &=&K^2{}_{20}=K^3{}_{30}=0, \nonumber \\
K^1{}_{01} &=&K^2{}_{02}=K^3{}_{03}=h, \nonumber \\
K^0{}_{11} &=&K^0{}_{22}=K^0{}_{22}={}a^2h, \nonumber \\
K^1{}_{23} &=&K^2{}_{31}=K^3{}_{12}=-af, \nonumber \\
K^1{}_{32} &=&K^2{}_{13}=K^3{}_{21}=af.\end{aligned}$$
The non-vanishing components of the curvature $R^\rho {}_{\sigma \mu \nu }$ and the Ricci curvature $R{}_{\mu \nu }$ are
$$\begin{aligned}
R^0{}_{101} &=&R^0{}_{202}=R^0{}_{303}=a^2\left( \stackrel{{\bf \cdot }}{H}+H^2+Hh+\stackrel{\cdot }{h}\right) , \nonumber \\
R^0{}_{123} &=&-R^0{}_{213}=R^0{}_{312}=2a^3f\left( H+h\right) , \nonumber
\\
R^1{}_{203} &=&-R^1{}_{302}=R^2{}_{301}=-a\left( Hf+\stackrel{\cdot }{f}\right) , \nonumber \\
R^1{}_{212} &=&R^1{}_{313}=R^2{}_{323}=a^2\left( \left( H+h\right)
^2-f^2\right) ,\end{aligned}$$
$$\begin{aligned}
R{}_{00} &=&-3\stackrel{\cdot }{H}-3\stackrel{\cdot }{h}-3H^2-3Hh, \nonumber
\\
R{}_{11} &=&R{}_{22}=R{}_{33}=a^2\left( \stackrel{\cdot }{H}+\stackrel{\cdot
}{h}+3H^2+5Hh+2h^2-f^2\right) ,\end{aligned}$$
$$R{}=6\stackrel{\cdot }{H}+6\stackrel{\cdot }{h}+12H^2+18Hh+6h^2-3f^2,$$
where $H=\stackrel{\cdot }{a}\left( t\right) /a\left( t\right) $ is the Hubble parameter. Using these results and supposing the matter source is a fluid characterized by the energy density $\rho $, the pressure $p$ and the spin $s_{IJ}{}^\mu $ we obtain four independent equations from (2) and (3): $$\begin{aligned}
&&\left( H+h\right) ^2-f^2-\frac \rho 3 \nonumber \\
&&-\left( \beta +3\alpha \right) [4\left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) ^2+8H\left( H+h\right) \left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) \nonumber \\
&&-4h\left( h+2H\right) \left( h+H\right) ^2+8\left( h+H\right) ^2f^2-4f^4]
\nonumber \\
&&+2\gamma \left( 3h^2+4f^2\right) =0,\end{aligned}$$ $$\begin{aligned}
&&2\left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right)
+3H^2+4Hh+h^2-f^2-p \nonumber \\
&&+\left( \beta +3\alpha \right) [4\left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) ^2+8H\left( H+h\right) \left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) \nonumber \\
&&-\allowbreak 4h\left( h+2H\right) \left( h+H\right) ^2+8\left( h+H\right)
^2f^2+4f^4] \nonumber \\
&&-2\gamma \left( 2\stackrel{\cdot }{h}+8Hh+h^2{}+4f^2\right) =0,\end{aligned}$$ $$\begin{aligned}
&&\left( \beta +6\alpha \right) \left( \stackrel{\cdot \cdot }{H}+\stackrel{\cdot \cdot }{h}\right) +6\left( \beta +4\alpha \right) \left( H+h\right)
\stackrel{\cdot }{H}+\left( \allowbreak 5\beta +18\alpha \right) \left(
H+h\right) \stackrel{\cdot }{h}-4\left( \beta +3\alpha \right) f\stackrel{{\bf \cdot }}{f} \nonumber \\
&&+3\left( \beta +4\alpha \right) hH^2+\left( \allowbreak 5\beta +18\alpha
\right) h^2H+2\left( \beta +3\alpha \right) h^3-2\left( \beta +3\alpha
\right) hf^2+\frac 14h+\frac 12s_{01}{}^1=0,\end{aligned}$$ $$\begin{aligned}
&&f\{2\left( \beta +6\alpha \right) \left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) +6\left( \beta +\allowbreak 4\alpha \right) H^2+2\left(
5\beta +18\alpha \right) Hh \nonumber \\
&&+\left( \beta \ +3\alpha \right) \left( 4h^2-4f^2\right) -4\gamma +\frac 12\}-\frac 12s_{12}{}^3=0.\end{aligned}$$ The system of the equations (14)–(17) has the similar structure as the system of gravitational equations for homogeneous isotropic cosmological models in \[31\] except the coefficients. However, it is the differences in coefficients that make the system of the equations (14)–(17) easy to handle and possible to obtain some exact or analytic solutions in several cases as will be shown in the next sections.
The equations (14) and (15) can be written as $$\stackrel{\cdot }{H}+\stackrel{\cdot }{h}=2\gamma \stackrel{\cdot }{h}-2H^2+\left( 8\gamma -3\right) Hh-\left( 2\gamma {}+1\right) h^2+f^2+\frac 16\left( \rho +3p\right) ,$$ and $$\begin{aligned}
&&\left( \beta +3\alpha \right) [-4\left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) ^2-8\left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) H\left( H+h\right) \nonumber \\
&&+4h\left( h+2H\right) \left( h+H\right) ^2-8\left( h+H\right) ^2f^2+4f^4]
\nonumber \\
&&+\gamma \left( 6h^2+8f^2\right) +H^2+2Hh+h^2-f^2-\frac 13\rho =0.\end{aligned}$$
Since the spin orientation of particles in ordinary matter is random, the macroscopic spacetime average of the spin vanishes, we suppose $s_{IJ}{}^\lambda =0$, henceforth. Then, the equations (16), (17) become $$\begin{aligned}
&&\left( \beta +6\alpha \right) \left( \stackrel{\cdot \cdot }{H}+\stackrel{\cdot \cdot }{h}\right) +6\left( \beta +4\alpha \right) \left( H+h\right)
\stackrel{\cdot }{H}+\left( \allowbreak 5\beta +18\alpha \right) \left(
H+h\right) \stackrel{\cdot }{h}-4\left( \beta +3\alpha \right) f\stackrel{\cdot }{f} \nonumber \\
&&+3\left( \beta +4\alpha \right) hH^2+\left( \allowbreak 5\beta +18\alpha
\right) h^2H+2\left( \beta +3\alpha \right) h^3-2\left( \beta +3\alpha
\right) hf^2+\frac 14h=0,\end{aligned}$$ and $$\begin{aligned}
2 &&f\{2\left( \beta +6\alpha \right) \left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) +6\left( \beta +\allowbreak 4\alpha \right) H^2 \nonumber
\\
&&+2\left( 5\beta +18\alpha \right) Hh+\left( \beta \ +3\alpha \right)
\left( 4h^2-4f^2\right) -4\gamma +\frac 12\}=0.\end{aligned}$$ (21) has the solutions $$f=0,$$ and
$$\begin{aligned}
f^2 &=&\frac{\left( \beta +6\alpha \right) }{2\left( \beta \ +3\alpha
\right) }\left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) +\frac{3\left( \beta +\allowbreak 4\alpha \right) }{2\left( \beta \ +3\alpha
\right) }H^2+\frac{\left( 5\beta +18\alpha \right) }{2\left( \beta \
+3\alpha \right) }Hh+h^2 \nonumber \\
&&-\frac \gamma {\left( \beta \ +3\alpha \right) }+\frac 1{8\left( \beta \
+3\alpha \right) }.\end{aligned}$$
We will solve the equations (18-20) in the cases (22) and (23), respectively in the next two sections.
$\text{Exact de Sitter solutions with scalar torsion function}$
===============================================================
In the case $f=0$, (18) and (19) can be written as $$\stackrel{\cdot }{H}=\left( 2\gamma -1\right) \stackrel{\cdot }{h}-2H^2+\left( 8\gamma -3\right) Hh-\left( 2\gamma {}+1\right) h^2+\frac 16\left( \rho +3p\right) ,$$ and $$\left( \stackrel{\cdot }{h}+\left( 4H-h\right) h-\frac 1{2\gamma }\left(
H+h\right) ^2+\frac{\rho +3p}{12\gamma }\right) ^2-\frac{12\left( \beta
+3\alpha \right) \left( H+h\right) ^4+3\left( H+h\right) ^2+18h^2\gamma
-\rho }{48\gamma ^2\left( \beta +3\alpha \right) }=0,$$ which have the solutions
$$\stackrel{\cdot }{h}=-\left( 4H-h\right) h+\frac 1{2\gamma }\left(
H+h\right) ^2-\frac{\rho +3p}{12\gamma }\pm \sqrt{\frac{12\left( \beta
+3\alpha \right) \left( H+h\right) ^4+3\left( H+h\right) ^2+18h^2\gamma
-\rho }{48\gamma ^2\left( \beta +3\alpha \right) }},$$
$$\begin{aligned}
\stackrel{\cdot }{H} &=&-\frac 1{2\gamma }\left( H+h\right) ^2\allowbreak
-H^2+3Hh-h^2+\allowbreak \frac{\rho +3p}{12\gamma } \nonumber \\
&&\pm \left( 2\gamma -1\right) \sqrt{\frac{12\left( \beta +3\alpha \right)
\left( H+h\right) ^4+3\left( H+h\right) ^2+18h^2\gamma -\rho }{48\gamma
^2\left( \beta +3\alpha \right) }},\end{aligned}$$
Differentiating (24) gives $$\stackrel{\cdot \cdot }{H}=\left( 2\gamma -1\right) \stackrel{\cdot \cdot }{h}-4H\stackrel{\cdot }{H}+\left( 8\gamma -3\right) \stackrel{\cdot }{H}h+\left( 8\gamma -3\right) H\stackrel{\cdot }{h}-2\left( 2\gamma {}+1\right)
h\stackrel{\cdot }{h}+\frac 16\left( \stackrel{\cdot }{\rho }+3\stackrel{\cdot }{p}\right) .$$ The equation (20) has the form $$\begin{aligned}
&&\left( \beta +6\alpha \right) \left( \stackrel{\cdot \cdot }{H}+\stackrel{\cdot \cdot }{h}\right) +6\left( \beta +4\alpha \right) \left( H+h\right)
\stackrel{\cdot }{H}+\left( \allowbreak 5\beta +18\alpha \right) \left(
H+h\right) \stackrel{\cdot }{h} \nonumber \\
&&+3\left( \beta +4\alpha \right) hH^2+\left( \allowbreak 5\beta +18\alpha
\right) h^2H+2\left( \beta +3\alpha \right) h^3+\frac 14h=0.\end{aligned}$$ (28) and (29) have the solutions $$\begin{aligned}
\stackrel{\cdot \cdot }{H} &=&-\frac{\left( 48\alpha \gamma +12\beta \gamma
-2\beta \right) H+\left( 4\beta \gamma -3\beta -6\alpha \right) h}{2\gamma
\left( \beta +6\alpha \right) }\stackrel{\cdot }{H} \nonumber \\
&&-\frac{\left( 2\beta \gamma -2\beta -12\alpha \gamma \right) H+\left(
14\beta \gamma +60\alpha \gamma -3\beta -6\alpha \right) h}{2\gamma \left(
\beta +6\alpha \right) }\stackrel{\cdot }{h} \nonumber \\
&&-\frac{3\left( 2\gamma -1\right) \left( \beta +4\alpha \right) }{2\gamma
\left( \beta +6\alpha \right) }hH^2-\frac{\left( 2\gamma -1\right) \left(
5\beta +18\alpha \right) }{2\gamma \left( \beta +6\alpha \right) }h^2H-
\nonumber \\
&&\frac{\left( 2\gamma -1\right) \allowbreak \left( \beta +3\alpha \right) }{\gamma \left( \beta +6\alpha \right) }h^3-\frac{2\gamma -1}{8\gamma \left(
\beta +6\alpha \right) }h+\frac 1{12\gamma }\left( \stackrel{\cdot }{\rho }-3\stackrel{\cdot }{p}\right) ,\end{aligned}$$ $$\begin{aligned}
\stackrel{\cdot \cdot }{h} &=&-\frac{2\beta H+\left( 8\gamma \beta +48\gamma
\alpha +3\beta +6\alpha \right) h}{2\gamma \left( \beta +6\alpha \right) }\stackrel{\cdot }{H} \nonumber \\
&&-\frac{\left( 2\beta +8\gamma \beta +48\gamma \alpha \right) H+\left(
3\beta +6\alpha -4\gamma \beta -24\gamma \alpha \right) h}{2\gamma \left(
\beta +6\alpha \right) }\stackrel{\cdot }{h} \nonumber \\
&&-\frac{3\left( \beta +4\alpha \right) h}{2\gamma \left( \beta +6\alpha
\right) }H^2-\frac{\allowbreak 5\beta +18\alpha }{2\gamma \left( \beta
+6\alpha \right) }h^2H-\frac{\beta +3\alpha }{\gamma \left( \beta +6\alpha
\right) }h^3 \nonumber \\
&&-\frac 1{8\gamma \left( \beta +6\alpha \right) }h-\frac 1{12\gamma }\left(
\stackrel{\cdot }{\rho }-3\stackrel{\cdot }{p}\right) .\end{aligned}$$
Letting $$\stackrel{\cdot }{H}=X,\stackrel{\cdot }{h}=Y,$$ we have the dynamical system $$\begin{aligned}
\stackrel{\cdot }{X} &=&-\frac{\left( 48\alpha \gamma +12\beta \gamma
-2\beta \right) H+\left( 4\beta \gamma -3\beta -6\alpha \right) h}{2\gamma
\left( \beta +6\alpha \right) }X \\
&&-\frac{\left( 2\beta \gamma -2\beta -12\alpha \gamma \right) H+\left(
14\beta \gamma +60\alpha \gamma -3\beta -6\alpha \right) h}{2\gamma \left(
\beta +6\alpha \right) }Y\left( H,h\right) \\
&&-\frac{3\left( 2\gamma -1\right) \left( \beta +4\alpha \right) }{2\gamma
\left( \beta +6\alpha \right) }hH^2-\frac{\left( 2\gamma -1\right) \left(
5\beta +18\alpha \right) }{2\gamma \left( \beta +6\alpha \right) }h^2H- \\
&&\frac{\left( 2\gamma -1\right) \allowbreak \left( \beta +3\alpha \right) }{\gamma \left( \beta +6\alpha \right) }h^3-\frac{2\gamma -1}{8\gamma \left(
\beta +6\alpha \right) }h+\frac 1{12\gamma }\left( \stackrel{\cdot }{\rho }-3\stackrel{\cdot }{p}\right) ,\end{aligned}$$ $$\stackrel{\cdot }{H}=X,$$ $$\begin{aligned}
\stackrel{\cdot }{h} &=&Y\left( H,h\right) =\frac 1{2\gamma }\left(
H+h\right) ^2-\allowbreak \allowbreak \left( 4H-h\right) h-\frac 1{12\gamma }\left( \rho -3p\right) \nonumber \\
&&\pm \sqrt{\allowbreak \frac{12\left( \beta +3\alpha \right) \left(
H+h\right) ^4+3\allowbreak \left( H+h\right) ^2+18\gamma h^2-\rho }{48\gamma
^2\left( \beta +3\alpha \right) }}.\end{aligned}$$
The critical point equations consist of $$\begin{aligned}
&&-3\left( \beta +4\alpha \right) hH^2-\left( 5\beta +18\alpha \right) h^2H-
\\
&&2\allowbreak \left( \beta +3\alpha \right) h^3-\frac 14h+\frac{\beta
+6\alpha }{6\left( 2\gamma -1\right) }\left( \stackrel{\cdot }{\rho }-3\stackrel{\cdot }{p}\right) \\
&=&0,\end{aligned}$$ $$X=0,$$ $$\begin{aligned}
&&\left( H+h\right) ^2-2\gamma \allowbreak \allowbreak \left( 4H-h\right) h-\frac 16\left( \rho -3p\right) \nonumber \\
&&\pm \sqrt{\allowbreak \frac{12\left( \beta +3\alpha \right) \left(
H+h\right) ^4+3\allowbreak \left( H+h\right) ^2+18\gamma h^2-\rho }{12\left(
\beta +3\alpha \right) }}=0.\end{aligned}$$ In order to discuss the stability of the critical points we need to calculate the matrix elements of the Jacobian:
$$\begin{aligned}
\frac{\partial \stackrel{\cdot }{X}}{\partial X} &=&-\frac{\left( 48\alpha
\gamma +12\beta \gamma -2\beta \right) H+\left( 4\beta \gamma -3\beta
-6\alpha \right) h}{2\gamma \left( \beta +6\alpha \right) }, \\
\frac{\partial \stackrel{\cdot }{X}}{\partial H} &=&-\frac{\left( 48\alpha
\gamma +12\beta \gamma -2\beta \right) }{2\gamma \left( \beta +6\alpha
\right) }X-\frac{\left( 2\beta \gamma -2\beta -12\alpha \gamma \right) }{2\gamma \left( \beta +6\alpha \right) }Y\left( H,h\right) \\
&&-\frac{\left( 2\beta \gamma -2\beta -12\alpha \gamma \right) H+\left(
14\beta \gamma +60\alpha \gamma -3\beta -6\alpha \right) h}{2\gamma \left(
\beta +6\alpha \right) }\frac{\partial Y\left( H,h\right) }{\partial H} \\
&&-\frac{3\left( 2\gamma -1\right) \left( \beta +4\alpha \right) }{\gamma
\left( \beta +6\alpha \right) }hH-\frac{\left( 2\gamma -1\right) \left(
5\beta +18\alpha \right) }{2\gamma \left( \beta +6\alpha \right) }h^2,\end{aligned}$$
$$\begin{aligned}
\frac{\partial \stackrel{\cdot }{X}}{\partial h} &=&-\frac{\left( 4\beta
\gamma -3\beta -6\alpha \right) }{2\gamma \left( \beta +6\alpha \right) }X-\frac{\left( 14\beta \gamma +60\alpha \gamma -3\beta -6\alpha \right) }{2\gamma \left( \beta +6\alpha \right) }Y\left( H,h\right) \\
&&-\frac{\left( 2\beta \gamma -2\beta -12\alpha \gamma \right) H+\left(
14\beta \gamma +60\alpha \gamma -3\beta -6\alpha \right) h}{2\gamma \left(
\beta +6\alpha \right) }\frac{\partial Y\left( H,h\right) }{\partial h} \\
&&-\frac{3\left( 2\gamma -1\right) \left( \beta +4\alpha \right) }{2\gamma
\left( \beta +6\alpha \right) }H^2-\frac{\left( 2\gamma -1\right) \left(
5\beta +18\alpha \right) }{\gamma \left( \beta +6\alpha \right) }hH \\
&&-\frac{3\left( 2\gamma -1\right) \allowbreak \left( \beta +3\alpha \right)
}{\gamma \left( \beta +6\alpha \right) }h^2-\frac{2\gamma -1}{8\gamma \left(
\beta +6\alpha \right) }\end{aligned}$$
$$\frac{\partial \stackrel{\cdot }{H}}{\partial X}=1,\frac{\partial \stackrel{\cdot }{H}}{\partial H}=0,\frac{\partial \stackrel{\cdot }{H}}{\partial h}=0,\frac{\partial \stackrel{\cdot }{h}}{\partial X}=0,$$
$$\begin{aligned}
\frac{\partial \stackrel{\cdot }{h}}{\partial H} &=&\frac{\partial Y\left(
H,h\right) }{\partial H}=\frac 1\gamma H+\left( \frac 1\gamma -\allowbreak
\allowbreak 4\right) h \\
&&\pm \frac{8\left( \beta +3\alpha \right) \left( H+h\right) ^3+\allowbreak
H+h}{4\gamma \sqrt{\allowbreak 4\left( \beta +3\alpha \right) ^2\left(
H+h\right) ^4+\allowbreak \left( \beta +3\alpha \right) \left( H+h\right)
^2+6\left( \beta +3\alpha \right) \gamma h^2-\frac 13\left( \beta +3\alpha
\right) \rho }},\end{aligned}$$
$$\begin{aligned}
\frac{\partial \stackrel{\cdot }{h}}{\partial h} &=&\frac{\partial Y\left(
H,h\right) }{\partial h}=\left( \frac 1\gamma -\allowbreak \allowbreak
4\right) H+\left( \frac 1\gamma +2\right) h \nonumber \\
&&\pm \frac{8\left( \beta +3\alpha \right) \left( H+h\right) ^3+H+h+6\gamma h}{4\gamma \sqrt{\allowbreak 4\left( \beta +3\alpha \right) ^2\left(
H+h\right) ^4+\left( \beta +3\alpha \right) \allowbreak \left( H+h\right)
^2+6\left( \beta +3\alpha \right) \gamma h^2-\frac 13\left( \beta +3\alpha
\right) \rho }}.\end{aligned}$$
$\allowbreak $
In order to stress the role of the torsion as the source of the accelerating expansion of the universe we concentrate on the vacuum solutions for some special choices of the parameters $\alpha $, $\beta $, and $\gamma $. The critical point equations (34) can be simplified and solved exactly when $\beta =4\alpha $ or $\beta =3\alpha $.
$\label{0}$When $\protect\beta =-4\protect\alpha $
--------------------------------------------------
In this case the gravitational Lagrangian is a special case of quadratic curvature gravities \[36\] when the torsion vanishes.
In vacuum the dynamical system (33) becomes $$\begin{aligned}
\stackrel{\cdot }{X} &=&\left( 4h-\frac 2\gamma H-\frac 3{2\gamma }h\right)
X+\left( 5H-h-\frac 2\gamma H-\frac 3{2\gamma }h\right) Y\left( H,h\right) \\
&&+\frac{\left( 2\gamma -1\right) }{16\gamma \alpha }h\left( 8hH\alpha
+8\alpha h^2-1\right) ,\end{aligned}$$ $$\stackrel{\cdot }{H}=X,$$ $$\begin{aligned}
\stackrel{\cdot }{h} &=&Y\left( H,h\right) =\frac 1{2\gamma }\left(
H+h\right) ^2-\allowbreak \allowbreak \left( 4H-h\right) h \nonumber \\
&&\pm \frac 1{4\gamma }\sqrt{\allowbreak \frac{4\alpha \left( H+h\right)
^4-\allowbreak \left( H+h\right) ^2-6\gamma h^2}\alpha },\end{aligned}$$ and the critical point equations (34) become
$$h\left( 8hH\alpha +8\alpha h^2-1\right) =0,$$
$$X=0,$$
$$\begin{aligned}
&&\left( H+h\right) ^2-2\gamma \allowbreak \allowbreak \left( 4H-h\right) h
\nonumber \\
&&\pm \frac 12\sqrt{\allowbreak \frac{4\alpha \left( H+h\right)
^4-\allowbreak \left( H+h\right) ^2-6\gamma h^2}\alpha }=0.\end{aligned}$$
The equation (37) $$h\left( 8hH\alpha +8\alpha h^2-1\right) =0,$$ leads to $$h=0,$$ or $$8hH\alpha +8\alpha h^2-1=0.$$ Then we have two cases.
In the first case the solution $$h=0,H=0,\stackrel{\cdot }{H}=X=0,$$ corresponds to a static Minkowski spacetime.
In the second case, $h$ and $H$ satisfy the equations
$$8hH\alpha +8\alpha h^2-1=0,$$ and $$\begin{aligned}
&&\left( H+h\right) ^2-2\gamma \allowbreak \allowbreak \left( 4H-h\right) h
\\
&&\pm \frac 12\sqrt{\allowbreak \frac{4\alpha \left( H+h\right)
^4-\allowbreak \left( H+h\right) ^2-6\gamma h^2}\alpha }=0,\end{aligned}$$ which can be written as $$H=-h+\frac 1{8\alpha h},$$ and
$$25600\alpha ^3h^6\gamma ^2+128\gamma \alpha ^2h^4\left( 3-40\gamma \right)
+16\gamma \alpha h^2\left( 16\gamma +5\right) -8\gamma +1=0.$$
In order to obtain a concrete results we give some specific value of $\gamma
$.
When $$\gamma =4,$$ the equations (41) become $$H=-h+\frac 1{8\alpha h},$$ $\allowbreak $and $$409600\alpha ^3h^6-80384\alpha ^2h^4+4416\alpha h^2-31=0,$$ which have a real solution $$H=\frac{1.40162}{\sqrt{\alpha }},h=\frac{0.0841322}{\sqrt{\alpha }}.$$ The Jacobian matrix of the dynamical system (36) given by (35) is
$$M=\left(
\begin{array}{lll}
\allowbreak -\frac{0.\,39583}{\sqrt{\alpha }} & -\frac{2.\,093}\alpha &
\allowbreak -\frac{33.\,382}\alpha \\
1 & 0 & 0 \\
0 & -\frac{.\,33904}{\sqrt{\alpha }} & -\frac{5.\,4102}{\sqrt{\alpha }}\end{array}
\right) ,$$ which has the eigenvalues: $-0.\,85371/\sqrt{\alpha },\allowbreak -4.\,951/\sqrt{\alpha },\allowbreak -1.\,3522\times 10^{-3}/\sqrt{\alpha }$. If $\alpha >0$, all the real parts of the eigenvalues are negative. This means that the critical point $$X_c=0,H_c=\frac{1.40162}{\sqrt{\alpha }},h_c=\frac{0.0841322}{\sqrt{\alpha }},$$ is asymptotically stable and then $$\stackrel{\cdot }{H}=X=0,$$ gives an asymptotically stable de Sitter solution.
By the same way we can compute for
$$\gamma =2,$$
the dynamical system (36) has a real critical point
$$X_c=0,H_c=\frac{0.824837}{\sqrt{\alpha }},h_c=\frac{0.130802}{\sqrt{\alpha }},$$
the Jacobian has the eigenvalues: $-1.\,665/\sqrt{\alpha }+0.\,26071i/\sqrt{\alpha },\allowbreak -1.\,665/\sqrt{\alpha }-0.\,26071i/\sqrt{\alpha },\allowbreak -3.\,6641\times 10^{-3}/\sqrt{\alpha }$.
For $$\gamma =1,$$ (36) has a real critical point $$X_c=0,H_c=\frac{0.488003}{\sqrt{\alpha }},h_c=\frac{0.185576}{\sqrt{\alpha }},$$ the Jacobian has the eigenvalues: $-0.\,8932/\sqrt{\alpha }+0.\,97558i/\sqrt{\alpha },\allowbreak -0.\,8932/\sqrt{\alpha }-0.\,97558i/\sqrt{\alpha },\allowbreak -0.0\,4017/\sqrt{\alpha }$.
For $$\gamma =\frac 12,$$ (36) has a real critical point $$X_c=0,H_c=\frac{0.389146}{\sqrt{\alpha }},h_c=\frac{0.208985}{\sqrt{\alpha }},$$ the Jacobian has the eigenvalues: $-2.\,6618/\sqrt{\alpha },\allowbreak
-0.\,22666/\sqrt{\alpha },\allowbreak -3.\,5927\times 10^{-5}/\sqrt{\alpha }$. All the three critical points are asymptotically stable.
For $$\gamma =\frac 14,$$ (36) has a real critical point $$X_c=0,H_c=\frac{0.488003}{\sqrt{\alpha }},h=\frac{0.185576}{\sqrt{\alpha }},$$ the Jacobian has the eigenvalues: $-5.\,1634/\sqrt{\alpha },\allowbreak
-0.\,68424/\sqrt{\alpha },\allowbreak 3.\,8689\times 10^{-2}/\sqrt{\alpha }$. One of the eigenvalues is positive, which means that the critical point (52) is unstable. These examples illustrate that the stability of the critical points depends on $\gamma $, i.e. on the term $\gamma T{}^\mu
{}_{\nu \rho }T{}_\mu {}^{\nu \rho }$ in the action (1).
When $\protect\beta =-3\protect\alpha $
---------------------------------------
This corresponds to conformal (Weyl) gravity (resent see \[37\]) and Critical Gravity \[38\] when the torsion vanishes.
In this case the system of the equations (14)–(17) has the form
$$3H^2+6Hh+3\left( 6\gamma +1\right) h^2+3\left( 8\gamma -1\right) f^2-\rho =0,$$
$$2\stackrel{\cdot }{H}+2\left( 1-2\gamma \right) \stackrel{\cdot }{h}+3H^2+4\left( 1-4\gamma \right) Hh+\left( 1-2\gamma \right) h^2{}-\left(
1+8\gamma \right) f^2-p=0,$$
$$3\alpha \left( \stackrel{\cdot \cdot }{H}+\stackrel{\cdot \cdot }{h}\right)
+6\alpha \left( H+h\right) \stackrel{\cdot }{H}+3\alpha \left( H+h\right)
\stackrel{\cdot }{h}+3\alpha Hh\left( H+h\right) +\frac 14h+\frac 12s_{01}{}^1=0,$$
$$f\{6\alpha \left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) +6\alpha
H^2+6\alpha Hh-4\gamma +\frac 12\}-\frac 12s_{12}{}^3=0.$$
When $s_{IJ}{}^\mu =0$, the equation (56) leads to
$$f=0,$$ or
$$6\alpha \left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) +6\alpha
H^2+6\alpha Hh-4\gamma +\frac 12=0.$$
$\newline
\allowbreak $
We deal with only the first case $f=0$ in this section, then the equations (53)–(55) can be written as$\allowbreak $$$h=\frac{-H\pm \sqrt{-6\gamma H^2+\frac 13\left( 1+6\gamma \right) \rho }}{1+6\gamma },$$ $$\stackrel{\cdot }{h}=\frac 1{2\gamma -1}\stackrel{\cdot }{H}+\frac 3{2\left(
2\gamma -1\right) }H^2+\frac{2-8\gamma }{2\gamma -1}Hh-\frac 12h^2{}-\frac 1{2\left( 2\gamma -1\right) }p,$$ and $$\stackrel{\cdot \cdot }{H}+\stackrel{\cdot \cdot }{h}+2\left( H+h\right)
\stackrel{\cdot }{H}+\left( H+h\right) \stackrel{\cdot }{h}+hH\left(
H+h\right) +\frac 1{12\alpha }h=0,$$ Differentiating (59) gives $$\stackrel{\cdot \cdot }{h}=\frac 1{2\gamma -1}\stackrel{\cdot \cdot }{H}+\frac 3{2\gamma -1}H\stackrel{\cdot }{H}+\frac{2-8\gamma }{2\gamma -1}h\stackrel{\cdot }{H}+\frac{2-8\gamma }{2\gamma -1}H\stackrel{\cdot }{h}-h\stackrel{\cdot }{h}{}-\frac 1{2\left( 2\gamma -1\right) }\stackrel{\cdot }{p}.$$ Substituting (58), (59) and (61) into (60) yields
$$\begin{aligned}
\stackrel{\cdot \cdot }{H} &=&-\allowbreak \frac{2\left( 12\gamma ^2-8\gamma
-3\right) }{\left( 2\gamma -1\right) \left( 1+6\gamma \right) }H\stackrel{\cdot }{H}+\frac 2{\left( 6\gamma +1\right) }P\stackrel{\cdot }{H}\allowbreak
\nonumber \\
&&+\frac{\left( 636\gamma ^2-48\gamma -7\right) }{2\left( 2\gamma -1\right)
\left( 1+6\gamma \right) ^2}H^3-\allowbreak \frac{\left( 6\gamma -1\right)
\left( 26\gamma -5\right) }{\left( 2\gamma -1\right) \left( 1+6\gamma
\right) ^2}H^2P-\allowbreak \allowbreak \frac{10\gamma -3}{4\gamma \left(
1+6\gamma \right) ^2}HP^2 \nonumber \\
&&+\frac{\left( 2\gamma -1\right) H}{24\gamma \alpha \left( 6\gamma
+1\right) }-\frac{\left( 2\gamma -1\right) P}{24\gamma \alpha \left( 6\gamma
+1\right) }-\allowbreak \frac{6\gamma -1}{4\gamma \left( 2\gamma -1\right) }Hp+\frac 1{4\gamma }\stackrel{\cdot }{p},\end{aligned}$$
where $$P=\pm \sqrt{-6\gamma H^2+\frac 13\left( 1+6\gamma \right) \rho }.$$
Letting $$\stackrel{\cdot }{H}=X,$$ we have the dynamical system$\allowbreak $$$\begin{aligned}
\stackrel{\cdot }{H} &=&X, \nonumber \\
\stackrel{\cdot }{X} &=&-\frac{2\left( 12\gamma ^2-8\gamma -3\right) }{\left( 2\gamma -1\right) \left( 6\gamma +1\right) }HX+\frac 2{\left( 6\gamma
+1\right) }PX \nonumber \\
&&+\frac{\left( 636\gamma ^2-48\gamma -7\right) }{2\left( 2\gamma -1\right)
\left( 6\gamma +1\right) ^2}H^3-\allowbreak \frac{\left( 6\gamma -1\right)
\left( 26\gamma -5\right) }{\left( 2\gamma -1\right) \left( 6\gamma
+1\right) ^2}PH^2-\frac{10\gamma -3}{4\gamma \left( 6\gamma +1\right) ^2}P^2H
\nonumber \\
&&+\frac{\left( 2\gamma -1\right) H}{24\gamma \alpha \left( 6\gamma
+1\right) }-\frac{\left( 2\gamma -1\right) P}{24\gamma \alpha \left( 6\gamma
+1\right) }-\frac{6\gamma -1}{4\gamma \left( 2\gamma -1\right) }Hp+\frac 1{4\gamma }\stackrel{{\bf \cdot }}{p},\end{aligned}$$ with the matrix elements of its Jacobian: $$\begin{aligned}
\frac{\partial \stackrel{\cdot }{H}}{\partial H} &=&0,\frac{\partial
\stackrel{\cdot }{H}}{\partial X}=1, \nonumber \\
\frac{\partial \stackrel{\cdot }{X}}{\partial X} &=&-\allowbreak \frac{2\left( 12\gamma ^2-8\gamma -3\right) }{\left( 2\gamma -1\right) \left(
1+6\gamma \right) }H+\frac 2{\left( 6\gamma +1\right) }P, \nonumber \\
\frac{\partial \stackrel{\cdot }{X}}{\partial H} &=&-\allowbreak \frac{2\left( 12\gamma ^2-8\gamma -3\right) }{\left( 2\gamma -1\right) \left(
1+6\gamma \right) }X-\frac{12\gamma }{\left( 6\gamma +1\right) }\frac HPX\allowbreak \nonumber \\
&&+\allowbreak \frac{6\gamma \left( 6\gamma -1\right) \left( 26\gamma
-5\right) }{\left( 2\gamma -1\right) \left( 1+6\gamma \right) ^2}\frac{H^3}P+\frac{3\left( 676\gamma ^2-80\gamma -1\right) }{2\left( 2\gamma -1\right)
\left( 1+6\gamma \right) ^2}\allowbreak H^2 \nonumber \\
&&-\allowbreak \frac{2\left( 6\gamma -1\right) \left( 26\gamma -5\right) }{\left( 2\gamma -1\right) \left( 1+6\gamma \right) ^2}HP-\allowbreak
\allowbreak \frac{10\gamma -3}{4\gamma \left( 1+6\gamma \right) ^2}P^2
\nonumber \\
&&+\frac{2\gamma -1}{24\gamma \alpha \left( 6\gamma +1\right) }+\frac{2\gamma -1}{4\alpha \left( 6\gamma +1\right) }\frac HP-\allowbreak \frac{6\gamma -1}{4\gamma \left( 2\gamma -1\right) }p.\end{aligned}$$
The fixed point equations are
$$\begin{aligned}
X &=&0, \nonumber \\
&&\frac{\left( 636\gamma ^2-48\gamma -7\right) }{2\left( 2\gamma -1\right)
\left( 6\gamma +1\right) ^2}H^3-\allowbreak \frac{\left( 6\gamma -1\right)
\left( 26\gamma -5\right) }{\left( 2\gamma -1\right) \left( 6\gamma
+1\right) ^2}PH^2-\frac{10\gamma -3}{4\gamma \left( 6\gamma +1\right) ^2}P^2H
\nonumber \\
&&+\frac{\left( 2\gamma -1\right) H}{24\gamma \alpha \left( 6\gamma
+1\right) }-\frac{\left( 2\gamma -1\right) P}{24\gamma \alpha \left( 6\gamma
+1\right) }-\frac{6\gamma -1}{4\gamma \left( 2\gamma -1\right) }Hp+\frac 1{4\gamma }\stackrel{\cdot }{p} \nonumber \\
&=&0.\end{aligned}$$
Using (63) we obtain the equation $$\begin{aligned}
&&\frac{348\gamma ^2-48\gamma +1}{\left( 2\gamma -1\right) \left( 1+6\gamma
\right) ^2}H^3+\frac{\left( 2\gamma -1\right) H}{24\gamma \alpha \left(
6\gamma +1\right) }-\left( \allowbreak \frac{\left( 6\gamma -1\right) \left(
26\gamma -5\right) }{\left( 2\gamma -1\right) \left( 6\gamma +1\right) ^2}H^2+\frac{\left( 2\gamma -1\right) }{24\gamma \alpha \left( 6\gamma
+1\right) }\right) P \nonumber \\
&&-\frac{10\gamma -3}{12\gamma \left( 6\gamma +1\right) }H\rho -\frac{6\gamma -1}{4\gamma \left( 2\gamma -1\right) }Hp+\frac 1{4\gamma }\stackrel{\cdot }{p} \nonumber \\
&=&0.\end{aligned}$$
In vacuum $$\rho =p=0,P=\nu H,\nu =\pm \sqrt{-6\gamma },$$ The fixed point equation (68) becomes $$\frac{348\gamma ^2-48\gamma +1}{\left( 2\gamma -1\right) \left( 1+6\gamma
\right) ^2}H^3+\frac{\left( 2\gamma -1\right) H}{24\gamma \alpha \left(
6\gamma +1\right) }-\left( \allowbreak \frac{\left( 6\gamma -1\right) \left(
26\gamma -5\right) }{\left( 2\gamma -1\right) \left( 6\gamma +1\right) ^2}H^2+\frac{2\gamma -1}{24\gamma \alpha \left( 6\gamma +1\right) }\right) \nu
H=0,$$ which leads to $$H=0,$$ or $$H^2=-\frac{\left( \nu -1\right) \left( 2\gamma -1\right) ^2\left( 1+6\gamma
\right) }{24\gamma \alpha \left( \left( 156\gamma ^2-56\gamma +5\right) \nu
-348\gamma ^2+48\gamma -1\right) }.$$
In the first cas, we have $$\stackrel{\cdot }{H}=X=0,H=0,h=0,$$ which correspond to a static Minkowski solution.
In the second case $$H^2=-\frac{\left( \nu -1\right) \left( 2\gamma -1\right) ^2\left( 1+6\gamma
\right) }{24\gamma \alpha \left( \left( 156\gamma ^2-56\gamma +5\right) \nu
-348\gamma ^2+48\gamma -1\right) },$$ according to (66) the Jacobian matrix has the form $$M=\left(
\begin{array}{ll}
0 & 1 \\
b & c\end{array}
\right) ,$$ with $$\begin{aligned}
b &=&\allowbreak \frac{6\gamma \left( 6\gamma -1\right) \left( 26\gamma
-5\right) }{\left( 2\gamma -1\right) \left( 1+6\gamma \right) ^2}\frac{H^2}\nu +\frac{3\left( 676\gamma ^2-80\gamma -1\right) }{2\left( 2\gamma
-1\right) \left( 1+6\gamma \right) ^2}\allowbreak H^2 \nonumber \\
&&-\allowbreak \frac{2\left( 6\gamma -1\right) \left( 26\gamma -5\right) }{\left( 2\gamma -1\right) \left( 1+6\gamma \right) ^2}\nu H^2-\allowbreak
\allowbreak \frac{10\gamma -3}{4\gamma \left( 1+6\gamma \right) ^2}\nu ^2H^2
\nonumber \\
&&+\frac{2\gamma -1}{24\gamma \alpha \left( 6\gamma +1\right) }+\frac{2\gamma -1}{4\alpha \left( 6\gamma +1\right) \nu }, \nonumber \\
c &=&-\frac{2\left( 12\gamma ^2-8\gamma -3-2\nu \gamma +\nu \right) }{\left(
2\gamma -1\right) \left( 1+6\gamma \right) }H.\end{aligned}$$ and has the eigenvalues: $\frac 12c+\frac 12\sqrt{\left( c^2+4b\right) },\allowbreak \frac 12c-\frac 12\sqrt{\left( c^2+4b\right) }$.
If $$\gamma =-\frac 1{24},$$ (69), (74) and (76) give, separately, $$\nu =-\frac 12,$$ $$H^2=\frac{169}{948\alpha },$$ $\allowbreak $ $$b=-\frac{408901}{30336\alpha },c=-\frac{13}7\sqrt{\frac{169}{948\alpha }}.$$ Then the Jacobian matrix has the eigenvalues:
$$\begin{aligned}
\frac 12c+\frac 12\sqrt{\left( c^2+4b\right) } &=&-\frac{13}{14\sqrt{\alpha }}\sqrt{\frac{169}{948}}+\frac i{2\sqrt{\alpha }}\sqrt{-\frac{19807661}{371616}},\allowbreak \\
\frac 12c-\frac 12\sqrt{\left( c^2+4b\right) } &=&-\frac{13}{14\sqrt{\alpha }}\sqrt{\frac{169}{948}}-\frac i{2\sqrt{\alpha }}\sqrt{-\frac{19807661}{371616}}.\end{aligned}$$
If $\alpha >0$, all the real parts of the two eigenvalues are negative. This means that the critical point $$X_c=0,H_c=\sqrt{\frac{169}{948\alpha }},$$ is asymptotically stable and then $$\stackrel{\cdot }{H}=X=0,$$ gives an asymptotically stable de Sitter solution. (58) gives $$h=-\frac{13}{711}\frac{\sqrt{237}}{\sqrt{\alpha }},$$ or
$$h=-\frac{13}{237}\frac{\sqrt{237}}{\sqrt{\alpha }}.$$
We have seen that by appropriate choices of $\gamma $, we can obtain asymptotically stable de Sitter solutions in both cases, $\beta =4\alpha $ and $\beta =3\alpha $, though the cosmological equations have different structure. This means that the structure of the cosmological equations depends on $\beta $, while the stability of the solutions depends on $\gamma
.$
In all of the solutions obtained the torsion function $h$ does not vanish in vacuum, which means that the torsion is an intrinsic geometric nature of the spacetime. It is the torsion that causes the accelerating expansion of the universe in vacuum.
$\text{Analytic solutions with pseudoscalar torsion function}$
==============================================================
Differentiating (23) gives
$$f\stackrel{\cdot }{f}=\frac{\beta +6\alpha }{4\left( \beta +3\alpha \right) }\left( \stackrel{\cdot \cdot }{H}+\stackrel{\cdot \cdot }{h}\right) +\frac{3\left( \beta +\allowbreak 4\alpha \right) }{2\left( \beta +3\alpha \right) }H\stackrel{\cdot }{H}+\frac{5\beta +18\alpha }{4\left( \beta +3\alpha
\right) }\stackrel{\cdot }{H}h+\frac{5\beta +18\alpha }{4\left( \beta
+3\alpha \right) }H\stackrel{\cdot }{h}+h\stackrel{\cdot }{h}.$$
Substituting (23) and (79) into (20) gives
$$h=0.$$
Then the equations (18), (19) and (23) become $$\stackrel{\cdot }{H}=-2H^2+f^2+\frac 16\left( \rho +3p\right) ,$$ $$\left( \beta +3\alpha \right) [-4\stackrel{\cdot }{H}^2-8\stackrel{\cdot }{H}H^2-8H^2f^2+4f^4]+\left( 8\gamma -1\right) f^2+H^2-\frac 13\rho =0,$$ $$f^2=\frac{\left( \beta +6\alpha \right) }{2\left( \beta \ +3\alpha \right) }\stackrel{\cdot }{H}+\frac{3\left( \beta +\allowbreak 4\alpha \right) }{2\left( \beta \ +3\alpha \right) }H^2-\frac \gamma {\left( \beta \ +3\alpha
\right) }+\frac 1{8\left( \beta \ +3\alpha \right) }.$$ They have the solutions $$H^2\allowbreak =\frac{\left( 8\gamma -1\right) ^2}{32\gamma \beta }+\frac 1{24\gamma }\rho -\frac{\left( 8\gamma -1\right) \left( \beta +4\alpha \right)
}{16\gamma \beta }\left( \rho +3p\right) +\frac{\left( \beta +3\alpha
\right) \left( \beta +4\alpha \right) }{24\gamma \beta }\left( \rho
+3p\right) ^2,$$
$$\stackrel{{\bf \cdot }}{H}=\allowbreak -\frac{\left( 8\gamma -1\right)
\left( 16\gamma -1\right) }{32\gamma \beta }-\frac 1{24\gamma }\rho +\frac{40\gamma \beta +144\alpha \gamma -3\beta -12\alpha }{48\gamma \beta }\left(
\rho +3p\right) -\frac{\left( \beta +3\alpha \right) \left( \beta +4\alpha
\right) }{24\gamma \beta }\left( \rho +3p\right) ^2$$
$$f^2=\allowbreak \frac{1-8\gamma }{32\gamma \beta }+\frac 1{24\gamma }\rho -\frac{16\gamma \left( \beta +3\alpha \right) -3\left( \beta +4\alpha \right)
}{48\gamma \beta }\left( \rho +3p\right) +\frac{\left( \beta +3\alpha
\right) \left( \beta +4\alpha \right) }{24\gamma \beta }\left( \rho
+3p\right) ^2.$$
$\allowbreak $
The equations (84) and (85) ply the roles of the Friedmann equation and the Raychaudhuri equation in General Relativity. The equation (86) indicates that even in vacuum the spacetime possesses the torsion $f=\sqrt{\allowbreak
\frac{1-8\gamma }{32\gamma \beta }}$, which has been found in \[32\]. Hence the conception of the vacuum as physical notion is changed essentially. Instead of it as passive receptacle of physical objects and processes, the vacuum assumes a dynamical properties as a gravitating object. The combination of (84) and (85) yields the acceleration equation
$$\frac{\stackrel{\cdot \cdot }{a}}a=-\frac{8\gamma -1}{4\beta }+\frac{\beta \
+3\alpha }{3\beta }\left( \rho +3p\right) .$$
Letting $$\beta =n\alpha ,$$ we have $$\frac{\stackrel{\cdot \cdot }{a}}a=\frac{1-8\gamma }{4n\alpha }+\frac{n\ +3}{3n}\left( \rho +3p\right) .$$ Some important consequences can be obtained from (88):
i\) The term $\frac{1-8\gamma }{4n\alpha }$ plies the role of the cosmological constant, which agrees with the result in \[31\]. If $\frac{1-8\gamma }{4n\alpha }>0,$ $\rho =p=0,$ then $\stackrel{\cdot \cdot }{a}>0$, the [*acceleration*]{} of cosmological expansion [*acquires the vacuum origin*]{}.
ii\) If $$n>0,\text{ or }n<-3$$ $\rho +3p$ accelerates the expansion of the universe. If $$-3<n<0,$$ $\rho +3p$ decelerates the expansion of the universe. Especially, when $n=-2,\gamma =1/8$, (88) becomes the acceleration equation in general relativity. In other words, the latter is only a special case of the former.
iii\) If $$n>0,\gamma >\frac 18,\text{ or }n<-3,\gamma <\frac 18,$$ the universe can undergo a phase transformation from an accelerating to a decelerating expansion.
iv\) If $$-3<n<0,\gamma >\frac 18,$$ the universe can undergo a phase transformation from a decelerating to an accelerating expansion.
We find this picture very appealing and physical since it seems to indicate that in metric-affine gravity as matter tells spacetime how to curve, matter will also tell spacetime how to twirl.
$\text{Analytic solutions in the case $\protect\gamma =0$}$
===========================================================
In the last two sections we have seen that the coefficient $\gamma $ plies a central role in determining the behavior of the scale factor and the evolution of the universe. In order to investigate this point thoroughly, we discuss a extreme case, $\gamma =0$. In this case the equations (14)–(17) take the form
$$\begin{aligned}
&&\left( H+h\right) ^2-f^2-\frac \rho 3 \nonumber \\
&&+\left( \beta +3\alpha \right) [-4\left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) ^2-8H\left( H+h\right) \left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) \nonumber \\
&&+4h\left( h+2H\right) \left( h+H\right) ^2-8\left( h+H\right)
^2f^2+4f^4]=0,\end{aligned}$$
$$\begin{aligned}
&&2\left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right)
+3H^2+4Hh+h^2-f^2-p \nonumber \\
&&-\left( \beta +3\alpha \right) [-4\left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) ^2-8\left( H^2+Hh\right) \left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) \nonumber \\
&&+\allowbreak 4h\left( h+2H\right) \left( h+H\right) ^2-8\left( h+H\right)
^2f^2+4f^4]=0,\end{aligned}$$
$$\begin{aligned}
&&\left( \beta +6\alpha \right) \left( \stackrel{\cdot \cdot }{H}+\stackrel{\cdot \cdot }{h}\right) +6\left( \beta +4\alpha \right) \left( H+h\right)
\stackrel{\cdot }{H}+\left( \allowbreak 5\beta +18\alpha \right) \left(
H+h\right) \stackrel{\cdot }{h}-4\left( \beta +3\alpha \right) f\stackrel{\cdot }{f} \nonumber \\
&&+3\left( \beta +4\alpha \right) hH^2+\left( \allowbreak 5\beta +18\alpha
\right) h^2H+2\left( \beta +3\alpha \right) h^3-2\left( \beta +3\alpha
\right) hf^2+\frac 14h+\frac 12s_{01}{}^1=0,\end{aligned}$$
$$\begin{aligned}
&&f\{2\left( \beta +6\alpha \right) \left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) +6\left( \beta +\allowbreak 4\alpha \right) H^2 \nonumber
\\
&&+2\left( 5\beta +18\alpha \right) Hh+\left( \beta \ +3\alpha \right)
\left( 4h^2-4f^2\right) +\frac 12\}-\frac 12s_{12}{}^3=0.\end{aligned}$$
\(93) and (94) can be written as $$\stackrel{\cdot }{H}+\stackrel{\cdot }{h}=-2H^2-3Hh-h^2+f^2+\frac 16\left(
\rho +3p\right) ,$$ $$\begin{aligned}
&&3\left( h+H\right) ^2-3f^2-\rho \nonumber \\
&&+4\left( \beta +3\alpha \right) \left( \left( H+h\right) ^2-f^2\right)
\left( \rho +3p\right) -\frac 13\left( \beta +3\alpha \right) \left( \rho
+3p\right) ^2 \nonumber \\
&=&0,\end{aligned}$$ Differentiating (97) gives $$\stackrel{\cdot \cdot }{H}+\stackrel{\cdot \cdot }{h}=-4H\stackrel{\cdot }{H}-3h\stackrel{\cdot }{H}-3H\stackrel{\cdot }{h}-2h\stackrel{\cdot }{h}+2f\stackrel{\cdot }{f}+\frac 16\left( \stackrel{\cdot }{\rho }+3\stackrel{\cdot }{p}\right) .$$ Substituting (97) and (99) into (95) and (96) yields
$$\begin{aligned}
&&-\beta \left( h+4H\right) \left( h+H\right) ^2+\allowbreak \beta \left(
h+2H\right) f^2-\allowbreak 2\beta f\stackrel{\cdot }{f}+\frac 14h\allowbreak +\frac 16\left( 2\beta H+6\alpha h+3\beta h\right) \left( \rho
+3p\right) \nonumber \\
&&+\allowbreak \frac 16\left( \beta +6\alpha \right) \left( \stackrel{\cdot }{\rho }+3\stackrel{\cdot }{p}\right) +\frac 12s_{01} \nonumber \\
&=&0.\end{aligned}$$
and $$f\{2\beta \left( h+H\right) ^2-2\beta f^2+\frac 13\left( \beta +6\alpha
\right) \left( \rho +3p\right) +\frac 12\}-\frac 12s_{12}{}^3=0.$$ If $$s_{IJ}{}^\mu =0,$$ (101) leads to $$f=0,$$ or $$2\beta \left( h+H\right) ^2-2\beta f^2+\frac 13\left( \beta +6\alpha \right)
\left( \rho +3p\right) +\frac 12=0.$$
When $f=0$
----------
The equations (98) and (100) become $$\left( H+h\right) ^2=\frac{\frac 13\left( \beta +3\alpha \right) \left( \rho
+3p\right) ^2+\rho }{4\left( \beta +3\alpha \right) \left( 3p+\rho \right) +3},$$ and $$\begin{aligned}
&&\frac 14h+\frac 16\allowbreak \left( 2\beta H+3\beta h+6\alpha h\right)
\left( \rho +3p\right) \allowbreak +\left( \beta +6\alpha \right) \left(
\stackrel{\cdot }{\rho }+3\stackrel{\cdot }{p}\right) \nonumber \\
&&-\beta \left( h+4H\right) \left( H+h\right) ^2=0,\end{aligned}$$ which have the solutions $$\begin{aligned}
H &=&\frac{\frac 34+\allowbreak \frac 32\beta \left( \rho +5p\right)
+\allowbreak 6\alpha \left( \rho +3p\right) +\allowbreak \frac 13\left(
\beta +3\alpha \right) \left( 5\beta +12\alpha \right) \left( \rho
+3p\right) ^2}{\frac 34+\allowbreak \frac 92\beta \left( \rho +p\right)
+\allowbreak 6\alpha \left( \rho +3p\right) +\allowbreak \frac 13\left(
\beta +3\alpha \right) \left( 5\beta +12\alpha \right) \left( \rho
+3p\right) ^2}J \\
&&+\frac{4\left( \beta +3\alpha \right) \left( \beta +6\alpha \right) \left(
\rho +3p\right) +\allowbreak 3\left( \beta +6\alpha \right) }{\frac 34+\allowbreak \frac 92\beta \left( \rho +p\right) +\allowbreak 6\alpha \left(
\rho +3p\right) +\allowbreak \frac 13\left( \beta +3\alpha \right) \left(
5\beta +12\alpha \right) \left( \rho +3p\right) ^2}\left( \stackrel{\cdot }{\rho }+3\stackrel{\cdot }{p}\right) , \nonumber\end{aligned}$$
$$\begin{aligned}
h &=&\frac{3\beta \left( \rho -p\right) }{\frac 34+\allowbreak \frac 92\beta
\left( \rho +p\right) +\allowbreak 6\alpha \left( \rho +3p\right)
+\allowbreak \frac 13\left( \beta +3\alpha \right) \left( 5\beta +12\alpha
\right) \left( \rho +3p\right) ^2}J \nonumber \\
&&-\frac{4\left( \beta +3\alpha \right) \left( \beta +6\alpha \right) \left(
\rho +3p\right) +\allowbreak 3\left( \beta +6\alpha \right) }{\frac 34+\allowbreak \frac 92\beta \left( \rho +p\right) +\allowbreak 6\alpha \left(
\rho +3p\right) +\allowbreak \frac 13\left( \beta +3\alpha \right) \left(
5\beta +12\alpha \right) \left( \rho +3p\right) ^2}\left( \stackrel{\cdot }{\rho }+3\stackrel{\cdot }{p}\right) ,\end{aligned}$$
with $$J=\pm \sqrt{\frac{\frac 13\left( \beta +3\alpha \right) \left( \rho
+3p\right) ^2+\rho }{4\left( \beta +3\alpha \right) \left( 3p+\rho \right) +3}}.$$ The equation (97) now becomes $$\stackrel{\cdot }{H}+\stackrel{\cdot }{h}=-\left( 2H+h\right) \left(
H+h\right) +\frac 16\left( \rho +3p\right) .$$
Letting $$\beta =n\alpha ,p=w\rho ,$$ (105), (106) and (107) can be written as
$$H=\frac{\left( \frac 34+\allowbreak \frac 32A\alpha \rho +\frac 13B\alpha
^2\rho ^2\right) J+\left( 4D\alpha \rho +\allowbreak G\right) \alpha \left(
1+3w\right) \stackrel{\cdot }{\rho }}{\frac 34+\allowbreak \frac 32C\alpha
\rho +\frac 13B\alpha ^2\rho ^2},$$
$$h=\frac{3n\left( 1-w\right) \alpha \rho J-\left( 4D\alpha \rho +G\right)
\alpha \left( 1+3w\right) \stackrel{\cdot }{\rho }}{\frac 34+\frac 32C\alpha
\rho +\allowbreak \frac 13B\alpha ^2\rho ^2},$$
$$J=\pm \sqrt{\frac{\frac 13E\alpha \rho +1}{4F\alpha \rho +3}\rho },$$
where $$\begin{aligned}
A &=&n+4+\left( 5n+12\right) w,B=\left( n+3\right) \left( 5n+12\right)
\left( 1+3w\right) ^2, \nonumber \\
C &=&3n+4+\left( 3n+12\right) w,D=\left( n+3\right) \left( n+6\right) \left(
1+3w\right) , \nonumber \\
E &=&\left( n+3\right) \left( 1+3w\right) ^2,F=\left( n+3\right) \left(
3w+1\right) ,G=3\left( n+6\right) .\end{aligned}$$ The equation (108) gives the acceleration equation
$$\frac{\stackrel{\cdot \cdot }{a}}a=-\stackrel{\cdot }{h}\allowbreak -\left(
H+h\right) ^2-hH+\frac 16\left( 1+3w\right) \rho ,$$
Let us consider two special cases.
i\) For the early universe, $$\alpha \rho \gg 1,$$ we compute using (110)– (114) and obtain approximately $$\frac{\stackrel{\cdot \cdot }{a}}a=\allowbreak \frac 1{12}\left( 1+3w\right)
\rho .$$ This represents an inflation universe.
ii\) For the later epoch $$\alpha \rho \ll 1,$$ we have $$\frac{\stackrel{\cdot \cdot }{a}}a=-\allowbreak \frac 16\rho \left(
1-3w\right) .$$ This represents a uniformly expanding universe if $w=1/3$ (radiation epoch), or an decelerating universe if $w<1/3$ (matter epoch).
when $f\neq 0$
--------------
The function $f$ satisfies the equation (102), which yields $$f^2=\left( H+h\right) ^2+\frac{\beta +6\alpha }{6\beta }\left( \rho
+3p\right) +\frac 1{4\beta },$$ and $$f\stackrel{\cdot }{f}=\left( H+h\right) \left( \stackrel{\cdot }{H}+\stackrel{\cdot }{h}\right) +\frac{\beta +6\alpha }{12\beta }\left( \stackrel{\cdot }{\rho }+3\stackrel{\cdot }{p}\right) .$$ Using (97) we have $$\stackrel{\cdot }{H}+\stackrel{\cdot }{h}=-H\left( H+h\right) +\frac{\beta
+3\alpha }{3\beta }\left( \rho +3p\right) +\frac 1{4\beta },$$ $$f\stackrel{\cdot }{f}=-H\left( H+h\right) ^2+\frac{\beta +3\alpha }{3\beta }\left( H+h\right) \left( \rho +3p\right) +\frac 1{4\beta }\left( H+h\right) +\frac{\beta +6\alpha }{12\beta }\left( \stackrel{\cdot }{\rho }+3\stackrel{\cdot }{p}\right) ,$$ substituting in to (98) and (100) we have $$-\frac{\beta +4\alpha }{3\beta }\left( \rho +3p\right) ^2-\frac{\beta
+4\alpha }{2\beta \left( \beta +3\alpha \right) }\left( \rho +3p\right) -\frac \rho {3\left( \beta +3\alpha \right) }-\frac 1{4\beta \left( \beta
+3\alpha \right) }=0,$$ $$0=0.$$ So the field equations have no definite solution.
The results obtain above indicates that if $\gamma =0$ in both cases, $f=0$, and $f\neq 0$, there exists no solution describing an accelerating universe. In other words, the term $\gamma T{}^\mu {}_{\nu \rho }T{}_\mu {}^{\nu \rho
} $ is necessary to the existence of the solutions describing an accelerating universe.
$\text{Conclusions}$
====================
Quadratic theories of gravity described by the Lagrangian $R+\alpha
R^2+\beta R_{\mu \nu }R^{\mu \nu }$ have been studied in many works in supergravity, quantum gravity, string theory and M-theory. However, the cosmology in these theories has not been explored extensively, especially, when the torsion of the spacetime is considered. In this paper we show that by only allowing the connection to be asymmetrical and adding a term $\gamma
T{}^\mu {}_{\nu \rho }T{}_\mu {}^{\nu \rho }$ to the Lagrangian $R+\alpha
R^2+\beta R_{\mu \nu }R^{\mu \nu }$ some meaningful cosmological solutions can be obtained. These solutions provide several possible explanations to the acceleration of the cosmological expansion without a cosmological constant or dark energy. One can find that although the field equation (2) returns to Einstein’s equation when $\alpha =\beta =\gamma =0$, the cosmological equations (18-21) are essentially different from the Friedmann equation and the Raychaudhuri equation and then give different description to the evolution of the universe. The acceleration equation of the universe in general relativity is only a special case of the equation (87). These equations involving higher-derivatives can be solved by appropriate choice of $\beta $ and $\gamma $. Not only numbers of asymptotically stable de Sitter solutions expressed by critical points of a dynamical system but also exact analytic solutions are obtained. These solutions indicate that the terms $\beta R_{\mu \nu }R^{\mu \nu }$ and $\gamma T{}^\mu {}_{\nu \rho
}T{}_\mu {}^{\nu \rho }$ ply the different roles: the former determines the structure of the equations while the latter determines the behavior and the stability of the solutions. To construct a model of cosmic acceleration the Lagrangian $R+\alpha R^2+\beta R_{\mu \nu }R^{\mu \nu }+$ $\gamma T{}^\mu
{}_{\nu \rho }T{}_\mu {}^{\nu \rho }$ is sufficient and necessary.
Owing to the solutions obtained some conceptions have to be changed essentially. According to these solutions, even in vacuum the spacetime can possesses torsion and curvature. Therefore, instead of vacuum as passive receptacle of physical objects and processes, the vacuum assumes a dynamical property as a gravitating object. It is the torsion of the spacetime that causes the acceleration of the cosmological expansion in vacuum. Both the torsion and the accelerating expansion possess geometrical nature and do not invoke any matter origin. Furthermore, the energy and pressure of the ordinary matter can produce the torsion of the spacetime and cause either the deceleration or the acceleration of the cosmological expansion depending on choices of $\beta $ and $\gamma $.
[\*]{}
Capozziello S and De Laurentis M, arXiv:1108.6266 \[gr-qc\]
De Felice A and Tsujikawa S 2010 [*Living Rev. Rel*]{}. [**13**]{} 3
Sotiriou T P and Faraoni V 2010 [*Rev. Mod. Phys*]{}. [**82**]{} 451
Nojiri S and Odintsov S D 2011 [*Phys. Rept.*]{} [**505**]{} 59
Sotiriou T P 2009 [*J. Phys. Conf. Ser.*]{}[**189**]{} 012039 (arXiv:0810.5594 \[gr-qc\])
Nojiri S and Odintsov S D 2007 [*Int. J. Geom. Meth. Mod. Phys*]{}. [**4**]{} 115
Buchdahl H A 1970 [*Mon. Not. Roy. Ast. Soc*]{}. [**150**]{} 1
Capozziello S and Francaviglia M 2008 [*Gen. Rel. Grav.*]{} [**40**]{} 357
Sotiriou T P 2009 [*Class. Quant. Grav*]{}. [**26**]{} 152001
Vitagliano V, Sotiriou T P and Liberati S 2010 [*Phys. Rev*]{}. D[**82**]{} 084007 [** **]{}
Barragan C and Olmo G J 2010[** **]{}[*Phys. Rev*]{}. D[**82**]{} 084015
Olmo G J, Sanchis-Alepuz H and Tripathi S 2009 [*Phys. Rev*]{}. D[**80**]{} 024013 [** **]{}
Barragan C, Olmo G J and Sanchis-Alepuz H 2009 [*Phys. Rev*]{}. D[**80**]{} 024016 [** **]{}
Olmo G J 2011 [*Int. J. Mod. Phys*]{}. D[**20**]{} 413 [* *]{}
Sotiriou T P and Liberati S 2007 [*Annals Phys.*]{} [**322**]{} 935
Sotiriou T P and Liberati S 2007 [*J. Phys. Conf. Ser.*]{} [**68**]{} 012022
Sotiriou T P 2009 [*Class. Quant. Grav.*]{} [**26**]{} 152001
Vitagliano[** **]{}V, Sotiriou T P and Liberati S 2010 [*Phys. Rev.*]{} D[**82**]{} 084007
Koivisto[** **]{}T S[** **]{}2011 [*Phys. Rev.* ]{}D[**83**]{} 101501
Olmo G J, Sanchis-Alepuz H and Tripathi S arXiv:1002.3920 \[gr-qc\].[** **]{}
Capozziello S, Cianci R, Stornaiolo C and Vignolo S 2007 [*Class. Quant. Grav*]{}. [**24**]{} 6417
Sotiriou T P 2009 [*Class. Quant. Grav*]{}. [**26**]{} 152001
Vitagliano V, Sotiriou T P and Liberati S 2011[* *]{}[*Annals Phys.*]{} [**326**]{} 1259 [** **]{}
Shapiro I 2002 [*Phys. Rept.*]{} [**357**]{} 113
Hammond R 2002 [*Rept. Prog. Phys*]{}. [**65**]{} 599
Hehl F W and Obukhov Y N arXiv:0711.1535 \[gr-qc\]
Hammond R 2010 [*Gen. Rel. Grav*]{}. [**42**]{} 2345
Jimenez J B and Koivisto T S arXiv:1201.4018 \[gr-qc\]
Aldrovandi R, Pereira J G and Vu K H 2004 [*Braz. J. Phys.*]{} [**34**]{} 1374 (arXiv:gr-qc/0312008)
Bengochea G and Ferraro R 2009 [*Phys. Rev.*]{} D[**79**]{} 124019
Linder E V 2010 [*Phys. Rev.*]{} D[**81**]{} 127301
Li B, Sotiriou T P and Barrow J D 2011 [*Phys. Rev*]{}. D[**83**]{} 104017
Harko T, Koivisto T S, Lobo F S N and Olmo G J arXiv:1110.1049 \[gr-qc\]
Nojiri S and Odintsov S D 2003 [*Phys. Rev.*]{} D[** 68**]{} 123512
M. Carroll M, Duvvuri V, Trodden M and Turner M S 2004 [*Phys. Rev.*]{} D[** 70**]{} 043528
Vollick D N 2003 [*Phys. Rev.*]{} D [**68**]{} 063510
Barrow J D and Hervik S 2006 [*Phys. Rev.* ]{}D[**74**]{} 124017
Olmo G J arXiv:1112.1572 \[gr-qc\]
Barragan C and Olmo G J 2010 [*Phys. Rev*]{}. D[**82**]{} 084015
Olmo[** **]{}G J, Sanchis-Alepuz H and Tripathi S 2009 [*Phys. Rev*]{}. D[**80**]{} 024013
Rubilar G F 1998 [*Class. Quantum Grav*]{}. [**15**]{} 239
Hehl F W, McCrea J D, Mielke E W and Ne’emanY 1995 [*Phys. Rep.*]{} [**258**]{} 1
Shie K-F, Nester J M and Yo H-J 2008 [*Phys. Rev*]{}. D [**78**]{} 023522
Chen H, Ho F-H, Nester J M, Wang C-H, and Yo H-J 2009 [*JCAP*]{} [**0910**]{} 027
Baekler P, Hehl F W and Nester J M 2011 [*Phys. Rev.*]{} D [**83**]{} 024001
Baekler P and Hehl F W arXiv:1105.3504 \[gr-qc\]
Li X-Z, Sun C-B and Xi P 2009 [*Phys. Rev.*]{} D [**79**]{} 027301
Ao X-C, Li X-Z and Xi P 2010 [*Phys. Lett.*]{} B[**694**]{} 186
Mielke E W and Romero E S 2006 [*Phys. Rev.*]{} D [**73**]{} 043521
Garkun A S, Kudin V I, Minkevich AV and Vasilevsky Y G arXiv:1107.1566 \[gr-qc\]
Minkevich A V arXiv:1102.0620 \[gr-qc\]
Minkevich A V 2011 [*Mod. Phys. Lett.*]{} A [**26**]{} 259
Minkevich A V 2009 [*Phys. Lett.*]{} B [**678**]{}, 423
Minkevich A V,Garkun A S and Kudin V I 2007 [*Class. Quant. Grav.*]{} [**24**]{} 5835
Capozziello S, Martin-Moruno P and Rubano C 2010 [*Phys. Lett*]{}. B [**689**]{} 117
Faraoni V and Gunzig E 1999 [* *]{}[*Int. J. Theor. Phys.*]{} [**38**]{} 217
Wainwright J and Ellis G F R 1997 [*Dynamical System in Cosmology,*]{} (Cambridge University Press, Cambridge)
Coley A A, arXiv:gr-qc/9910074
Carloni S and Dunsby P K S arXiv:gr-qc/0611122
Carloni S, Troisi A and Dunsby P K S arXiv:0706.0452
Faraoni V 2005 [*Annals Phys.*]{} [**317**]{} 366
Faraoni V 2005 [*Phys. Rev*]{}. D [**72**]{} 061501
Faraoni V and Nadeau S 2005 [*Phys. Rev.*]{} D [**72**]{} 124005
Wands D and Holden D J 1998 [*Class. Quant. Grav*]{}. [**15**]{} 3271
Copeland E J, Liddle A R and Wands D 1998 [*Phys. Rev.* ]{}D [**57**]{} 4686
Coley A A 1999 [*Gen. Relativ. Grav*]{}. [**31**]{} 1295
Gunzig E et al 2000 [*Class. Quant. Grav.*]{} [**17**]{} 1783
Tsamparlis M 1979 [*Phys. Lett*]{}. A [**75**]{} 27
Goenner H F M and Muller-Hoissen F 1984 [*Class. Quant. Grav*]{}. [**1**]{}, 651
Deser S and Tekin B 2002 [*Phys. Rev. Lett*]{}. [**89**]{} 101101
Mannheim P D arXiv:1101.2186 \[hep-th\]
Maldacena J arXiv:1105.5632 \[hep-th\]
Lu H and Pope C N 2011 [*Phys. Rev. Lett*]{}. [**106**]{}, 181302
Deser S, Liu H, Lu H, C N Pope, Sisman T C and B. Tekin 2011 [*Phys. Rev*]{}. D [**83**]{} 061502
Lu H, Pang Y and Pope C N arXiv:1106.4657 \[hep-th\]
Chen Y-X, Lu H and Shao K-N arXiv:1108.5184 \[hep-th\]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we present a multiresolution-based method for period determination that is able to deal with unevenly sampled data. This method allows us to detect superimposed periodic signals with lower signal-to-noise ratios than in classical methods. This multiresolution-based method is a generalization of the wavelet-based method for period detection that is unable to deal with unevenly sampled data, presented by the authors in [@otazu]. This new method is a useful tool for the analysis of real data and, in particular, for the analysis of astronomical data.'
author:
- |
X. Otazu$^{1}$[^1], M. Ribó$^{2\star}$, J. M. Paredes$^{3\star}$[^2], M. Peracaula$^{4\star}$ and J. Núñez$^{3,5\star}$\
$^1$Centre de Visió per Computador, Campus Universitat Autònoma de Barcelona, 08193 Cerdanyola del Vallès (Barcelona), Spain\
$^2$Service d’Astrophysique, CEA Saclay, Bât. 709, L’Orme des Merisiers, 91191 Gif-sur-Yvette, Cedex, France\
$^3$Departament d’Astronomia i Meteorologia, Universitat de Barcelona, Av. Diagonal 647, 08028, Barcelona, Spain\
$^4$Institut d’Informàtica i Aplicacions, Universitat de Girona, Campus de Montilivi s/n, 17071 Girona, Spain\
$^5$Observatori Fabra, Camí de l’Observatori sn, 08035 Barcelona, Spain
date: 'Accepted YYYY MMMMMMMM DD. Received YYYY MMMMMMMM DD'
title: Multiresolution approach for period determination on unevenly sampled data
---
\[firstpage\]
methods: data analysis – methods: numerical – stars: variables: other – X-rays: binaries.
Introduction {#intro}
============
The detection of periodic signals in astronomical data has been usually addressed by classical Fourier-based or epoch folding methods. These methods have different problems when dealing with non-sinusoidal periodic signals or with very low signal-to-noise ratios. When the analysed data set contains several periodic signals, the behavior of classical period determination methods highly depends on intrinsic signal characteristics (see, for example, @lafler; @jurkevich; @stellingwerf; @lomb; @scargleA, [-@scargleB]; @roberts; @press). The reader is referred to the introduction of [@otazu] (hereafter Paper I) for a general discussion. To avoid this problem, we presented there, as a preliminary step, a wavelet-based approach that only works with evenly time sampled data.
In astronomy, however, this is not a usual situation, since data are mostly acquired on irregular intervals of time. In such a case there are two possibilities: resample the data into a new evenly sampled data set, or use a method able to deal with the original unevenly sampled data set. In the first case we are forced to modify the original data, which necessarily implies a loss of information. Moreover, this is not always possible if the temporal gaps are larger than some of the periods present in the data. In order to avoid these problems, a technique capable to deal with unevenly sampled data is needed.
The present paper is a natural extension of Paper I that allows to work on unevenly time sampled data. We show how the methodology of multiresolution decomposition (similar to the wavelet decomposition philosophy) is very well suited to this problem, since it is completely oriented towards decomposing functions into several frequential characteristics.
As in Paper I, the main objective is to isolate every signal present in our data and to analyse them separately, avoiding their mutual influences. In Section \[multiresolution\] we outline some concepts in multiresolution analysis and their similarities with wavelet theory that are relevant to the stated problem. In Section \[period\] we propose an algorithm to detect each of the periodic signals present in a data set by combining multiresolution analysis decomposition with classical period determination methods. In Sections \[simulated\] and \[results\] we present some examples of synthetic data we used to test the algorithm and the results we obtained. We summarise our conclusions in Section \[conclusions\].
Multiresolution analysis {#multiresolution}
========================
Multiresolution decomposition introduces the concept of the presence of details between successive levels of resolution. Many wavelet decomposition algorithms are based on multiresolution analysis schemes (@chui; @daubechies; @meyer; @kaiser; @vetterli; @mallat; @starck), and some astronomical applications using wavelets for timing analysis have been reported (@szatmary; @foster; @ribo; @poly). In fact, all these methods use the same philosophy and the obtained results can be interpreted in the same way. However, wavelet theory presents some constraints on mathematical functions. These constraints are not respected in all multiresolution decomposition schemes. Therefore, when we are using a certain multiresolution decomposition algorithm that fulfills the wavelet constraints, we are obtaining a wavelet decomposition. In contrast, when a given multiresolution decomposition algorithm violates these wavelet constraints, we are obtaining a result similar to, but which is not, a wavelet decomposition.
As discussed in Paper I, we note that there are wavelet approaches that are based on approximations of a continuous wavelet transform and on the subsequent study of wavelet space coefficients (see, e.g., @szatmary), which are able to deal with unevenly sampled data sets. However, these algorithms present a [*non-direct*]{} inverse wavelet transform, in the sense that the search for periodicities is based on the fit between the wavelet base function profile and the signal one, and therefore on the values of the wavelet transform coefficients, which highly depend on the wavelet base used. Moreover, the period analysis has to be performed on the wavelet coefficients space but, since it is usually decimated, accurate period detection is a difficult task. The multiresolution decomposition scheme we use in this work performs the decomposition on the temporal space, which allows to find accurate values for periodicities.
A multiresolution analysis algorithm {#multires_alg}
------------------------------------
In order to obtain a multiresolution decomposition for signals, an algorithm to decompose the signal into frequency planes can be defined as follows. Given a signal $p$ we construct the sequence of approximations: $$p_1 = F_1(p),~~~~~
p_2 = F_2(p_1),~~~~~
p_3 = F_3(p_2),\cdots,$$ performing successive convolutions with Gaussian filters $F_i$.
It is important to note the difference between this sequence of convolutions and the one used in Paper I. In the latter we were dealing with a discrete convolution mask, and hence forced to work with evenly spaced data. In contrast, the continuous nature of the convolution functions used in the present paper, allows us to work with unevenly sampled data.
Similarly to the wavelet planes, the multiresolution frequency planes are computed as the differences between two consecutive approximations $p_{i-1}$ and $p_i$. Letting $w_i = p_{i-1} - p_i \;\;(i=1,\cdots,n)$, in which $p_0 =
p$, we can write the reconstruction formula: $$p =
\sum_{i=1}^{n}w_i + p_r\;\;. \label{wav_des}$$
In this representation, the signals $p_i~(i=0,\cdots,n)$ are versions of the original signal $p$ at increasing scales (decreasing resolution levels), $w_i~(i=1,\cdots,n)$ are the multiresolution frequency planes and $p_r$ is a residual signal (in fact $n=r$, but we explicitly substitute $n$ by $r$ to clearly express the concept of $residual$). In our case, we are using a dyadic decomposition scheme. This means that the standard deviation of the Gaussian function associated with the $F_i$ filter is $\sigma_i=2\sigma_{i-1}$. Thus, similarly to the wavelet approach, the original signal $p_0$ has double resolution compared to $p_1$, and so on. All these $p_i~(i=0,\cdots,n)$ signals have the same number of data points as the original signal $p$.
Since $\sigma_{i+1}$ depends on $\sigma_i$, a value for $\sigma_0$ has to be carefully chosen for every data set. It has to be fixed considering a likely minimum value for the time-duration of the features and the characteristic time sampling of the data set, in order to include a significant number of points on which to perform the convolutions. Too small values do not accurately describe feature profiles, and suffer from poor or noisy data. In contrast, too large values reduce the noise effect by integrating a lot of data points, but may ignore interesting high frequency features. Hence, a first analysis of data has to be performed in order to obtain a useful initial $\sigma_0$ value.
We have used the same notation as in the wavelet decomposition described in Paper I because, as explained above, the idea of these multiresolution planes is similar to the wavelet ones.
We note that this particular decomposition scheme that uses a Gaussian kernel can also be interpreted as a scale-space filtering (@baubaud, @sporring, @witkin94) or as a particular case of more general image diffusion approaches (@lindeberg, @perona). Smoothed data sets $p_i$ can be interpreted as diffused scale-space images, and the difference between them as the details at different scales.
Period detection algorithm {#period}
==========================
We propose to apply this multiresolution analysis algorithm to solve our initially stated problem: to isolate each of the periodic signals contained in a set of unevenly sampled data and study them separately.
In order to do so, we proceed as in Paper I, and the period detection algorithm we propose is as follows:
1. choose values for $\sigma_0$ and $n$; decompose the original signal $p$ into its multiresolution frequency planes $w_i~(i=0,\cdots,n)$.
2. detect periods in each of the $n$ obtained frequency planes $w_i$.
Phase Dispersion Minimization ([PDM]{}) [@stellingwerf] and [CLEAN]{} [@roberts] methods are used to detect periods in the original data and in every multiresolution frequency plane. In Paper I we described several undesirable effects of [PDM]{} and [CLEAN]{} methods on data with superimposed signals, as well as the advantages of using multiresolution-based methods over the classical ones.
Hereafter, and for notational convenience, the multiresolution-based [PDM]{} and [CLEAN]{} methods will be called and [MRCLEAN]{}, respectively.
Simulated data {#simulated}
==============
In order to check the benefit of applying [MRPDM]{} versus [PDM]{}, or [MRCLEAN]{} versus [CLEAN]{}, we proceeded similarly as in Paper I by generating and analysing several sets of simulated data containing two superimposed periodic signals. Each data set is composed of a high-amplitude primary sinusoidal function and a secondary low-amplitude Gaussian one. Finally, we added a white-Gaussian noise to this combination of signals. We increased the value of the noise standard deviation, $\sigma$, up to the value where detection of periodic signals became statistically insignificant in both the classical and the multiresolution-based methods.
The primary signal is intended to simulate variable sources with a pure sinusoidal intensity profile (like precession of accretion discs), and the secondary, burst-like events (like pulses or eclipses) superimposed to it. We note that in Paper I we also studied the superposition of two sinusoidal functions, the so called sine+sine case. However, here we have directly focused on the more realistic situation of the so called sine+Gaussian case, where the difference between the classical and the multiresolution-based methods is more critical (see Paper I).
The characteristics of the signals are:
1. Each signal is generated as an unevenly spaced data set. Its time sampling has been taken from observations of the X-ray binary LMC X-1 by the All Sky Monitor onboard the RXTE satellite (@levine). We have used an observation period which has 8270 measurements during 679 days.
2. The high-amplitude sinusoidal function has an amplitude equal to 1 and a period of 108.5 days.
3. The amplitudes of the low-amplitude periodic Gaussian function are 0.1 and 0.5.
4. The periods used for the secondary function are 13.13 and 23.11 days.
5. We have used two values for the full width at half-maximum (FWHM) of the Gaussian signal, corresponding to 2 and 6 days, respectively.
In the first four columns of Table \[gauss\_table\] we present the parameters used to generate each simulated data set.
Results
=======
We recall that the simultaneous use of two independent methods, such as [CLEAN]{} and [PDM]{}, is usually applied to discriminate false period detections from the true ones. A similar procedure can be used with each one of the multiresolution planes in the [MRPDM]{} and [MRCLEAN]{} methods. Therefore, when comparing the behavior of these different methods, we have to compare the usual [PDM-CLEAN]{} method combination for period estimation prior to the new [MRPDM-MRCLEAN]{} combination.
Taking into account the characteristics of the simulated data we have chosen a $\sigma_0=1$ day. As in Paper I, the primary period (108.5 days) is always detected by all methods, and does not appear in Table \[gauss\_table\]. In this table, the four last columns show the detected low-amplitude periods for each data set using [PDM]{}, [CLEAN]{}, [MRPDM]{} and [MRCLEAN]{} methods, respectively. A dash is shown when a period is not detected, and a question mark when the detection is difficult or doubtful. When a period is found in the multiresolution-based methods, we also show in parentheses the mutiresolution planes where it is detected.
FWHM Period Amplitude $\sigma_{\mathrm noise}$ [PDM]{} [CLEAN]{} [MRPDM]{} [MRCLEAN]{}
------ -------- ----------- -------------------------- --------- ----------- ------------------ ----------------
2 13.13 0.1 0.0 - 13.13 13.13 (1,2) 13.13 (1,2)
0.1 - 13.13 13.13 (1,2) 13.13 (1,2)
0.15 - 13.13 13.13 (1?,2) 13.13 (1?,2)
0.2 - 13.15? 13.13 (1?,2) 13.13 (2)
0.25 - - 13.14 (1?,2) -
0.3 - - 13.14 (2) -
0.35 - - 13.14 (2?) -
0.4 - - - -
0.5 0.0 - 13.13 13.13 (1,2,3) 13.13 (1,2,3)
0.5 - 13.13 13.13 (1?,2,3) 13.13 (1?,2,3)
1.0 - 13.14? 13.13 (2,3) 13.13 (2,3)
2.0 - - 13.14 (2,3) 13.14 (3?)
2.25 - - 13.14 (2?) -
2.5 - - - -
23.11 0.1 0.0 - 23.13 23.11 (1,2?) -
0.1 - 23.19 23.11 (2?) -
0.15 - - - -
0.5 0.0 - 23.12 23.11 (2?,3,4) 23.11 (3?)
0.5 - 23.11? 23.11 (2?,3,4) -
1.0 - - 23.11 (2?,3,4) -
1.25 - - 23.11 (2?,3,4) -
1.5 - - - -
6 13.13 0.1 0.0 - 13.13 13.13 (1,2,3) 13.13 (1,2,3)
0.1 - 13.14 13.13 (1,2,3) 13.13 (1,2,3)
0.2 - 13.14 13.13 (1?,2,3) 13.13 (1?,2,3)
0.3 - 13.14 13.13 (2,3) 13.13 (2,3)
0.4 - 13.15? 13.14 (2,3) 13.14 (3)
0.6 - - 13.14 (2,3) 13.14 (3?)
0.8 - - 13.15 (3) 13.15 (3?)
0.9 - - - -
0.5 0.0 13.13 13.13 13.13 (1,2,3) 13.13 (1,2,3)
0.5 13.13 13.12 13.13 (1,2,3) 13.13 (1,2,3)
1.0 13.13 13.14 13.13 (1?,2,3) 13.13 (1?,2,3)
2.0 13.13 13.13 13.13 (2,3) 13.13 (2,3)
3.0 13.13? 13.11? 13.13 (2?,3) 13.13 (3)
4.0 - - - -
23.11 0.1 0.0 - 23.12 23.11 (2,3) 23.11 (2,3)
0.1 - 23.12 23.11 (2,3) 23.11 (2,3)
0.2 - 23.12 23.11 (2?,3) 23.11 (2?,3)
0.3 - 23.12 23.11 (3) 23.11 (3)
0.4 - 23.12? 23.11 (3?) 23.11 (3)
0.5 - - 23.11 (3?) -
0.6 - - - -
0.5 0.0 23.11 23.11 23.11 (2?,3,4,5) 23.11 (3?,4,5)
0.5 23.11 23.11 23.11 (2?,3,4,5) 23.11 (4,5)
1.0 23.11 23.11 23.11 (3,4,5) 23.11 (4,5?)
2.0 23.11? 23.12 23.11 (3?,4,5) 23.11 (4,5?)
2.5 23.11? 23.12 23.11 (4,5?) 23.11 (4,5?)
3.0 - - 23.12 (4) 23.11 (4)
3.5 - - 23.12 (4?) -
4.0 - - - -
We must note that the use of two different FWHM for the Gaussian, combined with two different periods (13.13 and 23.11 days), gives 4 different profiles. Hence, the phase duration of the burst-like event ranges from very low to relatively high values in the following order: FWHM=2 and period=23.11, FWHM=2 and period=13.13, FWHM=6 and period=23.11, and finally FWHM=6 and period=13.13.
In view of the results displayed in Table \[gauss\_table\] we can make the following comments:
1. In all methods, with high noise-to-signal ratios the detected periods are slightly different from the simulated ones.
2. There is a better performance of [CLEAN]{} over [PDM]{}. We must note that when the FWHM is only 2 days, [PDM]{} never detects the secondary period. Only with FWHM=6 days and a relatively high amplitude (0.5), can [PDM]{} detect the low-amplitude periodic signals.
3. As the noise increases, the detection starts to fail in the lower multiresolution planes (higher frequencies), and only the higher ones (lower frequencies) are noise-free enough to allow period detection.
4. \[iv\] [MRPDM]{} and [MRCLEAN]{} perform better or similar than [PDM]{} and [CLEAN]{} methods (see exception below). When [CLEAN]{} marginally detects the secondary period, [MRPDM]{} and most of times [MRCLEAN]{} have no problems to detect it, and they work properly even with higher noise. In the [MRPDM]{} case, the results are always better than with [PDM]{}.
5. In all cases with FWHM=2, the [MRPDM]{} performance is much better than [MRCLEAN]{}, because the signal is clearly non-sinusoidal. In the FWHM=6 cases, [MRPDM]{} is only slightly better than [MRCLEAN]{}, since the signals are closer to a sinusoidal profile.
6. For a given amplitude of the Gaussian signal, the maximum noise-to-signal ratio achieved with [MRPDM]{} and [MRCLEAN]{} increases with the phase duration of the FWHM.
The only exception to \[iv\] is for the most extreme of the simulated cases, i.e., the one with the lower amplitude, lower FWHM value and longer period. However, we note that the period detected by CLEAN is slightly different than the simulated one.
All these results are very similar to those shown in Table 2 of Paper I. Nevertheless the maximum noise-to-signal ratios achieved in the present cases are around 2.5 times higher. This can be explained because, although the time span of the data sets used here is around 1.5 times smaller, the number of points per unit time is around 12 times higher than in Paper I.
Finally, and for illustrative purposes, we show in Fig. \[fig:simulated\] the simulated data set generated with the following: 13.13-day period, FWHM=6.0 days and amplitude=0.1 with $\sigma_{\mathrm noise}=0.6$. The outputs of [PDM]{} and [CLEAN]{}, are also shown. None of these methods is able to detect the 13.13-day period. We show in Fig. \[fig:mrpdm\] the outputs from [MRPDM]{}. The simulated period is detected, with its corresponding subharmonics, in the multiresolution planes $w_2$ and $w_3$. The [MRCLEAN]{} outputs, shown in Fig. \[fig:mrclean\], reveal a marginal detection of the 13.13-day period in multiresolution plane $w_3$. We note that we would not consider this detection as a true one when taken alone. However, since the same period is clearly detected in two multiresolution planes of [MRPDM]{}, we can establish the existence of this period in the analysed data set.
Conclusions
===========
In this paper we have presented a multiresolution-based method for period determination able to deal with unevenly sampled data. This constitutes a significant improvement with respect to the wavelet-based method presented in Paper I, which is unable to deal with unevenly sampled data. The overall performance of the present method is similar to the wavelet-based one, in the sense that it allows us to detect superimposed periodic signals with lower signal-to-noise ratios than in classical methods. We stress that one advantage of the present method over classical methods is the simultaneous detection of a period in more than one multiresolution plane, allowing to improve the confidence of a given detection. Moreover, since here we are not forced to lose or modify the information when averaging or interpolating the original data, we can reach higher noise-to-signal ratios than in the wavelet-based method described in Paper I.
We note that the multiresolution decomposition scheme that we have used can be interpreted as a particular case of scale-space filtering. In order to improve isolation of periodic features, more general approaches could be used to perform this decomposition. In this context, anisotropic diffusion schemes proposed by [@perona] could be useful if properly tuned.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank useful comments and suggestions from an anonymous referee. We acknowledge partial support by DGI of the Ministerio de Ciencia y Tecnología (Spain) under grant AYA2001-3092, as well as partial support by the European Regional Development Fund (ERDF/FEDER). This research has made use of facilities of CESCA and CEPBA, coordinated by C4 (Centre de Computació i Comunicacions de Catalunya). XO is a researcher of the programme [*Ramón y Cajal*]{} funded by the Spanish Ministery of Science and Technology and Centre de Visió per Computador. MR acknowledges support by a Marie Curie Fellowship of the European Community programme Improving Human Potential under contract number HPMF-CT-2002-02053. MP is a researcher of the programme [*Ramón y Cajal*]{} funded by the Spanish Ministery of Science and Technology and Universitat de Girona.
[1]{}
Baubaud J., Witkin A., Baudin M., Duda R.O., 1986, IEEE Trans. Pattern Anal. Mach. Intell. PAMI-8, 26 Chui C.K., 1992, An Introduction to Wavelets, Boston Ac. Press, Boston
Daubechies I., 1992, Ten Lectures on Wavelets, SIAM Press, Philadelphia
Foster G., 1996, AJ, 112, 1709
Jurkevich I., 1971, Ap&SS, 13, 154
Kaiser G., 1994, A Friendly Guide to Wavelets, Birkhäuser, Berlin
Lafler J., Kinman T.D., 1965, ApJS, 11, 216
Levine A.M., Bradt H., Cui W., Jernigan J.G., Morgan E.H., Remillard R., Shirey R.E., Smith D.A., 1996, ApJ, 469, L33
Lindeberg T., 1990, IEEE Trans. Pattern Anal. and Mach. Intell., 12, 234 Lomb N.R., 1976, Ap&SS, 39, 447
Mallat S., 1999, A Wavelet Tour of Signal Processing, 2nd ed., Academic Press, San Diego
Meyer Y., 1993, Wavelets. Algorithm and Applications, SIAM Press, Philadelphia
Otazu X., Ribó M., Peracaula M., Paredes J.M., Núñez J., 2002, MNRAS, 333, 365 (Paper I)
Perona P., Malik J., 1990, IEEE Trans. Pattern Anal. and Mach. Intell., 12, 629 Polygiannakis J., Preka-Papadema P., Moussas X., 2003, MNRAS, 343, 725
Press W.H., Rybicky G.B., 1989, ApJ, 338, 277
Ribó M., Peracaula M., Paredes J.M., Núñez J., Otazu X., 2001, in Giménez A., Reglero V., Winkler C., eds, Proc. Fourth INTEGRAL Workshop, Exploring the Gamma-Ray Universe. ESA Publications Division, Noordwijk, p. 333
Roberts D.H., Lehár J., Dreher J.W., 1987, AJ, 93, 968
Scargle J.D., 1982, ApJ, 263, 835
Scargle J.D., 1989, ApJ, 343, 874
Sporring J., Niselsen M., Florack L., Johansen P., 1997, Gaussian Scale-Space Theory, Kluwer Ac. Pub., Drodrecht Starck J.L., Murtagh F., 2002, Astronomical Image and Data Analysis, Springer, Berlin
Stellingwerf R.F., 1978, ApJ, 224, 953
Szatmáry K., Vinkó J., Gál J., 1994, A&AS, 108, 377
Vetterli M., Kovacevic J., 1995, Wavelets and Subband Coding, Prentice Hall, Englewood Cliffs, New Jersey
Witkin A.P., 1984, Image Understanding 1984, Ullman and Richards editors, Norwood Ablex, New Jersey
\[lastpage\]
[^1]: E-mail: [email protected] (XO); [email protected] (MR); [email protected] (JMP); [email protected] (MP); [email protected] (JN)
[^2]: CER on Astrophysics, Particle Physics and Cosmology. Universitat de Barcelona
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper presents an efficient object detection method from satellite imagery. Among a number of machine learning algorithms, we proposed a combination of two convolutional neural networks (CNN) aimed at high precision and high recall, respectively. We validated our models using golf courses as target objects. The proposed deep learning method demonstrated higher accuracy than previous object identification methods.'
address: '$^1$National Institute of Advanced Industrial Science and Technology, Tokyo, Japan'
bibliography:
- 'strings.bib'
- 'refs.bib'
title: 'Object detection in satellite imagery using 2-Step Convolutional Neural Networks'
---
remote sensing, object detection, convolutional neural networks, golf course, negative mining
Introduction {#sec:intro}
============
Earth observation satellites have been monitoring changes on the Earth’s surface over a long period of time. High resolution satellite imagery can detect small objects such as ships, cars, aircraft, and individual houses; whereas, medium resolution satellite imagery can detect relatively larger objects, such as ports, roads, airports and large buildings \[1\]\[6\]. Total data amount, however, would be too huge to be inspected by human eyes. Therefore, we need an efficient algorithm for automatic object detection on satellite imagery. Previous works employed higher-order local auto correlation \[1\], random forests \[2\], and deep learning \[3\]\[4\]\[5\]\[6\]. Among them, deep learning \[7\]\[8\] showed higher accuracy in object detection of images than other machine learning methods.
As an example of target object, we selected golf courses because they exist everywhere in the world, are typically of a recognizable size and shape with 30 meter resolution of Landsat 8 imagery. According to the R & A report \[9\], in 2016 there were 33161 golf courses, which provide more than enough data for training and detection purpose. Once we establish an accurate algorithm, we can continuously monitor the new construction and disappearance of all the golf courses on the Earth.
Among the general object detection framework, Faster R-CNN \[8\] is the state-of-the-art method. This method consists of a region proposal network for predicting candidate regions and region classification network for classifying object proposal. It is an end-to-end detector that outputs the location and category of object simultaneously. Similarly, we propose a model called “high recall network” specifically for detection of candidate golf course regions and a model called “high precision network” for further confirmation. We then compared the proposed method with other existing methods \[1\]\[6\].
METHOD {#sec:format}
======
The framework of our object detection method, which involved a two-step process, is illustrated in Fig. 1. The first step employs the high recall network (HRN) model to find candidate regions as much as possible. The second step uses the high precision network (HPN) model for binary classification (golf course or not) of the HRN output. Each step is customized for the purpose as described in 2.1 and 2.2.
High Recall Network (HRN)
--------------------------
The training and validation data set for HRN is derived from 117 clear Landsat 8 scenes taken between 2013 and 2015 all over Japan except five areas reserved for testing (see 3.1 for more detail). Each area is observed 3 to 4 times at different time to account for seasonal variation. Moreover, we have prepared Ground Truth (GT) polygons which outline all the golf courses in Japan. The Landsat scenes were gridded into tiles with 16$\times$16 pixels. A tile is classified as positive if the coverage of GT polygons is larger than 20% of the total area; tiles with no overlap with golf courses were classified as negative; and the remainder which fell between 0% $\sim$ 20% coverage were classified as neither positive nor negative (Fig. 2).
Ishii et al. \[6\] proposed an object detection method that applying classification to detection as like Fully Convolutional Neural Network (FCN) \[7\] for satellite imagery. They found the increase of recall performance as the relative abundance of negative image decreases in the training data set. Our HRN model is equivalent to their model (Fig. 2), but the percentage of negative data is adjusted to achieve higher recall (Table. 1). In addition, we focus on the recall performance rather than precision during the training. HRN training process output a learned snapshot model each 10 epoch. From the many snapshot models generated by the HRN training process, we selected the model with the highest recall and with at least over 50% precision. When selecting this model, we use validation dataset rather than training data set.
![High Recall Network(HRN) has FCN structure consisting of 4 convolution layers. []{data-label=""}](./fig_HighRecall_Network.png)
High Precision Network (HPN) {#ssec:subhead}
-----------------------------
The HPN structure and training process is illustrated in Fig. 3. The HPN performs binary classification on the candidate regions resulted from HRN. Input data used in this process are derived from cropping HRN output into tiles with 64$\times$64 pixels (red rectangles in Fig. 3).
For HPN model training, positive images are generated by cropping satellite image of golf course region at the centroid from GT polygons. We conducted data augmentation by rotating each positive image in the step of 90$^\circ$ and then flipped in horizontally and vertically respectively. Negative images are produced by negative mining that crop a centroid of false positive regions generated by HRN. The number of convolution layers is increased to 8 to give the highest precision.
![The High Precision Network (HPN) comprises 8 convolutional layers and a fully connected layer. []{data-label=""}](./fig_HighPrecisionNetwork.png)
EXPERIMENTS AND RESULTS {#sec:pagestyle}
=======================
In this section, we conducted golf course detection experiments using our method as well as previous methods (Ishii et al \[6\] and Uehara et al \[1\]). For direct comparison, we employed five scenes in Uehara et al \[1\] as testing data (Fig. 4). It should be noted again that these testing data are not included in the training data. After removing cloudy areas, we classified the output tiles as true positives (TP) if it contains GT polygons and as false positive (FP) otherwise. It would be natural to regard GT polygons with no overlapping detected tiles as false negatives (FN). In Fig. 5, red lines are golf course regions from GT and red rectangles indicate the output tiles: (a) represents an example of TP in which a golf course was correctly detected; (b) represents FP in which an area was erroneously identified as a golf course; and (c) represents FN in which a golf course was not detected due to its atypical structure (a narrow course, no observable trees, etc.).
Table. 2 compares the performance of three methods measured by recall \[ TP/(TP+FN) \], precision \[TP/(TP+FP)\] and F-measure \[2$\times$precision$\times$recall/(precision+recall)\]. When we use F-measure as a total performance proxy, our method showed 4% more improvement than previous two methods. This improvement could be attributed to collaboration between HRN and HPN. In actual applications, F-measure may not be necessarily the best index to estimate the performance. Some applications would require an exhaustive list of the possible target candidates even with the lower precision, while others may need higher precision at the cost of lower recall. Our framework could be easily adapted to diverse applications due to the explicit separation into two complementary networks.
------------ ---------- ---------- ---------- ----------
positive negative positive negative
training 52276 936000 110314 298945
validation 17620 4686120 35766 938271
------------ ---------- ---------- ---------- ----------
: Number of training data in experiments
\[dataset\]
A B C D E total
----------- ------- ------- ------- ------- ------- -------
Recall
Ishii 0.333 0.736 0.743 0.903 0.889 0.759
Uehara 0.428 0.714 0.848 0.841 0.897 0.753
Our 0.892 0.929 0.833 0.937 0.908 0.901
Precision
Ishii 0.443 1.000 1.000 0.981 0.986 0.924
Uehara 0.715 0.947 0.924 0.910 0.911 0.895
Our 0.456 0.956 0.977 0.953 0.959 0.874
F-measure
Ishii 0.380 0.848 0.852 0.940 0.935 0.833
Uehara 0535 0.814 0.884 0.874 0.904 0.818
Our 0.604 0.942 0.900 0.945 0.933 0.870
: Experimental results
\[japan\]
![An example of experimental results with the proposed method: (a) True Positives (b) False Positives (c) False Negatives. The red lines represent GT polygons of the golf course, while rectangles indicate the output tiles (64x64 pixels) with detection.[]{data-label=""}](./resultsample.png)
CONCLUSION {#sec:typestyle}
==========
In this paper, we proposed a method to detect arbitrary objects in satellite imagery . The method entailed an integration of two convolutional neural networks (CNN) devoted to high recall and high precision, respectively. We customized each CNN through network structure, selection of the input training data and target parameters for optimization. Our method showed a 4% overall improvement compared to previous methods. Furthermore, we can flexibly tune the relative importance of recall and precision by balancing two networks.
In the future work, we will (1) expand the target areas to a global level and evaluate the generalization performance. (2) increase the tile size for HPN to apply the state-of-the-art CNN for general images \[10\]\[11\]. (3) make a comparison with the state-of-the-art two step algorithm for general images, such as Faster R-CNN \[8\].
ACKNOWLEDGMENT {#sec:majhead}
==============
This paper is based on results obtained from a project commissioned by the New Energy and Industrial Technology Development Organization (NEDO).
[10]{} Kazuki Uehara, Hidenori Sakanashi, Nosato Hirokazu, Masahiro Murakawa, Hiroki Miyamoto, and Ryosuke Nakamura, “Object Detection on Satellite Images Using Multi Channel Higher-order Local Autocorrelation, ” in 2017 IEEE International conference on Sysytem, Man, and Cybernetics(SMC). Jordan M. Malof, Kyle Bradbury, Leslie M. Collins, and Richard G. Newell, “Automatic Detection of Solar Photovoltaic Arrays in High Resolution Aerial Imagery, ” in Applied Energy, vol. 183, pp229-240, 2016. Weijia Li, Haohuan Fu, Le Yu, and Arthur Cracknell, “ Deep Learning Based Oil Palm Tree Detection and Counting for High-Resolution Remote Sensing Images, ” in MDPI Remote Sens, 2017,9,22 Lei Liu, Zongxu Pan, and Bin Lei, “Learning a Rotation Invariant Detector with Rotatable Bounding Box, ” in arXiv 1711. 09405, 2017. Rodrigo F. Berriel, Andr’e Teixeira Lopes, Alberto F. de Souza, and Thiago Oliveira-Santos, “Deep Learning Based Large-Scale Automatic Satellite Crosswalk Classification, ” in IEEE Geoscience and Remote Sensing Society, vol. 14, pp1513-1517, 2017. Tomohiro Ishii, Edgar Simo-Serra, Satoshi Iizuka, Yoshihiko Mochizuki, Akihiro Sugimoto, Ryosuke Nakamura, and Hiroshi Ishikawa, “Detection by Classification of Buildings in Multispectral Satellite Imagery, ” in 2016 IEEE International conference on pattern recognition (ICPR).
Jonathan Long, Evan Shelhamer, and Trevor Darrell, “Fully Convolutional Networks for Semantic Segmentation, ” in 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun, “Faster R-CNN:Towards Real-Time Object Detection with Region Proposal Networks, ” in 2015 Neural Information Processing System (NIPS). The R & A, “Golf around the world 2017, ” in https://www.randa.org/ Kaiming He, Xiangyu Zhang, Shaoqong Ren, and Jin Sun, “Deep Residual Learning for Image Recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
Gao Huang, Zhuang Liu, and Laurens van Maaten, “Densely Connected Convolutional Networks,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
\
Departamento de Física Teórica, Universidad Autónoma de Madrid (UAM), 28049 Spain\
Instituto de Física Teórica UAM–CSIC, Campus de Cantoblanco UAM, 28049 Spain\
E-mail:
bibliography:
- 'munussm\_v2.bib'
title: |
Searching for SUSY and decaying gravitino DM\
at the LHC and Fermi-LAT with the $\mu\nu$SSM
---
Introduction
============
Supersymmetry (SUSY) is still the most compelling theory for physics beyond the standard model. SUSY not only solves several important theoretical problems of the standard model, such as the gauge hierarchy problem and others, but also has spectacular experimental implications. As is well known, the spectrum of elementary particles is doubled with masses of about 1 TeV, thus even the simplest SUSY model, the minimal supersymmetric standard model (MSSM, see Ref. [@Martin:1997ns] for a review), predicts a rich phenomenology. However, the LHC started operations several years ago and, with Run 1 already finished, SUSY has not been discovered yet. Because of this, it has been raised the question of whether SUSY is still alive. The question is fair of course, but in our opinion the answer is yes, and we think that there are several arguments in favor of this answer. Here there are some of them:
$\bullet$ The lower bounds on SUSY particle (sparticle) masses are smaller than 1 TeV or about that number, depending on the sparticle analyzed. Thus the SUSY masses are still reasonable, and in that sense we can keep in mind what happened before the discovery of the Higgs boson.
$\bullet$ Because of the complicated parameter space of SUSY, experimentalists use in their analyses simplified models that do not cover the full MSSM. For example, branching-ratio variations are not considered in much detail, and other assumptions are also made.
$\bullet$ Run 2 is going on, and for the moment with a low luminosity of about 20 fb$^{-1}$. Therefore, to (be prepared) wait for the results with higher luminosity seems to be a sensible strategy, since 100 fb$^{-1}$ are expected for the end of the Run 2.
$\bullet$ Most searches at the LHC assume $R$-parity conservation ($\rp$), with the lightest supersymmetric particle (LSP) stable, requiring therefore missing energy in the final state to claim for SUSY detection. But, if $R$ parity is violated ($\rpv$), sparticles can decay to standard model particles, and the bounds on their masses can become significantly weaker.
Nevertheless, despite all these arguments, it is also honest to recognize that SUSY has its own theoretical problems in its formulation at low energy. and, in particular, a crucial one is the so-called $\mu$ problem [@Kim:1983dt]. In the superpotential of the MSSM $$W =
\epsilon_{ab} \left(
Y_{u_{ij}} \, \hat H_u^b\, \hat Q^a_i \, \hat u_j^c +
Y_{d_{ij}} \, \hat H_d^a\, \hat Q^b_i \, \hat d_j^c +
Y_{e_{ij}} \, \hat H_d^a\, \hat L^b_i \, \hat e_j^c \right)
-
\epsilon{_{ab}} \mu \,\hat H_d^a \hat H_u^b
\ ,
\label{superpotential1}$$ the presence of the mass parameter $\mu$ is necessary, for example to generate Higgsino masses given the current experimental lower bound of about 100 GeV on chargino masses. In the presence of a high-energy theory like a grand unified theory (GUT) or a string theory, with a typical scale of the order of $10^{16}$ GeV or larger, and/or a gravitational theory at the Planck scale, one should be able to explain how to obtain a mass parameter in the superpotential of the order of the electroweak (EW) scale. The MSSM does not solve the $\mu$ problem. One takes for granted that the $\mu$ term is there and that is of the order of the EW scale, and that’s it. In this sense, the MSSM is a kind of effective theory.
Another theoretical problem of SUSY is to be able to build a model solving the $\nu$ problem: how to reproduce neutrino data [@Tortola:2012te], i.e. masses and mixing angles. Let us emphasize in this sense that in the MSSM, by construction, neutrinos are massless.
The ‘$\mu$ from $\nu$’ supersymmetric standard model ($\mu\nu$SSM [@LopezFogliani:2005yw; @Escudero:2008jg], see Refs. [@Munoz:2009an; @LopezFogliani:2010bf] for reviews), includes new couplings with the three families of right-handed (RH) neutrino superfields ($\hat \nu^c_i$ with $i=1,2,3$) in the superpontential in order to solve the $\mu$-problem, while simultaneously explains the origin of neutrino masses. The $SU(3)_c\times SU(2)_L\times U(1)_Y$ invariant couplings $\lambda_i \hat \nu^c_i \hat H_d\hat H_u$ generate an effective $\mu$ term through RH sneutrino vacuum expectation values (VEVs), $\langle \tilde \nu^c_i \rangle \equiv v_{\nu^c_i}$, after the successful electroweak symmetry breaking (EWSB): $\mu^{eff}=\lambda_i v_{\nu^c_i}$. In addition, other gauge invariant couplings $\frac{1}{3}\kappa{_{ijk}} \hat \nu^c_i\hat \nu^c_j\hat \nu^c_k$ generate effective Majorana masses for the RH neutrinos, $M_{ij}^{eff}=2\kappa_{ijk}v_{\nu^c_k}$, contributing to a generalized EW-scale seesaw mechanism which can reproduce the observed neutrino masses and mixing angles. We will review this solution to the $\mu$ problem and neutrino physics in Section \[model\].
Since sparticles do not appear in pairs in these couplings that solve the $\mu$ and $\nu$ problems, we say in the usual language that they produce explicit $\rpv$. This implies that the phenomenology of the $\mu\nu$SSM is very different from the one of the MSSM or the next-to-MSSM (NMSSM, see Ref. [@Ellwanger:2009dp] for a review). We will briefly review the phenomenology of the $\mn$ at the LHC in Section \[lhc\]. There we will see that since the LSP is not stable because of $\rpv$, it decays leading to prompt or displaced vertices, and producing final states with multi-leptons/taus/jets/photons and missing energy.
On the other hand, the usual sparticle candidates for the dark matter (DM) of the Universe in the case of $\rp$, the neutralino or the RH sneutrino, have very short lifetimes in $\rpv$ models, and therefore can no longer be used. Nevertheless, the gravitino can still be a candidate for DM since its lifetime is typically very long, being suppressed both by the gravitational interaction and by the small $\rpv$ couplings. In Section \[darkmatter\], we will discuss the feasibility of gravitino DM in the $\mu\nu$SSM, as well as its possible detection in gamma-ray satellite experiments such as the Fermi Large Area Telescope (LAT). Our conclusions are left for Section \[conclusions\].
The $\mu\nu$SSM {#model}
===============
The superpotential of the $\mu\nu$SSM contains in addition to the MSSM Yukawas for quarks and charged leptons, Yukawas for neutrinos and the two couplings discussed in the introduction that generate the effective $\mu$ term and Majorana masses [@LopezFogliani:2005yw; @Escudero:2008jg]: $$\begin{aligned}
W & = &
\ \epsilon_{ab} \left(
Y_{u_{ij}} \, \hat H_u^b\, \hat Q^a_i \, \hat u_j^c +
Y_{d_{ij}} \, \hat H_d^a\, \hat Q^b_i \, \hat d_j^c +
Y_{e_{ij}} \, \hat H_d^a\, \hat L^b_i \, \hat e_j^c +
Y_{\nu_{ij}} \, \hat H_u^b\, \hat L^a_i \, \hat \nu^c_j
\right)
\nonumber\\
& - &
\epsilon{_{ab}} \lambda_{i} \, \hat \nu^c_i\,\hat H_d^a \hat H_u^b
+
\frac{1}{3}
\kappa{_{ijk}}
\hat \nu^c_i\hat \nu^c_j\hat \nu^c_k \ .
\label{superpotential}\end{aligned}$$ Notice that in the limit $Y_{\nu_{ij}} \to 0$, $\hat \nu^c_i$ can be identified as pure singlet superfields without lepton number, similar to the case of the NMSSM, where one singlet is added to the spectrum and there is $\rp$. Thus $\rpv$ in the $\mu\nu$SSM is determined by the values of the neutrino Yukawa couplings, and as a consequence is going to be small.
The Solution to the $\mu$ Problem
---------------------------------
To confirm that the $5^{\rm th}$ term in the superpotential generates dynamically the $\mu$ term (and the $6^{\rm th}$ term the Majorana masses for neutrinos), as discussed in the Introduction, we have to probe that the VEVs of the RH sneutrinos are naturally of the order of the EWSB scale. Since only dimensionless trilinear couplings are present in (\[superpotential\]), the EWSB is determined by the usual soft SUSY-breaking terms of the scalar potential. Thus it is remarkable that all known particle physics phenomenology can be reproduced in the $\mu\nu$SSM with one scale, avoiding the introduction of [*ad-hoc*]{} high-energy scales like e.g. in the GUT-scale seesaw.
To carry out the minimization, let us remember that in addition to the soft terms the tree-level neutral scalar potential receives the $D$ and $F$ term contributions that can be found in Refs. [@LopezFogliani:2005yw; @Escudero:2008jg]. With the choice of CP conservation,[^1] after the EWSB the neutral scalars develop in general the following real VEVs: $$\langle H_d^0 \rangle = v_d\ , \, \quad \langle H_u^0 \rangle = v_u\ , \,
\quad \langle \widetilde \nu_i \rangle = v_{\nu_i}\ , \, \quad
\langle \widetilde \nu_i^c \rangle = v_{\nu^c_i}\ ,
\label{vevs}$$ where in addition to the usual VEVs of the MSSM Higgses, $H_u^0$ and $H_d^0$, the new couplings generate VEVs for left-handed (LH) sneutrinos, $\widetilde \nu_i$, as well as for the RH sneutrinos, $\widetilde \nu_i^c$. The eight minimization conditions can be written as [$$\begin{aligned}
m_{H_{d}}^{2}
& =& -\frac{1}{4}G^2\left(v_{\nu_{i}}v_{\nu_{i}}+v_{d}^{2}-v_{u}^{2}\right)
- \lambda_i\lambda_{j}v_{\nu_i^c} v_{\nu_j^c}
- \lambda_{i}\lambda_{i}
v_{u}^{2}
\nonumber\\
&&
+v_{\nu_i^c}\tan\beta\left(a_{\lambda_i}
+ \lambda_{j}\kappa_{ijk}
v_{\nu_k^c}
\right)
+ Y_{\nu_{ij}} \frac{v_{\nu_{i}}}{v_d}
\left(\lambda_k v_{\nu_k^c} v_{\nu_j^c} +
\lambda_{j}
v_{u}^2\right)
\ ,
\label{tadpoles1}
\\
\nonumber
\\
m_{H_{u}}^{2}
& =& \frac{1}{4}G^2\left(v_{\nu_{i}}v_{\nu_{i}}+v_{d}^{2}-v_{u}^{2}\right)
- \lambda_{i}\lambda_j
v_{\nu_{i}^c}v_{\nu_{j}^c}
-\lambda_{j}\lambda_{j}v_{d}^2
\nonumber\\
&&
+ 2\lambda_j Y_{\nu_{ij}}v_{\nu_{i}}v_{d}
- Y_{\nu_{ij}}Y_{\nu_{ik}}
v_{\nu_{k}^c}v_{\nu_{j}^c}
- Y_{\nu_{ij}}Y_{\nu_{kj}}v_{\nu_{i}}v_{\nu_{k}}
\nonumber \\
&&
+v_{\nu^c_i} \frac{1}{\mathrm{tan} \beta}\left(a_{\lambda_i}
+\lambda_{j}\kappa_{ijk}v_{\nu^c_k}\right)
-\frac{v_{\nu_i}}{v_u}\left(a_{\nu_{ij}}v_{\nu^c_j}+
Y_{\nu_{ij}}\kappa_{ljk}v_{\nu^c_l}v_{\nu^c_k}\right)
\ ,
\label{tadpoles2}
\\
\nonumber\\
m^2_{\widetilde{\nu}_{ij}^{c}}v_{\nu_{j}^{c}} &=&
-a_{\nu_{ji}}v_{\nu_{j}}v_{u}
+ a_{\lambda_i}v_u v_d - a_{\kappa_{ijk}}v_{\nu_{j}^c}v_{\nu_{k}^c}
- \lambda_i\lambda_{j}\left(v_{u}^{2}+v_{d}^{2}\right)v_{\nu_{j}^c}
+ 2\lambda_{j}\kappa_{ijk}v_{d}v_{u}v_{\nu_{k}^c}
\nonumber \\
&&
- 2\kappa_{lim}\kappa_{ljk}v_{\nu_{m}^c}v_{\nu_{j}^c}v_{\nu_{k}^c}
+ Y_{\nu_{ji}}\lambda_{k}v_{\nu_{j}}v_{\nu_{k}^c}v_{d}
+ Y_{\nu_{kj}}\lambda_{i}v_{d}v_{\nu_{k}}v_{\nu_{j}^c}
- 2 Y_{\nu_{jk}}\kappa_{ikl}v_{u}v_{\nu_{j}}v_{\nu_{l}^c}
\nonumber \\
&&
- Y_{\nu_{ji}}Y_{\nu_{lk}}v_{\nu_{j}}v_{\nu_{l}}v_{\nu_{k}^c}
- Y_{\nu_{ki}}Y_{\nu_{kj}}v_{u}^{2}v_{\nu_{j}^c}
\ ,
\label{tadpoles3}
\\
\nonumber\\
m^2_{\widetilde{L}_{ij}}v_{\nu_{j}}& =& -\frac{1}{4}G^2\left(v_{\nu_{j}}v_{\nu_{j}}+v_{d}^{2}-v_{u}^{2}\right)v_{\nu_{i}}
-a_{\nu_{ij}}v_{u}v_{\nu_{j}^c}
+ Y_{\nu_{ij}}\lambda_{k}v_{d}v_{\nu_{j}^c}v_{\nu_{k}^c}
+ Y_{\nu_{ij}}\lambda_{j}v_{u}^{2}v_{d}
\nonumber \\
&&
- Y_{\nu_{il}}\kappa_{ljk}v_{u}v_{\nu_{j}^c}v_{\nu_{k}^c}
- Y_{\nu_{ij}} Y_{\nu_{lk}}v_{\nu_{l}}v_{\nu_{j}^c}v_{\nu_{k}^c}
- Y_{\nu_{ik}}Y_{\nu_{jk}}v_{u}^{2}v_{\nu_{j}^c}
\ ,
\label{tadpoles4}\end{aligned}$$ ]{} where the low-energy soft masses $m_{H_{d}}^{2}$, $m_{H_{u}}^{2}$, $m^2_{\widetilde{\nu}_{ij}^{c}}$ and $m^2_{\widetilde{L}_{ij}}$ are calculated as functions of the VEVs $v_d$, $v_u$, $v_{\nu^c_i}$, $v_{\nu_i}$. Besides, inspired by the structure of supergravity, the soft trilinear parameters are taken directly proportional to the couplings, $a_{\lambda_i}= A_{\lambda_i}\lambda_i$, $a_{\kappa_{ijk}}= A_{\kappa_{ijk}} \kappa_{ijk}$, $a_{\nu_{ij}}= A_{\nu_{ij}} Y_{\nu_{ij}}$, etc., where the summation convention on repeated indices does not apply for these particular formulas.
As can be easily seen from Eq. (\[tadpoles3\]), the VEVs of the RH sneutrinos, $v_{\nu_{j}^{c}}$, are naturally of the order of the EWSB scale, confirming that the solution to the $\mu$ problem works.
The solution to the $\nu$ problem {#nuproblem}
---------------------------------
Using the same argument as above, we trivially confirm that the $6^{\rm th}$ term in the superpotential (\[superpotential\]) generates the effective Majorana masses for RH neutrinos, as discussed in the Introduction. Thus we can implement naturally an EW-scale seesaw in the $\mu\nu$SSM, asking for neutrino Yukawa couplings of the order of the electron Yukawa coupling or smaller, $Y_{\nu_{ij}} \sim 10^{-6} - 10^{-7}$ [@LopezFogliani:2005yw; @Escudero:2008jg; @Ghosh:2008yh; @Bartl:2009an; @Fidalgo:2009dm; @Ghosh:2010zi; @LopezFogliani:2010bf; @Ghosh:2010ig], i.e. we work with Dirac masses for neutrinos, $m_D\sim Y_{\nu}v_u\lsim 10^{-4}$ GeV. On the other hand, the VEVs of the LH sneutrinos, $v_{\nu_i}$, are much smaller than the other VEVs (\[vevs\]) in the $\mu\nu$SSM. Notice in this respect that in Eq. (\[tadpoles4\]), $v_{\nu}\to 0$ as $Y_{\nu}\to 0$. It is then easy to estimate the values of these VEVs as $v_{\nu}\lsim m_D$ [@LopezFogliani:2005yw].
As is well known, the couplings and Higgs VEVs present in the MSSM (determined by the superpotential (\[superpotential1\])) generate the mixing of neutral gauginos and Higgsinos, where the eigenstates are the so-called neutralinos. A similar situation occurs in the $\mu\nu$SSM, however in this model there are new couplings and VEVs (see Eqs. (\[superpotential\]) and (\[vevs\])), implying larger mass matrices than those of the MSSM or NMSSM. In particular, in the case of the neutralinos, they turn out to be also mixed with the LH and RH neutrinos. Besides, we have seen before that Majorana masses for RH neutrinos are generated dynamically, thus they will behave as the singlino components of the neutralinos. Altogether, in a basis where ${\chi^{0}}^T=(\tilde{B^{0}},
\tilde{W^{0}},\tilde{H_{d}},\tilde{H_{u}},\nu_{R_i},\nu_{L_i})$, one obtains the following $10\times 10$ neutral fermion (neutralino-neutrino) mass matrix [@LopezFogliani:2005yw; @Escudero:2008jg]: $${\mathcal M}_n=
\left(
\begin{array}{cc}
M & m\\
m^{T} & 0_{3\times3}
\end{array}
\right)\ ,
\label{mixing}$$ with
$$M=
\left(
\begin{array}{ccccccc}
M_{1} & 0 & -A v_{d} & A v_{u} & 0 & 0 & 0\\
0 & M_{2} &B v_{d} & -B v_{u} & 0 & 0 & 0\\
-A v_{d} & B v_{d} & 0 & -\lambda_{i}v_{\nu^c_{i}} & -\lambda_{1}v_{u} & -\lambda_{2}v_{u} & -\lambda_{3}v_{u}\\
A v_{u} & -B v_{u} & \: \: -\lambda_{i}v_{\nu^c_{i}} & 0 & -\lambda_{1}v_{d}+Y_{\nu_{i1}}v_{\nu_{i}} & -\lambda_{2}v_{d}+Y_{\nu_{i2}}v_{\nu_{i}} & -\lambda_{3}v_{d}+Y_{\nu_{i3}}v_{\nu_{i}}\\
0 & 0 & -\lambda_{1}v_{u} & \: \:-\lambda_{1}v_{d}+Y_{\nu_{i1}}v_{\nu_{i}} & 2\kappa_{11j}v_{\nu^c_{j}} & 2\kappa_{12j}v_{\nu^c_{j}} & 2\kappa_{13j}v_{\nu^c_{j}}\\
0 & 0 & -\lambda_{2}v_{u} & \: \: -\lambda_{2}v_{d}+Y_{\nu_{i2}}v_{\nu_{i}} & 2\kappa_{21j}v_{\nu^c_{j}} & 2\kappa_{22j}v_{\nu^c_{j}} & 2\kappa_{23j}v_{\nu^c_{j}}\\
0 & 0 & -\lambda_{3}v_{u} & \: \:-\lambda_{3}v_{d}+Y_{\nu_{i3}}v_{\nu_{i}} & 2\kappa_{31j}v_{\nu^c_{j}} & 2\kappa_{32j}v_{\nu^c_{j}} & 2\kappa_{33j}v_{\nu^c_{j}}
\end{array}
\right)\ ,
\label{neumatrix}
$$
where $A\equiv\frac{G}{\sqrt{2}} \sin\theta_W$, $B\equiv\frac{G}{\sqrt{2}} \cos\theta_W$, with $G^2\equiv g_{1}^{2}+g_{2}^{2}$, and $$m^{T}=
\left(\begin{array}{ccccccc}
-\frac{g_{1}}{\sqrt{2}}v_{\nu_{1}} \: & \: \frac{g_{2}}{\sqrt{2}}v_{\nu_{1}} & \: 0 & \: Y_{\nu_{1i}}v_{\nu^c_{i}} & \: Y_{\nu_{11}}v_{u} & \: Y_{\nu_{12}}v_{u} & \: Y_{\nu_{13}}v_{u}\\
\: -\frac{g_{1}}{\sqrt{2}}v_{\nu_{2}} & \: \frac{g_{2}}{\sqrt{2}}v_{\nu_{2}} & \: 0 & \: Y_{\nu_{2i}}v_{\nu^c_{i}} & \: Y_{\nu_{21}}v_{u} & \: Y_{\nu_{22}}v_{u} & \: Y_{\nu_{23}}v_{u}\\
\: -\frac{g_{1}}{\sqrt{2}}v_{\nu_{3}}\: & \: \frac{g_{2}}{\sqrt{2}}v_{\nu_{3}} & \: 0 & \: Y_{\nu_{3i}}v_{\nu^c_{i}} & \: Y_{\nu_{31}}v_{u} & \: Y_{\nu_{32}}v_{u} & \: Y_{\nu_{33}}v_{u}\end{array}\right)\ .
\label{mixing3}$$ The structure of this mass matrix is that of a generalized EW-scale seesaw, since it involves not only the RH neutrinos but also the neutralinos. Because of this structure, data on neutrino physics can easily be reproduced at tree level [@LopezFogliani:2005yw; @Escudero:2008jg; @Ghosh:2008yh; @Fidalgo:2009dm; @Ghosh:2010zi], even with diagonal Yukawa couplings $Y_{\nu_i}$ [@Ghosh:2008yh; @Fidalgo:2009dm]. Qualitatively, we can understand this in the following way. First of all, neutrino masses are going to be very small since the entries of the matrix $M$ are much larger than the ones of the matrix $m$. Notice in this sense that the entries of $M$ are of the order of the EW scale, whereas the ones in $m$ are of the order of the Dirac masses for neutrinos [@LopezFogliani:2005yw; @Escudero:2008jg]. Second, from the above matrices, in the limit of large $\tan\beta$ (where $\tan\beta\equiv v_u/v_d$) one can obtain a simplified formula for the effective neutrino mixing mass matrix [@Fidalgo:2009dm]: $$\begin{aligned}
(m^{eff}_{\nu})_{ij}
\simeq \frac{Y_{\nu_i}Y_{\nu_j}v_u^2}
{6 \kappa v_{\nu^c}}
(1-3 \delta_{ij})-\frac{v_{\nu_i} v_{\nu_j}}{2M} \ ,
\label{Limit no mixing Higgsinos gauginos}\end{aligned}$$ where $\kappa_{iii}\equiv \kappa_{i} \equiv \kappa$ and vanishing otherwise, $v_{\nu^c_i}\equiv v_{\nu^c}$, and ${M} \equiv \frac{M_1 M_2}{g_1^2 M_2 + g_2^2 M_1}$. Using this approximate formula it is easy to understand how diagonal Yukawas, $Y_{\nu_{ii}}=Y_{\nu_i}$ and vanishing otherwise, can give rise to off-diagonal entries in the mass matrix. One of the key points is the extra contribution given by the first term of Eq. (\[Limit no mixing Higgsinos gauginos\]) with respect to the ordinary seesaw where it is absent. Another extra contribution to the off-diagonal entries is the last term, which is generated through the mixing of LH neutrinos with gauginos.
In a sense, all these arguments give an answer to the question why the mixing angles are so different in the quark and lepton sectors: because no generalized seesaw exists for the quarks.
LHC Phenomenology {#lhc}
=================
We have already mentioned in the previous section that mass matrices in the $\mn$ are larger than in the MSSM or NMSSM, because of the presence of new couplings and VEVs. We also discussed the case of the $10\times 10$ neutral fermion (neutralino-neutrino) mass matrix. For the rest of the mass matrices, a similar situation occurs and new mixing of states are induced [@LopezFogliani:2005yw; @Escudero:2008jg]. Summarizing, there are five charged fermions (charginos-charged leptons), seven CP-odd and eight CP-even neutral scalars (Higgses-sneutrinos), and seven charged scalars (charged Higgses-sleptons). As a consequence, the phenomenology of the $\mu\nu$SSM is very different from the one of the MSSM or NMSSM.
Needless to mention, as in $\rpv$ models the LSP is no longer stable, and therefore not all SUSY chains must yield missing energy events at colliders. In particular, in the $\mu\nu$SSM the LSP decays leading to prompt or displaced vertices, depending on the value of the couplings, and producing final states with multi-leptons/taus/jets/photons and missing energy. This unusual phenomenology was explored first in Refs. [@Bartl:2009an; @Bandyopadhyay:2010cu; @Fidalgo:2011ky; @Ghosh:2012pq], discussing the decay properties of the LSP assumed to be the lightest neutralino, as well as novel Higgs decays. Further, detailed collider analyses for a Higgs-like scalar decaying into a pair of neutralinos was also discussed in Refs. [@Bandyopadhyay:2010cu; @Ghosh:2012pq], provided that these states lie below in the mass spectrum. More recently, this issue was revisited and, under the same assumption, a Higgs-like scalar decaying to a pair of scalars/pseudoscalars was also considered [@Ghosh:2014ida]. The case of non-standard on-shell decays of $W^\pm$ and $Z$ bosons to light singlet-like scalar(s), pseudoscalar(s) and neutralinos(s) was studied in Ref. [@Ghosh:2014rha].
On the other hand, all sparticles are potential LSP’s in $\rpv$ models, since the problem of stable charged particles as DM is not present. So to study the whole potential phenomenology of the $\mu\nu$SSM at the LHC, we should be prepared to analyze systematically not only the usual lightest neutralino as the LSP, but also the lightest stau, squark, chargino, and sneutrino as LSP’s with a wide range of masses. In a first detailed analysis [@sneutrino] we have concentrated in the LH sneutrino as the LSP. We have shown that for a sneutrino mass in the range about $95-145$ GeV, a diphoton signal plus leptons, or plus missing transverse energy (from neutrinos), is observable at the LHC, even at the current Run 2 with 100 fb$^{-1}$ of luminosity. The dominant sneutrino pair production channels are the direct production via a $Z$ boson, or through a $W^{\pm}$ decaying into a sneutrino and a LH charged slepton next-to-LSP, with the latter decaying into another sneutrino plus a very soft $W^{\pm}$. We think that these signals (where one of the sneutrinos decays in a way not very different from the Higgs) are worthy of attention by our experimental colleagues.
Gravitino Dark Matter {#darkmatter}
=====================
As already mentioned in the Introduction, the gravitino is an interesting candidate for DM in $\rpv$ models. This occurs when it becomes the LSP. The gravitino has an interaction term in the supergravity Lagrangian with the photon and the photino. Since the photino and the LH neutrinos are mixed in the neutral fermion mass matrix due to the $\rpv$, as discussed in Eq. (\[mixing3\]), the gravitino will be able to decay into a photon and a neutrino, as shown in Fig. \[decay\]. Nevertheless, this decay is suppressed both by the gravitational interaction and by the small $\rpv$ coupling, making the gravitino lifetime much longer than the age of the Universe [@Takayama:2000uz]. From the supergravity Lagrangian one obtains $$\Gamma\left(\Psi_{3/2}
\to \sum_i \gamma \nu_i\right)\simeq
\frac{1}{64 \pi} |U_{\widetilde{\gamma} \nu}|^2\,\frac{m^3_{3/2}}{M_P^2}\,,
\label{c1}$$ where $M_P=2.4\times 10^{18}$GeV is the reduced Planck mass, $m_{3/2}$ is the gravitino mass, and $|U_{\widetilde{\gamma}\nu}|^{2}$ determines the neutrino content of the photino: $$|U_{\widetilde{\gamma}\nu}|^{2}=\sum_{i=1}^{3}|N_{i1}\cos\theta_{W}+N_{i2}\sin\theta_{W}|^{2}.$$ Here $N_{i1}$ ($N_{i2}$) is the Bino (Wino) component of the $i$-th neutrino, and $\theta_{W}$ is the weak mixing angle. The same result for the decay width holds for the conjugated processes $\Psi_{3/2}\rightarrow\gamma\bar{\nu}_i$.
Assuming that this is the only relevant decay channel of the gravitino, its lifetime can then be written as $${\tau}
(\Psi_{3/2}\rightarrow\sum_i\gamma\nu_i
)
=\frac{1}{2\Gamma\left(\Psi_{3/2}\rightarrow\sum_i\gamma\nu_i\right)}
\simeq 3.8\times 10^{27}\, {s}
\left(\frac{10^{-16}}{|U_{\widetilde{\gamma}\nu}|^2}\right)
\left(\frac{10\, \mathrm{GeV}}{m_{3/2}}\right)^{3}\ ,
\label{lifetimegamma}$$ where the factor 2 takes into account the charged conjugated final states. If $|U_{\widetilde{\gamma}\nu}|^{2}$ is small enough, the gravitino can be very long lived compared to the current age of the Universe which is about $4\times 10^{17}$s.
[figgfn-eps-converted-to.pdf]{}
We can easily estimate the value of $|U_{\widetilde{\gamma}\nu}|^{2}$ in the $\mu\nu$SSM [@Choi:2009ng]. Using the mass insertion technique, from the entries in the neutral fermion mass matrix (\[mixing3\]) and Fig. \[decay\], we can deduce that the relevant coupling for the mixing between the photino and the neutrinos is given approximately by $g_{1}v_{\nu}$, and as a consequence $|U_{\widetilde{\gamma}\nu}|\sim |\frac{g_{1}v_{\tilde\nu}}{M_{1}}| \sim 10^{-6}$-$10^{-8}$, giving rise to $$10^{-16} \lesssim |U_{\widetilde{\gamma}\nu}|^{2} \lesssim 10^{-12}.
\label{representative}$$ One can confirm this estimation performing a scan of the low-energy parameter space of the $\mu\nu$SSM with the exact formulas above [@Choi:2009ng], imposing that neutrino data must be reproduced. As a result of the scan, typically the mass of the neutralino is above 20GeV, and since $m_{3/2}$ is constrained to be smaller than that value, as we will see, the gravitino can safely be used as the LSP. Let us remark then, that under this assumption of gravitino DM, each candidate for LSP mentioned in the previous section would in fact be the next-to-LSP (NLSP), since the gravitino would be the LSP. Nevertheless, the analysis of the phenomenology at the LHC would not be altered since the NLSP would also decay into ordinary particles using the same channels as if it were the LSP. Thus our analysis there can be applied exactly the same for the case of neutralino/sneutrino/stau/squark/chargino NLSP with the gravitino as the LSP.
On the other hand, for the gravitino to be a good DM candidate we still need to check that it can be present in the right amount to explain the observed relic density $\Omega_{DM} h^2 \simeq 0.1$. With the introduction of inflation, the primordial gravitinos are diluted during the exponential expansion of the Universe. Nevertheless, after inflation, in the reheating process, the gravitinos are reproduced again from the relativistic particles in the thermal bath. The yield of gravitinos from the scatterings is proportional to the reheating temperature, $T_R$, and estimated to be [@Bolz:2000fu] $$\Omega_{3/2} h^2 \simeq 0.27\left(\frac{T_R}{10^{10}\ GeV} \right)
\left(\frac{100\ GeV}{m_{3/2}} \right)\left(\frac{M_{\tilde g}}{1\ TeV}\right)^2,
\label{oh2buchmuller}$$ where $M_{\tilde g}$ is the gluino mass. As is well known, adjusting the reheating temperature one can reproduce the correct relic density for each possible value of the gravitino mass. For example for $m_{3/2}$ of the order of $1-1000$ GeV one obtains $\Omega_{3/2} h^2 \simeq 0.1$ for $T_R\sim 10^8-10^{11}$ GeV, with $M_{\tilde g}\sim 1$ TeV. Even with a high value of $T_R$ there is no cosmological gravitino problem, since the NLSP decays to standard model particles much earlier than Big Bang nucleosynthesis (BBN) epoch via $\rpv$ interactions.
Thus, the gravitino, which is a super-weakly interacting massive particle (superWIMP), represents a good DM candidate. Most importantly, as pointed out in Ref. [@Takayama:2000uz] for the case of $\rpv$, gravitino decays in the Milky Way halo would produce monochromatic gamma rays with an energy equal to half of the gravitino mass, and therefore its presence can, in principle, be inferred indirectly from gamma-ray observations. We will discuss this crucial issue in the next subsection.
Detection
----------
The detection of gravitino DM in several $\rpv$ scenarios has been studied in the literature considering the case of gravitinos emitting gamma rays when decaying in the smooth galactic halo and extragalactic regions at cosmological distances [@Takayama:2000uz; @Buchmuller:2007ui; @Bertone:2007aw; @Ibarra:2007wg; @Ishiwata:2009vx; @Buchmuller:2009xv; @Choi:2009ng; @Restrepo:2011rj; @Albert:2014hwa], and also in nearby extragalactic structures [@GomezVargas:2011ph]. In the interesting case of the galactic halo, the gamma-ray signal is an anisotropic sharp line and the flux is given by $$\label{halo}
\frac{d\Phi}{dE}(E)=\frac{\delta(E-\frac{m_{3/2}}{2})}{4 \pi \tau_{3/2} m_{3/2}}
\int_{\textrm{los}}\rho_{halo}(\vec{l})d\vec{l}\ .
$$ It is worth noting that this equation has two independent factors. The first one corresponds to the particle physics properties of the DM candidate. In particular, its lifetime, $\tau_{3/2}$, its mass, $m_{3/2}$, and a delta function associated to the fact that the gravitino decays into a photon (and a neutrino), producing therefore a line with an energy equal to $m_{3/2}/2$. The second factor corresponds to the astrophysics and is given by the integral along the line of sight $l$ of the halo DM density.
![ Skymap of the ROI used in our analysis [@Albert:2014hwa]; plotted in Galactic coordinates using the Hammer-Aitoff projection. The region ROI$_{pol}$ (blue) is optimized for the signal-to-background ratio in the case of DM decay, while the region ROI$_{cen}$ (red) is optimised for the signal-to-background ratio in the case of DM annihilation. The dashed line encloses the area for the control regions along the Galactic plane (light gray), while the gray region is an example of one of the 31 control regions used in the analysis. []{data-label="fig:gfn"}](roidm-eps-converted-to.pdf){width="90.00000%"}
A first analysis in the $\mu\nu$SSM of the possible detection of this kind of signal in the Fermi-LAT was carried out in Ref. [@Choi:2009ng]. Taking into account the data reported by Fermi at that time, from the non-observation of lines it was possible to constrain the lifetime and the mass of the gravitino. In particular, the mass has to be around 10 GeV or smaller. In a more recent work together with Fermi-LAT members [@Albert:2014hwa], a search for 100 MeV to 10 GeV gamma-ray lines was carried out using 62 months of Fermi-LAT data, and the implications for gravitino DM in the $\mu\nu$SSM were analyzed. In this category 2 paper of the Fermi-LAT collaboration we used an Einasto profile with a finite central density [@einasto; @Navarro:2003ew]: $$\rho_{{Ein}}(r)=\rho_{\odot}\exp\left( -\frac{2}{\alpha}\left( \left( \frac{r}{r_s}\right) ^\alpha-\left( \frac{R_{\odot}}{r_s}\right) ^\alpha\right) \right) ,$$ where we adopted $\alpha=0.17$ and $r_s=20\,$kpc for the case of the Milky Way and a local DM density of $\rho_{\odot}\simeq 0.4$ GeV cm$^{-3}$ [@Catena:2009mf; @Weber:2009pt; @Salucci:2010qr]. Other halo profiles as well as uncertainties on the halo parameters were also taken into account, but all these profiles behave similar in the outer part of the Milky Way, where is our region of interest (ROI), and therefore the results are similar. Concerning the ROI, we selected one that optimizes the signal-to-background ratio for searches for decay, where the Galactic poles are included, ROI$_{pol}: |b| > 60^o$. This is shown in Fig. \[fig:gfn\] The final result of the analysis is shown in Fig. \[fig:results\]. We did not find any statistically significant spectral lines and have set robust limits on DM interactions that would produce monochromatic gamma rays. When these limits are applied to the $\mu\nu$SSM, under the assumption that the gravitino is the DM, we find that the mass must be $m_{3/2}<4.8$ GeV and the lifetime $\tau_{3/2}>7.9\times 10^{27}$ s at 95% CL if we assume that all the DM in the Universe is in the form of gravitinos.
![Result of Ref. [@Albert:2014hwa], where the parameter space of decaying gravitino DM is given in terms of the gravitino lifetime and the gravitino mass. The diagonal band shows the allowed parameter space for gravitino DM in the $\mu\nu$SSM. The numbers on the solid and dashed lines show the corresponding value of the photino/neutrino mixing parameter, as discussed in section 4. The theoretically most favored region is colored in gray. We also show several 95% CL lower limits on the gravitino lifetime coming from gamma-ray observations. The blue shaded region is excluded by the limits derived in the paper. []{data-label="fig:results"}](results-eps-converted-to.pdf){width="90.00000%"}
In a work in preparation [@gravitino], we are performing a deeper exploration of the $\mu\nu$SSM parameter space, taking also into account 3-body final states in the computation. The preliminary result shows that in some regions of the parameter space is possible to increase the upper bound on the gravitino mass to about 20 GeV and to lower the lower bound on the lifetime to about $10^{25}$ s.
Conclusions
===========
The $\mu\nu$SSM solves the $\mu$ problem of SUSY models and reproduces neutrino data, simply using couplings with the three families of RH neutrinos. These new couplings produce $\rpv$, generating a phenomenology very different from the one of the MSSM or the NMSSM. We have shown that novel signatures of SUSY at the LHC are expected. In particular, all sparticles are potential candidates for the LSP, not only the usual lightest neutralino but also the lightest stau, squark, chargino, sneutrino. The LSP is not stable leading to prompt or displaced vertices, and producing final states with multi-leptons/taus/jets/photons and missing energy. On the other hand, the gravitino turns out to be an interesting candidate for DM, since its lifetime is longer than the age of the Universe. It can be searched through gamma-ray observations such as those of the Fermi-LAT. The non-observation of spectral lines allows to set robust limits on the parameters of the model. In particular, depending on the region of the parameter space, the gravitino mass must be smaller than about $5-20$ GeV and the lifetime larger than about $10^{25-28}$ s.
Acknowledgments {#acknowledgments .unnumbered}
===============
I gratefully acknowledge the local organizers of DSU 2015 for the fantastic atmosphere of the workshop, and the marvellous visit to Yukawa’s office where I am sure that all participants hope to have gotten some inspiration.
This work was supported in part by the Spanish grant FPA2015-65929-P MINECO/FEDER UE, and by the Programme SEV-2012-0249 ‘Centro de Excelencia Severo Ochoa’. The author also acknowledges specially the support of the Spanish MINECO’s Consolider-Ingenio 2010 Programme under grant MultiDark CSD2009-00064.
[^1]: $\mu\nu$SSM with spontaneous CP violation was studied in Ref. [@Fidalgo:2009dm].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In the context of global optimization and mixed-integer non-linear programming, generalizing a technique of D’Ambrosio, Fampa, Lee and Vigerske for handling the square-root function, we develop a virtuous smoothing method, using cubics, aimed at functions having some limited non-smoothness. Our results pertain to root functions ($w^p$ with $0<p<1$) and their increasing concave relatives. We provide (i) a sufficient condition (which applies to functions more general than root functions) for our smoothing to be increasing and concave, (ii) a proof that when $p=1/q$ for integers $q\geq 2$, our smoothing lower bounds the root function, (iii) substantial progress (i.e., a proof for integers $2\leq q\leq 10,000$) on the conjecture that our smoothing is a sharper bound on the root function than the natural and simpler “shifted root function”, and (iv) for all root functions, a quantification of the superiority (in an average sense) of our smoothing versus the shifted root function near 0.'
author:
- Jon Lee
- Daphne Skipper
bibliography:
- 'smooth.bib'
title: Virtuous smoothing for global optimization
---
Introduction {#sec:intro .unnumbered}
============
Important models for framing and attacking hard (often non-linear) combinatorial-optimization problems are GO (global optimization) and MINLP (mixed-integer non-linear programming). Virtually all GO and MINLP solvers (e.g., `SCIP` [@Achterberg2009], `Baron` [@sahinidis], `Couenne` [@Belotti09], `Antigone` [@misener-floudas:ANTIGONE:2014]) apply some variant of spatial branch-and-bound (see, [@Smith99], for example), and they rely on NLP (non-linear-programming) solvers, both to solve continuous relaxations (to generate lower bounds for minimization) and often to generate good feasible solutions (to generate upper bounds). Sometimes models are organized to have a convex relaxation, and then either outer approximation, NLP-based branch-and-bound, or some hybrid of the two is employed (e.g., `Bonmin` [@Bonami]; also see [@BLLW]). In such a case, NLP solvers are also heavily relied upon, and for the same uses as in the non-convex case.
Convergence of most NLP solvers (e.g. `Ipopt` [@WB06]) requires that functions be twice continuously differentiable. Yet many models naturally utilize functions with some limited non-differentiability. One approach to handle limited non-differentiability is smoothing.
Of course there is a vast literature on global optimization concerning convexification (see [@TawarSahinBook]). Such research aims at developing tractable lower bounds for non-convex formulations $z:=\min \{f_0(x) ~:~ x\in \mathcal{F}\}$. Even when the only non-convexities are integrality of some variables, improving the trivial convexification (relaxing integrality) is crucial to the success of algorithms such as outer-approximation (see [@Bonami], for example). But lower bounding $z$ is not the complete story. For example, spatial branch-and-bound (for formulations that are non-convex after relaxing integrality) and outer-approximation (for formulations that are convex after relaxing integrality) both require solutions of the NLPs (convex or not) obtained by relaxing integrality. They do this in an effort to find actual (incumbent) feasible solutions and hence upper bounds on $z$. The success of this step, requires close approximation of the MINLP and tractability of the NLPs. Our results for smoothing are aimed at getting tractable NLPs in the presence of functions with limited non-differentiability. For univariate concave functions, we aim for concave under-estimation which has the effect of forcing the approximation to be near the function that we approximate. Note that convexifying a univariate concave function via secant under-estimation can do a rather poor job of close approximation. This idea of using one version of a function for convexification in the context of lower bounding $z$ and another version of a function aimed at coming closer to the MINLP was implemented for objective functions in `Bonmin`, in the context of an application; see [@BDLLT06; @BDLLT12] (a study of optimal water-network refurbishment via MINLP). Additionally, it is possible to simply use our smoothing methodology as a formulation pre-processor for a spatial branch-and-bound algorithm. With such a use, the smoothing inherits the convexification structure of the MINLP; that is, our approximation, because of its concavity, allows for secant under-estimation and tangent over-estimation, and this has been implemented in `SCIP` (see §\[subsec:soft\]). Furthermore, in the context of smoothing as a pre-processing step, [@DFLV2014; @DFLV2015] used the under-estimation property of our smoothing to get an *a priori* upper-bound on how much the optimal value of their smoothed MINLP could be below $z$. Finally, also employing smooth concave under-estimators, [@Sergeyev1998] describes a global-optimization algorithm for univariate functions; so we can see another use of such under estimators in global optimization.
Additionally in [@BDLLT06; @BDLLT12], an ad-hoc smoothing method is used to address non-differentiability near 0 of the Hazen-Williams (empirical) formula for the pressure drop of turbulent water flow in a pipe as a function of the flow. Choosing a small positive $\delta$ and fitting an odd homogeneous quintic on $[-\delta,\delta]$, so as to match the function and its first and second derivatives at $\pm\delta$ and the function value at 0, the resulting piecewise function is smooth enough for NLP solvers. However on $(-\delta,\delta)$ the quintic is neither an upper bound nor a lower bound on the function it approximates.
In [@GMS13] (a study of the TSP with “neighborhoods”), $\sqrt{w}$ is smoothed (near 0) by choosing again a small positive $\delta$ and then using a linear extrapolation of $\sqrt{w}$ at $w=\delta$ to approximate $\sqrt{w}$ on $[0,\delta)$. Shortcomings of this approach are that the resulting piecewise function is: (i) not twice differentiable at $w=\delta$, and (ii) over-estimates $\sqrt{w}$ on $[0,\delta)$. Regarding (ii), in many formulations (see [@DFLV2014; @DFLV2015], for example), we need an under-estimate of the function we are approximating to get a valid relaxation.
To address the identified shortcomings of the methodology of [@GMS13] for smoothing square roots, motivated by developing tractable mixed-integer non-linear-optimization models for the Euclidean Steiner Problem (see [@ITOR:ITOR12207]), [@DFLV2014; @DFLV2015] developed a virtuous method. On the interval $[0,\delta]$, they fit a homogeneous cubic, to match the function value of the square root at the endpoints, and matching the first and second derivatives at $w=\delta$. They demonstrate that the resulting smooth function, though piecewise defined, is increasing, strictly concave, under-estimates $\sqrt{w}$ on $(0,\delta)$, and is a stronger under-estimate of $\sqrt{w}$ than the more elementary “shift” $\sqrt{w+\lambda}-\sqrt{\lambda}$, for some $\lambda>0$. To make a fair comparison, $\delta$ is chosen as a function of $\lambda$ so that the two approximating functions have the same derivative at 0. This makes sense because, for each approximation, we would choose the value of the smoothing parameter ($\delta$ or $\lambda$) as small as the numerics will tolerate — that is, we would have an upper bound for the derivative of the approximation (at zero, where it is greatest).
For the $\sqrt{\cdot}$ function, we have depicted all of these smoothings in Figure \[fig:smooth\]. From top to bottom: The “linear extrapolation” follows the dot-dashed line ($\cdot\,\text{-}\cdot\text{-}\,\cdot$) below $\delta=0.1$. The solid curve ($\textbf{-----}$) is the true $\sqrt{\cdot}$. The “smooth under-estimation”, which we advocate, follows the dotted curve ($\cdot\!\cdot\!\cdot\!\cdot\!\cdot\!\cdot$) below $\delta=0.1$. The “shift”, chosen to have the same derivative as our preferred smoothing at 0, follows the weaker under-estimate (on the entire non-negative domain) given by the dashed curve (- - -).
![Behavior of all smoothings of the square root[]{data-label="fig:smooth"}](smoothSteiner.pdf){width=".80\textwidth"}
Most of our results are for root functions ($w^p$ with $0<p<1$) and their increasing concave relatives. Smoothing roots (not just square roots) can play an important role in working with $\ell_q$ norms ($q>1$) and quasi-norms ($0<q<1$): $\ell_q(x):=(\sum_j x_j^q)^{1/q}$. Depending on $q$ (less than 1, or greater than 1), each inner power or the outer power is a root function. Of course the use of norms is quite common. Additionally, in sparse optimization, $\ell_q$ quasi-norms are used to induce sparsity; see, for example, [@Wotao] and the references therein. So, our work can be used to apply global-optimization techniques in this important setting. Additionally, root functions are natural for fitting nearly-smooth functions to data that follows an increasing concave trend; for example, cost functions with economies of scale, and the well-known Cobb-Douglas production function (and generalizations), relating production to labor and capital inputs, where the exponents of the inputs are the output elasticities (see [@douglas] and [@arrow]). After such a data-analysis step, fitted functions can be incorporated into optimization models, and our results would then be applicable; also see [@Cozad] for a modern data-driven integrated function-fitting/optimization methodology. Additionally, roots occur in other signomial functions besides the Cobb-Douglas production function (see [@Duffin1973]). Finally, besides root functions, there are other simple univariate building-block functions that our scheme applies to; for example $\log(1+w)$ and ${\rm ArcSinh}(\sqrt{w})$ (see Examples \[example:log\] and \[example:sinh\]).
In §\[sec:general\], we provide a sufficient condition (which applies to functions more general than root functions) for our smoothing to be increasing and concave. Moreover, we give an interesting example to illustrate that when our condition is not satisfied, the conclusion need not hold. In §\[sec:roots\], we establish that when $p=1/q$ for integers $q\geq 2$, our smoothing lower bounds the root function. Having such control over the root function is important in the context of global optimization — in fact, this was a key motivation of [@DFLV2014; @DFLV2015]. In §\[sec:better\], we present substantial progress (i.e., a proof for integers $2\leq q\leq 10,000$) on the conjecture that our smoothing is a sharper bound on the root function than the natural and simpler “shifted root function”. In §\[sec:avg\] we quantify the average relative performance of our smoothing and of the shifted root function near 0. We demonstrate that our smoothing is much better with respect to this performance measure. Finally, in §\[sec:future\], we make some concluding remarks: describing alternatives, available software, some extended use, and our ongoing work.
General smoothing via a homogenous cubic {#sec:general}
========================================
Construction of our smoothing {#sec:cubic}
-----------------------------
We are given a function $f$ defined on $[0, +\infty)$ having the following properties: $f(0)=0$, $f$ is increasing and concave on $[0, +\infty)$, $f'(w)$ and $f''(w)$ are defined on all of $(0,+\infty)$, but $f'(0)$ is undefined. For example, the *root function* $f(w):=w^p$, with $0<p<1$, has these properties. Our goal is to find a function $g$ that mimics $f$ well, but is differentiable everywhere (in particular at 0). In addition, because our context is global optimization, we want $g$ to lower bound $f$ on $[0, +\infty)$. In this way, we can develop smooth relaxations of certain optimization problems involving $f$.
The definition of our function $g$ depends on a parameter $\delta>0$. Our function $g$ is simply $f$ on $[\delta,+\infty)$. This parameter $\delta$ allows us to control the derivative of $g$ at 0. Essentially, lowering $\delta$ increases the derivative of $g$ at 0, and so in practice, we choose $\delta$ as low as the numerics will tolerate.
We extend $g$ to $[0,\delta]$, as a homogeneous cubic, so that $g(0)=f(0)=0$, $g(\delta)=f(\delta)$, $g'(\delta)=f'(\delta)$ and $g''(\delta)=f''(\delta)$. The homogeneity immediately gives us $g(0)=0$, and such a polynomial is the lowest-degree one that allows us to match $f$, $f'$ and $f''$ at $\delta$. We choose the three coefficients of $g(w) := Aw^3 + Bw^2 + Cw$ so that the remaining three conditions are satisfied.
The constants $A$, $B$ and $C$ are solutions to the system: $$\begin{aligned}
{3}
(g(\delta)=)\quad& \delta^3 A + \delta^2 B + \delta C && = f(\delta)\\
(g'(\delta)=)\quad& 3\delta^2 A + 2\delta B + C && = f'(\delta)\\
(g''(\delta)=)\quad& 6\delta A + 2B && = f''(\delta).\end{aligned}$$
We find that $$\begin{aligned}
A &~=~& \frac{f(\delta)}{\delta^3}~-~\frac{f'(\delta)}{\delta^2}~+~\frac{f''(\delta)}{2\delta}~, \\
B &~=~& -~\frac{3f(\delta)}{\delta^2} ~+~ \frac{3f'(\delta)}{\delta} ~-~ f''(\delta)~, \\
C &~=~& \frac{3f(\delta)}{\delta} ~-~ 2f'(\delta) ~+~ \frac{\delta f''(\delta)}{2}~.\end{aligned}$$
By construction, we have the following result.
The constructed function $g$ has $g(0)=0$, $g(\delta)=f(\delta)$, $g'(\delta)=f'(\delta)$ and $g''(\delta)=f''(\delta)$.
Increasing and concave {#sec:properties}
----------------------
Mimicking $f$ should mean that $g$ is increasing and concave on all of $[0,+\infty)$. Next, we give a sufficient condition for this. The condition is a bound on the amount of negative curvature of $f$ at $\delta$.
\[thm:inc\_conv\] Let $\delta>0$ be given. On $[\delta,+\infty)$, let $f$ be increasing and differentiable, with $f'$ non-increasing (decreasing). Let $f(0)=0$, and let $f$ be twice differentiable at $\delta$. If $$f''(\delta) \geq \frac{2}{\delta}\left( f'(\delta)-\frac{f(\delta)}{\delta}\right), \tag{$T_\delta$}\label{Tdelta}$$ the associated function $g$ is increasing and concave (strictly concave) on $[0,+\infty)$.
For $w\in [0,\delta]$, the third derivative of $g$ is the constant $$g'''(w) = \frac{6}{\delta^3}\left(f(\delta) - \delta f'(\delta) + \frac{\delta^2}{2}f''(\delta)\right).$$ The first factor is clearly positive. The inequality requirement on $f''(\delta)$ in our hypothesis makes the second factor non-negative. We conclude that the third derivative of $g$ is non-negative, implying that the second derivative of $g$ is non-decreasing to a non-positive (negative) value, $g''(\delta)=f''(\delta)$, on the interval $[0,\delta]$. Consequently, $g'(w)$ is non-increasing (decreasing) to $g'(\delta)=f'(\delta)>0.$
Note that the assumptions on $f$ imply that for $w\in [\delta,+\infty)$, $g'(w)=f'(w)$ is non-increasing (decreasing) and $g'(w)=f'(w)>0$. Therefore, $g$ is concave (strictly concave) and increasing on $[0, +\infty)$.
Root functions, that is power functions of the form $f(w)=w^p$ with $0<p<1$, fit our general framework: $f(0)=0$, $f$ is increasing and concave on $[0, +\infty)$, $f'(w)$ and $f''(w)$ are defined on all of $(0,+\infty)$, but $f'(0)$ is undefined. Indeed, our work was inspired by the construction for $p=1/2$ in [@DFLV2014; @DFLV2015]. Next, we verify that Theorem \[thm:inc\_conv\] applies to root functions.
For $f(w)=w^p$, $0<p<1$, we have that $f$ satisfies \[Tdelta\] for all $\delta>0$.
For $f(w)=w^p$, the inequality \[Tdelta\] is $$p(p-1)\delta^{p-2} \geq \frac{2}{\delta}\left(p\delta^{p-1}-\frac{\delta^p}{\delta}\right),$$ which simplifies to $(p-1)(p-2) \geq 0$, and which is satisfied because $p<1$
So, by Theorem \[thm:inc\_conv\], we have the following result.
\[cor:pow\_concave\] For $f(w)=w^p$, $0<p<1$, the associated $g$ is increasing and strictly concave on $[0,+\infty)$.
The following very useful fact is easy to see.
\[lemma\_cone\] The set of $f$ with domain $[\delta,+\infty)$ satisfying any of
- $f(0)=0$,
- $f$ is increasing,
- $f$ is differentiable,
- $f'$ is non-increasing or decreasing,
- $f$ is twice differentiable at $\delta$,
- \[Tdelta\]
is a (blunt) cone in function space.
As a consequence of Lemma \[lemma\_cone\] and Corollary \[cor:pow\_concave\], adding a root function to any $f$ that is differentiable at 0 and satisfies *all* of the properties listed in Lemma \[lemma\_cone\], we get such a function that is non-differentiable at 0 and has decreasing first derivative.
Next, we give a couple of natural examples to demonstrate that Theorem \[thm:inc\_conv\] applies to other functions besides root functions.
\[example:log\] Let $f(w):=\log(1+w)$, which is clearly concave and increasing on $[0,+\infty)$, and has $f(0)=0$. To verify that \[Tdelta\] is satisfied for $\delta > 0$, we consider the expression $f''(\delta)-\frac{2}{\delta}\left(f'(\delta) - \frac{f(\delta)}{\delta}\right)$, which simplifies to $$\frac{2(1+\delta)^2 \log(1+\delta) -3\delta^2-2\delta}{\delta^2(1+\delta)^2}.$$ The denominator of this expression is positive so we focus on the numerator, which we define to be $k(\delta)$. The second derivative of the numerator, $k''(\delta)=4\log(1+\delta)$, is positive for $\delta > 0$, implying that the $k'(\delta)=4(1+\delta)\log(1+\delta)-6\delta$ increases from $k'(0)=0$. Therefore, $k(\delta)$ likewise increases from $k(0)=0$. We conclude that $T_\delta$ is satisfied for $\delta > 0$. Note that by Lemma \[lemma\_cone\], we can add $\sqrt{w}$ to $f$ to get an example that is not differentiable at 0. $\diamondsuit$
\[example:sinh\] Let $f(w):={\rm ArcSinh}(\sqrt{w})=\log(\sqrt{w}+\sqrt{1+w})$ on $[0,+\infty)$. Clearly $f(0)=0$. We have $f'(w)=1/(2 \sqrt{w} \sqrt{w+1})$, which is non-negative on $(0,+\infty)$, so $f$ is increasing, but it is not differentiable at 0. Additionally, $f''(x)= (-2 x-1)/(4 x^{3/2} (x+1)^{3/2})$, which is clearly negative on $(0,+\infty)$, so $f$ is strictly concave.
To verify that \[Tdelta\] is satisfied for $\delta > 0$, we consider the expression $f''(\delta)-\frac{2}{\delta}\left(f'(\delta) - \frac{f(\delta)}{\delta}\right)$, which simplifies to $$\frac{8 (\delta+1)^{3/2} {\rm ArcSinh}\left(\sqrt{\delta}\right)-\sqrt{\delta} (6 \delta+5)}{4 \delta^2
(\delta+1)^{3/2}}.$$ The denominator of this expression is positive so we focus on the numerator, which we define to be $k(\delta)$. We have that $k(0)=0$, so we will be able to conclude that $k$ is non-negative if we can show that it is non-decreasing. Note that this $k$ is not concave, so we cannot follow the method of the previous example. Rather, we calculate $$k'(\delta)=\frac{3}{2 \sqrt{\delta}}-5 \sqrt{\delta}+12 \sqrt{\delta+1} \log
\left(\sqrt{\delta}+\sqrt{\delta+1}\right).$$ We will seek to demonstrate $k'(\delta)\geq 0$ by showing $k'(\delta)-3/2 \sqrt{\delta}\geq 0$. The derivative of $k'(\delta)-3/2 \sqrt{\delta}$ is $$\frac{7}{2 \sqrt{\delta}}+\frac{6\, {\rm ArcSinh}\left(\sqrt{\delta}\right)}{\sqrt{\delta+1}},$$ which is clearly non-negative, and so \[Tdelta\] is satisfied for $\delta > 0$. $\diamondsuit$
It is natural to wonder whether $T_\delta$ is really needed in Theorem \[thm:inc\_conv\]. Next, we give an example, where all conditions of Theorem \[thm:inc\_conv\] hold, except for $T_\delta$, and the conclusion of Theorem \[thm:inc\_conv\] does *not* hold.
For $\epsilon>0$, let $$f(w):= \left\{
\begin{array}{ll}
\sqrt{w-1}-\sqrt{\epsilon}+\frac{1+\epsilon}{2\sqrt{\epsilon}}, & w\geq 1+\epsilon; \\
\frac{1}{2\sqrt{\epsilon}}w, & w\leq 1+\epsilon.
\end{array}
\right.$$ This function $f$, solid in Figure \[fig:graph\_convex\_g\], has $f(0)=0$, is differentiable, increasing and concave on $[0,+\infty)$ and is twice differentiable for $w>1$.
![$g$ is convex and decreasing near 0[]{data-label="fig:graph_convex_g"}](example3.pdf){width=".80\textwidth"}
Now, let $\delta = 1 + \epsilon + \phi$, for $\phi>0$. For $\epsilon = 1/10$ and $\phi = 1/100$, $f''(\delta)-\frac{2}{\delta}\left(f'(\delta) - \frac{f(\delta)}{\delta}\right) \approx -6.7$, so our condition $T_\delta$ is *not* satisfied. In fact, a few calculations reveal that the associated cubic $g$, dotted in Figure \[fig:graph\_convex\_g\], is *convex and decreasing* for $0 < w < \epsilon$. The issue is that $f$ has too much negative curvature at $\delta$ to be concave on $[0,\delta]$ and have $f(0)=0$. Note that by Lemma \[lemma\_cone\], we can add a small positive multiple of $\sqrt{w}$ to $f$ to get an example that is strictly concave and not differentiable at 0. $\diamondsuit$
Lower bound for roots {#sec:roots}
=====================
For root functions, applying our general construction, direct calculation gives us the coefficients of $g(w) := Aw^3 + Bw^2 + Cw$ : $$\begin{aligned}
A &=& \delta^{p-3}(p^2-3p+2)/2,\\
B &=& -\delta^{p-2}(p^2-4p+3),\\
C &=& \delta^{p-1}(p^2-5p+6)/2.\end{aligned}$$
For use in global optimization, we want control over the relationship between $f(w)$ and $g(w)$. We believe that $g(w)\leq f(w)$ on $[0,\delta]$ for all root functions $f$. For now, we can only establish this for $f(w):=w^p$, with $p:=1/q$ for *integer* $q \geq 2$. The case of $q=2$ was established in [@DFLV2014; @DFLV2015] via a much easier argument (see the proof of Part 5 of Theorem 1 in the Appendix of [@DFLV2014]).
\[thm:lower\] For $f(w):=w^p$, with $p=1/q$ for *integer* $q \geq 2$, we have $g(w) \leq f(w)$ for all $w \in [0,+\infty)$
Clearly we can confine our attention to $[0,\delta]$. Our strategy is to express $f-g$ as the product of positive factors. It is convenient to make several substitutions before we factor. Starting with $$(f-g)(w) = w^p - \frac{\delta^{p-3}}{2}(p^2-3p+2)w^3 + \delta^{p-2}(p^2-4p+3)w^2 - \frac{\delta^{p-1}}{2}(p^2-5p+6)w,$$ for $0 \leq w \leq \delta$, we introduce the change the variables $t:=w^p$, $q:=1/p$ and $L:=\delta^p$ to arrive at $$t - \frac{1}{2L^{3q-1}}\left ( \frac{1}{q^2}-\frac{3}{q}+2\right)t^{3q}+\frac{1}{L^{2q-1}}\left (\frac{1}{q^2}-\frac{4}{q}+3\right )t^{2q} - \frac{1}{2L^{q-1}}\left (\frac{1}{q^2}-\frac{5}{q}+6\right )t^{q},$$ for $0 \leq t \leq L$.
We begin factoring by removing a positive monomial, $$\begin{aligned}
&=& \frac{t}{2q^2 L^{3q-1}}\left ( \phantom{\biggl(} 2q^2 L^{3q-1} - (6q^2-5q+1) L^{2q}t^{q-1} \right. \\
&& \left. + (6q^2-8q+2)L^{q}t^{2q-1} - (2q^2-3q+1)t^{3q-1} \phantom{\biggl(} \right ),\end{aligned}$$ so we can restrict our focus to the expression on the right, which we further simplify by making one last series of substitutions: $$\begin{aligned}
a &=& 2q^2,\\
b &=& 6q^2-5q+1,\\
c &=& 6q^2-8q+2,\\
d &=& 2q^2-3q+1.\end{aligned}$$ We claim this last expression, $$\tag{$P_q$}
a L^{3q-1} - b L^{2q}t^{q-1}+cL^q t^{2q-1}-d t^{3q-1},$$ factors into $Q_q (L-t)^3$, where $Q_q$ is a polynomial in $L$ and $t$. Because $0 \leq t \leq L$, $(L-t)^3$ is positive. With the Lemmas \[lem:terms\] and \[app:lemma2\] in the Appendix, we show that $Q_q$ is positive for integer $q \geq 2$, implying that $g(w) \leq f(w)$ for $w \in [0,\delta]$ as desired.
Better bound {#sec:better}
============
We return, temporarily, to our general setting, where we are given a function $f$ defined over the interval $[0, +\infty)$ having the following properties: $f(0)=0$, $f$ is increasing and concave on $[0, +\infty)$, $f'(w)$ and $f''(w)$ are defined on all of $(0,+\infty)$, but $f'(0)$ is undefined.
A natural and simple lower bound on $f$ is to choose $\lambda>0$, and define the *shifted* $f$ as $h(w):= f(w+\lambda)-f(\lambda)$. It is easy to see that $$h(w):= f(w+\lambda)-f(\lambda) \leq f(w),$$ because $f$ is concave and non-negative at 0, which implies that $f$ is subadditive on $[0,+\infty)$.
On the interval $[0,\delta]$, we wish to compare this $h$ (the shifted $f$) to our smoothing $g$. But $g$ is defined based on a choice of $\delta$ and $h$ is defined based on a choice of $\lambda$, a fair comparison is achieved by making these choices so that the derivative at 0 is the same. In this way, both smoothings of $f$ have the same numerical properties: they both have the same maximum derivative (maximized at zero where $f'$ blows up).
At $w=0$, the first derivative of $g$ is $$g'(0)= 3 f(\delta)/\delta - 2f'(\delta) +\delta f''(\delta)/2.$$ We have that $$h'(0)=f'(\lambda).$$ For each $\delta>0$, there is a $\lambda>0$ so that $g'(0)=h'(0)$. Now, suppose that $f'$ is decreasing on $[0,+\infty)$. Then $(f')^{-1}$ exists, and $$\hat{\lambda} := (f')^{-1}
\left(
3 f(\delta)/\delta - 2f'(\delta) +\delta f''(\delta)/2
\right)$$ is the value of $\lambda$ for which $g'(0)=h'(0)$.
So, in general, we want to check that for each $\delta>0$, $$f(w+\hat{\lambda})-f(\hat{\lambda})\leq g(w),
\tag{$*$}\label{star}$$ for all $w\in (0,\delta)$. To go further, we now confine our attention, once again, to root functions.
Already, [@DFLV2014; @DFLV2015] established this for the square-root function, though their proof has a certain weakness (see the proof of Proposition 3 in the Appendix of [@DFLV2014]), relying on some numerics, which our proof does not suffer from. Our goal is to establish this property for all root functions. This seems to be quite difficult, and so we set our focus now on root functions of the form $f(w):=w^p$, with $p:=1/q$ for *integer* $q \geq 2$. We have a substantial partial result, which as a by product provides an air-tight proof of the previous result of [@DFLV2014; @DFLV2015] for $q=2$.
\[thm:better\] For root functions of the form $f(w):=w^p$, with $p:=1/q$, (\[star\]) holds for integers $2\leq q \leq 10,000$.
The function $(g-h)(w)$, which we wish to prove is non-negative on the interval $[0,\delta]$, is $$\frac{\delta^{p-3}}{2}(p^2-3p+2)w^3 -\delta^{p-2}(p^2-4p+3)w^2 +\frac{\delta^{p-1}}{2}(p^2-5p+6)w
-(w+\hat{\lambda})^p+\hat{\lambda}^p ,$$ where the shift constant $\hat{\lambda}$ for which $h'(0)=g'(0)$ is $$\hat{\lambda}
= (f')^{-1}(g'(0))
= \delta \left(\frac{p^2-5 p+6}{2p}\right)^{\frac{1}{p-1}}.$$
With a few substitutions, we simplify the function and express it in polynomial form. For the first substitution, set $q := 1/p$, $\gamma := \delta^{1/q}$, and $t: = \frac{w}{\gamma^q}$. We obtain a function of $t$ over $[0,1]$ that has $\gamma$ as a factor: $$\begin{array}{lcl}
(g-h)_t(t) &=& \gamma \left[ \frac{2 q^2-3 q+1}{2 q^2} t^3+\frac{-6 q^2+8 q-2}{2 q^2} t^2+\frac{6 q^2-5 q+1}{2 q^2} t \right.\\
&&\left. -\left( \left(\frac{2q}{6 q^2-5 q+1}\right)^{\frac{q}{q-1}}+t\right)^{1/q}+\left(\frac{2q}{6 q^2-5 q+1}\right)^{\frac{1}{q-1}} \right].\\
\\
\end{array}$$ Next, we set $Q :=\left(\frac{2q}{6 q^2-5 q+1}\right)^{\frac{1}{q-1}}$ and $u :=(t+Q^q)^{1/q}$ (so $t \rightarrow u^q-Q^q$). The resulting polynomial in $u$ is $$\begin{array}{lcl}
(g-h)_u(u) &= \frac{\gamma}{2q^2} & \left[
\left(2 q^2-3 q+1\right) \left(u^q-Q^q\right)^3+\left(-6 q^2+8 q-2\right) \left(u^q-Q^q\right)^2 \right. \\
&& \left. +\left(6 q^2-5 q+1\right) \left(u^q-Q^q\right)+2q^2Q-2q^2u \right],
\end{array}$$ for $Q \leq u \leq (1+Q^q)^{1/q}.$ Since $\gamma >0$, our task is reduced to proving that the second factor, $$K_u(u) := d \left(u^q-Q^q\right)^3-c\left(u^q-Q^q\right)^2 +b \left(u^q-Q^q\right)-a(u-Q), \\$$ where $$\begin{aligned}
a &:=& 2q^2,\\
b &:=& 6q^2-5q+1,\\
c &:=& 6q^2-8q+2, \text{ and}\\
d &:=& 2q^2-3q+1,\end{aligned}$$ is non-negative for $Q \leq u \leq (1+Q^q)^{1/q}.$
It is obvious that $K_u$ has a root at $Q$. In fact, $K_u$ has a double root at $Q$, which we can verify by showing that the first derivative of $K_u$, $$\begin{aligned}
K_u'(u)=3 d q \left(u^q-Q^q\right)^2 u^{q-1}-2 c q \left(u^q-Q^q\right)u^{q-1}+b q u^{q-1}-a,\end{aligned}$$ also has a root at $Q$. This is easily accomplished by noticing that $$\begin{aligned}
b q u^{q-1} &=& b q \left(\left(\frac{2q}{6 q^2-5 q+1}\right)^{\frac{1}{q-1}}\right)^{q-1} = 2q^2 = a.\end{aligned}$$
By construction of $g$ and $h$, $$(g-h)_u((1+Q^q)^{1/q})=(g-h)(\delta)=(f-h)(\delta)>0,$$ which means that $K_u((1+Q^q)^{1/q})>0$. In order to prove that $K_u(u)\geq 0$ for $u \in (Q,(1+Q^q)^{1/q})$, it suffices to show that there are no roots in the interval $(Q,(1+Q^q)^{1/q})$. In fact, we prove that the only root in the interval $(0,(1+Q^q)^{1/q})$ $\supseteq (Q,(1+Q^q)^{1/q})$ is the double root at $Q$.
Using a known technique (e.g., see [@Sagraloff]), we apply the Möbius transformation $$K_u\left(\frac{(1+Q^q)^{1/q}}{v+1}\right)$$ to express $K_u$, $u \in (0,(1+Q^q)^{1/q}))$, as a rational function in $v$ over the interval $(0, \infty)$. Note that when $v=0$, $K_u\left(\frac{(1+Q^q)^{1/q}}{v+1}\right)=K_u\left((1+Q^q)^{1/q}\right),$ and as $v \rightarrow \infty$, $K_u\left(\frac{(1+Q^q)^{1/q}}{v+1}\right)\rightarrow K_u(0)$.
Next, we calculate expressions for the coefficients of the polynomial $$K_v(v):=(v+1)^{3q} K_u\left(\frac{\beta}{v+1}\right),$$ where $\beta:=(1+Q^q)^{1/q}$, and the domain is $(0,\infty)$.
Expanding the binomials in $u^q$ and $Q^q$ and multiplying by $(v+1)^{3q}$, we have $$\begin{aligned}
&&K_v(v) = (aQ - bQ^q -cQ^{2q}-dQ^{3q})(v+1)^{3q} -a \beta (v+1)^{3q-1} \\
&& ~+(3d\beta^q Q^{2q} +2c\beta^q Q^q + b \beta^q)(v+1)^{2q} -(3d\beta^{2q}Q^q + c\beta^{2q})(v+1)^q + d\beta^{3q}.\end{aligned}$$ Expanding binomials and collecting like terms, we find that $$\begin{aligned}
&& K_v(v) = V + W + X + Y + Z \\
&&~ + \sum_{i=1}^q \left[{q \choose q-i}W + {2q \choose 2q-i}X + {3q-1 \choose 3q-1-i}Y + {3q \choose 3q-i}Z\right]v^i \\
&& ~ + \sum_{i=q+1}^{2q} \left[{2q \choose 2q-i}X + {3q-1 \choose 3q-1-i}Y+ {3q \choose 3q-i}Z\right]v^i\\
&&~ + \sum_{i=2q+1}^{3q-1} \left[{3q-1 \choose 3q-1-i}Y + {3q \choose 3q-1}Z\right] v^i
~+~ Z v^{3q}, \\\end{aligned}$$ where $$\begin{aligned}
V & := & d\beta^{3q}, \\
W &:=& -3d\beta^{2q}Q^q - c\beta^{2q},\\
X &:=& 3d\beta^q Q^{2q} +2c\beta^q Q^q + b \beta^q,\\
Y &:=& -a \beta, \text{ and}\\
Z &:=& -d Q^{3q} -c Q^{2q} -b Q^q + aQ.\end{aligned}$$
Armed with these expressions for the coefficients of the polynomials $K_v(v)$, we verified (with `Mathematica`) that for integers $2 \leq q \leq 10,000$, there are exactly two sign changes in each coefficient sequence. By Descartes’ Rule of Signs, we conclude that there are at most two positive roots of $K_v(v)$ (for these values of $q$), and therefore at most two roots of $K_u(u)$ in the interval $(0, (1+Q^q)^{1/q})$ (the double root at $Q$).
Our proof technique can work for any *fixed* integer $q \geq 2$. In carrying out the technique, there is some computational burden for which we employ `Mathematica`. We only carried this out for integers $2 \leq q \leq 10,000$, but in principle we could go further. It is important to point out that the calculations were done exactly and only truncated to finite precision at the end.
The remaining challenge is to make a proof for all integers $q \geq 2$. But the coefficients of $K_u\left(\frac{(1+Q^q)^{1/q}}{v+1}\right)$ are rather complicated for general $q$, so it is difficult to analyze their signs in general.
Average performance for roots {#sec:avg}
=============================
On $[\delta,+\infty)$, $g$ coincides with $f$ by definition (we are assuming that $f'(\delta)$ and $f''(\delta)$ are defined so that $g$ is well defined), while $h$ strictly under-estimates $f$ for increasing concave functions $f$ for which $f(\delta)>0$. So it becomes interesting to examine the performance of $g$ and $h$ near 0, that is on the interval $[0,\delta]$. Here, we focus on average relative performance: $$\frac{1}{\delta}\int_0^\delta \frac{g(w)}{f(w)} dw~,\qquad
\frac{1}{\delta}\int_0^\delta \frac{h(w)}{f(w)} dw~,$$ where we are further assuming that $f(0)=0$ and $f$ is increasing. In what follows, we compare these performance measures for root functions.
\[thm:indep\] For $f(w) = w^p$, $0<p<1$, and $\lambda$ chosen as a function of $\delta >0$ so that $g'(0) = h'(0)$, we have that $$\frac{1}{\delta}\int_0^\delta \frac{g(w)}{f(w)} dw ~=~ \frac{3}{4-p},$$ notably independent of $\delta$, and also $$\frac{1}{\delta}\int_0^\delta \frac{h(w)}{f(w)} dw$$ is independent of the choice of $\delta$.
We apply the change of variable $w \rightarrow \delta v$ to rewrite each expression without $\delta$. For the first integral, letting $a := \frac{1}{2}(p^2 -3p+2)$, $b:= p^2-4p+3$, and $c := \frac{1}{2}(p^2-5p+6)$, we have $$\begin{aligned}
\frac{1}{\delta}\int_0^\delta \frac{g(w)}{f(w)} dw &~= \frac{1}{\delta} \int_0^\delta \frac{\delta^{p-3}aw^3 - \delta^{p-2}bw^2 + \delta^{p-1}cw}{w^p} dw \\
&~= \int_0^1 \left(av^{3-p}-bv^{2-p} +cv^{1-p}\right) dv. \\
&~= \frac{a}{4-p} -\frac{b}{3-p} +\frac{c}{2-p} ~=~ \frac{3}{4-p}~.\end{aligned}$$ Letting $P := \left(\frac{p^2-5 p+6}{2p}\right)^{\frac{1}{p-1}}$ in the second integral, the change of variable produces $$\begin{aligned}
\frac{1}{\delta}\int_0^\delta \frac{h(w)}{f(w)} dw &~= \frac{1}{\delta} \int_0^\delta \frac{(w+\delta P)^p-(\delta P)^p}{w^p} dw \\
&~= \int_0^1 \frac{(v+P)^p - P^p}{v^p} dv,\end{aligned}$$ and again $\delta$ disappears from the expression.
We note that $\frac{3}{4-p}$ is increasing in $p$, and so its infimum on $(0,1)$ is $\frac{3}{4}$ at $p=0$. Additionally, we note that there is no closed-form expression for the last integration of the proof, though it can be expressed in terms of an evaluation of a Gaussian/ordinary hypergeometric function $\tensor*[_2]{\rm F}{_1}$ (see [@andrews1999]). Specifically $$\left(\frac{p^2-5 p+6}{2p}\right)^{\frac{p}{p-1}}
\left( \frac{-1+\tensor*[_2]{\rm F}{_1}(1-p\, ,\, -p\, ;\, 2-p\, ;\, -1/p)}{1-p}\right)~.$$ In the key special case of $p=1/2$, we do get the closed-form expression $$\frac{8}{15}\left(
-1+
\frac{3615+16\sqrt{241}\,{\rm ArcSinh}\!\left(\frac{15}{4}\right)}{120\sqrt{241}}
\right)\approx 0.646125$$ (where ${\rm ArcSinh}(x)=\ln\left(x+\sqrt{x^2+1}\right)$), which is significantly less than $\left.\frac{3}{4-p}\right\rvert_{p=\frac{1}{2}}= \frac{6}{7} \approx 0.857143$.
In Figure \[fig:averagefor roots\] we have plotted the two performance measures, varying $p$ on $(0,1)$; Theorem \[thm:indep\] allows us to do this without separate curves for different $\delta$. We can readily see that $g$ outperforms $h$ bigly, with the performance gap being most extreme as $p\to 0$, and decreasing in $p$ on $(0,1)$.
![Average relative performance on $[0,\delta]$ as a function of $p$\
($f(w):=w^p$, $0<p<1$)[]{data-label="fig:averagefor roots"}](computations_for_ref2.pdf){width=".95\textwidth"}
Alternatives, software, extended use, and ongoing work {#sec:future}
======================================================
Alternatives
------------
We do not mean to imply that our smoothing ideas are the only viable or even preferred way to handle all instances of the functions to which our ideas apply. For example, there is the possibility to take a function $f(w)$ and replace it with a new variable $y$ and the constraint $f^{-1}(y)=w$. For example, $f(w):=w^p$, with $0<p<1$, can be replaced with $y$ and the constraint $w=y^{\frac{1}{p}}$, which is now smooth at $w=0$. But there is the computational cost of including an additional non-linear equation in a model. In fact, this could turn an unconstrained model into a constrained model. For other functions, we may not have a nice expression for the inverse readily available. Even when we do, there can be other issues to consider. Taking again $f(w):=w^p$, with $0<p<1$, suppose that we have $\psi:\mathbb{R}^m\rightarrow \mathbb{R}_+$ and $\phi:\mathbb{R}^n\rightarrow \mathbb{R}$. Now suppose that we have a model with the constraint $f(\psi(y))\geq \phi(x)$. Of course $y$ and $x$ may be involved in other constraints as well. Now, suppose further that the range of $\phi$ on the feasible region is $(-\infty,+\infty)$. There could then be a difficulty in trying to work with $\psi(y) \geq f^{-1}(\phi(x))$, which we could repair by instead working with $\psi(y) \geq f^{-1}(\phi_+(x))$, where $\phi_+(x):=\max\{\phi(x),0\}$. But this means working now with the piecewise-defined function $\phi_+$, which can involve a lot of pieces (e.g., consider the univariate $\phi(x):=\sin(x)$).
In the end, we do not see our technique as a panacea, but rather as a viable method with nice properties that modelers and solvers should have in their bags.
Software {#subsec:soft}
--------
In the context of the square-root smoothing of [@DFLV2014; @DFLV2015], a new and exciting experimental (“$\alpha$”) feature was developed for `SCIP` Version 3.2 (see [@SCIP32]) to handle univariate piecewise functions that are user-certified (through `AMPL` suffixes) as being globally convex or concave. At this writing, `SCIP` is the only global solver that can accommodate such functions. Such a feature is extremely useful for taking advantage of the results that we present here, because our smoothings (like those of [@DFLV2014; @DFLV2015]) are piecewise-defined, and so the user must identify global concavity to the solver. This is accomplished through the new `SCIP` operator type `SCIP_EXPR_USER`. A bit of detail about this feature, from [@SCIP32], is enlightening:
> “Currently, the following callbacks can or have to be provided: computing the value, gradient, and Hessian of a user-function, given values (numbers or intervals) for all arguments of the function; indicating convexity/concavity based on bounds and convexity information of the arguments; tighten bounds on the arguments of the function when bounds on the function itself are given; estimating the function from below or above by a linear function in the arguments; copy and free the function data. Currently, this feature is meant for low-dimensional (few arguments) functions that are fast to evaluate and for which bound propagation and/or convexification methods are available that provide a performance improvement over the existing expression framework.”
Extended use
------------
Our techniques have broader applicability than to functions that are purely concave (or symmetrically, convex). In general, given a univariate piecewise-defined function that is concave or convex on each piece, and possibly non-differentiable at each breakpoint, we can seek to find a smoothing that closely mimics and approximates the function. It is not at all clear how to accommodate such functions in global-optimization software like `SCIP`; because such functions are not, in general, *globally* concave or convex, they cannot be correctly handled with the new `SCIP` feature (see §\[subsec:soft\]). Still, such functions and their smoothings can be useful within the common paradigm of seeking (good) local optima for non-linear-optimization formulations.
For example, for $0<p<1$, consider the function $$f(w):= \left\{
\begin{array}{ll}
w^p,\quad & w\geq 0; \\
-(-w)^p,\quad & w\leq 0.
\end{array}
\right.$$ This function is continuous, increasing, convex on $(-\infty,0]$, concave on $[0,+\infty)$ and of course not differentiable at 0. We would like to replace it with a function that has all of these properties but is somewhat smooth at 0. If we apply our smoothing to $f$ separately, for $w\geq 0$ and for $w\leq 0$, we would arrive at a function of the form $$g(w):= \left\{
\begin{array}{ll}
w^p,\quad & w\geq \delta; \\
Aw^3+Bw^2+Cw,\quad & 0\leq w \leq \delta;\\
Aw^3-Bw^2+Cw,\quad & -\delta \leq w \leq 0;\\
-(-w)^p,\quad & w\leq -\delta,
\end{array}
\right.$$ for appropriate $A,B,C$ (see §\[sec:general\]). It is easy to check that the resulting $g$ is continuous, *differentiable at 0*, twice differentiable everywhere but at 0, increasing, convex on $(-\infty,0]$, and concave on $[0,+\infty)$. In short, $g$ mimics $f$ very well, but is smoother. Moreover, for $p=1/q$, with integer $q\geq 2$, $g$ upper bounds $f$ on $(-\infty,0]$ and lower bounds $f$ on $[0,+\infty)$.
Note that the obvious “double shift” $$h(w):= \left\{
\begin{array}{ll}
(w+\lambda)^p,\quad & w> 0; \\
?, \quad & w= 0; \\
-(-w+\lambda)^p,\quad & w< 0
\end{array}
\right.$$ is not even continuous at 0.
Additionally, for $p=1/q$, with integer $2\leq q\leq 10,000$, in the sense of §\[sec:better\], $g$ is a better upper bound on $f$ than $h$ on $(-\infty,0]$, and $g$ is a better lower bound on $f$ than $h$ on $[0,+\infty)$.
Ongoing work
------------
To extend the applicability of our results, we are pursuing two directions:
- We would like to generalize Theorem \[thm:lower\] to all root functions $w^p$ with $0<p<1$. A strategy that we are exploring is to try to make a similar proof to what we have, for the case in which $p$ is rational, and then employ a continuity argument to establish the result for all real exponents.
- We would like to generalize Theorem \[thm:better\] for all root functions $w^p$ with $0<p<1$. For now, that seems like a rather ambitious goal, and what is more in sight is generalizing Theorem \[thm:better\] for *all* integer $q \geq 2$. To do this we are trying to sharpen our arguments employing Descartes’ Rule of Signs, or, alternatively, to develop a sum-of-squares argument.
The authors gratefully acknowledge the anonymous referee who proposed the performance measure studied in §\[sec:avg\]. J. Lee gratefully acknowledges partial support from ONR grant N00014-14-1-0315.
Appendix {#appendix .unnumbered}
========
\[lem:terms\] The polynomial $P_q$ as defined above for integer $q \geq 2$ can be expressed as $Q_q (L-t)^3$, where polynomial $Q_q$ has the following $3q-3$ terms: $$\begin{aligned}
{i+2 \choose 2}a L^{3q-4-i} t^{i}, &\mbox{for}& i = 0,1, \dots, q-2;\\
\left[ {i+2 \choose 2}a - {i-q+3 \choose 2}b \right] L^{3q-4-i} t^{i}, &\mbox{for}& i = q-1,q, \dots, 2q-2;\\
\left[{i+2 \choose 2}a - {i-q+3 \choose 2}b +{i-2q+3 \choose 2}c \right] L^{3q-4-i} t^{i}, &\mbox{for}& i = 2q-1,2q, \dots, 3q-4.\end{aligned}$$ Note that for $q=2$, $2q-2 = 3q-4 = 2$, so there are no terms of the third type.
We expand $Q_q (L-t)^3$ to see that it is equivalent to $P_q$, first for specific cases $q = 2, 3 \mbox{ and } 4$, and finally for general $q \geq 5$. The following are easily verified: $$\begin{aligned}
&&\underline{Q_q (L-t)^3 = P_q}\\
q=2:&&(8L^2+9Lt+3t^2) (L-t)^3 = 8L^5-15L^4t+10L^2t^3-3t^5;\\
q=3:&&(18L^5+54L^4t+68L^3t^2+60L^2t^3+30Lt^4+10t^5) (L-t)^3 \\
&&\quad = 18L^8-40L^6t^2+32L^3t^5-10t^8;\\
q=4:&& (32L^8+96L^7 t+192L^6t^2+243L^5t^3+249L^4t^4+210L^3t^5 \\
&&\quad +126L^2t^6+63L t^7+21t^8) (L-t)^3 = 32L^{11}-77L^8t^3+66L^4t^7-21t^{11}.\end{aligned}$$
Now considering general $q \geq 5$, most of the terms of $Q_q(L-t)^3$ cancel out due to the following equations: $$\begin{aligned}
(-3){2 \choose 2} + {3 \choose 2} &= -3+3&= 0;\\
(3){2 \choose 2} +(-3){3 \choose 2} + {4 \choose 2}&=3-9+6&= 0; \\\end{aligned}$$ and for $i \geq 3,$ $$\begin{aligned}
(-1){i-1 \choose 2} + (3){i \choose 2} +(-3){i+1 \choose 2} + {i+2 \choose 2}&&=0.\\\end{aligned}$$ If the expression for the coefficient of $t^j$, $0\leq j \leq 3q-1$, has more than one term involving $a$, $b$, or $c$, that variable ($a$, $b$, or $c$) has a coefficient of one of the forms above and cancels out. The only time the variable remains is when it has only a single term in the expression. The terms of $Q_q(L-t)^3 = Q_q(-t^3+3Lt^2-3L^2t+L^3)$ for $q \geq 5$ increasing in the degree $j$ of $t$ are as follows: $$\begin{aligned}
j=0:&& \left(aL^{3q-4}\right)\left(L^3\right)=aL^{3q-1} \\
j=1:&& \left(aL^{3q-4}\right)\left(-3L^2t\right) + \left(3aL^{3q-5}t\right)\left(L^3\right) \\
&& =(-3 + 3)aL^{3q-2}t \\
&& = 0\\
j=2:&& \left(aL^{3q-4}\right)\left(3Lt^2\right) + \left(3aL^{3q-5}t\right)\left(-3L^2t\right) + \left(6aL^{3q-6}t^2\right)\left(L^3\right)\\
&& = (3-9+6)aL^{3q-3}t^2\\
&& = 0 \\
3 \leq j \leq q-2:&& {j-1 \choose 2}aL^{3q-1-j}t^{j-3}\left(-t^3\right) + {j \choose 2}aL^{3q-2-j}t^{j-2}\left(3Lt^2\right) + \\
&&\quad {j+1 \choose 2}aL^{3q-3-j}t^{j-1}\left(-3L^2t\right) + {j+2 \choose 2}aL^{3q-4-j}t^{j}\left(L^3\right)\\
&& = \left[(-1){j-1 \choose 2} + (3){j \choose 2} +(-3){j+1 \choose 2} + {j+2 \choose 2}\right]aL^{3q-1-j}t^j\\
&&= 0\\
j=q-1:&& {q-2 \choose 2}aL^{2q}t^{q-4}\left(-t^3\right) + {q-1 \choose 2}aL^{2q-1}t^{q-3}\left(3Lt^2\right) + \\
&&\quad {q \choose 2}aL^{2q-2}t^{q-2}\left(-3L^2t\right) + \left[{q+1 \choose 2}a-b\right]L^{2q-3}t^{q-1} \left(L^3\right)\\
&&= \left[(-1){q-2 \choose 2} + (3){q-1 \choose 2} +(-3){q \choose 2} + {q+1 \choose 2}\right]aL^{2q}t^{q-1} \\
&& \quad -bL^{2q}t^{q-1}\\
&&= -bL^{2q}t^{q-1}\\
j=q: &&{q-1 \choose 2}aL^{2q-1}t^{q-3}\left(-t^3\right) + {q \choose 2}aL^{2q-2}t^{q-2}\left(3Lt^2\right) + \\
&&\quad \left[{q+1 \choose 2}a-b\right]L^{2q-3}t^{q-1}\left(-3L^2t\right) + \left[{q+2 \choose 2}a-3b\right]L^{2q-4}t^{q} \left(L^3\right)\\
&& =\left[(-1){q-1 \choose 2} + (3){q \choose 2} +(-3){q+1 \choose 2} + {q+2 \choose 2}\right]aL^{2q-1}t^{q} \\
&& \quad +(3-3)bL^{2q-1}t^{q}\\
&& = 0\\
j=q+1: &&{q \choose 2}aL^{2q-2}t^{q-2}\left(-t^3\right) \\
&&\quad + \left[{q+1 \choose 2}a-b\right]L^{2q-3}t^{q-1}\left(3Lt^2\right) \\
&&\quad + \left[{q+2 \choose 2}a-3b\right]L^{2q-4}t^{q}\left(-3L^2t\right) \\
&&\quad + \left[{q+3 \choose 2}a-6b\right]L^{2q-5}t^{q+1} \left(L^3\right) \\
&& = 0\\
q+2 \leq j \leq 2q-2: &&\left[{j-1 \choose 2}a-{j-q \choose 2}b\right]L^{3q-1-j}t^{j-3}\left(-t^3\right) \\
&& \quad + \left[{j \choose 2}a-{j-q+1 \choose 2}b\right]L^{3q-2-j}t^{j-2}\left(3Lt^2\right) \\
&&\quad + \left[{j+1 \choose 2}a-{j-q+2 \choose 2}b\right]L^{3q-3-j}t^{j-1}\left(-3L^2t\right) \\
&&\quad + \left[{j+2 \choose 2}a-{j-q+3 \choose 2}b\right]L^{3q-4-j}t^j \left(L^3\right)\\
&& = 0\\
j=2q-1:&&\left[{2q-2 \choose 2}a-{q-1 \choose 2}b\right]L^{q}t^{2q-4}\left(-t^3\right) \\
&& \quad + \left[{2q-1 \choose 2}a-{q \choose 2}b\right]L^{q-1}t^{2q-3}\left(3Lt^2\right) \\
&&\quad + \left[{2q \choose 2}a-{q+1 \choose 2}b\right]L^{q-2}t^{2q-2}\left(-3L^2t\right) \\
&&\quad + \left[{2q+1 \choose 2}a-{q+2 \choose 2}b+c\right]L^{q-3}t^{2q-1} \left(L^3\right)\\
&& = cL^qt^{2q-1} \\
j=2q:&&\left[{2q-1 \choose 2}a-{q \choose 2}b\right]L^{q-1}t^{2q-3}\left(-t^3\right) \\
&& \quad + \left[{2q \choose 2}a-{q+1 \choose 2}b\right]L^{q-2}t^{2q-2}\left(3Lt^2\right) \\
&&\quad + \left[{2q+1 \choose 2}a-{q+2 \choose 2}b+c\right]L^{q-3}t^{2q-1}\left(-3L^2t\right) \\
&&\quad + \left[{2q+2 \choose 2}a-{q+3 \choose 2}b+3c\right]L^{q-4}t^{2q} \left(L^3\right)\\
&& = 0 \\
j=2q+1:&&\left[{2q \choose 2}a-{q+1 \choose 2}b\right]L^{q-2}t^{2q-2}\left(-t^3\right) \\
&& \quad + \left[{2q+1 \choose 2}a-{q+2 \choose 2}b+c\right]L^{q-3}t^{2q-1}\left(3Lt^2\right) \\
&&\quad + \left[{2q+2 \choose 2}a-{q+3 \choose 2}b+3c\right]L^{q-4}t^{2q}\left(-3L^2t\right) \\
&&\quad + \left[{2q+3 \choose 2}a-{q+4 \choose 2}b+6c\right]L^{q-5}t^{2q+1} \left(L^3\right)\\
&& = 0 \\
2q+2 \leq j \leq 3q-4:&&\left[{j-1 \choose 2}a-{j-q \choose 2}b+{j-2q \choose 2}c\right]L^{3q-1-j}t^{j-3}\left(-t^3\right) \\
&& \quad + \left[{j \choose 2}a-{j-q+1 \choose 2}b+{j-2q+1 \choose 2}c\right]L^{3q-2-j}t^{j-2}\left(3Lt^2\right) \\
&&\quad + \left[{j+1 \choose 2}a-{j-q+2 \choose 2}b+{j-2q+2 \choose 2}c\right]L^{3q-3-j}t^{j-1}\left(-3L^2t\right) \\
&&\quad + \left[{j+2 \choose 2}a-{j-q+3 \choose 2}b+{j-2q+3 \choose 2}c\right]L^{3q-4-j}t^{j} \left(L^3\right)\\
&& = 0 \\\end{aligned}$$ The cancellation pattern above fails for the last three terms. It is necessary to replace $a$, $b$, and $c$ with the equivalent expressions involving $q$ to verify each of the following. $$\begin{aligned}
j=3q-3:&& \left[{3q-4 \choose 2}a - {2q-3 \choose 2}b + {q-3 \choose 2}c\right]L^2t^{3q-6}(-t^3) \\
&& \quad + \left[{3q-3 \choose 2}a - {2q-2 \choose 2}b + {q-2 \choose 2}c\right]Lt^{3q-5}(3Lt^2) \\
&& \quad + \left[{3q-2 \choose 2}a - {2q-1 \choose 2}b + {q-1 \choose 2}c\right]t^{3q-4}(-3L^2t) \\
&& =0\\
j=3q-2:&& \left[{3q-3 \choose 2}a - {2q-2 \choose 2}b + {q-2 \choose 2}c\right]Lt^{3q-5}(-t^3) \\
&& \quad + \left[{3q-2 \choose 2}a - {2q-1 \choose 2}b + {q-1 \choose 2}c\right]t^{3q-4}(3Lt^2) \\
&& =0\\
j=3q-1:&& \left[{3q-2 \choose 2}a - {2q-1 \choose 2}b + {q-1 \choose 2}c\right]t^{3q-4}(-t^3) =-dt^{3q-1}
$$
\[app:lemma2\] The polynomial $Q_q$ as defined in the previous lemma for integers $q \geq 2$ has all positive coefficients.
We consider each of the three types of coefficients of $Q_q$ separately. The first type of coefficients, $${i+1 \choose 2}2q^2, \mbox{ for } i = 1,2, \dots, q-1,$$ are obviously all positive.
Coefficients of the second type have the form $${i+1 \choose 2}a - {i-q+2 \choose 2}b, \mbox{ for } i = q, q+1, \dots, 2q-1.$$ The real function $C_2(x) = \frac{1}{2}(x+1)(x)a - \frac{1}{2}(x-q+2)(x-q)b, ~x \in [q,2q-1]$, has second derivative $C_2''(x)=-4q^2+5q-1$, which is negative for $q>1$. Therefore, $C_2$ is concave on the interval $[q,2q-1]$. Evaluating $C_2$ at the ends of the interval, we find that $C_2(q)=q^4+q^3-6q^2+5q-1$ and $C_2(2q-1)=q^4-\frac{5}{2}q^3+2q^2-\frac{1}{2}q$, both of which are positive for $q>1$. We conclude that $C_2$ is positive over the interval $[q,2q-1]$, and all of the type-two coefficients are positive.
Finally, the third type of coefficients have the form $${i+1 \choose 2}a - {i-q+2 \choose 2}b +{i-2q+2 \choose 2}c, \mbox{ for } i = 2q, 2q+1, \dots, 3q-3.$$ As above, we consider the real extension of this function, $C_3(x) = \frac{1}{2}(x+1)(x)a - \frac{1}{2}(x-q+2)(x-q+1)b +\frac{1}{2}(x-2q+2)(x-2q+1)c, ~x \in [2q,3q-3]$. The first derivative of this function, $C_3'(x) = (2q^2-3q+1)x -(6q^3-12q^2+\frac{15}{2} q -\frac{3}{2})$, is linear in $x$ and has positive slope for $q>1$. Furthermore, the $x$ intercept of $C_3'(x)$ is $x = 3q-\frac{3}{2}$. This means that $C_3'(x)<0$ for $x < 3q-\frac{3}{2}$. In particular, $C_3(x)$ is decreasing on the interval $[2q, 3q-3]$. The right end of this interval is $C_3(3q-3)=2q^2-3q+1$, which is positive for $q>1$, and all of the type-three terms are positive.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A class of generalized exclusion processes with symmetric nearest-neighbor hopping which are parametrized by the maximal occupancy, $k\geq 1$, is investigated. For these processes on hyper-cubic lattices we compute the diffusion coefficient in all spatial dimensions. In the extreme cases of $k=1$ (symmetric simple exclusion process) and $k=\infty$ (non-interacting symmetric random walks) the diffusion coefficient is constant, while for $2\leq k<\infty$ it depends on the density and $k$. We also study the evolution of the tagged particle, show that it exhibits a normal diffusive behavior in all dimensions, and probe numerically the coefficient of self-diffusion.'
author:
- Chikashi Arita
- 'P. L. Krapivsky'
- Kirone Mallick
title: 'Generalized exclusion processes: Transport coefficients'
---
Introduction
============
Exclusion processes constitute an important class of lattice gases that plays a prominent role in numerous subjects including non-equilibrium statistical mechanics, soft matter, traffic models, biophysics, combinatorics and probability theory [@KLS; @Spohn; @SZ95; @D98; @L99; @KL99; @KRB10; @Schad; @CKZ]. By definition, exclusion processes are lattice gases supplemented with stochastic hopping and obeying the constraint that at most one particle per site is allowed. In simple exclusion models, only hops to nearest-neighboring sites are allowed. These models are exactly solvable in low dimensions [@D98; @S00; @BE07], and they have become benchmarks to test general theories for non-equilibrium behaviors [@DerrReview; @Bertini; @Bodineau; @DCairns; @JonaSeoul; @MFTReview].
Because of its ubiquity and usefulness, numerous more complicated variants of the basic exclusion process have been investigated (see [@SZ95; @S00; @CKZ] and references therein). One natural generalization is to alleviate the exclusion constraint by allowing $k$ particles per site ($k \ge 1$ is a fixed integer). More precisely, this process is defined in $d$ dimensions, say on the hyper-cubic lattice $\mathbb{Z}^d$, as follows: (i) each particle attempts to hop to its $2d$ neighbors with the same unit rate to each neighbor (symmetric hopping), (ii) every hopping attempt is successful when the target site is occupied by less than $k$ particles, otherwise the hopping attempt is rejected (Fig. \[pic\_fig\]). The symmetric exclusion process (SEP) is recovered when $k=1$, whereas for $k=\infty$ the model reduces to independent random walks undergoing symmetric nearest-neighbor hopping. Letting $k$ vary from 1 to $\infty$ allows us to interpolate between a strongly interacting to a non-interacting system. (In quantum systems, restricting the maximum number of particles in a given quantum state to an integer $k$, with $1 < k < \infty$, leads to the so-called Gentile statistics [@Gentile; @Khare] interpolating between Fermi-Dirac statistics and Bose-Einstein statistics. In this paper we consider only classical lattice gases.)
Generalized exclusion processes (GEPs) with maximal occupation number $1<k<\infty$ have been studied in [@KLO94; @Kipnis2; @Timo]; see also Refs. [@SandowSchutz; @Barma_drop; @Barma; @Schutz; @PK; @Kurchan; @CM; @CGRS; @Ryabov; @Becker_13; @Becker_preprint] for other versions of GEPs. Some of these models can be mapped onto multi-species exclusion processes [@MMR; @EFM; @FerrariMartin] but with non-conserving species. Overall, the GEPs are considerably less understood than the ordinary exclusion process—the integrability properties of the SEP do not carry over to the GEPs and a different perspective is required.
![GEP with $k=3$ in one dimension. Hopping into a site occupied by three particles is forbidden; other hopping events occur with the same (unit) rate. Thus the total hopping rate from every site is equal to the number of particles at the site times the number of neighboring sites which are not fully occupied. []{data-label="pic_fig"}](pic.pdf){width="8.7cm"}
The objective of this work is to investigate the GEP at a coarse-grained level and to calculate the transport coefficients that underlie the hydrodynamic description. Section \[sec:HB\] reviews known properties of the GEP and outlines its macroscopic (i.e., hydrodynamic) regime which is governed by a diffusion equation. We also formulate our main result, namely a parametric representation of the diffusion coefficient. In Sec. \[sec:DC\], we present the derivation of the diffusion coefficient. In Sec. \[sec:SDP\], stationary density profiles are computed and compared with simulations. In Sec. \[sec:SDC\], we examine the evolution of a tagged particle in the GEP in various spatial dimensions. We derive the self-diffusion coefficient using a mean-field approximation. We also probe the self-diffusion coefficient numerically and show a reasonable qualitative agreement with mean-field predictions. We summarize our results, and mention remaining challenges and possible extensions, in Sec. \[sec:Sum\].
Hydrodynamic Behavior {#sec:HB}
=====================
For the generalized exclusion process with symmetric hopping, steady states are remarkably simple and are given by a product measure [@Spohn; @L99; @KL99; @PK]. In other words, one only needs to know the probabilities $P_j$ to have $j$ particles per site and the stationary weight of any configuration factors into the product of these basic probabilities. The basic probabilities are given by an elementary formula [@Barma; @KL99; @PK] $$\label{probs}
P_j = \frac{\lambda^j}{j!}\,\frac{1}{Z_k(\lambda)}\, .$$ To justify it suffices to use the factorization and verify that the flow $(i,j)\Longrightarrow (i-1,j+1)$, which is given by $iP_iP_j$, is equal to the flow $(i-1,j+1)\Longrightarrow (i,j)$, which is given by $(j+1)P_{i-1}P_{j+1}$. With the choice , we indeed get $iP_iP_j = (j+1)P_{i-1}P_{j+1}$. The ‘partition function’ $Z_k(\lambda)$, which appears in Eq. , is fixed by the normalization requirement $\sum_{0\leq j\leq k}P_j=1$. It is equal to an incomplete exponential function: $$\label{partition}
Z_k(\lambda) = \sum_{j=0}^k \frac{\lambda^j}{j!}\, .$$ The ‘fugacity’ parameter $\lambda$ is implicitly determined by the density $\rho$: $$\label{density}
\rho = \sum_{j=0}^k jP_j = \lambda\,\frac{Z_{k-1}(\lambda)}{Z_k(\lambda)}\,.$$ Furthermore, GEPs with rather general hopping rates depending on the number of particles on the exit site have been also studied; see, e.g. [@KL99; @KLO94; @Barma_drop; @Barma; @PK; @YN_06]. In all these models the steady-state probabilities are also given by a product measure. We emphasize that this product measure structure is akin to that of the zero-range process [@MartinZRP; @DeMasi] although the two processes are fundamentally different (in the GEP, the jump rate of a particle does depend on the state of the target site, contrary to what is assumed in the zero-range process).
In order to study the dynamics of the system, the knowledge of the steady-state distribution is not sufficient and a full description of the evolution requires the complete spectrum and eigenstates of the evolution matrix. Yet the large scale “hydrodynamic” behavior is conceptually simple. The only relevant hydrodynamic variable is density and it evolves according to a diffusion equation. In one dimension, for instance, it reads $$\label{DE}
\frac{\partial \rho}{\partial t} = \frac{\partial}{\partial x}
\!\left[D_k(\rho)\frac{\partial \rho}{\partial x}\right] .$$ This generic result is valid for lattice gases with symmetric hopping [@Spohn; @KL99; @L99]. Thus the detailed microscopic rules underlying the dynamics of the lattice gas play little role; namely, they are all encapsulated in a single density-dependent function, the diffusion coefficient.
The determination of the diffusion coefficient is in principle a very difficult problem as we do not assume the lattice gas to be dilute. For the GEP, the diffusion coefficient $D_k$ is known in the extreme cases, namely for symmetric random walks ($k=\infty$) and for the SEP ($k=1$). In both these cases the diffusion coefficient is constant; with our choice of the hopping rates, we have $$\label{Diff}
D_1 = D_\infty = 1.$$ For other maximal occupancies ($1<k<\infty$), the diffusion coefficient is density-dependent. This already follows from the asymptotic behaviors $$\label{Diff_asymp}
D_k(\rho) =
\begin{cases}
1 & \rho\to 0 ,\\ k & \rho\to k .
\end{cases}$$ The small-density asymptotic corresponds to the diffusion of a single particle in the empty system, while the behavior in the $\rho\to k$ limit can be understood by considering a single vacancy in the fully occupied system.
The computation of $D_k(\rho)$ for all $k$ will be presented in section \[sec:DC\]. We will show that $$\label{Dk}
D_k = \Lambda_k - \rho\,\frac{d\Lambda_k}{d\rho}\,, \quad
\Lambda_k(\rho)=1-P_k(\rho) .$$ Using and one obtains the parametric representation of $\Lambda_k(\rho)$: $$\label{P_max}
\rho = \lambda\,\Lambda_k(\lambda)\,,\quad \Lambda_k(\lambda) =
\frac{ Z_{k-1}(\lambda)}{ Z_k(\lambda)} .$$ These formulas apply to all $1\leq k\leq \infty$ including the extreme cases. For the SEP we have $\Lambda_1=1-\rho$, while for random walks $\Lambda_\infty=1$; in both cases we recover . In all other cases ($2\leq k<\infty$), the diffusion coefficients $D_k(\rho)$ are monotonically increasing convex functions of the density (as illustrated in Fig. \[D5\_fig\] for $k=2,3,4,5$).
![Diffusion coefficient as a function of density for the GEP with $k=2,3,4,5$.[]{data-label="D5_fig"}](GEP-D_k-n.pdf){width="88mm"}
For lattice gases in higher dimensions, the density generally satisfies a diffusion equation $$\label{rho:eq}
\frac{\partial \rho}{\partial t} = \sum_{a,b=1}^d \frac{\partial
}{\partial x_a} \left[D^{ab}(\rho) \frac{\partial \rho}{\partial
x_b}\right]$$ with a $d\times d$ diffusion matrix $D^{ab}(\rho)$. An ordinary diffusion process (e.g., a symmetric random walk) is macroscopically isotropic, so the diffusion matrix is scalar: $D^{ab}(\rho)=\delta^{ab}D(\rho)$. Generally the diffusion matrix is symmetric, $D^{ab}(\rho)=D^{ba}(\rho)$, and for lattice gases on $\mathbb{Z}^d$ the symmetry of the lattice limits the number of independent matrix elements to two: All diagonal elements are equal \[we denote them by $D(\rho)$\], and all off-diagonal are also equal \[we denote them by $\widehat{D}(\rho)$\]. In three dimensions, for instance, the diffusion matrix is $$\label{matrix}
{\bf D}(\rho) = \left[ \begin{array}{ccc} D(\rho)
&\widehat{D}(\rho) & \widehat{D}(\rho)\\ \widehat{D}(\rho) & D(\rho)
& \widehat{D}(\rho)\\ \widehat{D}(\rho) & \widehat{D}(\rho) &
D(\rho)
\end{array} \right].$$
For lattice gases on $\mathbb{Z}^d$ in which each particle occupies a single lattice site one expects the diffusion matrix to be scalar and the diffusion coefficient to be strictly positive [@critical]. As far as we know these physically obvious assertions have not been proved in full generality. In the next section, we calculate $D_k(\rho)$ for all $k$. Our derivation implies that for the GEP the diffusion process is indeed macroscopically isotropic, and our explicit results show that the diffusion coefficients satisfy $D_k(\rho)>1$ and they are independent of the spatial dimension.
Diffusion Coefficient {#sec:DC}
=====================
In this section, we calculate the diffusion coefficient for the GEP that appears in Eq. (\[rho:eq\]). As a warm-up, we recall the well-known case of the SEP. Then we show the crucial difference between the GEPs with $2\leq k<\infty$ and the SEP, namely the presence of the higher-order correlators, which makes impossible an elementary derivation of the diffusion coefficient. Fortunately, the understanding of the equilibrium in the GEP and a perturbative expansion around the equilibrium gives a method for deriving the diffusion coefficient for the GEPs with $2\leq k<\infty$. We present a detailed derivation of $D_2(\rho)$ in one and higher dimensions. The case of arbitrary $k$ is outlined at the end of this section.
A configuration of the SEP on a one-dimensional lattice is fully described by binary variables $n_j(t)$: If the site $j\in\mathbb{Z}$ is empty, $n_j(t)=0$; if it is occupied, $n_j(t)=1$. In an infinitesimal time interval $dt$, the particle hops from site $j$ to site $j+1$ with probability $n_j(1-n_{j+1})dt$. This choice assures that the hopping event happens only when the site $j$ is occupied and the site $j+1$ is empty. Taking into account all possible hops one finds that the average density evolves according to $$\begin{aligned}
\begin{split}
\frac{d \langle n_j\rangle}{dt} &= \langle n_{j-1}(1-n_j)
+n_{j+1}(1-n_j)\rangle\\ &- \langle
n_j(1-n_{j-1})+n_j(1-n_{j+1})\rangle \, ,
\end{split}\end{aligned}$$ which simplifies to the discrete diffusion equation $$\label{DDE}
\frac{d \langle n_j\rangle}{dt} = \langle n_{j-1}\rangle - 2\langle
n_j\rangle + \langle n_{j+1}\rangle \, .$$ The remarkable cancellation of the higher-order correlation functions allows one to prove the validity of the hydrodynamic limit without further assumptions—no need to use the absence of correlations in the steady state. By definition, in the hydrodynamic limit the average density varies on the scales greatly exceeding the lattice spacing. Therefore we write $\langle n_j(t)\rangle =\rho(x,t)$; the notation $x=j$ emphasizes that we are switching to the continuum description. We then expand $\langle n_{j\pm 1}\rangle$ in Taylor series $$\label{n11}
\langle n_{j\pm 1}\rangle = \rho \pm \rho_x + \tfrac{1}{2}\rho_{xx}+\cdots$$ and recast the set of difference-differential equations into a classical diffusion equation, namely Eq. with $D_1=1$. In higher dimensions, the cancellation still holds; in two dimensions, for instance, $$\begin{aligned}
\begin{split}
\frac{d \langle n_{i,j}\rangle}{dt} & = \langle n_{i,j-1}\rangle +
\langle n_{i,j+1}\rangle + \langle n_{i-1,j}\rangle + \langle
n_{i+1,j}\rangle\\ & - 4\langle n_{i,j}\rangle\, .
\end{split}\end{aligned}$$ Therefore the hydrodynamic description is again the classical diffusion equation $\rho_t = \rho_{xx} + \rho_{yy}$.
Consider now the simplest GEP different from the SEP, namely the GEP with $k=2$ on a one-dimensional lattice. The occupation number $n_j$ is either 0, or 1, or 2 when $k=2$. The process $(n_j, n_{j+1})\Longrightarrow (n_j - 1, n_{j+1} + 1)$ proceeds with rate $$\label{rate}
n_jF(n_{j+1}), \quad F(n) =1-\frac{n(n-1)}{2} .$$ Therefore the average density evolves according to $$\begin{aligned}
\begin{split}
\label{nj_av}
\frac{d \langle n_j\rangle}{dt} &= \langle [n_{j-1}+ n_{j+1}]
F(n_j)\rangle \\ &- \langle n_j
[F(n_{j-1})+F(n_{j+1})]\rangle\, .
\end{split}\end{aligned}$$ In contrast with the case of the SEP, higher-order correlation functions do not cancel as it is obvious from an explicit representation of the right-hand side of : $$\begin{aligned}
\begin{split}
\frac{d \langle n_j\rangle}{dt} &= \langle n_{j-1}\rangle - 2\langle
n_j\rangle + \langle n_{j+1}\rangle \\ &+ \tfrac{1}{2}\langle
n_j [n_{j-1}^2+n_{j+1}^2]-[n_{j-1}+ n_{j+1}] n_j^2\rangle\, .
\end{split}\end{aligned}$$
It is often possible to advance for lattice gases of the *gradient* type [@Spohn; @KL99]. These are lattice gas models in which the current $J_{j,j+1}$ of particles moving from any site $j$ to $j+1$ can be written as a discrete gradient. For instance, the SEP is the gradient lattice gas since $ J_{j,j+1} = n_j-n_{j+1}$. For the GEP with $k=2$, the current $$J_{j,j+1}=n_j-n_{j+1} + \tfrac{1}{2}[n_{j+1}n_j^2-n_{j}n_{j+1}^2]$$ is obviously not a discrete gradient. Generally all GEPs with $2\leq k<\infty$ are non-gradient lattice gases.
We now outline the idea of a perturbative approach which we shall use to establish the diffusion coefficient in non-gradient lattice gases, and then return to the GEP.
Perturbative Approach
---------------------
For non-gradient lattice gases it is sometimes possible to study the hydrodynamic regime in the realm of a perturbative approach. The idea is to ignore correlations. This is true in the equilibrium. In the evolving state, the presence of local density gradients induces long-ranged correlations, but in numerous lattice gases these correlations vanish to first order in the density difference. This has been rigorously established (in all spatial dimensions) for lattice gases with hard-core exclusion [@Spohn83], and it is expected to apply to a much larger class of models. There can be appreciable correlations in the earlier time regime, but we are interested in the hydrodynamic limit which, by definition, describes the evolution close to equilibrium. The vanishing of correlations in the first order in the density gradient implies that in the hydrodynamic regime our perturbative treatment leads to exact predictions for the diffusion coefficient.
To appreciate the validity of a perturbative approach it is useful to compare the situation with kinetic theory [@Spohn; @KRB10; @fluid]. Recall that the Boltzmann equation, even though it is mean-field in nature (as it is based on the assumption of molecular chaos), is asymptotically exact in the hydrodynamic regime, so the emerging transport coefficients are exact. In the context of kinetic theory the main challenge is technical—even for dilute monoatomic gases (e.g., for hard spheres gas), it has not been possible to *extract* transport coefficients analytically [@fluid]. Lattice gases are much more tractable, so even for dense lattice gases the computing of the diffusion coefficient is occasionally feasible. The crucial ingredient is the understanding of the equilibrium state. We emphasize that for lattice gases of gradient type (e.g., for the Katz-Lebowitz-Spohn model with symmetric hopping [@KLS; @Krug01] and for repulsion processes [@RP]) when computations using a Green-Kubo formula become feasible, the results for the diffusion coefficient *agree* with predictions derived using the perturbative approach. Further, for lattice gases of non-gradient type whenever it was possible to apply the perturbative approach (see [@SOC:sing; @EPA]), the predictions for the diffusion coefficient were again exact as it was evidenced through rigorous analyses, mappings to gradient type lattice gases, and comparisons with simulations.
GEP with $k=2$
--------------
To implement the perturbative approach for the GEP with $k=2$ on the one-dimensional lattice, we first replace by $$\begin{aligned}
\begin{split}
\label{nj_av_1d}
\frac{d \langle n_j\rangle}{dt} & = [\langle n_{j-1}\rangle + \langle
n_{j+1}\rangle] \langle F(n_j)\rangle \\ & - \langle
n_j\rangle [\langle F(n_{j-1}) \rangle +\langle F(n_{j+1})\rangle] .
\end{split}\end{aligned}$$ In the hydrodynamic limit we write $\langle n_j(t)\rangle =\rho(x,t)$ and we use Eq. for $\langle n_{j\pm 1}(t)\rangle$ to yield $$\label{nn}
\langle n_{j-1}\rangle + \langle n_{j+1}\rangle = 2\rho+\rho_{xx} .$$ Hereinafter, we keep the terms which survive in the hydrodynamic limit, e.g., in Eq. we have dropped $\tfrac{1}{12}\rho_{xxxx}$ and the following terms with higher derivatives.
The average $\langle F(n)\rangle$ has a neat form $$\label{F_av}
\langle F(n)\rangle = 1- P_2(\rho) ,$$ which is obvious from the definition of the process (the hopping can occur only when the target site hosts less than two particles). We shall use the shorthand notation $1- P_2(\rho)=\Lambda_2(\rho)$.
In the hydrodynamic limit $\langle F(n_{j-1}) \rangle +\langle
F(n_{j+1})\rangle$ turns into $\Lambda_2[\rho(x-1)]+\Lambda_2[\rho(x+1)]$, which is expanded to yield $$\label{FF_hydro}
2 \Lambda_2(\rho) + \Lambda_2'(\rho)\, \rho_{xx} +
\Lambda_2''(\rho)\,\rho_x^2\, .$$ Inserting all these expansions into we arrive at $$\label{rho_1d}
\rho_t = \left[\Lambda_2(\rho) - \rho\Lambda_2'(\rho)\right] \rho_{xx}
-\rho \Lambda_2''(\rho)\, \rho_x^2\, .$$ This equation can be re-written as the diffusion equation with diffusion coefficient $$\label{D2}
D_2 = \Lambda_2(\rho) - \rho\Lambda_2'(\rho)\, .$$ Recall that, for $k=2$, we have $$\label{rP2}
\rho = \frac{\lambda+\lambda^2}{1+\lambda+\tfrac{1}{2}\lambda^2}\,,
\quad \Lambda_2 = \frac{1+\lambda}{1+\lambda+\tfrac{1}{2}\lambda^2}$$ from which we find an explicit expression for $\Lambda_2(\rho)$: $$\label{L2}
\Lambda_2(\rho) =
\frac{ 1 - \rho + \sqrt{1+2\rho-\rho^2} }{2}\, .$$ Inserting this into yields an explicit expression of the diffusion coefficient $$\label{D2_explicit}
D_2(\rho) = \frac{1 + \rho +
\sqrt{1+2\rho-\rho^2}}{2\sqrt{1+2\rho-\rho^2}}\, .$$
We now consider the GEP with $k=2$ in arbitrary dimension. In two dimensions, for instance, the average density satisfies $$\begin{aligned}
&&\frac{d \langle n_{i,j}\rangle}{dt} =\langle ( n_{i-1, j} +
n_{i+1,j} + n_{i,j-1}+ n_{i,j+1} ) F(n_{i,j})\rangle\\
&&-\langle n_{i,j} [F(n_{i-1, j})\! + \!F(n_{i+1,j})\! +\! F(n_{i,j-1})\!+\!
F(n_{i,j+1})]\rangle\end{aligned}$$ which in the hydrodynamic limit becomes $$\rho_t = \partial_x(D_2\,\rho_x)+\partial_y(D_2\,\rho_y)$$ with $D_2$ given by Eq. as in one dimension. The same holds in any spatial dimension, namely the GEP is described by the diffusion equation $$\rho_t = \nabla\cdot (D_2 \nabla \rho),$$ where the diffusion coefficient is given by a universal formula valid in arbitrary dimension. The symmetric GEP is therefore *isotropic* on the hydrodynamic scale; namely, it is described by the scalar diffusion coefficient.
GEP with arbitrary $k$
----------------------
For the GEP with arbitrary $k$ the analysis is similar to the one presented above. The process $(n_j, n_{j+1})\Longrightarrow (n_j - 1, n_{j+1} + 1)$ proceeds with rate , where we only need to modify $F(n)$ to $$\label{rate_k}
F(n) =1-\frac{n(n-1)\cdots (n-k+1)}{k!}\, .$$
It suffices to consider the one-dimensional case as the results for the diffusion coefficient are independent of the spatial dimensionality. Equations and , with $F(n)$ given by , remain valid. Equations – also hold if we replace $\Lambda_2$ by $\Lambda_k$, the probability that a site is not fully occupied; for instance, Eq. becomes $\langle F(n)\rangle = 1-P_k(\rho)\equiv \Lambda_k(\rho)$. Thus the diffusion coefficient is indeed given by the announced expression . For $k\geq 5 $ an explicit expression for $\Lambda_k(\rho)$ is apparently impossible to deduce, but we can use a parametric expression which follows from , , and the definition $\Lambda_k(\rho)=1- P_k(\rho)$.
Stationary density profiles {#sec:SDP}
===========================
In the previous section we calculated the diffusion coefficient for the GEP using a perturbative approach. In this section, we present a non-direct test of our predictions. Specifically, we shall calculate stationary density profiles in one and two dimensions and compare these theoretical predictions with simulation results. We will show that the diffusion equation with the diffusion coefficient given by Eqs. and provides an accurate description of the system at a macroscopic scale.
One-dimensional density profiles
--------------------------------
Consider the GEP on the interval $(0, L)$ with boundary conditions $$\label{BC}
\rho(0) = \rho_0, \quad \rho(L) = \rho_1 \, .$$ Solving $D_k(\rho)\frac{d\rho}{dx}=\text{const}\,$, subject to , we obtain $$\label{formal}
\frac{\int_{\rho_0}^\rho dr\,D_k(r)}{\int_{\rho_0}^{\rho_1}
dr\,D_k(r)} = \frac{x}{L} \, .$$ Let us focus on a special case when the right boundary is a sink: $\rho_1=0$. To simplify formulas we write $\rho_0=n$. When $k=2$, the integrals on the left-hand side of can be explicitly determined to yield an implicit representation of the stationary density profile $\rho(x)$: $$\frac{1 + \frac{\pi}{2} + \rho - \sqrt{1+2\rho-\rho^2} +
2\arcsin\left(\frac{\rho-1}{\sqrt{2}}\right)} {1 + \frac{\pi}{2} + n
- \sqrt{1+2n-n^2} + 2\arcsin\left(\frac{n-1}{\sqrt{2}}\right)} =
1-\frac{x}{L}\, .$$ In the case of the maximal density on the left boundary, $\rho_0=n=2$, we get $$\label{profile:k=2}
\frac{1}{2} - \frac{\rho - \sqrt{1+2\rho-\rho^2} +
2\arcsin\left(\frac{\rho-1}{\sqrt{2}}\right)} {\pi + 2} =
\frac{x}{L} \, ,$$ see Fig. \[profile\_fig\].
In the general case of arbitrary $k$ we use and and establish the following parametric representation $$\label{Dk_profile}
\frac{\int_0^\lambda d\mu\,[\Lambda_k(\mu)]^2} {\int_0^\ell
d\mu\,[\Lambda_k(\mu)]^2} = 1 - \frac{x}{L}\, , \quad n =
\ell\,\Lambda_k(\ell) \, .$$ The maximal density on the left boundary, $\rho_0=n=k$, corresponds to $\ell = \infty$. The density profiles in this situation, $$\label{Dk_profile_extreme}
\frac{\int_0^\lambda d\mu\,[\Lambda_k(\mu)]^2} {\int_0^\infty
d\mu\,[\Lambda_k(\mu)]^2} = 1 - \frac{x}{L}\, , \quad \rho =
\lambda\,\Lambda_k(\lambda)$$ are plotted in Fig. \[profile\_fig\].
![ Stationary density profiles versus $x/L$ for the GEP with $k=1,2,3,4,5$ on a segment with $L=10^3$. The solid lines are theoretical predictions in the case of extremal boundary densities, $\rho_0=k$ and $\rho_1=0$. For the SEP ($k=1$), the density profile is linear; for $k=2$, the density profile is given by , and generally it is extracted from Eq. . Simulation results (shown by $\bullet$) were obtained by averaging over the time window $
5\times 10^6 \leq t\le 10^7$.[]{data-label="profile_fig"}](GEP-1d-profile.pdf){width="85mm"}
The GEP in an annulus
---------------------
For the GEP in the annulus $a\leq R\leq L$, we solve $RD_k(\rho)\frac{d\rho}{dR}=\text{const}$, subject to the boundary conditions $\rho(a)=\rho_0$ and $\rho(L) = \rho_1$, and get $$\label{formal_2d}
\frac{\int_{\rho_0}^\rho dr\,D_k(r)}{ \int_{\rho_0}^{\rho_1}
dr\,D_k(r)} = \frac{\ln(R/a)}{\ln(L/a)} \,.$$
Let us look more carefully at the case of $k=2$ with boundary densities $\rho_1=0, \rho_0=2$ (the density on the inner circle is maximal). We use dimensionless variables $\alpha = a/L$ and $\xi =
R/L$, so that $0<\alpha\leq \xi\leq 1$. With these choices, Eq. becomes $$\label{profile_2d:a=O(L)}
\frac{1}{2} +
\frac{\rho-\sqrt{1+2\rho-\rho^2}+2\arcsin\left(\frac{\rho-1}{\sqrt{2}}\right)}
{\pi + 2} =\frac{\ln \xi}{\ln \alpha } \,.$$ This density profile is compared with simulation results on Fig. \[profile\_D2\_fig\].
![Stationary density profiles versus $R$ for the GEP with $k=2$ in the annulus with external radius $L=100$. Dots are simulation results which were obtained by averaging over $5\times 10^6\le t\le
10^7$. Solid lines are theoretical predictions given by with $\alpha=1/4$ in the case of $a=25$ and by when $a=1$. []{data-label="profile_D2_fig"}](GEP-profile-2d.pdf){width="85mm"}
For the GEP in the annulus $a\leq R\leq L$, the usage of the continuum (diffusion equation) approach is somewhat questionable near the inner circle if $a=O(1)$. Indeed, we cannot even talk about a circle on a lattice if its radius is comparable with the lattice spacing. Nevertheless, let us use Eq. , again with $k=2$ and $(\rho_0,\rho_1)=(2,0)$, in the extreme case of $a=1$. Equation becomes $$\label{profile_2d}
\frac{1}{2} - \frac{\rho - \sqrt{1+2\rho-\rho^2} +
2\arcsin\left(\frac{\rho-1}{\sqrt{2}}\right)} {\pi + 2} =\frac{\ln
R }{\ln L } \,.$$ Choosing the inner radius equal to lattice spacing is essentially equivalent to the simplest lattice setting with reservoir connected to the origin and postulating that whenever a particle leaves the origin, a particle from reservoir is immediately added, so the density at the origin remains maximal $\rho_0=2$. There is also a sink at the circle $R=L$; that is, whenever a particle at a site on distance $<L$ hops and gets outside this circle, it leaves the system forever. On distances $R\gg 1$ the profile should become asymptotically exact. Figure \[profile\_D2\_fig\] shows an excellent agreement between theory and simulations over the entire range $1\leq
R\leq L$.
To emphasize the difference between one and two dimensions let us consider the GEP with $k=2$ and boundary densities $(\rho_0,\rho_1)=(2,0)$ and compare the density profiles and . In one dimension, the intermediate density $\rho_* = (\rho_0+\rho_1)/2$, i.e., $\rho_* = 1$ in our case, is reached at $$\frac{x_*}{L} = \frac{1}{2} + \frac{\sqrt{2}-1}{\pi + 2} =
0.580561\ldots ,$$ while in two dimensions this happens at $$\frac{\ln R_*}{\ln L} = \frac{1}{2} + \frac{\sqrt{2}-1}{\pi + 2} ,$$ which is much closer to the source, $R_*\sim L^{0.580561}$.
Second, we compare the total (average) number of particles. In one dimension we integrate by part to get $$N=\int_0^L dx\,\rho(x)= \int_0^2 d\rho\, x(\rho) .$$ Using we perform the integration and find $$N = \frac{3\pi + 2}{2\pi+4}\,L .$$ In two dimensions, we similarly find $$\label{N_2d}
N=\int_0^L dR\,2\pi R \rho(R)= \pi\int_0^2 d\rho\, R^2(\rho) \,.$$ The dominant part of the integral in is gathered near $\rho=0$. Expanding the left-hand side of we find $$\frac{\ln R}{\ln L} = 1 - \frac{2}{\pi + 2}\,\rho - \frac{1}{3(\pi +
2)}\,\rho^3 + \frac{1}{2(\pi + 2)}\,\rho^4+\cdots \,.$$ Equation becomes $$\frac{N}{\pi L^2}\simeq \int_0^\infty d\rho\, \exp\!
\left[- \ln L\, \frac{4\rho +\tfrac{2}{3}\rho^3 - \rho^4 }{\pi + 2}\right] ,$$ which gives $$\label{N_2d_asymp}
N =\frac{\pi(\pi + 2)}{4}\,\frac{L^2}{\ln L}
\left[1 + \frac{C_2}{(\ln L)^2}+\frac{C_3}{(\ln L)^3}+\cdots\right]$$ with $C_2=-\left(\frac{\pi+2}{4}\right)^3,
C_3=6\left(\frac{\pi+2}{4}\right)^4$, etc. Thus the convergence to the leading asymptotic behavior is slow in two dimensions.
For arbitrary $k$, let us choose again $\rho_1=0$ and $\rho_0=k$. the density profile is implicitly given by $$\frac{\int_0^\lambda d\mu\,[\Lambda_k(\mu)]^2} {\int_0^\infty
d\mu\,[\Lambda_k(\mu)]^2} = 1-\frac{\ln R}{\ln L} \,.$$ In the small $\rho$ limit, we get $$\frac{\ln R}{\ln L} = 1 - \frac{\rho}{I_k}+\cdots, \quad
I_k=\int_0^\infty d\mu\,[\Lambda_k(\mu)]^2 \,,$$ and the leading asymptotic behavior of the total average number of particles is $$N \simeq \frac{\pi I_k}{2}\,\frac{L^2}{\ln L}$$ where the coefficients $I_k$ can be evaluated numerically (e.g. $I_3=4.29139\ldots$).
Self-Diffusion Coefficient {#sec:SDC}
==========================
Even an equilibrium situation (in which the density is spatially uniform) possesses interesting non-equilibrium features. One important example is the phenomenon of self-diffusion. In this section we investigate the evolution of a tagged particle in the GEP at equilibrium. We assume that the tagged particle is identical to the host particles, so it merely carries a tag. Asymptotically, the tagged particle exhibits a diffusive behavior, so it suffices to compute the coefficient of self-diffusion. This problem is easy to pose, but there has been little progress even for simplest lattice gases. For instance, the coefficient of self-diffusion is unknown for the SEP in two and higher dimensions, it is only known [@LOV_01] that the coefficient of self-diffusion is a smooth function of the density.
Consider first the one-dimensional case. We tag a particle which is initially at $x(0)=0$ (without loss of generality) and we look at its position $x(t)$ in the long time limit. Generically, we expect a diffusive behavior. Thus the first two averages are $\langle x\rangle = 0$ and $\langle x^2\rangle\sim t$, and it suffices to determine the self-diffusion coefficient $$\label{SD:def}
\lim_{t\to\infty}\frac{\langle x^2\rangle}{2t} \, = \,
\mathcal{D}_k(\rho).$$ The self-diffusion coefficient $\mathcal{D}_k$ generally differs from the diffusion coefficient $D_k$. We have $\mathcal{D}_\infty =
D_\infty=1$ for non-interacting random walks. For $k<\infty$, the inequality $\mathcal{D}_k<D_k$ is physically apparent, although it may be difficult to prove.
For the SEP in one dimension, the self-diffusion coefficient vanishes: $\mathcal{D}_1=0$. Indeed, the ordering between the particles is conserved and this leads to anomalously slow sub-diffusive behavior [@Harris_65; @Levitt_73; @PMR_77; @AP_78; @Arr_83]: $\langle x^2\rangle_{\text{SEP}, \,\, d=1}\sim t^{1/2}$. This is an exceptional feature; the normal diffusion is recovered for the SEP in dimensions higher than 1. For the GEP with $k\geq 2$, the phenomenon of self-diffusion is not pathological even in one dimension, viz. the self-diffusion coefficient $\mathcal{D}_k(\rho)$ is positive. Moreover, $\mathcal{D}_k(\rho)$ is a monotonically decreasing function of $\rho$ in the interval $0<\rho<k$ with asymptotic behaviors $$\label{SD_asymp}
\mathcal{D}_k(\rho) =
\begin{cases}
1 & \rho\to 0 ,\\
0 & \rho\to k .
\end{cases}$$
For lattice gases in $d>1$ dimensions, the spread of the tagged particle is generically described by a matrix. For the GEP on the hyper-cubic lattice, and generally for lattice gases on $\mathbb{Z}^d$ where each particle occupies only one site, one expects the self-diffusion process to be isotropic on the hydrodynamic scales. This has been proved only for the SEP, see [@KV_86], so it remains conjectural for the GEPs with $2\leq k<\infty$.
![Mean-square displacement of the tagged particle vs time for the GEP in one dimension. The maximal occupancy varies between $k=2$ and $k=5$. Simulations were performed on the ring of length $10^3$.[]{data-label="GEP-Xt2_fig"}](GEP-Xt2.pdf){width="85mm"}
There is one important feature which distinguishes the self-diffusion coefficient from the diffusion coefficient, viz. the self-diffusion coefficient certainly depends on the dimensionality: $\mathcal{D}_k(\rho, d)$. The extreme density behaviors of the self-diffusion coefficient $\mathcal{D}_k(\rho, d)$ with $k\geq 2$ are universal and given by in all dimensions.
We performed simulations to probe the self-diffusion coefficient on lattices with $L^d$ sites, in $d=1,2,3$ dimensions, with periodic boundaries. The sizes of the simulated systems are $L=10^3,~32,~10$ for $d=1,2,3$, respectively. The number of simulation runs for each set of parameters $ (k,d,\rho)$ is $2\times 10^5/(L^d \rho)$, and we tagged all the $L^d \rho$ particles in each run. Thus $\langle \cdot \rangle $ is the average over effectively $2\times 10^5$ tagged particles. We checked the validity of the diffusive scaling up to $t=5\times 10^4$, as shown in Fig. \[GEP-Xt2\_fig\] for $\rho = 3k/5$ and $d=1$. (As long as $t\ll L^{2} = 10^6$, finite size effects can be safely ignored.) We calculated $\frac{\langle \mathbf{r}^2 \rangle}{2d\cdot t}$ by using data in the $0\le t\le 5\times 10^4$ time window. The results are shown in Figs. \[GEP-self-1d\_fig\] and \[SD-123d\_fig\].
![Coefficient of self-diffusion vs density for the GEP in one dimension; the maximal occupancy varies between $k=2$ and $k=5$. Dots represent simulation results for the GEP in one dimension. Solid lines are the mean-field predictions, Eq. .[]{data-label="GEP-self-1d_fig"}](GEP-SD-1d-n.pdf){width="88mm"}
As a reference point, it is useful to have a mean-field prediction. To derive the mean-field prediction for the self-diffusion coefficient of the GEP with an arbitrary maximal occupancy $k$ we recall that a site is occupied by $k$ particles with probability $P_k$, so it can be a destination site with probability $1-P_k$. Therefore for a tagged particle, the hopping rate to each neighboring site appears to be $1-P_k(\rho)\equiv \Lambda_k(\rho)$, which tells us that the self-diffusion coefficient is $$\label{SD:MF}
\mathcal{D}_k^\text{MF} = \Lambda_k =
\frac{ Z_{k-1}(\lambda) }{ Z_k(\lambda)} =
\frac{\rho}{\lambda}\,.$$ In particular, $\mathcal{D}_1^\text{MF}=1-\rho$ for the SEP, while for $k=2$ the mean-field prediction $\mathcal{D}_2^\text{MF}=\Lambda_2(\rho)$ is given by the explicit formula . For non-interacting random walks Eq. yields $\mathcal{D}_\infty^\text{MF}=1$, and this is the only case when the prediction is exact. In all other cases ($1\leq k<\infty$), the prediction of Eq. is *not* exact.
To appreciate the mean-field nature of the prediction we first note that for every site, the probability that any neighboring site contains less than $k$ particles is indeed $1-P_k$, and these probabilities are uncorrelated. So if we pick a particle and mark it with a tag, it appears that this particle is indeed diffusing with the coefficient equal to $1-P_k$. But we must keep the identity of the tagged particle. This already causes the problem—immediately after the tagged particle has undergone the first jump, the site from which it jumped will be surely occupied by less than $k$ particles. In the $d\to\infty$ limit this is irrelevant, but for any finite dimension the derivation of Eq. involves an uncontrolled approximation. We thus realize that Eq. only provides a mean-field approximation. To summarize, the prediction satisfies the following properties:
1. It agrees with the limiting behaviors .
2. It appears to be an upper bound for all $d\geq 1$.
3. It becomes exact in the $d\to\infty$ limit. This justifies calling a mean-field prediction.
4. It is also exact for non-interacting random walks ($k=\infty$).
The validity of the first property easily follows from Eq. , and it is also seen (in the one-dimensional case) from Fig. \[GEP-self-1d\_fig\]. The second property seems very plausible, but has not been proved; it is supported by simulation results; see Figs. \[GEP-self-1d\_fig\] and \[SD-123d\_fig\]. The validity of the third and fourth properties is obvious.
Figures \[GEP-self-1d\_fig\] and \[SD-123d\_fig\] indicate that the disagreement between the actual behaviors and the mean-field predictions is most pronounced in one dimension, so the mean-field estimate provides a good approximation in two and three dimensions.
![Coefficient of self-diffusion vs density for the SEP in two and three dimensions, and for the GEP with $k=2$ in dimensions $d=1, 2, 3$. Dots represent simulation results. Solid lines are the mean-field predictions: $\mathcal{D}_1^\text{MF} = 1 - \rho$ for the SEP, while for the GEP with $k=2$ the mean-field prediction $\mathcal{D}_2^\text{MF} = \Lambda_2(\rho)$ acquires the explicit form .[]{data-label="SD-123d_fig"}](GEP-SD-123d-n.pdf){width="88mm"}
Assuming that the mean-field estimate provides qualitatively correct small and large $\rho$ behaviors also in finite dimensions, we anticipate that $$\lim_{\rho\downarrow 0} \frac{1-\mathcal{D}_k(\rho, d)}{\rho^k} =
A_k(d), \ \ \lim_{\rho\uparrow k} \frac{\mathcal{D}_k(\rho,
d)}{k-\rho} = B_k(d)\,.$$
Little is known about these amplitudes. For $d=\infty$, $$A_k(\infty) = \frac{1}{k!}\,, \ \ B_k(\infty) = 1\,,$$ because the mean-field prediction becomes exact when $d=\infty$. Since the mean-field estimate apparently provides an upper bound, we expect that $A_k(d)>\frac{1}{k!}$ and $B_k(d)<1$.
Summary {#sec:Sum}
=======
We investigated generalized exclusion processes with symmetric nearest-neighbor hopping parametrized by an integer $k$, the maximal occupancy. Specifically, we studied a class of such processes interpolating between the symmetric exclusion process ($k=1$) and non-interacting random walkers ($k=\infty$). For these lattice gases the hydrodynamic behavior is governed by a diffusion equation. We computed the diffusion coefficient $D_k$ and showed that for every $k$, it does not depend on the spatial dimension, but it does depend on $k$. We showed that, apart from the extreme cases of $k=1$ and $k=\infty$, the diffusion coefficient depends on the density.
We studied numerically the self-diffusion phenomenon for the GEPs in one, two, and three dimensions. An interesting challenge is to compute the self-diffusion coefficient for the GEPs with $k\geq 2$ in one dimension. In two and higher dimensions this problem seems intractable, even for the SEP in two dimensions the coefficient of self-diffusion is unknown; it has only been established that $\mathcal{D}_1(\rho, d)$ is a smooth function of the density [@LOV_01]. In the one-dimensional setting, the tagged particle in the case of the SEP undergoes an anomalously slow sub-diffusive behavior which is well understood, and even large deviations of the displacement of the tagged particle have been probed (see e.g. [@SV_13; @KMS_14] and references therein). Therefore there is a hope that the self-diffusion phenomenon in the one-dimensional GEP is also analytically tractable.
In addition to the diffusion coefficient, a second transport coefficient, the mobility (or conductance) $\sigma(\rho)$, plays an important role in the macroscopic fluctuation theory [@MFTReview]. The knowledge of $\sigma(\rho)$ is required if one wants to understand fluctuations around the (deterministic) hydrodynamic behaviors, including large deviations. We leave the determination of the mobility $\sigma(\rho)$ for future studies of fluctuations and large deviations in the GEPs.
In this article we considered only the GEP with symmetric hopping. One would like to understand the asymmetric version of the GEP. The problem is that in the driven case the structure of the steady states is unknown: it is not a product measure anymore, even on a ring [@Timo]. One possibility is to modify the rules of the GEP to make the structure of the steady states more accessible. An interesting variant is to employ a drop-push dynamics when a hopping attempt to a fully occupied neighboring site is not rejected, but instead the particle proceeds in the same direction and lands at the closest site which is not fully occupied [@Barma_drop; @Barma]; a similar process was suggested, and studied for $k=1$, in the context of self-organized criticality [@SOC:sing]. For these GEPs with drop-push dynamics the structure of the steady states is known in the general case of asymmetric hopping [@Barma_drop; @Barma]. The symmetric version of the GEPs with drop-push dynamics also deserves further analysis, e.g., one would like to compute the diffusion coefficient for this class of lattice gases.
**Acknowledgments.** One of the authors (CA) thanks Chihiro Matsui for a stimulating discussion. KM and PLK are grateful to the Galileo Galilei Institute for Theoretical Physics for the hospitality and the INFN for partial support during the completion of this work. The work of P.L.K. was partly supported by grant No. 2012145 from the US–Israel Binational Science Foundation (BSF).
[99]{}
S. Katz, J. L. Lebowitz, and H. Spohn, J. Stat. Phys. [**34**]{}, 497 (1984).
H. Spohn, [*Large Scale Dynamics of Interacting Particles*]{} (Springer-Verlag, New York, 1991).
B. Schmittmann and R. K. P. Zia, *Statistical Mechanics of Driven Diffusive Systems*, in: [*Phase Transitions and Critical Phenomena*]{}, edited by C. Domb and J. L. Lebowitz (Academic Press, London, 1995), Vol. 17.
B. Derrida, Phys. Rep. [**301**]{}, 65 (1998).
T. M. Liggett, [*Stochastic Interacting Systems: Contact, Voter, and Exclusion Processes*]{} (Springer, New York, 1999).
C. Kipnis and C. Landim, [*Scaling Limits of Interacting Particle Systems*]{} (Springer, New York, 1999).
P. L. Krapivsky, S. Redner, and E. Ben-Naim, [*A Kinetic View of Statistical Physics*]{} (Cambridge University Press, Cambridge, UK, 2010).
D. Chowdhury, L. Santen, and A Schadschneider, Phys. Rep. [**329**]{}, 199 (2000).
T. Chou, K. Mallick, and R. K. P. Zia, Rep. Prog. Phys. [**74**]{}, 116601 (2011).
G. M. Schütz, *Exactly Solvable Models for Many-Body Systems Far From Equilibrium*, in [*Phase Transitions and Critical Phenomena*]{}, edited by C. Domb and J. L. Lebowitz (Academic Press, London, 2000), Vol. 19.
R. A. Blythe and M. R. Evans, J. Phys. A [**40**]{}, R333 (2007).
B. Derrida, J. Stat. Mech. P07023 (2007).
L. Bertini, A. De Sole, D. Gabrielli, G. Jona-Lasinio, and C. Landim, J. Stat. Phys. [**107**]{}, 635 (2002).
T. Bodineau and B. Derrida, C. R. Physique [**8**]{}, 540 (2007).
B. Derrida, J. Stat. Mech. P01030 (2011).
G. Jona-Lasinio, J. Stat. Mech. P02004 (2014).
L. Bertini, A. De Sole, D. Gabrielli, G. Jona-Lasinio, and C. Landim, arXiv:1404.6466.
G. Gentile, Nuovo. Cim. [**17**]{}, 493 (1940).
A. Khare, [*Fractional Statistics and Quantum Theory*]{} (World Scientific, Singapore, 1997).
C. Kipnis, C. Landim, and S. Olla, Comm. Pure Appl. Math. [**47**]{}, 1475 (1994).
C. Kipnis, C. Landim, and S. Olla, Ann. Inst. H. Poincaré Probab. Statist. [**31**]{}, 191 (1995).
T. Seppäläinen, Ann. Prob. [**27**]{}, 361 (1999).
G. M. Schütz and S. Sandow, Phys. Rev. E [**49**]{}, 2726 (1994).
M. Barma and R. Ramaswamy, in: *Nonlinearity and Breakdown in Soft Matter*, eds. B. K. Chakrabarti, K. K. Baradhan and A. Hansen (Springer, Berlin, 1994).
G. M. Schütz, R. Ramaswamy, and M. Barma, J. Phys. A: Math. Gen. [**29**]{}, 837 (1996).
F. Tabatabaei and G. M. Schütz, Phys. Rev. E [**74**]{}, 051108 (2006).
U. Basu and P. K. Mohanty, Phys. Rev. E [**82**]{}, 041117 (2010).
J. Tailleur, J. Kurchan, and V. Lecomte, J. Phys. A: Math. Theor. [**41**]{}, 505001 (2008).
C. Matsui, arXiv 1311.7473 (2013).
G. Carinci, C. Giardinà, F. Redig, and T. Sasamoto, arXiv:1407.3367 (2014).
A. Ryabov, Phys. Rev. E [**89**]{}, 022115 (2014).
T. Becker, K. Nelissen, B. Cleuren, B. Partoens, and C. Van den Broeck, Phys. Rev. Lett. [**111**]{}, 110601 (2013).
T. Becker, K. Nelissen, B. Cleuren, B. Partoens, and C. Van den Broeck, arXiv:1410.3360 (2014).
K. Mallick, S. Mallick, and N. Rajewsky, J. Phys. A: Math. Gen. [**32**]{}, 8399 (1999).
M. R. Evans, P. A. Ferrari, and K. Mallick, [J. Stat. Phys.]{} **135**, 217 (2009).
P. A. Ferrari and J. B. Martin, Ann. Probab. **35**, 807 (2007).
Y. Nagahata, Stoch. Process Appl. [**116**]{}, 957 (2006).
M. R. Evans and T. Hanney, J. Phys. A: Math. Gen. [**38**]{}, R195 (2005).
A. De Masi and P. A. Ferrari, J. Stat. Phys. **36**, 81 (1984).
More precisely, $D(\rho)>0$ away from a critical density where the equilibrium lattice gas undergoes a phase transition. Such a phase transition may occur in $d\geq 2$ dimensions; e.g., it occurs in exclusion processes with infinitely strong repulsion between particles in adjacent sites. However, there are no phase transitions in GEPs which we study. In these lattice gases a stronger bound, $D_k(\rho)\geq 1$, is actually valid for all $k$ and $\rho$.
H. Spohn, J. Phys. A: Math. Gen. [**16**]{}, 4275 (1983).
P. Resibois and M. De Leener, [*Classical Kinetic Theory of Fluids*]{} (Wiley, New York, 1977).
J. S. Hager, J. Krug, V. Popkov, and G. M. Schütz, Phys. Rev. E [**63**]{}, 056110 (2001).
P. L. Krapivsky, J. Stat. Mech. P06012 (2013).
J. M. Carlson, J. T. Chayes, E. R. Grannan, and G. H. Swindle, Phys. Rev. Lett. [**65**]{}, 2547 (1990).
U. Bhat and P. L. Krapivsky, Phys. Rev. E [**90**]{}, 012133 (2014).
C. Landim, S. Olla, and S. R. S. Varadhan, Commun. Math. Phys. [**224**]{}, 307 (2001).
T. E. Harris, J. Appl. Prob. [**2**]{}, 323 (1965).
D. G. Levitt, Phys. Rev. A [**8**]{}, 3050 (1973).
P. M. Richards, Phys. Rev. B [**16**]{}, 1393 (1977).
S. Alexander and P. Pincus, Phys. Rev. B [**18**]{}, 2011 (1978).
R. Arratia, Ann. Probab. [**11**]{}, 362 (1983).
C. Kipnis and S. R. S. Varadhan, Commun. Math. Phys. [**104**]{}, 1 (1986).
S. Sethuraman and S. R. S. Varadhan, Ann. Probab. [**41**]{}, 1461 (2013).
P. L. Krapivsky, K. Mallick, and T. Sadhu, Phys. Rev. Lett. [**113**]{}, 078101 (2014).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Spin-current is an important physical quantity in present day spintronics and it might be very usefull in the physics of quantum plasma of spinning particles. Thus it is important to have an equation of the spin-current evolution. This equation naturally appears as a part of a set of the quantum hydrodynamics (QHD) equations. Consequently, we present the set of the QHD equations derived from the many-particle microscopic Schrodinger equation, which consists of the continuity equation, the Euler equation, the Bloch equation and equation of the spin-current evolution. We use these equations to study dispersion of the collective excitations in the three dimensional samples of the magnetized dielectrics. We show that dynamics of the spin-current leads to formation of new type of the collective excitations in the magnetized dielectrics, which we called spin-current waves.'
author:
- 'P. A. Andreev'
- 'L. S. Kuz’menkov'
title: 'Waves of spin-current in magnetized dielectrics '
---
\[sec:level1\]I. Introduction
=============================
The spin-current is a very important characteristic for various physical systems. For instance it is useful for description of such processes as spin injection [@Takahashi; @JMMM; @04] and other process, where spin transport is involved. Most of them are accumulated in a grate field of physics: spintronics [@Zutic; @RMP; @04], [@Sinova; @Nat; @Mat; @12]. There are different methods of the spin-current generation, such as spin pumping by sound waves [@Uchida; @JAP; @12], optical spin injection [@Pezzoli; @PRL; @12], and at using of junctions as Polarizers [@Rashidian; @IJTP; @12], spin flip in the result of electron interaction with electromagnetic wave [@Hammond; @APL; @12]. Spin-polarized currents are used in spin lasers development [@Lee; @PRB; @12] and in spin-diode structures [@Li; @JAP; @12]. It is interesting to admit that in some devises the spin-current exhibits the sine wave-like behavior [@Zhang; @P; @B; @12]. New modification of spin-field effect transistors have been suggested (see for example [@Ma; @JAP; @12], [@Bagraev; @P; @C; @10]), one of the spintronic devices utilize the electron’s spin properties in addition to its charge properties. Effects in junction play key role in spintronic devises. The spin-dependent Peltier effect was observed experimentally [@Flipse; @Nat; @Nano; @12]. It is based on the ability of the spin-up and spin-down channels to transport heat independently. The use of graphene has been involved in this field either [@Pesin; @Nat; @Mat; @12], [@Vera-Marun; @Nat; @phys; @12]. Processes of spin transport and relaxation in graphene have also been considered [@Han; @JMMM; @12]. It has been expected that silicon spintronics has potential to change information technology, this possibility discussed in Ref. [@Jansen; @Nat; @Mat; @12].
We have described a number of examples there the spin-current plays important role. Thus, it is necessary to have analytical definition of the spin-current and equation of the spin-current evolution. In Refs. [@An; @Sci; @Rep; @12]- [@Shi; @PRL; @06] authors discussed the definition of the spin-current in terms of the one-particle wave function describing the quantum state of the spinning particle. Knowledge of the wave function allows to calculate the spin-current and make conclusions on the system behavior. We keep the spin-current as an independent macroscopic variable defined via the many-particle wave function governs system behavior. We use the spin-current along with the particle concentration, the velocity field, and the magnetization. We consider the spin-current evolution due to interparticle interaction. For this purpose we derive the spin-current evolution equation as a part of the set of quantum hydrodynamic equations. Total angular momentum current was presented in Ref. [@An; @Sci; @Rep; @12] as a generalization of the spin current, which conserve even when the spin-current does not conserve. In our paper we consider the spin-current only, since the spin-current has appeared in the quantum hydrodynamic equations. It is not necessary to have a conserved quantity for the spin-current. We study the spin-current evolution due to interparticle interaction.
It can be expected that the spin-current evolution also gives influence on the spin waves, which are well-known phenomenon in many different physical systems, first of all they exist in the ferromagnetics and other structures with strong magnetization. For example quantum dots show interesting spin waves behavior [@Aliev; @PRB; @09]- [@Guslienko; @PRB; @10]. Dynamics of the magnetic moments in the quantum plasma has also been studied. In Ref. [@Andreev; @VestnMSU; @2007] existence of the self-consistent spin waves in the magnetized plasma was shown. Thus new branches of the wave dispersion appear in the magnetized plasma due to the dynamics of magnetic moment of electrons and ions. It was demonstrated by means of the quantum hydrodynamics (QHD) method [@MaksimovTMP; @1999]- [@Andreev; @arxiv; @11; @3]. Later, the generalization of the Vlasov equation for the plasma of spinning charged particles was used to study the same problem [@Brodin; @PRL; @08]. Magnetic moment dynamics also leads to existence of the effect of resonances interaction of the neutron beam with the magnetized plasma, which gives new method of wave generation in the magnetized plasma [@Andreev; @IJMP; @12], [@Andreev; @AtPhys; @08], [@Andreev; @PIERS; @2011]. Method of the QHD has become very popular and powerful method of studying influence of the magnetic moment dynamics on various processes in the magnetized plasma [@Andreev; @IJMP; @12], [@Maksimov; @VestnMSU; @2000]- [@Shukla; @RMP; @11]. The method of the QHD can be used for systems of the neutral particles with the magnetic moment. Such systems are more preferable to demonstrate results given by the QHD method at description of the physical effects caused by evolution of the magnetic moments. The QHD description of the self-consistent spin waves in a system of neutral particles in the one-, two-, and three-dimensional dielectrics was described in Ref. [@Andreev; @PIERS; @2012; @Spin]. Usually the set of QHD equations for the spinning charged particles consists of the three equations for evolution of the material fields, these are the continuity equation for the particles concentration, the Euler equation for the velocity field, and the generalized Bloch equation for the magnetic moment evolution, and the Maxwell equations for the electromagnetic field description.
The spin-current and an equation of its evolution are in the center of attention of this paper. The spin-current for many-particle systems of charged spinning particles appears in the QHD equations [@Andreev; @IJMP; @12], [@MaksimovTMP; @2001]. Usually we need to find approximate relation between the spin-current and other hydrodynamic variables to get closed set of the QHD equations. In this paper we go further, we derive an equation for the spin-current evolution by means of the QHD method. Recently, an analogous equation was derived for graphene electrons which have also been obtained by the QHD method [@Andreev; @arxiv; @12; @01]. The spin-current definition and corresponding equations appears in the QHD in the semi-relativistic approximation. The spin-current have been universally defined according to the quantum electrodynamics in Ref. [@An; @Sci; @Rep; @12].
The first step in the development of the QHD was made after the Schrodinger equation had been suggested. E. Madelung represented the Schrodinger equation for the one charged particle in an external field as a set of two equations [@Madelung; @26]. These equations are the continuity equation and the Euler equation for the eddy-free motion. Later T. Takabayasi considered the QHD representation of the Pauli equation [@Takabayasi; @PTP; @55].
It is well-known that the one-particle Pauli equation can be represented as the set of the QHD equations. This set of equations is almost equivalent to the first three equations of the chain of the many-particle QHD equations taken in the self-consistent field approximation. The differences are the follows. One-particle equations do not contain the thermal pressure and the thermal contribution in the spin-current. Hence, the one-particle approximation may be used in semiconductors at low temperatures. However, the hydrodynamical equations should be coupled with the set of Maxwell equations.
If we derive the QHD equations from the Pauli equation for the one-particle in an external field we find three equation of material field evolution mentioned above. However, if we have deal with many-particle system the set of the QHD equations contains some new functions, for example the kinetic pressure $p^{\alpha\beta}$ caused by the thermal motion of particles, and the spin-current $J^{\alpha\beta}$. Two-particle macroscopic functions also appear in the terms describing inter-particle interaction. In the self-consistent field approximation, containing no contribution of the exchange interaction, the terms describing interaction have same form as the terms caused by the external fields.
To continue the comparison of the many-particle QHD with the one-particle one we admit that in the one-particle case $p^{\alpha\beta}=0$ and $J^{\alpha\beta}=M^{\alpha}v^{\beta}$, where $M^{\alpha}$ is the density of magnetic moments, and $v^{\beta}$ is the velocity field. We can also notice that in the one-particle case the kinetic energy field $\varepsilon$ easily related with the particle concentration $n$, magnetization $M^{\alpha}$ and velocity field $v^{\beta}$, so we have $\varepsilon=mnv^{2}/2$. For the many-particle system we have that $p^{\alpha\beta}$, $J^{\alpha\beta}$ and $\varepsilon$ are independent material fields additional to $n$, $v^{\alpha}$, and $M^{\alpha}$. In the many-particle system $J^{\alpha\beta}$ and $\varepsilon$ are partly connected with $n$, $v^{\alpha}$, and $M^{\alpha}$. Let us make an example to describe the last statement. For the energy we have $\varepsilon=mnv^{2}/2+n\epsilon$, where the last term corresponds to internal energy caused by the thermal motion of particles [@MaksimovTMP; @1999]. A set of QHD equations including the energy evolution equation and the non-zero thermal pressure in the Euler equation can be derived from a quantum kinetic equation, but derivation of the quantum kinetic equation needs some additional assumption to construct a distribution function [@Balescu; @book; @Eq; @Stat; @mech], [@Wigner; @PR; @84]. Thus, to make more detailed study of magnetic moment dynamics we are going to present the straightforward derivation of equation for the magnetic moment current (or the spin current) $J^{\alpha\beta}$ evolution from the many-particle Schrodinger and study its influence on the spin wave dispersion. During several last decades a lot of different ways of the quantum kinetic equation derivation were suggested, but there are a lot of open questions in this field. Derivation of a kinetic equation via the Wigner distribution function is the most popular and actively used in recent publications. Moreover we keep developing the method of many-particle quantum hydrodynamics which is more direct way of derivation of equations for collective quantum dynamics [@Andreev; @IJMP; @12], [@Andreev; @Qkin; @12], [@Andreev; @Asenjo; @13].
Waves in systems of neutral and charged spinning particles have been considered by means of many-particle quantum hydrodynamics with no account of the spin-current equation. It has been assumed that the spin-current $J^{\alpha\beta}$ appearing in the magnetic moment evolution equation can be approximately considered as $J^{\alpha\beta}=M^{\alpha}v^{\beta}$. In this paper studying waves in the magnetized dielectrics we consider two kind of equilibrium states. One of them corresponds to the case when the equilibrium spin current $J^{\alpha\beta}$ equals to zero, in second case we consider a non zero equilibrium spin current $J^{\alpha\beta}$, but we suppose that the equilibrium velocity field equals to zero. Such structure might be realized by means two currents (flows of neutral particles) directed in opposite directions and having opposite equilibrium spin.
We also need to accent the fact that the many-particle QHD method has been used for the different physical systems, thus the sets of the QHD equations have been obtained for graphene [@Andreev; @arxiv; @12; @01], the neutral ultracold quantum gases [@Andreev; @arxiv; @12; @transv; @dip; @BEC], the Bose-Einstein condensate of excitons in graphene [@Andreev; @arxiv; @12; @GEBEC], along with the physics of plasma described above.
Presented here results are also important for the physics of magnetized ultracold quantum gases. Used where models are equivalent to the first three equations of the QHD, they are the continuity equation, the Euler equation, and the Bloch equation [@Machida; @JPSJ; @98], [@Ho; @PRL; @98].
This paper is organized as follows. In Sec. II we present and describe the set of the QHD equations derives in the paper. In Sec. III dispersion of the spin waves is considered, a contribution of the spin-current in the spin wave properties is studied. In Sec. IV brief summary of obtained results are presented.
II. The model
=============
The many-particle QHD equations are derived from the microscopic many-particle Schrodinger equation $\imath\hbar\partial_{t}\Psi(R,t)=\hat{H}\Psi(R,t)$, where $\Psi(R,t)$ is the wave function of $N$ interacting particles. $\Psi(R,t)$ depends on coordinate of all particles. We present it shortly by means $R$, which is $R=(\textbf{r}_{1},\textbf{r}_{2},...,\textbf{r}_{N})$, where $\textbf{r}_{i}$ is coordinate of $i$ th particle. The structure of the QHD equations depends on the explicit form of the Hamiltonian of considered system of particles. We do not described here the method of derivation of the QHD equations, a lot of paper are dedicated to this topic [@MaksimovTMP; @1999]- [@Andreev; @arxiv; @11; @3]. However, to be certain we present the Hamiltonian $$\hat{H}=\sum_{p}\biggl(\frac{1}{2m_{p}}\textbf{D}^{2}_{p}+e_{p}\varphi^{ext}_{p}-\gamma_{p}\sigma^{\alpha}_{p}B^{\alpha}_{p(ext)}\biggr)$$ $$\label{SC ham gen}+\frac{1}{2}\sum_{p,n\neq p}(e_{p}e_{n}G_{pn}-\gamma_{p}\gamma_{n}G^{\alpha\beta}_{pn}\sigma^{\alpha}_{p}\sigma^{\beta}_{n})$$ used for derivation of equations presented below, where $D_{p}^{\alpha}=-\imath\hbar\partial_{p}^{\alpha}-e_{p}A_{p,ext}^{\alpha}/c$, $\varphi_{p,ext}$, $A_{p,ext}^{\alpha}$ are the potentials of the external electromagnetic field, $\partial_{p}^{\alpha}=\nabla_{p}^{\alpha}$ is the derivatives on space variables, and $G_{pn}=1/r_{pn}$ is the Green functions of the Coulomb interaction, $G^{\alpha\beta}_{pn}=4\pi\delta^{\alpha\beta}\delta(\textbf{r}_{pn})+\partial^{\alpha}_{p}\partial^{\beta}_{p}(1/r_{pn})$ is the Green function of spin-spin interaction, $\gamma_{p}$ is the gyromagnetic ratio, $\sigma^{\alpha}_{p}$ is the Pauli matrix, a commutation relations for them is $$[\sigma^{\alpha}_{p},\sigma^{\beta}_{n}]=2\imath\delta_{pn}\varepsilon^{\alpha\beta\gamma}\sigma^{\gamma}_{p},$$ $e_{n}$, $m_{n}$ are the charge and the mass of particle, $\hbar$ is the Planck constant and $c$ is the speed of light. For electrons $\gamma_{p}$ reads $\gamma_{p}=e_{p}\hbar/(2m_{p}c)$, $e_{p}=-|e|$.
This Hamiltonian (\[SC ham gen\]) consists of two parts. The first of them is presented by the first three terms, which describe motion of independent particle in an external electromagnetic field. The first term is the kinetic energy, which contains vector potential via covariant derivative $\textbf{D}_{n}$. Therefore, it contains action of an external magnetic field and a rotational electric field on particle charge. The second term in formula (\[SC ham gen\]) describes interaction of particle charge with an external potential electric field. And the third term presents the action of an external magnetic field on the magnetic moments. The second group of terms consists of the two last terms. They describe the interparticle interaction. In this paper we consider the Coulomb and the spin-spin interactions presented by the fourth and fifth terms correspondingly.
Using explicit form of the Hamiltonian, we obtain the chain of equations, we truncate the chain of the QHD equations including the four equations only. These are the evolution equations for the particle concentration $n$, the velocity field $v^{\beta}$, the density of magnetic moment or spin $M^{\alpha}$, and the spin-current $J^{\alpha\beta}$.
The first step in derivation of the many-particle QHD equations is the definition of particle concentration. Which is the first collective quantum mechanical observable in our model. The particle concentration is the quantum mechanical average of the microscopic concentration $$\label{SC concentration}n(\textbf{r},t)=\int \Psi^{+}(R,t)\sum_{p}\delta(\textbf{r}-\textbf{r}_{p})\Psi(R,t) dR,$$ where we integrate over the 3N dimensional configurational space, and $dR=\prod_{p=1}^{N}d\textbf{r}_{p}$. The formula (\[SC concentration\]) is more than the first collective quantum mechanical observable. Using of the particle concentration operator $\hat{n}=\sum_{p}\delta(\textbf{r}-\textbf{r}_{p})$ gives the projector of the 3N dimensional configurational space in the three dimensional physical space. Waves propagation, the charge-current flow, the spin-current flow happen in the three dimensional physical space. Consequently it is worthwhile to have a model, which explicitly describes the dynamic of quantum many-particle system in the physical space. The QHD is an example of such model.
We now differentiate the particle concentration with respect to time and find the continuity equation, where the particles current $\textbf{j}=n\textbf{v}$ emerges.
Hence, the first equation of the QHD set of equations is the continuity equation $$\label{SC continuity equation}\partial_{t}n+\nabla (n\textbf{v})=0.$$
At derivation of the continuity equation (\[SC continuity equation\]) the explicit form of the particles current appears as $$\label{SC current}\textbf{j}=\int \sum_{p}\delta(\textbf{r}-\textbf{r}_{p})\frac{1}{2m_{p}}\biggl(\Psi^{+}(R,t)\textbf{D}_{p}\Psi(R,t)+h.c.\biggr) dR,$$ where h.c. means the hermitian conjugation.
Differentiating the function of current with respect to time, we obtain the momentum balance equations, this equation is an analog of the Euler equation $$mn(\partial_{t}+\textbf{v}\nabla)v^{\alpha}+\partial_{\beta}p^{\alpha\beta}$$ $$-\frac{\hbar^{2}}{4m}\partial^{\alpha}\triangle
n+\frac{\hbar^{2}}{4m}\partial^{\beta}\Biggl(\frac{\partial^{\alpha}n\cdot\partial^{\beta}n}{n}\Biggr)$$ $$\label{SC momentum balance eq}=enE^{\alpha}+\frac{e}{c}\varepsilon^{\alpha\beta\gamma}nv^{\beta}B^{\gamma}+M^{\beta}\nabla^{\alpha}B^{\beta},$$ where $\textbf{E}$ and $\textbf{B}$ are the electric and magnetic fields, $\textbf{M}$ is the density of magnetic moments, $\varepsilon^{\alpha\beta\gamma}$ is the antisymmetric symbol (the Levi-Civita symbol), $p^{\alpha\beta}$ is the kinetic pressure tensor. The momentum balance equation (\[SC momentum balance eq\]) has usual form, we see that evolution of the velocity field caused by momentum current on thermal velocities $p^{\alpha\beta}$, the quantum Bohm potential specific for quantum kinematics (two terms in the left-hand side of equation (\[SC momentum balance eq\]), which proportional to $\hbar^{2}$), and interaction, which is presented in the right-hand side of the momentum balance equation (\[SC momentum balance eq\]). We derive this equation for charged spinning particles, thus the force field contains the density of the Lorentz force describing action of electromagnetic field on charges presented by two first terms and force acting on the magnetic moment density from the magnetic field presented by the last term. We present the force field in the self-consistent field approximation. General form of the force field appearing in the Euler equation and introducing of the self-consistent field approximation are presented in Appendix A.
The second-fifth terms in the left-hand side of equation (\[SC momentum balance eq\]) appear due to representation of the momentum flux $\Pi^{\alpha\beta}$. At derivation of the Euler equation the momentum flux $\Pi^{\alpha\beta}$ emerges in the following explicit form $$\Pi^{\alpha\beta}=\int \sum_{p}\delta(\textbf{r}-\textbf{r}_{p})\frac{1}{2m_{p}}\times$$ $$\times\biggl(\Psi^{*}(R,t)\hat{D}_{p}^{\beta}\hat{D}_{p}^{\alpha}\Psi(R,t)+h.c.\biggr) dR.$$ To represent the momentum flux via the hydrodynamic variables we need to introduce the velocity of a quantum particle $v_{i}^{\alpha}(R,t)$. It appears via the phase $S(R,t)$ of the many-particle wave function $\Psi(R,t)=a(R,t)e^{\imath S(R,t)}$. The velocity of i th particle is $v_{i}^{\alpha}(R,t)=\frac{\hbar}{m}\nabla_{i}S(R,t)$. We can also introduce the thermal velocity of i th particle $u_{i}^{\alpha}=v_{i}^{\alpha}(R,t)-v^{\alpha}(\textbf{r},t)$ as difference of the velocity of i th particle and the the center of mass velocity (the velocity field) $\textbf{v}(\textbf{r},t)=\textbf{j}(\textbf{r},t)/n(\textbf{r},t)$ (for details see Refs. [@MaksimovTMP; @1999], [@Andreev; @arxiv; @11; @3]).
The definition of magnetization $$\label{SC magnetization}M^{\alpha}(\textbf{r},t)=\int \sum_{p}\delta(\textbf{r}-\textbf{r}_{p})\Psi^{+}(R,t)\widehat{\sigma}^{\alpha}\Psi(R,t) dR,$$ appears at derivation of the Euler equation (\[SC momentum balance eq\])
The equation of evolution of the magnetic moments $$\label{SC magn mom balance eq}\partial_{t}M^{\alpha}+\nabla^{\beta}J^{\alpha\beta}=\frac{2\gamma}{\hbar}\varepsilon^{\alpha\beta\gamma}M^{\beta}B^{\gamma}$$ is derived at differentiating of the magnetization with respect to time and using of the Schrodinger equation for the time derivatives of the wave function. This equation is a generalization of the Bloch equation. From equation (\[SC magn mom balance eq\]) we see that evolution of magnetic moment density caused by both the spin current $J^{\alpha\beta}$ and interaction of the magnetic moments with the magnetic field. Charge of particles gives no interference in dynamics of the magnetic moment density.
The explicit form of spin-current in the many-particle system is $$J^{\alpha\beta}=\int \sum_{p}\delta(\textbf{r}-\textbf{r}_{p})\frac{1}{2m_{p}}\times$$ $$\label{SC spin current}\times\biggl(\Psi^{+}(R,t)\hat{D}_{p}^{\beta}\widehat{\sigma}_{p}^{\alpha}\Psi(R,t)+h.c.\biggr) dR.$$
Next equation is the equation of spin-current evolution. The spin-current appears in the Bloch equation, and if we include the spin-orbit and the spin-current interaction we get that the spin-current gives contribution in the force field in the Euler equation (\[SC momentum balance eq\]), for example see Ref. [@Andreev; @IJMP; @12]. The spin-current evolution equation for system of neutral particles is $$\partial_{t}J^{\alpha\beta}+\partial^{\gamma}(J^{\alpha\beta}v^{\gamma})$$ $$\label{SC magn mom current balance eq}=\frac{\gamma^{2}}{m}n\partial^{\beta}B^{\alpha}-\frac{2\gamma}{\hbar}\varepsilon^{\alpha\gamma\delta}B^{\gamma}J^{\delta\beta}.$$ for flux of the spin-current $J^{\alpha\beta\gamma}$ which emerges in the second term in the left-hand side of equation (\[SC magn mom current balance eq\]). Its explicit form is $$J^{\alpha\beta\gamma}=\int \sum_{p}\delta(\textbf{r}-\textbf{r}_{p})\frac{1}{4m_{p}^{2}}\times$$ $$\label{SC spin current flux} \times\biggl(\Psi^{+}(R,t)\hat{D}_{p}^{\gamma}\hat{D}_{p}^{\beta}\widehat{\sigma}_{p}^{\alpha}\Psi(R,t)+h.c.\biggr) dR.$$ Definition of the flux of spin-current contains two operators of the long derivative $\hat{D}_{p}^{\alpha}$ when the spin-current contains one operator of the long derivative. Since $J^{\alpha\beta\gamma}$ is the flux of the spin-current we use an approximate formula $J^{\alpha\beta\gamma}=J^{\alpha\beta}v^{\gamma}$. In general case $J^{\alpha\beta\gamma}$ has more complex structure and contains additional contribution of both the thermal motion and quantum kinematics as the quantum Bohm potential. The last one shows in the form of a term analogous to the quantum Bohm potential. We do not consider these contributions and pay attention to the spin current evolution caused by inter-particle interaction. We should pay special attention to equation (\[SC magn mom current balance eq\]) because all this paper is dedicated to consideration of the influence of the spin-current evolution on dynamics of particles system. Equation (\[SC magn mom current balance eq\]) is presented for the chargeless spinning particles, and this form will be used in the paper. However, we now present it for the charged spinning particles $$\partial_{t}J^{\alpha\beta}+\partial^{\gamma}(J^{\alpha\beta}v^{\gamma})=\frac{e}{m}M^{\alpha}E^{\beta}$$ $$\label{SC magn mom current balance eq charge}+\frac{e}{mc}\gamma\varepsilon^{\beta\gamma\delta}J^{\alpha\gamma}B^{\delta}+\frac{\gamma^{2}}{m}n\partial^{\beta}B^{\alpha}-\frac{2\gamma}{\hbar}\varepsilon^{\alpha\gamma\delta}B^{\gamma}J^{\delta\beta}.$$
Equations (\[SC continuity equation\])-(\[SC magn mom current balance eq\]) take place for each species of particles. The electric $\textbf{E}$ and magnetic $\textbf{B}$ fields appearing in the equations (\[SC continuity equation\])-(\[SC magn mom current balance eq\]) are caused by charges, electric currents, and magnetic moments of medium and satisfy to the Maxwell equations. Thus the QHD equations for each species of particles connect by means of the Maxwell equations $$\begin{array}{ccc} \nabla\textbf{B}=0 ,& \nabla\textbf{E}=4\pi\sum_{a}e_{a}n_{a} \end{array},$$ $$\nabla\times\textbf{E}=0,$$ $$\label{SC Maxwell eq}\nabla\times\textbf{B}=\frac{4\pi}{c}\sum_{a}e_{a}n_{a}\textbf{v}_{a}+4\pi \sum_{a}\nabla\times\textbf{M}_{a},$$ where subindex “a” describes the species of particles. The Maxwell equations (\[SC Maxwell eq\]) presented here do not contain time derivatives of the electric and the magnetic fields, because we have derived the QHD equations from the non-relativistic theory.
III. Dispersion equation
========================
We consider the collective eigen-waves in a system of the neutral spinning particles being in an external uniform magnetic field. Hence, we have deal with the paramagnetic and diamagnetic dielectrics. There are the two fundamental collective excitations in such physical systems, they are the sound waves and the spin waves. Following the QHD description of the three dimensional magnetized dielectrics [@Andreev; @PIERS; @2012; @Spin] one can show that there is one type of the spin waves in such systems. These waves have a constant eigen-frequency $\omega=2\gamma B_{0}/\hbar$, which is the cyclotron frequency, where $B_{0}$ is an external magnetic field. Here we have no dependency on wave vector, consequently the group velocity $\partial\omega/\partial k$ of these waves equal to zero.
We are interested in an interference of the spin-current evolution on the dispersion properties of the medium. To find the dispersion dependence of eigen-waves in the described systems we consider small amplitude excitations around an equilibrium state of the medium. We consider two different equilibrium states. In the first case we suppose that an equilibrium spin-current equals to zero, and in the second case we consider a medium with an equilibrium spin-current under condition that an equilibrium velocity field equals to zero.
For getting of solution we consider hydrodynamic variables as the sum of an equilibrium part and a small perturbation $$\begin{array}{ccc}n=n_{0}+\delta n, &\textbf{E}=0+\textbf{E} \end{array},$$ $$\begin{array}{ccc}\textbf{B}=B_{0}\textbf{e}_{z}+\delta \textbf{B},&
\textbf{v}=0+\delta\textbf{v} \end{array},$$ $$\begin{array}{ccc}M^{\alpha}=M_{0}^{\alpha}+\delta M^{\alpha},& M_{0}^{\alpha}=\chi
B_{0}^{\alpha},& J^{\alpha\beta}=J^{\alpha\beta}_{0}+\delta J^{\alpha\beta},\end{array}$$ $$\label{SC linearisation}\begin{array}{ccc}& p^{\alpha\beta}=p\delta^{\alpha\beta}, &
\delta p=mv_{F}^{2}\delta n_{a},
\end{array}$$ where $\delta^{\alpha\beta}$ is the Kronecker symbol, $v_{F}$ is the Fermi velocity, $\chi_{a}=\kappa_{a}/\nu_{a}$ is the ratio between the equilibrium magnetic susceptibility $\kappa_{a}$ and the magnetic permeability $\nu_{a}=1+4\pi\kappa_{a}$. In the case $\kappa_{a}\ll 1$ we have $\chi_{a}\simeq\kappa_{a}$. Substituting relations (\[SC linearisation\]) in the set of equations (\[SC continuity equation\]), (\[SC momentum balance eq\]), (\[SC magn mom balance eq\]) and (\[SC Maxwell eq\]) and neglecting by the nonlinear terms, we obtain a system of linear homogeneous equations in the partial derivatives with constant coefficients. Passing to the following representation for the small perturbations $\delta f$ $$\delta f =f(\omega, k) exp(-\imath\omega t+\imath k x)$$ yields a homogeneous system of algebraic equations.
Even when we consider an equilibrium spin-current the linear set of QHD equations splits on four independent sets. One of them contains $n$, $\delta v_{x}$, $\delta M_{z}$, $J_{zx}$, the second one contains $\delta M_{x}$, $\delta M_{y}$, $\delta J_{xx}$, $\delta J_{yx}$, the third set includes $\delta J_{xy}$, $\delta
J_{yy}$, and fourth includes $\delta J_{xz}$, $\delta J_{yz}$. Two last sets have the same solution, which is the cyclotron frequency $\omega=2\gamma B_{0}/\hbar$. Hence, we have to consider solutions of the first and second sets.
A. The first group of dispersion branches
-----------------------------------------
The first set give the following dispersion equation $$\omega^{4}+\omega^{2}\biggl(\frac{4\pi n_{0}\gamma^{2}}{m}-\upsilon^{2}\biggr)k^{2}$$ $$\label{SC disp curr wave}+\frac{4\pi k^{3}M_{0}}{mn_{0}}J_{0}^{zx}\omega-\frac{4\pi n_{0}\gamma^{2}}{m}\upsilon^{2}k^{4}=0,$$ where $$\label{SC quantum vel}\upsilon^{2}=v_{F}^{2}+\frac{\hbar^{2}k^{2}}{4m^{2}},$$ and $$v_{F}=(3\pi^{2}n)^{1/3}\frac{\hbar}{m}.$$ We can see that a non-zero equilibrium spin-current leads to existence of the additional term in the dispersion equation.
In the absence of an equilibrium spin-current we have two solutions of the equation (\[SC disp curr wave\]) $$\label{SC new disp sol damp}\omega^{2}=-4\pi\gamma^{2}n_{0}k^{2}/m,$$ and $$\label{SC sound}\omega^{2}=\upsilon^{2}k^{2},$$ where solution (\[SC sound\]) is the well-known sound wave. Dispersion of the sound wave consists of two parts (\[SC quantum vel\]). The first of them is the usual linear term, which appears due to the Fermi pressure. The second term is the contribution of the quantum Bohm potential giving dispersion of the de Broglie wave. Moreover the evolution of spin-current leads to existence of the second solution (\[SC new disp sol damp\]). This solution can be rewritten as $$\label{SC new disp sol damp lin fr}\omega=\pm\imath\sqrt{\frac{4\pi n_{0}}{m}}\gamma k.$$ It shows faster damping and gives no wave behavior.
The sound wave solution (\[SC sound\]) and spin wave $\omega=2\gamma B_{0}/\hbar$ appear in the simpler model without account of the spin-current evolution. In this case a system of neutral spinning particles can be described by the continuity, Euler, and magnetic moment evolution equations, where we can put $J^{\alpha\beta}=M^{\alpha}v^{\beta}$ to close the set of equations. It has been mostly used for systems of charged spinning particles [@Andreev; @VestnMSU; @2007], [@Andreev; @IJMP; @12], [@Maksimov; @VestnMSU; @2000].
Presence of the equilibrium spin-current in the equation (\[SC disp curr wave\]) makes it rather complicate. So, we are going to solve it numerically. It seems reasonable to chouse $\Omega\equiv\omega/\lambda=\sqrt{m/(\pi n_{0})}\omega/(2\gamma k)$ as dimensionless frequency, where we introduced $$\label{SC lambda}\lambda^{2}=\frac{4\pi\gamma^{2}n_{0}}{m}k^{2}.$$ In this case we get equation (\[SC disp curr wave\]) in the form of $$\label{SC disp curr wave dimless} \Omega^{4}+(1-\alpha)\Omega^{2}+\beta\Omega-\alpha=0,$$ where $\alpha=m\upsilon^{2}/(4\pi n_{0}\gamma^{2})$ is the parameter describing contribution of the quantum Bohm potential, $\beta=kM_{0}J_{0}^{zx}/(\gamma^{2}n_{0}^{2})$ is the parameter describing contribution of the equilibrium spin-current. On figures we present a limit case: de-Broglie regime when $\alpha\simeq\alpha_{\hbar}=\hbar^{2}k^{2}/(16\pi mn_{0}\gamma^{2})$. The both parameters $\alpha$ and $\beta$ depend on module of the wave vector $k$. Consequently equation (\[SC disp curr wave dimless\]) allows to get $\Omega(k)$ dependence. Equation (\[SC disp curr wave dimless\]) gives two solutions. One of them is stable solution, which dispersion presented on Fig. (\[SC waves 1\]). The second solution of equation (\[SC disp curr wave dimless\]) shows an instability, which exists due to the existence of the equilibrium spin-current $J_{0}^{zx}$. This solution is presented on Fig. (\[SC waves 2\]).
Solutions (\[SC new disp sol damp\]) and (\[SC sound\]) can be briefly written in terms of reduced frequency $\Omega$ and $\alpha$, so we have $\Omega_{1}^{2}=-1$ and $\Omega_{2}^{2}=\alpha$.
![\[SC waves 1\] (Color online) The figure describes the dispersion dependence of the stable collective excitation described by equation (\[SC disp curr wave dimless\]). This figure is obtained in the de-Broglie regime. We see that $\Omega$ approximately proportional to $k$. So this figure reveals quadratic dependence of the frequency $\omega$ of wave vector $k$.](SCnewwsol01.eps){width="8cm"}
![\[SC waves 2\] (Color online) The figure describes the second solution of equation (\[SC disp curr wave dimless\]) in the de-Broglie regime. This solution is unstable.) The dispersion dependence and corresponding instability increment of the collective excitation are presented.](SCnewwsol02withinst.eps){width="8cm"}
Let us consider approximate solutions of equation (\[SC disp curr wave dimless\]) at nonzero equilibrium spin-current under assumption that $\beta$ gives small contribution in dispersion of waves (\[SC sound\]) and (\[SC new disp sol damp lin fr\]). Thus solutions of equation (\[SC disp curr wave dimless\]) emerge as $$\label{SC big disp first group sound}\omega=\upsilon k+\zeta_{1}$$ and $$\label{SC big disp first group damp}\omega=\pm\imath\sqrt{\frac{4\pi n_{0}}{m}}\gamma k+\zeta_{2},$$ where $$\zeta_{i}=\frac{\beta\omega_{i}}{2\omega_{i}(2(\omega_{i}/\lambda)^{2}+1-\alpha)+\lambda\beta}$$ $$\label{SC zeta} \simeq\frac{\beta}{2(2(\omega_{i}/\lambda)^{2}+1-\alpha)}.$$ $\omega_{i}$ is used for short representation of solutions (\[SC sound\]) and (\[SC new disp sol damp lin fr\]). Including $\alpha\lambda^{2}=\upsilon k$ we come to the following dispersion relation for the sound wave $$\omega=\upsilon k+ \frac{\beta\lambda^{2}}{2(\upsilon^{2}k^{2}+\lambda^{2})}\equiv U_{1}^{*}k,$$ where we introduced a modified sound velocity $U_{1}^{*}$, which slightly depends on the wave vector $k$ via the quantum Bohm potential. $$\label{SC sound vel new} U_{1}^{*}=\upsilon+\frac{2\pi\gamma^{2}n_{0}(\beta/k)}{m\upsilon^{2}+4\pi\gamma^{2}n_{0}}.$$ Since $\beta >0$ and $\beta\sim k$ we see that $U_{1}^{*} > \upsilon$.
Contribution of the spin-current in unstable branch (\[SC big disp first group damp\]) in considering limit $$\omega=\pm\imath\lambda-\frac{\beta\lambda^{2}}{2(\upsilon^{2}k^{2}+\lambda^{2})}$$ $$\label{SC big disp first group damp explicit}= \pm\imath\sqrt{\frac{4\pi n_{0}}{m}}\gamma k -\frac{2\pi\gamma^{2}n_{0}\beta}{m\upsilon^{2}+4\pi\gamma^{2}n_{0}}$$ appears to be real and negative. We see that real part of solution (\[SC big disp first group damp explicit\]) is almost a linear function of the wave vector $k$, but it contains small additional dependence on the wave vector via $\upsilon$ containing contribution of the quantum Bohm potential. This results differs from previously considered case depicted on Fig.2. Formula (\[SC big disp first group damp explicit\]) is obtained in the limit of small contribution of the equilibrium spin-current. When Fig.2 is obtained for a finite value of the equilibrium spin-current, so it reveals an interesting dependence of real and imaginary parts of the frequency on the wave vector discussed above.
B. Wave dispersion for the second group of variables
----------------------------------------------------
The second set of equations containing evolution of $\delta
M_{x}$, $\delta M_{y}$, $\delta J_{xx}$, $\delta J_{yx}$ gives the following dispersion equation $$\omega^{2}+8\pi\frac{\Omega_{\gamma}\gamma}{\hbar\omega}k J^{zx}_{0}$$ $$\label{SC disp eq for MMJJ}+4\pi\frac{k}{\omega}(\omega^{2}+\Omega^{2}_{\gamma})\frac{\omega k n_{0}\gamma^{2}\hbar+\Omega_{\gamma}2\gamma m J^{zx}_{0}}{m\hbar(\omega^{2}-\Omega^{2}_{\gamma})}-\Omega_{\gamma}^{2}(1-4\pi\xi)=0,$$ where $\Omega_{\gamma}=\frac{2\gamma B_{0}}{\hbar}$ and $\xi=M_{0}/B_{0}$. For the ferromagnetic samples $\xi$ is larger than one $\xi\gg 1$, while $\xi$ is rather small $\xi\ll 1$ for the para- and the dia-magnetics.
In the absence of the equilibrium spin-current this equation is simplified to $$\label{SC disp eq for MMJJ simp}\omega^{4}+\biggl(\lambda^{2}-2\Omega^{2}_{\gamma}(1-2\pi\xi)\biggr)\omega^{2}+\Omega^{2}_{\gamma}\biggl(\lambda^{2}+\Omega^{2}_{\gamma}(1-4\pi\xi)\biggr)=0.$$ Solving this equation we get $$\omega^{2}=\Omega^{2}_{\gamma}(1-2\pi\xi)-\frac{1}{2}\lambda^{2}$$ $$\label{SC spin new waves gen} \pm\sqrt{\frac{1}{4}\lambda^{4}-2\lambda^{2}\Omega^{2}_{\gamma}(1-\pi\xi)+16\pi^{2}\xi^{2}\Omega^{4}_{\gamma}}.$$ These are new solutions. Their occurrence is caused by the account of the spin-current evolution, so we called them spin-current waves.
In the small $k$ limit we come to $$\label{SC spin new waves small k} \omega^{2}=\Omega^{2}_{\gamma}\left(\begin{array}{ccc}1+2\pi\xi & &\\ 1-6\pi\xi & &\\\end{array}\right)+\lambda^{2}\left(\begin{array}{ccc} -\frac{1}{4}-\frac{1}{4\pi\xi} & &\\ -\frac{3}{4}+\frac{1}{4\pi\xi} & &\\\end{array}\right).$$ In the paramagnetic limit formula (\[SC disp eq for MMJJ simp\]) gives $$\label{SC spin new waves param lim} \omega_{\xi}^{2}=\Omega^{2}_{\gamma}-\frac{1}{2}\lambda^{2}\pm\lambda\sqrt{\frac{1}{4}\lambda^{2}-2\Omega^{2}_{\gamma}}.$$ At large $k$ at small magnetic field ($\lambda^{2}\gg \Omega^{2}_{\gamma}$) only one solution exists, which corresponds to the sign plus in front of the square root in formula (\[SC spin new waves gen\]). This solution appears as $$\label{SC } \omega_{\lambda, +}=\Omega_{\gamma}\sqrt{16\pi^{2}\xi^{2}\frac{\Omega^{2}_{\gamma}}{\lambda^{2}}-1},$$ since $\frac{\Omega^{2}_{\gamma}}{\lambda^{2}}\ll 1$ we have that this solution exists at $\xi >\frac{\lambda}{4\pi\mid\Omega_{\gamma}\mid}\gg 1$. It corresponds to large magnetization $M_{0}$.
Equation (\[SC disp eq for MMJJ\]) can be rewritten as an equation of fifth degree. One of its solution is $\omega=0$. So we have equation of fourth degree $$\omega^{4}+\biggl(\lambda^{2}-2\Omega_{\gamma}^{2}(1-2\pi\xi)\biggr)\omega^{2}$$ $$\label{SC disp eq for MMJJ reduced}+\vartheta\omega+\Omega_{\gamma}^{2}\biggl(\Omega_{\gamma}^{2}(1-4\pi\xi)+\lambda^{2}\biggr)=0,$$ where $$\vartheta=16\pi\gamma\Omega_{\gamma}J_{0}^{zx}k/\hbar.$$ This equation (\[SC disp eq for MMJJ reduced\]) differs from (\[SC disp eq for MMJJ simp\]) by one term only. It is the third term $\vartheta\omega$. Thus we have changing of solutions (\[SC spin new waves gen\]) caused by the equilibrium spin-current $\vartheta\sim J_{0}^{zx}$. Solutions of equation (\[SC disp eq for MMJJ reduced\]) we calculated approximately under assumption that the equilibrium spin-current gives small contribution in dispersion dependence. Designating solutions of equation (\[SC disp eq for MMJJ simp\]) presented by formula (\[SC spin new waves gen\]) as $\varpi$. Then solutions of equation (\[SC disp eq for MMJJ reduced\]) appear as $$\omega=\varpi+\frac{\vartheta\varpi}{2\varpi(2\varpi^{2}+\lambda^{2}-2\Omega_{\gamma}^{2}(1-2\pi\xi))+\vartheta}$$ $$\label{SC disp eq for MMJJ appr fprm 1}\simeq\varpi+\frac{\vartheta}{2(2\varpi^{2}+\lambda^{2}-2\Omega_{\gamma}^{2}(1-2\pi\xi))}.$$ Using explicit form of $\varpi$ in the second term of formula (\[SC disp eq for MMJJ appr fprm 1\]) we come to the following formula $$\label{SC disp eq for MMJJ appr fprm 2}\omega=\varpi\pm\frac{\vartheta}{2\sqrt{\lambda^{4}-8\lambda^{2}\Omega_{\gamma}^{2}(1-\pi\xi)+64\pi^{2}\xi^{2}\Omega_{\gamma}^{4}}}.$$ Signs plus and minus in front of the second term of formula (\[SC disp eq for MMJJ appr fprm 2\]) correspond to signs in formula (\[SC spin new waves gen\]).
C. Discussion
-------------
Formulas (\[SC new disp sol damp\]), (\[SC sound\]), (\[SC spin new waves gen\]) and the cyclotron frequency $\omega=2\gamma
B_{0}/\hbar$ appear as solutions of a dispersion equation obtained at account of the spin-current evolution in the absence of the equilibrium spin-current.
Solution (\[SC new disp sol damp\]) reveals two branches (\[SC new disp sol damp lin fr\]). One of them has an increasing amplitude. Another one has a decreasing amplitude. Since we have no source of the energy in the system we have no mechanism for the amplitude increasing. So, we conclude that the decreasing branch takes place in considering case. This decreasing solution gives no contribution in the spectrum as it reveals monotonic decreasing of the amplitude (non oscillating solution). Thus we have got left the two unchanged solutions: the sound wave (\[SC sound\]) and the spin wave $\omega=2\gamma
B_{0}/\hbar$, which can be found with no account of the spin-current evolution. We have also found solutions (\[SC spin new waves gen\]), which present the two spin-current waves reaching the wave spectrum of magnetized dielectrics.
At this step we can conclude that the account of the spin-current evolution, without changing of conditions system being at, we obtained additional information about processes happen in the system. In considering case we have got the two additional wave branches (the spin-current waves) described by formula (\[SC spin new waves gen\]). Formula (\[SC spin new waves small k\]) reveals that at small $k$ and large enough external magnetic field $B_{0}$ the cyclotron frequency $\Omega_{\gamma}=2\gamma B_{0}/\hbar$ gives main contribution in the dispersion of the spin-current waves.
Consideration of the spin-current as an independent physical variable gives us possibility to consider some conditions, which can not be included in more simple model. One of these conditions is the existence of an equilibrium spin-current with the zero equilibrium velocity field $v_{0}=0$. Presence of an equilibrium spin-current leads to changes of solutions (\[SC new disp sol damp\]) and (\[SC sound\]), and to complication of dispersion equation (\[SC disp eq for MMJJ simp\]) (degree of this equation increases, so we get equation (\[SC disp eq for MMJJ\])). However it gives no change in the dispersion of spin waves with the cyclotron frequency $\omega=2\gamma B_{0}/\hbar$.
Presence of an equilibrium spin-current $J_{0}^{zx}$ leads to change of the solutions (\[SC sound\]) and (\[SC new disp sol damp lin fr\]). Real part of solution (\[SC new disp sol damp lin fr\]) appears due to $J_{0}^{zx}$. In the de-Brolie regime it is pictured by lowest curve on Fig. (2). It reveals as a curve with two linear areas, one at small $k$, and the second one at $k\geq 4\cdot 10^{5}$ cm$^{-1}$. They connect smoothly around $k=3\cdot 10^{5}$ cm$^{-1}$. Thus, we can assume $\Omega=\nu_{i}k$, with different $\nu_{i}$ for each area. Consequently, we have $Re\omega=\sqrt{\frac{\pi n_{0}}{m}}2\gamma\nu_{i}k^{2}$. Imaginary part of solution $Im\omega$ is pictured by upper curve on Fig.(2). Its form is similar to the form of the real part of the frequency $Re\omega$, but it has larger value. Limit of small contribution of the equilibrium spin-current allows to obtain some analytical solution (\[SC big disp first group damp explicit\]). In this approximation the equilibrium spin-current $J_{0}^{zx}$ gives rise to appearance of the negative real part of the frequency. It gives no changes in the imaginary part of the solution.
Let us discuss the sound wave. In the absence of an equilibrium spin-current it is described by formula (\[SC sound\]). In the presence of an equilibrium spin-current we numerically consider a limit case: the de-Broglie regime. In the de-Broglie regime we find that the equilibrium spin-current does not change form of the dispersion curve (formula (\[SC sound\]) gives us $\omega=\frac{\hbar k^{2}}{2m}$ in the de-Broglie regime). Thus we see that $\Omega$ linearly depends on the wave vector $k$, and $\omega\sim k^{2}$. The approximation of small equilibrium spin-current lets to trace contribution of the equilibrium spin-current on the sound wave analytically. In the de-Broglie regime an addition to spectrum of the free quantum particles as $$\triangle\omega_{S}=\frac{8\pi mn_{0}\gamma^{2}\beta}{\hbar^{2}k^{2}+16\pi mn_{0}\gamma^{2}}\sim\frac{k}{k^{2}+\chi^{2}},$$ where $\chi$ does not depend on the wave vector $k$. One can see that the frequency shift is positive and decreases at the increasing of the wave vector. In the Fermi regime we find that the sound velocity increases on a constant value defined by formula (\[SC sound vel new\]).
Let us discuss now contribution of the equilibrium spin-current in the dispersion of the spin-current waves. Formula (\[SC disp eq for MMJJ appr fprm 2\]) shows that the spin-current wave having plus (minus) in front of the square root in formula (\[SC spin new waves gen\]) gets positive (negative) contribution of the equilibrium spin-current in the dispersion dependence. So we find an increasing (a decreasing) of the frequency. Including the fact that $\vartheta\sim k$ and $\lambda\sim k$ we have the following dependence of the last term in formula (\[SC disp eq for MMJJ appr fprm 2\]) (a frequency shift caused by the equilibrium spin-current) on the wave vector appears as $$\triangle\omega_{J}=\frac{8\sqrt{\pi m}\Omega_{\gamma}J_{0}^{zx}}{\hbar\sqrt{n_{0}}}\frac{k}{\sqrt{k^{4}-\tilde{a}k^{2}+\tilde{b}}},$$ where $\tilde{a}=2m(1-\pi\xi)\Omega_{\gamma}^{2}/(n_{0}\gamma^{2})$ and $\tilde{b}=4m^{2}\xi^{2}\Omega_{\gamma}^{4}/(n_{0}^{2}\gamma^{4})$ are positive constants. At $k\rightarrow 0$ we obtain $\triangle\omega_{J}\rightarrow 0$. $\triangle\omega_{J}$ increases with increasing of the wave vector. However $\triangle\omega_{J}$ reaches its maximum at an intermediate wave vector $k_{0}=\frac{\sqrt{2m\xi}\Omega_{\gamma}}{\gamma\sqrt{n_{0}}}$. This maximum value of the frequency shift is $$\triangle\omega_{J}(k_{0})=\frac{8\sqrt{\pi m}\Omega_{\gamma}J_{0}^{zx}}{\hbar\sqrt{n_{0}}}\frac{1}{\sqrt{2\sqrt{\tilde{b}}-\tilde{a}}}.$$ At large $k$ the frequency shift $\triangle\omega_{J}$ decreases as $1/k$.
IV. Conclusion
==============
To get influence of the spin-current on dynamics of magnetized dielectrics we have derived equation of the spin-current evolution as a part of the set of the QHD equations. In the result we have set of four equations: the continuity equation (particle number evolution equation), the Euler equation (momentum balance equation), the Bloch equation (magnetic moment balance equation), and equation of the spin-current evolution. These equations are used in the self-consistent field approximation.
With no account of the spin-current evolution equation we find two wave solutions: the sound wave and one spin wave solution. Including the spin-current evolution equation leads to account of new solutions. We found three new wave solutions. Two of them have frequencies near the cyclotron frequency. These solutions make spectrum of spin waves richer. It appears as a splitting of one spin wave branch on three branches. The third solution has negative square of the frequency and reveal monotonic damping of perturbation amplitude. These solutions emerge when the equilibrium spin-current equals to zero. Account of an equilibrium spin-current does not give new solution, but a change of wave dispersion was obtained.
Appendix A: General form of the force field and the self-consistent field approximation
=======================================================================================
In the Euler equation (\[SC momentum balance eq\]) the force field $$\label{SC Force SCF 1} F^{\alpha}=enE^{\alpha}+\frac{e}{c}\varepsilon^{\alpha\beta\gamma}nv^{\beta}B^{\gamma}+M^{\beta}\nabla^{\alpha}B^{\beta}$$ is presented in the self-consistent field approximation. Here we are going to present general form of this force field and explain how we got formula (\[SC Force SCF 1\]) from the general formula.
Force field consists of two parts $$\label{SC Force sum} F^{\alpha}=F^{\alpha}_{ext}+F^{\alpha}_{int}.$$ The first of them is the force of the particle interaction with an external field $$\label{SC Force ext} F^{\alpha}_{ext}=enE^{\alpha}_{ext}+\frac{e}{c}\varepsilon^{\alpha\beta\gamma}nv^{\beta}B^{\gamma}_{ext}+M^{\beta}\nabla^{\alpha}B^{\beta}_{ext},$$ and the second part is the inter-particle interactions $$F^{\alpha}_{int}=-e^{2}\int (\nabla^{\alpha} G(\textbf{r}-\textbf{r}')) n_{2}(\textbf{r},\textbf{r}',t) d\textbf{r}'$$ $$\label{SC Force int}
+\int (\nabla^{\alpha} G^{\beta\gamma}(\textbf{r}-\textbf{r}')) M_{2}^{\beta\gamma}(\textbf{r},\textbf{r}',t) d\textbf{r}',$$ where $$\label{SC two part conc} n_{2}(\textbf{r},\textbf{r}',t)=\int \sum_{p,n\neq p} \delta(\textbf{r}-\textbf{r}_{p})\delta(\textbf{r}-\textbf{r}'_{n}) \Psi^{*}(R,t)\Psi(R,t) dR$$ is the two-particle concentration, and $$M_{2}^{\alpha\beta}(\textbf{r},\textbf{r}',t)=\int \sum_{p,n\neq p} \delta(\textbf{r}-\textbf{r}_{p})\delta(\textbf{r}-\textbf{r}'_{n})\times$$ $$\label{SC two part magn} \times\mu_{B}^{2}\Psi^{*}(R,t)\sigma^{\alpha}_{p}\sigma^{\beta}_{n}\Psi(R,t) dR$$ is the two-particle magnetization.
It has been shown that a two-particle function $f_{2}(\textbf{r},\textbf{r}',t)$ (see formulas (\[SC two part conc\]) and (\[SC two part magn\])) appears as a sum of two terms $f_{2}(\textbf{r},\textbf{r}',t)=f(\textbf{r},t)f(\textbf{r}',t)+g_{2}(\textbf{r},\textbf{r}',t)$. The first of the two terms is the product of corresponding one-particle functions. It corresponds to the self-consistent field approach suitable for the long-range interaction. The second term gives contribution of quantum correlations, particularly the exchange correlation.
Thus, in the self-consistent field approximation we have that the two-particle concentration represents as the product of the concentrations in points $\textbf{r}$ and $\textbf{r}'$: $n_{2}(\textbf{r},\textbf{r}',t)=n(\textbf{r},t)n(\textbf{r}',t)$; and the two-particle magnetization gives us the product of the one-particle magnetization $M^{\alpha\beta}_{2}(\textbf{r},\textbf{r}',t)=M^{\alpha}(\textbf{r},t)M^{\beta}(\textbf{r}',t)$. Putting these representations in formula (\[SC Force int\]) we can introduce electric field caused by the electric charges and the magnetic field caused by magnetic moments. These fields emerge as $$E^{\alpha}=-e\nabla^{\alpha}\int G(\textbf{r}-\textbf{r}') n(\textbf{r}',t) d\textbf{r}',$$ and $$B^{\alpha}=\int G^{\alpha\beta}(\textbf{r}-\textbf{r}') M^{\beta}(\textbf{r}',t) d\textbf{r}'.$$ They satisfy the Maxwell equations (\[SC Maxwell eq\]). Using these fields we come to the force field (\[SC Force SCF 1\]).
[99]{}
S. Takahashi, S. Maekawa, J. Magnetism and Magnetic Materials **272**, 1423 (2004).
I. Zutic, J. Fabian, S. Das Sarma, Rev. Mod. Phys. **76**, 323 (2004).
J. Sinova and I. Zutic, Nature Materials **11**, 368 (2012).
K. Uchida, H. Adachi, T. An, H. Nakayama, M. Toda, B. Hillebrands, S. Maekawa, and E. Saitoh, J. Appl. Phys. **111**, 053903 (2012).
F. Pezzoli, F. Bottegoni, D. Trivedi, F. Ciccacci, A. Giorgioni, P. Li, S. Cecchi, E. Grilli, Y. Song, M. Guzzi, H. Dery, and G. Isella, Phys. Rev. Lett. **108**, 156603 (2012).
Z. Rashidian, F. Kheirandish, Int. J. Theor. Phys. **51**, 1989 (2012).
R. T. Hammond, Appl. Phys. Lett. **100**, 121112 (2012).
J. Lee, R. Oszwaldowski, C. Gothgen, and I. Zutic, Phys. Rev. B **85**, 045314 (2012).
Y. Li, M. B. A. Jalil, and S. Ghee Tan, J. Appl. Phys. **111**, 093717 (2012).
Rong Zhang, Long Bai, Chen-Long Duan, Physica B **407**, 2372 (2012).
M. J. Ma, M. B. A. Jalil, and Z. B. Siu, J. Appl. Phys. **111**, 07C326 (2012).
N. T. Bagraev, O. N. Guimbitskaya, L. E. Klyachkin, A. A. Kudryavtsev, A. M. Malyarenko, V. V. Romanov, A. I. Ryskin, I. A. Shelykh, A. S. Shcheulin, Physica C **893**, 470 (2010).
J. Flipse, F. L. Bakker, A. Slachter, F. K. Dejene and B. J. van Wees, Nature Nanotechnology **7**, 166 (2012).
D. Pesin and A. H. MacDonald, Nature Materials **11**, 409 (2012).
I. J. Vera-Marun, V. Ranjan and B. J. van Wees, Nature Physics **8**, 313 (2012).
Wei Han, K. M. Mc Creary, K. Pi , W. H. Wang, Yan Li, H. Wen, J. R. Chen, R. K. Kawakami, J. Magnetism and Magnetic Materials **324**, 369 (2012).
R. Jansen, Nature Materials **11**, 400 (2012).
Z. An, F. Q. Liu, Y. Lin, C. Liu, Sci. Rep. **2**, 388 (2012).
Q. F. Sun, X. C. Xie, Phys. Rev. B **72**, 245305 (2005).
Q. F. Sun, X. C. Xie, J. Wang, Phys. Rev. B **77**, 035327 (2008).
P. Jin, Y. Li, F. Zhang, J. Phys. A **39**, 7115 (2006).
J. Shi, P. Zhang, D. Xiao, Q. Niu, Phys. Rev. Lett. **96**, 076604 (2006).
Farkhad G. Aliev, Juan F. Sierra, Ahmad A. Awad, Gleb N. Kakazei, Dong-Soo Han, Sang-Koog Kim, Vitali Metlushko, Bojan Ilic, and Konstantin Y. Guslienko, Phys. Rev. B **79**, 174433 (2009). A. A. Awad, K. Y. Guslienko, J. F. Sierra, G. N. Kakazei, V. Metlushko, and F. G. Aliev, Appl. Phys. Lett. **96**, 012503 (2010). G. N. Kakazei, P. E. Wigen, K. Yu. Guslienko, V. Novosad, A. N. Slavin, V. O. Golub, N. A. Lesnik, and Y. Otani, Appl. Phys. Lett. **85**, 443 (2004). Ki-Suk Lee, Konstantin Y. Guslienko, Jun-Young Lee, and Sang-Koog Kim, Phys. Rev. B **76**, 174410 (2007). K. Y. Guslienko, R. H. Heredero, O. Chubykalo-Fesenko , Phys. Rev. B **82**, 014402 (2010). P. A. Andreev, L.S. Kuz’menkov, Moscow University Physics Bulletin **62**, N.5, 271 (2007).
L. S. Kuz’menkov and S. G. Maksimov, Teor. i Mat. Fiz., **118** 287 (1999) \[Theoretical and Mathematical Physics **118** 227 (1999)\].
P. A. Andreev, L. S. Kuz’menkov, Phys. Rev. A **78**, 053624 (2008).
M. I. Trukhanova, Int. J. Mod. Phys. B, **26**, 1250004 (2012).
P. A. Andreev, L. S. Kuz’menkov, M. I. Trukhanova, Phys. Rev. B **84**, 245401 (2011).
G. Brodin, M. Marklund, J. Zamanian, B. Ericsson and P. L. Mana, Phys. Rev. Lett. **101**, 245002 (2008).
P. A. Andreev, L. S. Kuz’menkov, Int. J. Mod. Phys. B **26** 1250186 (2012).
P. A. Andreev, L. S. Kuz’menkov, Physics of Atomic Nuclei **71**, N.10, 1724 (2008).
P. A. Andreev and L. S. Kuz’menkov, PIERS Proceedings, p.1047, March 20-23, Marrakesh, MOROCCO 2011.
L. S. Kuz’menkov, S. G. Maksimov, and V. V. Fedoseev, Vestn. Mosk. Univ., Ser. 3: Fiz., Astron., No. 5, 3 (2000) (Moscow Univ. Phys. Bull., No. 5, 1 (2000)\].
M. Marklund and G. Brodin, Phys. Rev. Lett. **98**, 025001 (2007).
S. M. Mahajan and F. A. Asenjo, Phys. Rev. Lett. **107**, 195003 (2011).
F. Haas, G. Manfredi, M. Feix, Phys. Rev. E **62**, 2763(2000).
P. K. Shukla, B. Eliasson, Rev. Mod. Phys. **83**, 885 (2011).
P. A. Andreev and L. S. Kuz’menkov, PIERS Proceedings, p. 1055, august 19-23, Moscow, Russia 2012.
L. S. Kuz’menkov, S. G. Maksimov, and V. V. Fedoseev, Theor. Math. Fiz. **126** 136 (2001) \[Theoretical and Mathematical Physics, **126** 110 (2001)\].
P. A. Andreev, arXiv: 1201.0779.
E. Madelung, Z. Phys., **40**, 332 (1926).
T. Takabayasi, Progr. Theor. Phys., **14**, 283 (1955).
R. Balescu, *Equilibrium and nonequilibrium statistical mechanics*, (Wiley, New York, 1975).
M. Hillery, R.F. O’Connell, M.O. Scully, E.P. Wigner, Physics Reports, **106**, 121 (1984).
P. A. Andreev, arXiv: 1212.0099.
P. A. Andreev, F. A. Asenjo, and S. M. Mahajan, arXiv: 1304.5780.
P. A. Andreev and L. S. Kuz’menkov, arXiv:1208.1000.
P. A. Andreev, arXiv: 1201.6553.
T. Ohmi and K. Machida, J. Phys. Soc. Jpn. **67**, 1822 (1998).
T. L. Ho, Phys. Rev. Lett. **81**, 742 (1998).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The photoelectric cross section of $^{16}$O($\gamma$,$\alpha$)$^{12}$C is estimated to be larger than the radiative capture cross section of $^{12}$C($\alpha$,$\gamma$)$^{16}$O. The predicted cross section and the angular distribution of $\alpha$-particle are illustrated for the future experiment. The cross section just above the $\alpha$-particle threshold is found to be dominated by the $E$2 excitation.'
author:
- 'M. Katsuma'
title: 'Photoelectric disintegration of $^{16}$O'
---
The low-energy $^{12}$C($\alpha$,$\gamma$)$^{16}$O reaction plays the crucial role in the nucleosynthesis of element in a star. However, the cross section is very small at the energies corresponding to helium burning temperatures, $E_{c.m.}\approx 0.3$ MeV, because of the Coulomb barrier, and it is far beyond out of reach by the present laboratory technologies [@Rol88]. ($E_{c.m.}$ is the center-of-mass energy of the $\alpha$+$^{12}$C system.) To cope with the difficulties, the experimental challenges have continuously been developed (e.g. [@Gai14; @Gai13; @Bru13; @Sch12; @Pla12; @Mak09; @Ass06; @Kun01]), as well as the theoretical predictions [@Kat08; @Kat14b]. The current experimental projects [@Gai14; @Ahm13] of photonuclear reactions with high intensity laser are the one of the precise measurements for the tiny cross section.
The probability of the photodisintegration of $^{16}$O just above the $\alpha$-particle threshold restricts the reaction rates of the $^{12}$C($\alpha$,$\gamma$)$^{16}$O reaction. In particular, the angular distribution of the emitted $\alpha$-particle is expected to reduce the large uncertainties of the cross section. The subthreshold 1$^-_1$ state at the excitation energy $E_x=7.12$ MeV is believed to couple strongly with the 1$^-_2$ state at $E_x=9.585$ MeV. The strong interference with two 1$^-$ states lead to the large enhancement of the low-energy cross section, and it has been presumed to play the important role in the derived reaction rates at helium burning temperatures. In contrast, the $E$1 cross section is predicted not to be enhanced by the high energy tail of the subthreshold 1$^-_1$ state because the $\alpha$+$^{12}$C system can be described by the weak coupling [@Kat08; @Kat14b; @Kat10b]. In our previous studies, the low-energy cross section is dominated by the $E$2 transition. This has been endorsed by the $\gamma$-ray angular distribution at $E_{c.m.}=1.254$ – 1.34 MeV [@Kun01; @Ass06] and the transparency of the $\alpha$+$^{12}$C system at low energies [@Kat10b; @Kat14b; @Tis09; @Pla87].
In the present report, we study the photoelectric disintegration of $^{16}$O. Using the reciprocity theorem, we illustrate the expected photoelectric cross section of the $^{16}$O($\gamma$,$\alpha$)$^{12}$C reaction.
Let us recall the relation between the photonuclear reaction and the capture reaction. The cross section of the $^{16}$O($\gamma$,$\alpha$)$^{12}$C reaction is given by the inverse reaction in the following expression, $$\begin{aligned}
\sigma_{\gamma\alpha} (E_\gamma) &=& \frac{k_{\rm c}^2}{2k_\gamma^2} \sigma_{\alpha\gamma} (E_{c.m.})
\label{eq:sig}
\end{aligned}$$ where $k_{\rm c}$ is the wavenumber of relative motion between $\alpha$-particle and $^{12}$C nuclei; $k_\gamma$ is the wavenumber of photon $k_\gamma=E_\gamma/(\hbar c)$, $E_\gamma=E_{c.m.}+7.162$ MeV; $\sigma_{\alpha\gamma}$ is the capture cross section for the $^{12}$C($\alpha$,$\gamma$)$^{16}$O reaction. To obtain the predicted values, we use the same potential model as [@Kat08].
Owing to the Coulomb barrier of the $\alpha$+$^{12}$C channel, the $\alpha$-particle cannot be emitted easily above the threshold energy. The resulting cross section is very small. To compensate for the rapid energy variation, we illustrate the cross section multiplied by the Gamow factor, $\exp(2\pi\eta)$, in the present report. $\eta$ is the Sommerfeld parameter, $\eta=Z_1Z_2 e^2/(\hbar v)$. $v$ is the velocity of relative motion between $\alpha$-particle and $^{12}$C nuclei. $Z_1$ and $Z_2$ are the charges of the interacting nuclei.
The cross section of the photodisintegration is expected to be larger than that of the $^{12}$C($\alpha$,$\gamma$)$^{16}$O reaction. From Eq. (\[eq:sig\]), we find $\sigma_{\gamma\alpha}\approx 50 \times \sigma_{\alpha\gamma}$ at $E_\gamma\approx 8.41$ MeV, corresponding to $E_{c.m.}= 1.25$ MeV close to the current lowest energy of the angular distribution measurement of $^{12}$C($\alpha$,$\gamma$)$^{16}$O [@Kun01]. This makes it possible to determine the $E$2/$E$1 ratio of the cross section more accurately.
![\[fig:sig\] Photoelectric cross section for the $^{16}$O($\gamma$,$\alpha$)$^{12}$C reaction as a function of the $\gamma$-ray energy. The solid curve is the result obtained from the potential model. The dashed and dotted curves are the $E$1 and $E$2 components, respectively. The cross section is multiplied by $\exp(2\pi\eta)$ to compensate for the rapid energy variation in the vicinity of the $\alpha$-particle threshold. The vertical thin line at $E_\gamma=7.162$ MeV indicates the energy position of the threshold. ](fig1.eps){width="0.98\linewidth"}
![\[fig:ang\] Differential cross sections for the $^{16}$O($\gamma$,$\alpha$)$^{12}$C reaction at $E_\gamma=8.0$, 8.5, 9.0 and 9.5 MeV. The solid curves are the results obtained from the potential model. The dashed and dotted curves are the pure $E$1 and $E$2 components. ](fig2.eps){width="0.85\linewidth"}
In Fig. \[fig:sig\], we illustrate the calculated photoelectric cross section of the $^{16}$O($\gamma$,$\alpha$)$^{12}$C reaction. The solid curve is the result obtained from the potential model. The dashed and dotted curves are the $E$1 and $E$2 components, respectively. The vertical thin line at $E_\gamma=7.162$ MeV indicates the energy position of the $\alpha$-particle threshold in $^{16}$O. The $\alpha$-particle is emitted above this energy. The peak at the 1$^-_2$ state can be seen at $E_\gamma\approx9.5$ MeV. The $E$1 component is approximately constant below $E_\gamma= 8$ MeV. The $E$2 cross section is enhanced as $E_\gamma$ decreases. The photoelectric disintegration is found to be dominated by the $E$2 excitation below $E_\gamma\approx 8$ MeV. This is due to the high energy tail of the subthreshold 2$^+_1$ state at $E_x=6.92$ MeV, which has the well-developed $\alpha$+$^{12}$C cluster structure [@Kat10b].
The $\sigma_{\gamma\alpha}\exp(2\pi\eta)$ in Fig. \[fig:sig\] appears to resemble the astrophysical $S$-factor in the $^{12}$C($\alpha$,$\gamma$)$^{16}$O reaction [@Kat08]. The energy variation, in fact, corresponds to $\sigma_{\gamma\alpha}\exp(2\pi\eta)\propto S/E_\gamma^2$.
The photonuclear reaction may remind us of the discussion about the dipole excitation. However, the photomagnetic dipole excitation ($s$-wave) is forbidden. The electric dipole transition ($p$-wave) is hindered by the isospin selection rule. Thus, the $E$2 excitation ($d$-wave) could be dominant mode of the transition in the vicinity of the $\alpha$-particle threshold. The $E$3 and $E$4 transitions ($f$- and $g$-waves) are negligible.
The angular distributions of the $\alpha$-particle at $E_\gamma=8.0$, 8.5, 9.0 and 9.5 MeV are shown in Fig. \[fig:ang\]. The solid curves are the results obtained from the potential model. $\theta_{c.m.}$ is defined with respect to the beam direction. At $E_\gamma=9.5$ MeV, the angular distribution appear to be the single peak due to the 1$^-_2$ state at $E_x=9.585$ MeV. The angular distribution becomes the double peaks at $E_\gamma=8.0$ MeV. The dashed and dotted curves are the pure $E$1 and $E$2 components. The $E$2 contribution caused by the subthreshold 2$^+_1$ state interferes with the dipole component.
The differential cross section at $\theta_{c.m.}=90^\circ$ can be made only from the $E$1 component, because the $E$2 component vanishes at this angle. Using the $90^\circ$ cross section, the integrated $E$1 cross section is expressed as $$\begin{aligned}
\sigma^{E1}_{\gamma\alpha} (E_\gamma) &=& \frac{8\pi}{3} \left. \frac{d\sigma_{\gamma\alpha}}{d\Omega} \right|_{\theta_{c.m.}=90^\circ}.
\label{eq:sig90}
\end{aligned}$$ The dashed curve in Fig. \[fig:sig\] can be confirmed from this relation. The numerical values are listed in Table \[tb:sig\].
$E_\gamma$ (MeV) d$\sigma_{\gamma\alpha}$/d$\Omega$(90$^\circ$) (b/sr) $\sigma^{E1}_{\gamma\alpha}$ (b) $\exp(2\pi\eta)$
------------------ ------------------------------------------------------- ---------------------------------- ------------------
8.0 3.11$\times10^{-12}$ 2.60$\times10^{-11}$ 5.72$\times10^9$
8.5 4.22$\times10^{-10}$ 3.53$\times10^{-9}$ 5.27$\times10^7$
9.0 1.19$\times10^{-8}$ 9.98$\times10^{-8}$ 3.88$\times10^6$
9.5 3.47$\times10^{-7}$ 2.91$\times10^{-6}$ 6.94$\times10^5$
: \[tb:sig\] Photoelectric dipole cross section and $90^\circ$ cross section. The value of $\exp(2\pi\eta)$ is also listed.
The 90$^\circ$ cross section basically determines the $E$1 contribution of the photodisintegration of $^{16}$O. In addition, the $E$2 contribution is predicted to be large at low energies, so the angular distribution has the minimum at $\theta_{c.m.}\approx90^\circ$. In this circumstance, the $E$1 component may be susceptible to the background noise. The absolute value of the $E$1 cross section could not be determined without precise measurement of the angular distribution. We expect that the $E$2/$E$1 ratio will be provided more accurately by the future experiment.
In summary, the $^{16}$O($\gamma$,$\alpha$)$^{12}$C reaction has been studied with the potential model. We have shown the calculated photoelectric cross section and the angular distribution of $\alpha$-particle. The cross section just above the $\alpha$-particle threshold is dominated by the $E$2 excitation. The cross section of the photoelectric disintegration is found to be larger than that of the $^{12}$C($\alpha$,$\gamma$)$^{16}$O reaction. We expect that the forthcoming experimental projects [@Gai14; @Ahm13] will determine the $^{12}$C($\alpha$,$\gamma$)$^{16}$O cross section more accurately.
The author is grateful to Prof. Moshe Gai and Prof. M. Fujiwara for their encouragement. He thanks Y. Kondō for early days of collaboration and Y. Ohnita and Y. Sakuragi for their hospitality. He also thanks M. Arnould, A. Jorissen, K. Takahashi, and H. Utsunomiya for their hospitality at his stay in Université Libre de Bruxelles.
[19]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ** (, ).
, ****, ().
, ****, ().
, ****, ().
, , , , ****, ().
, , , , , , , ****, ().
, , , , , , , , ****, ().
, , , , , , , , , , , ****, ().
, , , , , , , ****, ().
, ****, (); ****, (); ****, ().
, ().
, , , , , , , ().
, ****, (); ****, (); ****, ().
, , , , , , , , , , , ****, ().
, , , , , , ****, ().
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report the first results obtained from our campaign to characterize the intranight-optical variability (INOV) properties of [[*Fermi*]{}]{} detected blazars, using the observations from the recently commissioned 1.3 m J C Bhattacharya telescope (JCBT). During the first run, we were able to observe 17 blazars in the Bessel $R$ filter for $\sim$137 hrs. Using $C$ and scaled $F$-statistics, we quantify the extent of INOV and derive the duty cycle (DC) which is the fraction of time during which a source exhibits a substantial flux variability. We find a high DC of 40% for BL Lac objects and the flat spectrum radio quasars are relatively less variable (DC $\sim15$%). However, when estimated for blazars sub-classes, a high DC of $\sim$59% is found in low synchrotron peaked (LSP) blazars, whereas, intermediate and high synchrotron peaked objects have a low DC of $\sim$11% and 13%, respectively. We find evidences about the association of the high amplitude INOV with the $\gamma$-ray flaring state. We also notice a high polarization during the elevated INOV states (for the sources that have polarimetric data available), thus supporting the jet based origin of the observed variability. We plan to enlarge the sample and utilize the time availability from the small telescopes, such as 1.3 m JCBT, to strengthen/verify the results obtained in this work and those existed in the literature.'
author:
- 'Vaidehi S. Paliya$^{1}$, C. S. Stalin$^{2}$, M. Ajello$^{1}$, and A. Kaur$^{1}$'
bibliography:
- 'Master.bib'
title: 'Intra-Night Optical Variability Monitoring of [*Fermi*]{} Blazars: First Results from 1.3 J C Bhattacharya Telescope'
---
Introduction
============
[\[sec:Intro\]]{} Blazars are a special class of active galactic nuclei (AGN) with their relativistic jets pointed towards the observer [@1995PASP..107..803U]. Due to the peculiar orientation of the jet that enhances the radiation because of the relativistic boosting , blazars are known to emit over entire electromagnetic spectrum, from low energy radio waves to very high energy [$\gamma$]{}-rays. In addition to that, another defining characteristic feature displayed by blazars is the rapid and high amplitude flux variations observed at all the accessible wavelengths . Based on the equivalent width (EW) of the optical emission lines, blazars are classified as flat spectrum radio quasars (FSRQs) and BL Lac objects with FSRQs exhibit broad emission lines (EW$>$5Å). The observation of strong emission lines from FSRQs indicates for the presence of a luminous broad line region (BLR), which in turn, suggests a high and efficient accretion process that illuminate the BLR [e.g., @2012MNRAS.421.1764S]. In fact, the infrared (IR) to ultraviolet (UV) spectrum of many FSRQs is found to be dominated by thermal emission from the accretion disk [e.g., @2010MNRAS.402..497G]. The presence/absence of narrow emission lines in the optical spectrum of BL Lac objects, on the other hand, hints a relatively low and inefficient accretion and/or the dominance of the non-thermal synchrotron radiation originated from the plasma moving along the relativistic jet. @2010ApJ...716...30A introduced another blazar classification based on the location of the synchrotron peak in the broadband spectral energy distribution (SED) of the sources. A blazar is known as low synchrotron peaked or LSP if the rest-frame synchrotron peak frequency ($\nu^{\rm peak}_{\rm syn}$) is $\lesssim10^{14}$ Hz. On the other hand, for $10^{14}\lesssim\nu^{\rm peak}_{\rm syn}\lesssim10^{15}$ and $\nu^{\rm peak}_{\rm syn}\gtrsim10^{15}$ Hz, the sources are classified as intermediate synchrotron peaked (ISP) or high synchrotron peaked (HSP) blazars.
The intensity variations in blazars are generally explained with the ‘shock-in-jet’ model [@1985ApJ...298..114M] and this suggests that the observed temporal variability is associated with the radiative processes occurring in the AGN jets . However, the detection of the low amplitude optical variability from radio-quiet [see, e.g., @2016MNRAS.461..666K and references therein] hints that such optical flux variations can also originate from the accretion disk instabilities or perturbations [@1993ApJ...406..420M]. In the widely accepted scenario of the blazar emission, the high amplitude, rapid optical variability is associated with the non-thermal jetted synchrotron emission. This is also supported from the observations of the correlated multi-wavelength flux variations [e.g., @2011MNRAS.410..368B; @2015ApJ...803...15P] that demands the origin of the optical to $\gamma$-ray emission to be co-spatial. The observations of the high degree of variable optical polarization during the flaring activity of the quasars provide further supporting evidences to it [e.g., @2014PASJ...66..108I; @2016ApJ...833...77I].
The causes responsible for the observed flux variability from blazars are not well understood. In the framework of the one-zone leptonic emission models, a simultaneous multi-wavelength flux enhancement is possible due to injection of the fresh, highly energetic particles into the emission region or/and due to the sudden acceleration of the jet [e.g., @2015ApJ...803...15P]. On the other hand, there is evidence that flaring behavior was seen only at optical/UV energies [e.g., @2013ApJ...763L..11C; @2014ApJ...789..143P]. Such anomalous outbursts are explained on the basis of the changes in the magnetic field strength or the location of the emission region [@2007MNRAS.382L..82G; @2014ApJ...789..143P].
The intra-night optical variability (INOV), or microvariability (i.e., variations with durations extending from minutes to hours) of blazars has been studied for more than two decades . In fact, the study of the INOV properties using small 1$-$2 m class optical telescopes has been proved quite successful in determining the nature of the short-term optical flux variations in blazars and their connection with the high energy [$\gamma$]{}-ray flares [e.g., @2015ApJ...809..130C; @2016MNRAS.455..680A]. It is found in recent studies that [$\gamma$]{}-ray blazars have faster jets or they have higher relativistic beaming factors compared to their non [$\gamma$]{}-ray detected counterparts . Therefore, it is more likely to detect a relatively high amplitude INOV from [$\gamma$]{}-ray emitting blazars, thereby making them an important target to study the fast optical flux variability.
The release of the third catalog of the [$\gamma$]{}-ray emitting AGNs detected by the [*Fermi*]{}-Large Area Telescope [3LAC; @2015ApJ...810...14A] represents a uniform, flux-limited sample of [$\gamma$]{}-ray emitters, of which blazars forms a major fraction. Motivated by the availability of the [$\gamma$]{}-ray information for a large number of blazars and also the availability of telescope time on the recently commissioned 1.3 m J C Bhattacharya Telescope (JCBT) at the Vainu Bappu Observatory (VBO), Kavalur, India, we have started an optical monitoring campaign to systematically study the INOV properties of [$\gamma$]{}-ray emitting blazars. In this work, we present the results obtained from the first run of the observing campaign. In Section \[sec:sample\] we discuss our sample selection procedure and provide a brief introduction of the JCBT in Section \[sec:JCBT\]. We outline the s.pdf of the adopted data reduction in Section \[sec:data\] and present the results in Section \[sec:results\]. The results associated with the individual objects are highlighted in Section \[sec:objects\] and we briefly discuss and summarize our findings in Section \[sec:summary\].
Sample
======
[\[sec:sample\]]{} We start with 3LAC blazars that have a redshift measurements reported in the catalog. We select those sources that are observable from VBO (geographical coordinates: 12$^{\circ}$ 34$^{\prime}$ 0$^{\prime\prime}$ N, 78$^{\circ}$ 50$^{\prime}$ 0$^{\prime\prime}$ E). Since we use a relatively small telescope (1.3 m diameter), we retain only those blazars that have apparent $R$ band magnitude $<18$. We further look into the sky images[^1] and remove those sources where the field is crowded or/and if there are relatively bright stars close to the target object (e.g., 3FGL J1104.4+3812 or Mrk 421). This is required to avoid the CCD saturation for long exposures ($>5$ min). At this stage, we are left with about 100 blazars[^2]. Our observing campaign depends on the availability of the telescope time and we adopt an ‘as and when’ approach to fully utilize the monitoring opportunity. Therefore, the observing duration varies from $\sim$1 hr to $\sim$9 hrs on any particular night. Our first observing run started in 2015 January and ended in 2016 January. During this periods, we were able to observe 17 blazars (7 FSRQs and 10 BL Lac objects) for a total of $\sim$137 hrs in 27 nights. The general properties of these sources are given in Table \[tab:basic\_info\].
J C Bhattacharya Telescope
==========================
[\[sec:JCBT\]]{} The JCBT is a 1.3m F/8 Double Horseshoe telescope at VBO, Tamilnadu, India. It was inaugurated in 2014 April. There are two types of charge coupled devices (CCDs) mounted on the telescope: (i) 1k $\times$ 1k proEM CCD and (ii) 2k $\times$ 4k normal CCD. The proEM CCD is Peltier cooled with water circulation (with Ethylene Glycol mixture) and the liquid Nitrogen is used as a coolent for the normal CCD. The advantage of the ProEM CCD is that the maximum speed is 10 MHz for EM mode and 5 MHz for normal mode, both with variable gain settings. However, it has a small field of view (FOV). Since our primary objective is to perform the differential photometry, we need a suitable pair of non-variable stars present in the same exposure frame. This demands a relatively large FOV and therefore, we choose the 2k $\times$ 4k normal CCD. In all the observations, its central 2k $\times$ 2k region is used. Furthermore, all the observations are carried out using the Bessel $R$ filter ($\lambda_{\rm eff}\sim 6400$ Å) in which CCD has a high response (see the quantum efficiency plot in Figure \[fig\_ccd\]). The main characteristics of both CCDs are presented in Table \[tab:ccd\].
Data Reduction
==============
[\[sec:data\]]{} We use IRAF[^3] to perform the data reduction. The preliminary cleaning of the image frames is done by preforming the bias subtraction, flat fielding, and cosmic-ray removal. The bias frames are taken at regular intervals throughout the night and flat-field images are taken at dusk, i.e., at the beginning of the observations. We follow the aperture photometry method to extract the instrumental magnitudes of the target blazar and nearby comparison stars present in the same image frame. Among the various comparison stars, we select the pair with the steadiest differential light curve (DLC), i.e., with the minimum standard deviation for a given aperture [taken as full width at half maximum of the point spread function of stars, @1989PASP..101..616H]. Once the steadiest pair of stars is determined, the optimum aperture size is derived by considering a range of aperture radii, calculating the standard deviation of the DLC for the selected stars pair and choosing the aperture radius that minimizes the scatter. Using this optimized aperture radius, we extract the instrumental magnitudes of the target blazar and the selected pair of the comparison stars. In Table \[tab:com\_star\], we report the USNO-B1 cataloged positions and apparent $R$ band magnitudes for the selected comparison stars associated with each blazar. It should be noted that the possible uncertainties in the quoted magnitudes could be up to $\sim$0.25 magnitudes [@2003AJ....125..984M].
The IR-optical spectrum of the bright blazar 3FGL J1653.9+3945 (or Mkn 501) is known to be contaminated from the host galaxy radiation. However, since we are interested in measuring the differential flux rather than the absolute values, we neglect it. Furthermore, the INOV is not expected to be affected from the host galaxy compared to the fast variation originates from the relativistic jet and therefore it is a reasonable choice [see., e.g., @2016ApJS..222...24X for a discussion on the host galaxy effect on the long term variability behavior of Mkn 501].
Results
=======
[\[sec:results\]]{} We generate DLCs from the photometric instrumental magnitudes of the target source and a pair of the steady comparison stars. The star-star DLC represent the observational uncertainties and the intrinsic variability of the stars, whereas, quasar-stars DLCs indicate the intrinsic variability of the blazar with respect to steady stars. A blazar is considered to be variable when it shows correlated flux variations both in time and amplitude relative to the pair of comparison stars. On the other hand, the source is considered as non-variable if it does not show flux variability with respect to both the comparison stars. It should be noted that the uncertainty in the comparison stars magnitudes, as given in Table \[tab:com\_star\], will not have any effect on the DLCs as we look for the variability in the differential instrumental magnitudes between the source of interest and the comparison stars. The derived results are presented in Figure \[fig\_inov1\]-\[fig\_inov6\]. To statistically confirm the variability, we employ the two criteria, $C$ and $F$ statistics.
C-Statistics
------------
$C$ parameter is the commonly used statistical tool to decide the INOV nature of the sources [@1997AJ....114..565J]. It is defined as follows $$\label{eq:c-stat}
C_1 = \frac{SD_{\rm qso-s1}}{SD_{\rm s1-s2}},~ \\
C_2= \frac{SD_{\rm qso-s2}}{SD_{\rm s1-s2}},$$ where $SD_{\rm qso-s_i}$ is the standard deviation of the DLC of the target quasar and the $i^{\rm th}$ comparison star and $SD_{\rm s1-s2}$ is the standard deviation of the star-star DLC. Following @1997AJ....114..565J, we consider the source to be variable if both $C_1$ and $C_2$ are $\geq2.576$, which corresponds to the 99% confidence level.
Scaled F-statistics
-------------------
Another statistical tool to quantify the INOV behavior is the $F$-test and this has been shown to be a more reliable and robust method to characterize the micro-variability [@2010AJ....139.1269D]. The $F$ parameter is calculated as $$\label{eq:f-test}
F_1 = \frac {V_{\rm qso-s1}}{V_{\rm s1-s2}},~\\
F_2 = \frac {V_{\rm qso-s2}}{V_{\rm s1-s2}}$$ where $V_{\rm qso-s}$ and $V_{\rm s1-s2}$ are the variances of the quasar-star and star-star DLCs. These $F$ values are compared with the critical $F$ value, $F^{\alpha}_{\nu}$, where $\alpha$ is the significance level (taken as 0.01 corresponding to a confidence level of $>99$%) and $\nu$ (=$N_{\rm p}-1$) is the degree of freedom for the DLC. A source is considered to be variable if both $F$ values are found to exceed $F^{\alpha}_{\nu}$.
It is noted in various recent studies [e.g., @2011MNRAS.412.2717J; @2014AJ....148...93D] that the $F$ test does not estimate the micro-variability adequately if the comparison stars have brightness significantly different from the target quasar [see also @2007MNRAS.374..357C; @2015AJ....150...44D]. To compensate for the differences, we need to scale the star-star variance by a factor $\zeta$ [@2011MNRAS.412.2717J; @2014AJ....148...93D]. Since our analysis procedure is similar to that of @2011MNRAS.412.2717J, we follow their scaling factor to take into account the brightness differences of the comparison stars. It is defined as follows
$$\zeta=\left[\displaystyle{\frac{\sum_\mathbf{i=1}^{N_{\rm p}}\sigma^2_{i, {\rm err}}({\rm qso-s})/N_{\rm p}}{\sum_{i=1}^{N_{\rm p}}\sigma^2_{i, {\rm err}}({\rm s1-s2})/N_{\rm p}}}\right] \equiv \frac{\langle\sigma^2({\rm qso-s})\rangle}{\langle\sigma^2(s{\rm 1-s2})\rangle},
\label{eq:omega}$$
where $\sigma^2_{\rm i,err}({\rm qso-s})$ and $\sigma^2_{\rm i,err}({\rm s1-s2})$, respectively, are the errors on individual points of the quasar-star and star-star DLCs. With this, the scaled F-value, $F^{s}$, can be computed as,
$$\label{eq.fstest}
F_{1}^{s}=\frac{V_{\rm qso-s1}}{\zeta~V_{\rm s1-s2}},~ \\
F_{2}^{s}=\frac{V_{\rm qso-s2}}{\zeta~V_{\rm s1-s2}}.$$
In this work, we use this scaled $F$ test to ascertain the INOV nature of the target blazars.
It has been reported in recent works that claiming the presence of the INOV using $C$ statistics is not correct as it is almost too conservative [e.g., @2010AJ....139.1269D; @2014AJ....148...93D]. However, it is also important to consider the fact that this method might be a more compelling measure of the INOV, especially when the comparison stars are not steady [@2017MNRAS.tmp...73Z]. We also adopt this tool to compare our results with that obtained in earlier INOV studies of blazars where only $C$ statistics was employed [see, e.g., @2001AJ....122.2901D; @2004MNRAS.348..176S]. Accordingly, we report the results of both $C$ and scaled $F$-statistics in Table \[tab:stat\].
Amplitude of Variability
------------------------
The INOV variability amplitude ($\psi$) quantifies the actual variation exhibited by the target source on any given night, after correcting for the errors in the measurements. It is defined as follows
$$\label{eq:amplitude}
\psi = 100 \sqrt{A_{\rm max} - A_{\rm min})^{2} - 2V_{\rm s1-s2}}\%$$
with\
$A_{\rm max}$ = maximum in $\gamma$-NLSy1 differential light curve,\
$A_{\rm min}$ = minimum in $\gamma$-NLSy1 differential light curve,\
$V_{\rm s1-s2}$ = variance in the star$-$star DLC
The results of this exercise are presented in Table \[tab:stat\].
Duty Cycle
----------
Duty cycle (DC) is computed to determine the fraction of the time an object shows the variability. This is because a blazar may not show flux variations on all observing nights. The DC is evaluated by taking the ratio of the time over which the target blazar exhibits variability to the total monitoring time. Following , it is defined as
$$\label{eq:dc}
DC = 100\frac{{\sum_{k=1}^{N_{\rm p}}} Q_k(1/\Delta t_k)}{\sum_{k=1}^{N_{\rm p}} (1/\Delta t_k)} {\rm ~\%}$$
where $\Delta t_{k}$ = $\Delta t_{k, {\rm obs}}(1 + z)^{-1}$ is the redshift corrected time interval of the monitoring session of a source on the $i^{\rm th}$ night. $Q_{k}$ is set equal to 1 if the source is found to be variable, otherwise $Q_{k}=0$.
We find a DC of 26% and 25% for the entire sample, when considering the $C$-statistics and the scaled $F$-statistics, respectively. It is found to be 14% according to the $C$-statistics and 16% following the $F$-test, for FSRQs. For BL Lac objects, DC is derived as 40% using both the $C$ and the $F$-tests. Alternatively, if we define the variability criteria with the detection of the high amplitude of variability [$\psi>3$%, e.g., @2004JApA...25....1S], the overall DC increases to 66%. Considering FSRQs and BL Lac objects separately, it is estimated as 52% and 86%, respectively. These results confirms the earlier findings that BL Lac objects show a larger INOV compared to FSRQs. On dividing the sources according to the location of their $\nu^{\rm peak}_{\rm syn}$, we find that LSP blazars exhibit a substantially larger INOV (DC$\sim$59%) compared to ISP and HSP sources that have a DC of $\sim$11% and 13%, respectively.
Discussion on Individual Sources
================================
[\[sec:objects\]]{} Blazars are known to vary at all timescales [e.g., @2011ApJS..194...29R]. If we assume a co-spatial origin for the observed multi-wavelength radiation from blazars, a flaring state in one band should be reflected at other frequencies, though with different timescales due to different scales of the Doppler boosting [@1995ApJ...446L..63D]. Since the sources monitored under our observing campaign are known $\gamma$-ray emitters, we expect to observe a high amplitude INOV during the elevated $\gamma$-ray activity state. On the other hand, an optical flux variability without a $\gamma$-ray counterpart indicates the change in the magnetic field that enhances synchrotron emission or/and the location of the emission region being close to the black hole where the jet bulk Lorentz factor is small, as a possible causes of the orphan optical flares [e.g., @2007MNRAS.382L..82G; @2013ApJ...763L..11C]. The latter becomes a major factor for FSRQs whose $\gamma$-ray emission is known to be dominated by the external Compton processes [@2009ApJ...704...38S]. In other words, it is possible to detect a short term INOV from a non-thermal synchrotron dominated jet, even during the $\gamma$-ray quiescence.
To understand the optical-$\gamma$-ray connection in deeper perspective, we plot the publicly available, monthly binned, aperture photometric $\gamma$-ray light curves[^4] for all 17 blazars in Figure \[fig\_gamma1\]. It should be noted that the $\gamma$-ray aperture photometry may not be suitable for a detailed scientific investigation, however, since our primary purpose is just to get an idea about the $\gamma$-ray activity state during the epoch of the optical monitoring, these results can be used. In Figure \[fig\_gamma1\], the x-axis covers the time period since 2015 January 1 (MJD 57023) till 2016 February 5 (MJD 57423). On the other hand, range on the y-axes are chosen to cover the minimum and the maximum of the observed $\gamma$-ray fluxes since the launch of [[*Fermi*]{}]{} satellite. Therefore, it is possible that we may not see the highest or the lowest $\gamma$-ray flux in the covered time period. The advantage of this approach is that it shows us the relative activity state of the source during our optical monitoring campaign. In Figure \[fig\_gamma1\], we denote the epoch of the night observations with vertical lines. In order to understand the optical-[$\gamma$]{} ray connection for the covered period, we also utilize the optical $R$ band data taken from the Steward observatory [@2009arXiv0912.3621S], the Katzman Automatic Imaging Telescope [KAIT; @2003PASP..115..844L], and the SMARTS observatory [@2012ApJ...756...13B]. Out of 17 sources, we could find the long term optical data for 10 sources. We correct the data for the galactic reddening following @2011ApJ...737..103S and convert to flux units using the zero points of . The derived fluxes are shown in Figure \[fig\_gamma1\] with blue empty circles. It should be noted that adding the long term optical data was necessary since our INOV monitoring epochs are comparatively smaller for individual sources. We, then, perform a correlation test using the [*z-transformed discrete correlation function*]{} tool of @2013arXiv1302.1508A to quantify the strength of the correlation in the optical and the [$\gamma$]{}-ray bands. The derived results are presented in Table \[tab:dcf\]. As can be seen that a strong claim cannot be made owing to the large errorbars and the small number of the data points and the results are consistent with the zero lead/lag. However, visual inspection of the light curves in Figure \[fig\_gamma1\] supports the idea that a high [$\gamma$]{}-ray state is indeed accompanied by an optical flare and probably reflected in our INOV observations. Below, we individually discuss the results obtained for every blazar.
[**3FGL J0608.0$-$0835:**]{}\
This source was observed twice, first on 2015 March 7 and then later on 2016 January 3 (Figure \[fig\_inov1\]). Both $C$ and $F$ tests have confirmed the absence of flux variations on the former epoch but a significant INOV of the order of $\sim$0.1 mag is noticed during the later observations. Along with this, an extremely fast ($<10$ minutes), small amplitude flare is also noticed at the beginning of the observation. To our knowledge, this is the first report of the detection of INOV from this object. The interesting fact to note is that during both the observing epochs, the source has not shown any significant $\gamma$-ray variability (see the top panel of Figure \[fig\_gamma1\]).
[**3FGL J0656.4+4232:**]{}\
We have observed this object on 2016 January 5 but barring a couple of flickering events, no significant variability is detected.
[**3FGL J0710.5+4732:**]{}\
This is the highest redshift source in our sample ($z=1.29$) and we have monitored it on three nights. Other than a few flickering, we have not found any significant INOV on any night, which is further confirmed from $C$ and $F$ statistics. During our monitoring campaign, the source remained quiescent in $\gamma$-rays , thus indicating the lack of the jet activity as a probable cause of the non-detection of INOV.
[**3FGL J0721.9+7120:**]{}\
This source is a well-known TeV blazar and it was observed on 2015 January 21, 22 and a month after on February 23. Our first two observations in 2015 January were coincident with the exceptional multi-wavelength flaring activity of this source [Figure \[fig\_gamma1\]; see also, @2015ApJ...809..130C; @2015ATel.6999....1M; @2015MNRAS.452L..11W; @2016MNRAS.455..680A; @2016MNRAS.458.2350W]. Such a high activity state is also reflected in its INOV behavior (Figure \[fig\_inov2\]). As can be seen in Figure \[fig\_inov2\], a fast intra-night flare is noticed on 2015 January 21, that lasted for $\sim$1.5 hrs with the optical magnitude varies by about 0.04 mag. Similar behavior is also noticed during other two nights of the monitoring. Moreover, the source was also found to be in a high optical state as revealed from the Steward observatory monitoring. The fast optical variability from this source can be well explained from the beamed synchrotron emission which is further supported from the detection of the high optical polarization, $\sim5-10\%$, during the flaring episode from Steward observatory (see Table \[tab:stat\]).
[**3FGL J0739.4+0137:**]{}\
This is a FSRQ and is known to show fast INOV [@2003AJ....126...37C]. @2012MNRAS.424.2625B suggested that this blazar shows high amplitude variability when close to the optical maximum. We find various small amplitude flares during its observation on 2015 March 7 (Figure \[fig\_inov2\]). Interestingly, during this period of the elevated activity, a high optical polarization of $\sim5\%$ was recorded at Steward observatory. In the conventional one-zone leptonic emission models, a high activity in the optical band is also reflected at high energy $\gamma$-rays. Therefore, we also looked into the monthly binned aperture photometric $\gamma$-ray light curve of this source to determine the association of the optical activity with $\gamma$-rays. As can be seen in Figure \[fig\_gamma1\], the blazar was flaring in $\gamma$-rays in the month of 2015 March (see the first vertical line around MJD 57100), coinciding with the detection of the INOV. The similar results are also noticed in the long term optical data from the Steward observatory, though it appears that the source brightens even further after our INOV epoch. On the other hand, during the second night of the observation in 2016 January, when no significant optical variability was detected, it was in quiescence in the $\gamma$-ray band also. Therefore, it gives a strong indication that the probability of detecting a significant INOV from a blazar is high during the elevated $\gamma$-ray activity state.
[**3FGL J0809.8+5218:**]{}\
This source is a known TeV BL Lac object [@2009ApJ...690L.126A] and its long term optical monitoring has been studied by @2014AJ....148..110M. They did not find any significant INOV during the period spanned over $\sim$6 years. Our one night of the optical monitoring has not revealed any significant flux variability and we find this object to be in a very low $\gamma$-ray activity state during the epoch of the optical observation. The long term optical behavior of this object, as monitored by KAIT, also indicates the source to be in quiescence.
[**3FGL J0831.9+0430:**]{}\
3FGL J0831.9+0430 or PKS 0829+046 is a $\gamma$-ray luminous BL Lac object and it is observed at three different epochs. A significant INOV is detected during all three nights (Figure \[fig\_inov3\]) and this is probably the first report of the intra-night flux variations from this source. It is important to note that the source was $\gamma$-ray quiescent during our monitoring campaign. However, its IR-optical-UV spectrum is dominated by the non-thermal synchrotron emission [@2011MNRAS.414.2674G] and hence the observed INOV could be originated from it. The long term KAIT monitoring reveals a moderate activity in the optical band during our observing run.
[**J0854.8+2006:**]{}\
OJ 287 is one of the best studied BL Lac objects in our sample and is known to show violent optical variability [see, e.g., @2009ApJS..181..466F; @2011MNRAS.416..101G]. We have observed this source on 2015 March 10 and a significant INOV was detected. In particular, the source was found to be in a relatively moderate $\gamma$-ray activity state and we noted a small amplitude ($\sim$0.05 mag) flare that lasted less than an hour (see Figure \[fig\_inov3\]). The optical polarimetric monitoring from Steward observatory on 2015 March 17, closest to the epoch of our observation, reveals the detection of a high polarization of 13.4% (Table \[tab:stat\]), thus indicating the synchrotron based origin of the INOV.
[**3FGL J0922.4$-$0529, 3FGL J0927.9$-$2037, 3FGL J1006.7$-$2159, 3FGL J1015.0+4925, and 3FGL J1204.3$-$0708:**]{}\
The blazars 3FGL J0922.4$-$0529, 3FGL J0927.9$-$2037, and 3FGL J1015.0+4925 were monitored for one night each and we observed 3FGL J1006.7$-$2159 and 3FGL J1204.3$-$0708 for two and three nights, respectively. None of them have shown any significant INOV (Figure \[fig\_inov3\], \[fig\_inov4\]) and they were also found to be in $\gamma$-ray quiescence during the observing run (Figure \[fig\_gamma1\]).
[**3FGL J1229.1+0202:**]{}\
We have monitored 3C 273 for three nights but could not detect any significant INOV (see Figure \[fig\_inov5\]). In fact, the source was in one of its lowest $\gamma$-ray and optical states during our observing campaign (see Figure \[fig\_gamma1\] and <http://quasar.square7.ch/fqm/1226+023.html>). It is also important to note that the optical spectrum of this object is dominated by a luminous accretion disk [e.g., @2015MNRAS.452.1303P] which is not expected to vary over short timescales ($\sim$few hrs). The detection of the almost negligible optical polarization ($<1$%) recorded at Steward observatory and its historical micro-variability measurements [@2001AJ....122.2901D] indicates that the lack of the INOV detection could be due to the dominance of the luminous accretion disk emission and the low activity of the jet.
[**3FGL J1404.8+0401:**]{}\
We have observed this object on 2015 May 20 for about 4.5 hrs and did not see any significant INOV.
[**3FGL J1512.8$-$0906:**]{}\
We have monitored PKS 1510$-$08 during its exceptional flare in 2015 May (May 21 and 22). During this period of the high activity, it was significantly detected at very high energies (VHE, $E>100$ GeV) by MAGIC telescopes and the elevated activities were seen across the electromagnetic spectrum [@2016arXiv161009416M; @2016arXiv161102098Z], including the optical monitoring from the Steward observatory. The source is found to exhibit a large amplitude variability ($\psi>15$%) on the second night of the observation (see Table \[tab:stat\]). On the other hand, the length of the observations ($<2$ hrs) during the first night was not enough to detect a significant INOV. The coincidence of the detection of the INOV with a high $\gamma$-ray activity state (Figure \[fig\_gamma1\]) again indicates a close association of detecting a high amplitude intra-night temporal variability during the high $\gamma$-ray state [see also @2014ApJ...789..143P; @2016ApJ...819..121P for similar results]. Similar to other blazars, where INOV was detected, the optical polarimetric monitoring from Steward observatory during the flaring period reports the detection of the high polarization. Interestingly, on 2015 May 21, i.e., when we could not detect INOV, the noticed optical polarization was 7.4% and next day it increases to 15.3%. This indicates that on 2015 May 22, PKS 1510$-$08 was in even higher optically flaring state which resulted in the detection of a significant micro-variability in our $\sim$2.5 hrs of the night observations.
[**3FGL J1653.9+3945:**]{}\
Mkn 501 is a bright BL Lac object and it is known to display extremely fast TeV flux variations [@2007ApJ...669..862A]. The source was in a low $\gamma$-ray and optical activity stated during our observing run. It is found to exhibit a significant flux variability on one of the monitoring nights. In fact, we detected an intriguing feature, a dip in the brightness of the source on $\sim$30 min timescale, on 2015 May 20 (Figure \[fig\_inov6\]). Such features have also been observed from another TeV blazar PG 1553+113 [@2011MNRAS.416..101G] and are interpreted as possibly due to absorption in a dusty gas cloud, occulting the emission region responsible for the optical emission. The optical polarization of the source was quite low ($<3$%) during both epochs.
Discussion and Summary
======================
[\[sec:summary\]]{} We present the first results of our ongoing campaign to characterize the INOV properties of [[*Fermi*]{}]{} detected blazars. This consists of the monitoring of 17 objects, 7 FSRQs and 10 BL Lac objects, for a total of $\sim$137 hrs. In our sample, BL Lac objects are found to display significantly higher DC of (40%) than FSRQs ($\sim$15%), considering $C$ and $F$ statistics. The conclusion that BL Lac objects show high INOV compared to FSRQs is also reported in various earlier studies [e.g., @2004MNRAS.348..176S]. Altogether, the overall DC is found to be $\sim$25% which is well below than that known from blazars [$\sim$70%, @2004JApA...25....1S]. However, this is most probably due to the small sample size and less number of observations and it is likely that DC may further increase as we will acquire more observations. Alternatively, we assume a blazar to be variable when its amplitude of variability $\psi>3$%, the overall DC increases to 66% and it is 52% and 86% for FSRQs and BL Lac objects, respectively.
On classifying the sources according to the location of the synchrotron peaks, we find that LSP blazars exhibit significantly larger variability compared to ISP and HSP objects. Though a strong claim cannot be made due to the small size of the sample, the following discussion can provide a possible physical explanation. In the left panel of Figure \[fig\_syn\], we show a typical synchrotron spectrum [e.g., @2008ApJ...686..181F] of LSP (red), ISP (blue), and HSP (black) blazars and also plot the frequency range covered by the Bessel $R$ filter (grey strip), using which we have carried out the observations. As can be seen, for the case of LSP sources, $R$ band covers the falling part of the spectrum, whereas, ISP objects have their peak located at these energies. For HSP blazars, on the other hand, a rising synchrotron spectrum is covered by the $R$ band. In the conventional leptonic emission scenario, the broadband emission from blazars originates from the relativistic electron population whose energy distribution follows a broken power law spectral shape (see the right panel of Figure \[fig\_syn\]). This means that a rising synchrotron spectrum is produced by the low energy electrons, whereas, the high energy particles contribute to the falling part of the SED. In other words, for ISP and HSP blazars, a relatively low energy electron population radiates in the $R$ band compared to LSP sources. Since the cooling time is inversely proportional to the energy of the radiating particles ($t_{\rm cool}\propto \gamma^{-1}$, $\gamma$ is the Lorentz factor of the electrons), it is obvious that the high energy electrons should exhibit faster variability, implying LSP sources to be more variable compared to ISP and HSP blazars. In fact, HSP blazars display violent flux variability at hard X-rays and at TeV energies [see, e.g., @1996Natur.383..319G; @2015ApJ...811..143P] where the highest energy electrons, forming the tail of the synchrotron and inverse Compton processes, radiates.
Our observing results provide evidences about a close association of the detection of the high amplitude INOV and a high optical polarization (monitored from Steward observatory) with the elevated $\gamma$-ray activity in blazars. This probably indicates a high chance of detecting large amplitude INOV when a blazar goes into the $\gamma$-ray flaring state. This feature has already been observed in a few $\gamma$-ray emitting narrow line Seyfert 1 galaxies [@2014ApJ...789..143P; @2016ApJ...819..121P] and therefore, it appears reasonable to generalize it to the beamed $\gamma$-ray emitting AGNs. This also supports the jet based origin of the flares rather than from the accretion disk. Furthermore, we also find extremely fast flux variations of the orders of minutes (e.g., see the light curves of 3FGL J0608.0$-$0835). Such rapid flares are typically observed from blazars at VHE $\gamma$-rays and they are probably originated from the magnetic reconnection events [@2009MNRAS.395L..29G], structured jets [@2008AIPC.1085..431T], due to turbulence in the plasma flow [e.g., @2012MNRAS.420..604N; @2014ApJ...780...87M], or possibly due to rapid conversion of the magnetic energy to radiation by a process called [*magnetoluminescence*]{} [@2015AAS...22521407B]. It is, therefore, of utmost interest to detect extremely rapid optical flares within the night to further test these theories of rapid flux variability detected from blazars.
The easily accessible small 1-2 m class optical telescopes definitely play an important role in studying the fundamental problems of blazar physics. In future, apart from 1.3 m JCBT, we also plan to extend the observations of $\gamma$-ray blazars from various other small telescopes, such as 0.9 m SARA-North at Kitt Peak, Arizona and 0.6 m SARA-South at Cerro Tololo, Chille, to verify/confirm the findings reported in this article and those existed in literature.
We are grateful to the referee for providing a constructive report. VSP and CSS are thankful to V. Moorthy, M. Appakutty, G. Selvakumar, S. Venkatesh, B. Rahul, and P. Anbazhagan, and the staff members of the VBO, for their assistance during the observing run. VSP acknowledge the warm hospitality received at VBO where this work was completed. This research has made use of data obtained from the Leicester Database and Archive Service at the Department of Physics and Astronomy, Leicester University, UK. Data from the Steward Observatory spectropolarimetric monitoring project were used. This program is supported by Fermi Guest Investigator grants NNX08AW56G, NNX09AU10G, NNX12AO93G, and NNX15AU81G. This paper has made use of up-to-date SMARTS optical/near-infrared light curves that are available at www.astro.yale.edu/smarts/glast/home.php.
---------------- ------------------------- ------------- ---------------- ------- ------- ------------
3FGL name Other name RA (2000) Dec (2000) $z$ $R$ Type
(1) (2) (3) (4) (5) (6) (7)
J0608.0$-$0835 PKS 0605$-$08 06 07 59.62 $-$08 34 52.16 0.872 15.41 ISP FSRQ
J0656.4+4232 4C +42.22 06 56 10.67 42 37 02.90 0.059 11.11 HSP BL Lac
J0710.5+4732 S4 0707+47 07 10 46.31 47 32 13.46 1.292 13.44 ISP BL Lac
J0721.9+7120 S5 0716+71 07 21 53.39 71 20 36.65 0.127 14.14 LSP BL Lac
J0739.4+0137 PKS 0736+01 07 39 18.03 01 37 04.61 0.189 16.03 ISP FSRQ
J0809.8+5218 1ES 0806+524 08 09 49.20 52 18 58.35 0.138 14.63 HSP BL Lac
J0831.9+0430 PKS 0829+046 08 31 48.88 04 29 38.73 0.174 13.99 LSP BL Lac
J0854.8+2006 OJ 287 08 54 48.87 20 06 30.83 0.306 14.35 LSP BL Lac
J0922.4$-$0529 TXS 0919$-$052 09 22 24.09 $-$5 29 08.14 0.974 14.75 ISP FSRQ
J0927.9$-$2037 PKS 0925$-$203 09 27 51.83 $-$20 34 51.47 0.348 15.98 LSP FSRQ
J1006.7$-$2159 PKS 1004$-$217 10 06 46.41 $-$21 59 20.38 0.330 15.23 ISP FSRQ
J1015.0+4925 1H 1013+498 10 15 4.13 49 26 00.87 0.212 15.13 HSP BL Lac
J1204.3$-$0708 1RXS J120417.0$-$070959 12 04 16.66 $-$7 10 09.16 0.184 14.59 HSP BL Lac
J1229.1+0202 3C 273 12 29 06.70 02 03 08.59 0.158 14.09 LSP FSRQ
J1404.8+0401 MS 1402.3+0416 14 04 50.92 04 02 02.38 0.344 17.08 HSP BL Lac
J1512.8$-$0906 PKS 1510$-$08 15 12 50.54 $-$09 05 59.69 0.360 15.70 LSP FSRQ
J1653.9+3945 Mkn 501 16 53 52.21 39 45 36.31 0.034 9.07 HSP BL Lac
---------------- ------------------------- ------------- ---------------- ------- ------- ------------
Characteristics proEM CCD normal CCD
------------------------------------------------ -------------------------- --------------------
Array (pixels) 1024 $\times$ 1024 2048 $\times$ 4096
Pixel size (micron) 13 15
Plate scale (arcsec pixel$^{-1}$) 0.26 0.30
Full well capacity (e$^{-}$) 8 $\times$ 10$^4$ 2 $\times$ 10$^5$
Operating Temperature $-$68$^{\circ}$ C $-$100$^{\circ}$ C
Dark current (e$^{-}$ pixel$^{-1}$ sec$^{-1}$) 0.04 $4\times10^{-4}$
Read out noise (e$^{-}$) variable 3.8
Gain (e$^{-}$ ADU$^{-}$) variable 0.74
proEM CCD Gain
Frequency High gain Medium gain
100 kHz 0.699 1.38
1 MHz 0.693 1.37
5 MHz 1.19 2.26
proEM CCD read out noise
Frequency High gain Medium gain
100 kHz 3.87 4.74
1 MHz 6.24 10.35
5 MHz 13.0 14.20
---------------- ------ ------------- ---------------- -------
Source Star RA (J2000) Dec (J2000) $R$
(mag)
J0608.0$-$0835 S1 06 07 54.01 $-$08 33 37.23 15.22
S2 06 08 00.59 $-$08 35 51.63 15.18
J0656.4+4232 S1 06 56 05.95 42 38 52.16 14.71
S2 06 56 17.26 42 35 52.35 14.63
J0710.5+4732 S1 07 10 53.52 47 30 13.35 12.89
S2 07 10 38.60 47 33 14.51 12.80
J0721.9+7120 S1 07 21 52.29 71 18 17.92 13.11
S2 07 22 17.99 71 23 34.83 13.55
J0739.4+0137 S1 07 39 16.14 01 37 34.93 15.02
S2 07 39 23.38 01 35 28.24 15.15
J0809.8+5218 S1 08 10 03.24 52 18 57.82 14.56
S2 08 10 17.08 52 19 43.77 15.17
J0831.9+0430 S1 08 32 04.43 04 25 52.38 14.44
S2 08 31 47.41 04 33 11.46 14.55
J0854.8+2006 S1 08 54 55.20 20 05 42.43 14.91
S2 08 54 53.36 20 04 45.12 14.03
J0922.4$-$0529 S1 09 22 26.71 $-$05 29 32.06 15.27
S2 09 22 26.88 $-$05 29 57.14 14.76
J0927.9$-$2037 S1 09 27 49.35 $-$20 33 10.49 16.96
S2 09 27 57.76 $-$20 36 45.89 16.06
J1006.7$-$2159 S1 10 06 45.09 $-$21 59 22.18 15.45
S2 10 06 33.05 $-$21 58 29.60 15.25
J1015.0+4925 S1 10 15 08.03 49 25 42.33 13.74
S2 10 14 53.85 49 25 32.43 13.83
J1204.3$-$0708 S1 12 04 06.14 $-$07 07 56.82 15.41
S2 12 04 32.23 $-$07 13 24.71 16.12
J1229.1+0202 S1 12 29 03.21 02 03 18.88 12.79
S2 12 29 08.40 02 00 18.71 11.79
J1404.8+0401 S1 14 04 49.21 04 03 37.55 16.22
S2 14 05 00.89 04 00 12.71 16.61
J1512.8$-$0906 S1 15 12 52.82 $-$09 06 59.46 14.12
S2 15 12 44.30 $-$09 06 40.14 13.60
J1653.9+3945 S1 16 53 45.82 39 44 09.14 12.49
S2 16 53 28.51 39 46 59.20 13.12
---------------- ------ ------------- ---------------- -------
---------------- ------------ ----- ---------- ---------- ---------- -------- ---------- ---------- ---------------- -------- -------- ---------------------------------
Source Date N Duration [*C1*]{} [*C2*]{} Status [*F1*]{} [*F2*]{} $F_{\rm crit}$ Status $\psi$ Pol.
dd mm yyyy (hr) ($\%$) ($\%$)
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13)
J0608.0$-$0835 07 03 2015 15 1.29 1.87 2.15 NV 1.57 2.08 3.69 NV 5.25 —
03 01 2016 35 4.51 4.32 4.49 V 8.47 9.53 2.26 V 8.97 —
J0656.4+4232 05 01 2016 31 5.06 1.64 1.92 NV 0.77 1.06 2.39 NV 3.25 —
J0710.5+4732 26 02 2015 59 2.52 1.89 2.17 NV 1.36 2.06 1.86 NV 5.07 —
11 03 2015 136 4.31 1.37 1.35 NV 0.77 0.79 1.50 NV 2.25 —
04 01 2016 52 2.39 1.29 1.36 NV 0.68 0.81 1.94 NV 3.41 —
J0721.9+7120 21 01 2015 213 5.98 2.75 2.69 V 6.22 3.98 1.38 V 6.17 9.42$\pm$0.02
22 01 2015 92 4.21 2.86 2.99 V 8.39 5.73 1.63 V 5.06 5.48$\pm$0.02
23 02 2015 238 6.44 5.71 5.70 V 28.90 22.44 1.35 V 11.88 —
J0739.4+0137 17 03 2015 50 4.40 4.13 4.19 V 8.92 8.54 1.96 V 10.07 4.94$\pm$0.13
02 01 2016 25 5.05 3.77 3.65 V 1.38 1.26 2.72 NV 5.98 0.61$^{*}\pm$0.20 (12 01 2016)
J0809.8+5218 18 03 2015 28 4.82 1.34 1.57 NV 1.10 1.44 2.46 NV 2.02 —
J0831.9+0430 08 03 2015 14 1.27 4.66 4.55 V 9.36 8.35 3.70 V 5.57 —
09 03 2015 30 3.52 5.19 4.83 V 14.60 11.72 2.39 V 13.87 —
14 03 2015 76 5.41 4.04 3.88 V 7.34 6.52 1.71 V 9.17 —
J0854.8+2006 10 03 2015 160 4.27 3.50 3.39 V 8.53 12.15 1.45 V 9.65 13.38$^{*}\pm$0.07 (17 03 2015)
J0922.4$-$0529 12 03 2015 133 5.50 1.13 0.98 NV 1.27 1.13 1.50 NV 4.34 —
J0927.9$-$2037 13 03 2015 35 6.55 1.09 1.26 NV 0.99 1.54 2.26 NV 3.71 —
J1006.7$-$2159 24 02 2015 48 8.59 0.95 1.10 NV 0.72 1.17 1.99 NV 3.43 —
15 03 2015 10 2.61 1.08 1.29 NV 0.93 1.60 5.35 NV 2.21 —
16 03 2015 36 5.43 0.98 1.14 NV 0.76 1.27 2.23 NV 5.24 —
J1015.0+4925 20 03 2015 143 7.02 1.79 1.80 NV 1.14 1.17 1.48 NV 3.86 —
J1204.3$-$0708 25 02 2015 36 6.65 1.17 1.41 NV 1.34 1.30 2.23 NV 2.82 —
17 03 2015 30 4.05 1.62 1.78 NV 2.39 2.03 2.42 NV 4.24 —
J1229.1+0202 26 02 2015 15 0.36 1.26 0.86 NV 1.39 1.10 3.70 NV 2.30 0.28$^{*}\pm$0.06 (20 02 2015)
10 03 2015 68 1.60 1.15 1.27 NV 1.08 1.92 1.78 NV 2.53 0.31$^{*}\pm$0.04 (17 03 2015)
18 03 2015 300 4.56 1.20 1.05 NV 1.07 1.15 1.31 NV 6.46 0.31$^{*}\pm$0.04 (17 03 2015)
J1404.8+0401 20 05 2015 24 4.44 1.72 1.70 NV 1.86 1.58 2.72 NV 3.49 —
J1512.8$-$0906 21 05 2015 25 1.89 1.37 1.24 NV 1.19 1.28 2.66 NV 4.39 7.43$\pm$0.08
22 05 2015 39 2.33 5.22 5.24 V 21.48 31.97 2.16 V 15.66 15.32$\pm$0.05
J1653.9+3945 20 05 2015 151 4.20 5.17 5.20 V 14.11 12.71 1.46 V 11.89 2.20$\pm$0.05
14 07 2015 154 5.41 1.38 1.36 NV 0.93 0.83 1.46 NV 5.15 1.28$^{*}\pm$0.04 (15 07 2015)
---------------- ------------ ----- ---------- ---------- ---------- -------- ---------- ---------- ---------------- -------- -------- ---------------------------------
3FGL name Time (in days) Observatory
---------------- --------------------------- -------------
J0721.9+7120 $-1.67^{+11.33}_{-10.54}$ Steward
J0739.4+0137 $+1.24^{+8.96}_{-9.30}$ Steward
J0809.8+5218 $+39.36^{+3.93}_{-56.20}$ KAIT
J0831.9+0430 $+1.17^{+4.46}_{-4.60}$ KAIT
J0854.8+2006 $-2.29^{+13.37}_{-3.91}$ Steward
J1006.7$-$2159 $-4.01^{+29.67}_{-5.15}$ SMARTS
J1015.0+4925 $+18.4^{+11.35}_{-13.35}$ KAIT
J1229.1+0202 $-0.04^{+28.91}_{-4.37}$ Steward
J1512.8$-$0906 $+4.21^{+11.21}_{-10.54}$ Steward
J1653.9+3945 $-1.95^{+7.40}_{-13.95}$ Steward
{width="\linewidth"}
[^1]: http://www.ledas.ac.uk/DSSimage
[^2]: We note that the choice of observing a particular source is subjective and therefore our sample is not complete. It depends on the night sky condition, field of view (uncongested or crowded), and the brightness of nearby stars compared to the target quasar. Therefore, it is likely that we may be able to accommodate those blazars that were excluded from the first observing run.
[^3]: IRAF is distributed by the National Optical Astronomy Observatories, which is operated by the Association of Universities for Research in Astronomy Inc. under cooperative agreement with the National Science Foundation.
[^4]: http://fermi.gsfc.nasa.gov/ssc/data/access/lat/4yr\_catalog/ap\_lcs.php
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present some computer simulations run on a stochastic CA (cellular automaton). The CA simulates a gas of particles which are in a channel, the interval $[1,L]$ in $\mathbb Z$, but also in “reservoirs” $\mathcal R_1$ and $\mathcal R_2$. The evolution in the channel simulates a lattice gas with Kawasaki dynamics with attractive Kac interactions; the temperature is chosen smaller than the mean field critical one. There are also exchanges of particles between the channel and the reservoirs and among reservoirs. When the rate of exchanges among reservoirs is in a suitable interval the CA reaches an apparently stationary state with a non zero current; for different choices of the initial condition the current changes sign. We have a quite satisfactory theory of the phenomenon but we miss a full mathematical proof.'
author:
- 'M. Colangeli [^1], A. De Masi[^2], E. Presutti [^3]'
title: Particle models with self sustained current
---
[***Keywords:—*** Stochastic cellular automata, Kac potential, Fourier law and phase transition.]{}
Introduction {#sec.0}
============
In this paper we introduce models of macroscopic dissipative systems made of interacting particles which move stochastically in a circuit and exhibit a very surprising behavior. Despite the fact that there is no external bias we see, after a transient, an apparently stationary state with a non zero current, with suitably different initial conditions we may select another state with the opposite value of the current. We speculate that on much longer times there is a “dynamical phase transition” with the two states alternating one after the other. To make an analogy with equilibrium phase transitions, consider the 2D Ising model in a large but finite box with nearest neighbor ferromagnetic interactions. Running the Glauber dynamics at a temperature below the critical value we typically see long time intervals where the magnetization density has approximately the plus equilibrium value alternating via tunneling with those where it is close to the minus equilibrium value. The analogue of the equilibrium magnetization in our model is the current as we have two states with opposite values of the current. However we observe our circuit for times long but much smaller than those for tunneling so that we only see one of the two currents (selected by the initial condition) which then looks stationary. Our analysis relies mostly on computer simulations, we have theoretical explanations but we miss a mathematical proof.
There is a huge literature on the more general question of existence of periodic motions or oscillations especially in the context of biological systems and chemical reactions, the classical reference is the book by Kuramoto, [@kuramoto]. We just quote here a few examples selected with the purpose of introducing what we will be doing in this paper.
In [@tass] P. Tass discusses a simple system of rotators which interact attractively with each other and are subject to white noise forces. For small interactions the stationary state is homogeneous and even though each particle rotates there is no macroscopic change. However if the interaction increases the rotators form a macroscopic cluster which then moves periodically. This is a simplified model for neural activities, the angle of the rotator is related to the neuron potential and the crossing from $2\pi$ to 0 is interpreted as the neuron discharging its potential (“firing”): the appearance of a cluster causes a great potential change when the cluster crosses $2\pi$ which could explain some diseases related to anomalous neuron firing.
A quantum analogue of the rotator model has been studied by Wilczek in [@wilczek1] where it is shown that there are ground states with a localized cluster which rotates, this phenomenon called a “time crystal”. Comments on time crystals can be found in [@bruno]. Experimental evidence of “time crystals” are presented in [@zhang]. Time crystals in a classical (i.e. non quantum) context have been considered in [@wilczek2].
A rotators model is also considered in [@giacomin] where an additional external force is present. The main point in the paper is to show that for a critical set of values of the parameters there is a cluster which is however blocked (by the external force). However if the white noise strength is increased then the cluster starts moving and performs a periodic motion, this being a nice example of noise-induced periodicity. Also in our models noise is the fuel which makes the system run.
In the above models each particle by itself rotates: the macroscopic rotations arise from a “phase synchronization” of the rotators. Instead in the FitzHugh Nagumo class of models for the firing cycles of a neuron, the appearance of periodic motions is due to a different, more intrinsic mechanism. For what follows it is convenient to consider a particular model in the class which can and will be read in a statistical mechanics language. In such a context the model is defined by two (macroscopic) variables, the magnetization $m$ and the magnetic field $h$. $m$ is the “fast” and $h$ the “slow variable” as the evolution is defined by the equations: $$\label{0.1}
\eps \frac{dm}{dt} = -m + \tanh\{\beta (m + h)\},\quad \frac{dh}{dt} =-m$$ where $\eps>0$ is the “small parameter” and $\beta>1$ the inverse temperature. It can be seen that has a (stable) periodic solution which in the limit $\eps\to 0$ becomes the hysteresis cycle: $m= m_\pm(h)$, $m_{+}(h)$ the positive solution of $m = \tanh\{\beta (m + h)\}$ which exists for $h>-h_c$, $h_c>0$; $m_{-}(h)=-m_{+}(-h)$, $h<h_c$, see Fig. \[fig:1\]. The transition from the upper curve $m_{+}(\cdot)$ to the lower one $m_{-}(\cdot)$ (and viceversa) is discontinuous and hence very sharp for $\eps>0$ small, a fact which catches the main feature of the neuron voltage cycle namely that at the firing the potential changes very abruptly. Observe that $m_+(h)$ is metastable for $h<0$ as well as $m_-(h)$ for $h>0$, the metastable values of the magnetization will play a fundamental role also in this paper.
Dai Pra et al., [@daipra], derived similar patterns in a macroscopic limit from a Ising spin model with mean field interactions giving nice examples of “intrinsic” periodic oscillations in the stochastic Ising model. In this paper we will consider the relaxed version of mean field as defined by Kac potentials.
All the above examples can be interpreted in terms of a current in a circuit but in all of them there is a more or less hidden bias because the current can flow only in one direction and not in the opposite one, so that they do not fit in what we are looking for. However they have all a common feature with our models, namely the presence of a phase transition, responsible in the rotator models for the formation of a cluster and in the FitzHugh Nagumo models for the presence of a hysteresis cycle. The way phase transitions appear in our analysis is the following. In a first order phase transition there is a spontaneous separation of phases which gives rise to gradients of the order parameter without currents being present. The Fourier law associates to a gradient a current (in the opposite direction) so that the phase transition generates “effective forces” which prevent the gradients to give rise to currents. Our idea is to exploit such forces to construct a “battery” which allows for a non zero current in a circuit.
Our battery is a cellular automaton which simulates the Kawasaki dynamics in a lattice gas with interactions given by an attractive Kac potential which in the Lebowitz-Penrose limit has a van der Waals phase transition. Therefore we can distinguish between stable, metastable and unstable values of the density. The main and somehow unexpected feature of the system is that if we connect the endpoints of the channel to “infinite” (i.e. true) reservoirs which fix the density at values $\rho_- $ and $\rho_+= 1-\rho_->\rho_-$ with $\rho_{\pm}$ metastable densities we observe numerically a current which goes through the channel from the reservoir with smaller density $\rho_-$ to the one with the larger density $\rho_+$. We have a theoretical explanation of the phenomenon in terms of properties of the solution of an integral equation obtained from the process in the “mesoscopic limit” where the scaling parameter $\ga$ of the Kac potential vanishes, but we could verify these properties only numerically.
In [@CDP] we have presented numerical evidence that the current in the CA flows from the reservoir with smaller density to the one with larger density. In this paper we present a more complete set of simulations from where a very complex structure emerges for which we have a theoretical explanation, but we miss a complete mathematical proof. The other main point in this paper is that we can exploit the above to construct a circuit with a self sustained current without an external bias, as claimed in the first sentence of this Introduction. This is obtained by making the reservoirs finite and allowing also particles exchanges among the reservoirs. We show (via the simulations) that for suitable values of the parameters there are initial conditions which give rise to a steady non zero current (stationary for the times of our simulations); there are also other initial conditions where the current flows in the opposite direction and still others where there is no current at all. The state with zero current seems unstable while those with a non zero current seem locally stable.
As suggested by a referee, similar phenomena have also been studied in other models, e.g. the Bunimovich’s mushroom billiard model [@Bunim01] in which the presence of peculiar transport regimes can be traced back to the lack of ergodicity of the microscopic dynamics; but we have not yet explored this issue.\
The paper is organized as follows. In Section \[sec.j1\] we define two different versions of the CA used in the simulations, namely: one describing a single (open) channel in contact with two reservoirs (hereafter called OS-CA), and another mimicking the particle dynamics in a closed circuit (called CC-CA).\
In Section \[sec.j1.4\] we present the results of the simulations obtained by running the OS-CA and also explain how to run the CC-CA by exploiting the results first obtained with the OS-CA.\
In Section \[sec.j1.1\] we illustrate the behavior of the particle current in the CC-CA and comment on the dependence of this quantity on the parameters of the model.\
In Section \[sec.j1.3\] we study the continuum (mesoscopic) limits of both the OS-CA and the CC-CA, which are described by an integro-differential equation; proofs are deferred to the Appendix \[A1\].\
In Section \[sec.j1.5\] we discuss the adiabatic limit of the model, and check the consistency of our simulations of the CC-CA with the predicted adiabatic behavior.\
In Section \[sec6\] we consider the case where the reservoirs have stable densities and in Section \[sec7\] where the densities are not stable.\
In Section \[sec.j9\] we study the stability of a stationary density profile, referred to below as *the “bump” solution*, close to the boundary.\
Concluding remarks are finally drawn in Section \[sec:concl\].
The cellular automata {#sec.j1}
=====================
In this section we define two cellular automata: the first one, called “open system cellular automaton”, OS-CA in short, has been first introduced in [@LOP] and then used in [@CDP] to simulate a system in contact with reservoirs. The second one, simply called “closed circuit cellular automaton”, CC-CA, is a modification of the first one obtained by making finite the reservoirs and adding direct exchanges between them, so that it simulates a closed circuit.
[**The OS-CA.**]{} The OS-CA describes the evolution of particles in a “channel" $ \{1,2,..,L\}$, $L>1$ a positive integer. Besides moving in the channel particles may also leave from or enter into the channel through $L$ and $1$ (we then say that they are absorbed or released from the reservoir $\mathcal R_2$ if this happens at $L$ and from reservoir $\mathcal R_1$ if it happens at $1$). The two reservoirs are “infinite" in the sense that they do not have memory of the particles which are absorbed or released.
The CA in the channel is a parallel updating version of a weakly asymmetric simple exclusion process, designed for computer simulations. The $d=1$ symmetric simple exclusion process is a system of random walks jumping to the right and left with equal probability, the jump being suppressed if the arrival site is occupied. The weak asymmetry that we add is a small bias to jump in the direction where the density is higher. If the channel was a torus this would produce a phase separation into a region where the density is higher and another where it is smaller. But our channel is open as particles may leave or enter into the channel in a setup typical of the Fourier law but in a context where phase transitions are present.
Let us now go back to the definition of the CA. The phase space is $\mathcal S=\{(x,v), x\in\{1,..,L\}, v\in\{-1,1\}\}$, particles configurations are functions $\eta:\mathcal S\to \{0,1\}$, $\eta(x,v) \in \{0,1\}$ denotes the occupation variable at $(x,v)$ and $v$ will be interpreted as a velocity. $\eta(x)=\eta(x,-1)+\eta(x,1)\in\{0,1,2\}$ denotes the total number of particles at $x$. We may add a suffix $t$ when the occupation variables are computed at time $t$.
The definition of the OS-CA involves four more parameters: $\ga^{-1}\in \mathbb N$, $C>0$ and $\rho_{\pm}\in [0,1]$. In the simulations presented in this paper we have fixed $\ga^{-1}=30$, $C=1.25$, while the length of the channel is set equal to $L=600$. $\rho_{\pm}$ are referred to as the density of reservoir $\mathcal R_2$, respectively $\mathcal R_1$, they are fixed during a simulation but they may be changed in different simulations. In the definition of the CA we will use the notation $$\label{j1.2}
N_{+,x,\ga}
= \sum_{y=x+1}^{ x+\ga^{-1}}\eta^{(+)}(y),\;
N_{-,x,\ga}
= \sum_{y= x-\ga^{-1}}^{x-1}\eta^{(-)}(y),\quad x \in [1,L]$$ where $\eta^{(+)}(y)= \eta(y)$ if $y \in [1,L]$ and $\eta^{(+)}(y)= 2\rho_{+}$ if $y >L$; similarly $\eta^{(-)}(y)= \eta(y)$ if $y \in [1,L]$ and $\eta^{(-)}(y)= 2\rho_{-}$ if $y <1$. We want $N_{+,x,\ga}$ to be the total number of particles to the right of $x$ within distance $\ga^{-1}$ from $x$, however it may happen that if $x$ is close to the right boundary then there are not $\ga^{-1}$ sites in the channel to the right of $x$. Suppose that there are only $\ga^{-1}-m$ such sites, we then add fictitiously $2m$ phase points $(y,v)$, $v=\pm 1$ and $y$ takes $m$ values to be thought as $m$ physical sites to the right of the channel. The occupation number $\eta(y,v)$ is then set equal to $\rho_+$ so that the contribution to $N_{+,x,\ga}$ of the extra $m$ sites is $2\rho_+ m$, which explains the factor 2 in the definition of $\eta^{(+)}(y)$. Analogous interpretation applies to $\eta^{(-)}(y)$.
We are now ready to define how the OS-CA operates: the unit time step updating (from $t$ to $t+1$) is obtained as the result of three successive operations, we denote by $\eta$ the configuration at time $t$, by $\eta'$ and $\eta''$ two consecutive updates starting from $\eta$ and by $\eta'''$ the final update which gives the configuration at time $t+1$.
1. [*velocity flip*]{}. At all sites $x\in \{1,..,L\}$ where there is only one particle we update its velocity to become $+1$ with probability $\frac 12
+ \eps_{x,\ga}$ and $-1$ with probability $\frac 12
- \eps_{x,\ga}$, $\eps_{x,\ga}= C\ga^2[N_{+,x,\ga}-N_{-,x,\ga}]$ (the definition is well posed because $(2\ga^{-1}) C\ga^2= 2.5/30<\frac 12$, $(2\ga^{-1}) $ being an upper bound for $|N_{+,x,\ga}-N_{-,x,\ga}|$). At all other sites the occupation numbers are left unchanged. We denote by $\eta'$ the occupation numbers after the flip.
2. [*advection*]{}. After deleting the particles in the channel at $(1,-1)$ and $(L,1)$ (if present) we let each one of the remaining particles in the channel move by one lattice step in the direction of its velocity. We denote by $\eta''$ the occupation numbers after this advection step.
3. [*exchanges with the reservoirs*]{}. With probability $\rho_+$ we put a particle at $(L,-1)$ and with probability $1-\rho_+$ we leave $(L,-1)$ empty. We do independently the same operations at $(1,1)$ but with $\rho_-$ instead of $\rho_+$. The final configuration is then denoted by $\eta'''$.
[**The CC-CA.**]{} We now turn to the second CA which describes the evolution of particles in a “closed circuit". The phase space is the disjoint union $\mathcal S \cup \mathcal R_1\cup \mathcal R_2$, where $\mathcal S$ is as before while the two reservoirs $\mathcal R_1$ and $\mathcal R_2$ are finite sets both with cardinality $R$, $R$ a positive, even integer. $R$ is interpreted as the number of phase points in the reservoir, thus there will be $R/2$ sites with velocity $1$ and $R/2$ sites with velocity $-1$: the velocities in the reservoirs however do not play any role in the evolution, they are used only to have a symmetric description of the channel and the reservoirs. Unlike in the OS-CA now the total number of particles (i.e. those in the channel and in the reservoirs) is constant in time. In the CC-CA the densities $\rho_{\pm}$ in the two reservoirs are no longer constant but given by $N_{\mathcal R_1}/R$ and $N_{\mathcal R_2}/R$ where $$\label{j1.1}
N_{\mathcal R_1} = \sum_{(x,v)\in \mathcal R_1}\eta(x,v),\quad
N_{\mathcal R_2} = \sum_{(x,v)\in \mathcal R_2}\eta(x,v)$$ Accordingly we define $N_{\pm,x,\ga}$ in the CA as in but with $\rho_{\pm}$ replaced by the instantaneous values $N_{\mathcal R_1}/R$ and $N_{\mathcal R_2}/R$ of the density in $\mathcal R_1$ and $\mathcal R_2$. With these notation the first two steps of the evolution in the CA are the same as in the OS-CA. We call again $\eta'$ and $\eta''$ the configurations in the system after the first and the second step, with $\eta''=\eta'=\eta$ in $\mathcal R_1 \cup \mathcal R_2$ (i.e. the occupation numbers in the reservoirs are unchanged in the first two steps). In the third step instead they may change as we are going to see.
[*3. The new third step, (reservoirs exchanges).*]{} Its definition involves a new, suitably small parameter $\ga p > 0$. We first select with uniform probability a phase point $(x_1,v_1)\in \mathcal R_1$ and $(x_2,v_2)\in \mathcal R_2$: if $\eta(x_1,v_1)=0$ we set $\eta'''(1,1)=0$, if instead $\eta(x_1,v_1)=1$ we set $\eta'''(1,1)=1$. Analogously $\eta'''(L,-1)=0,1$ if $\eta(x_2,v_2)=0,1$. This concludes the definition of $\eta'''$ in the channel while in the reservoirs $\eta''' = \theta'''$, with $\theta'''$ defined as follows. We first define $\theta'$ by setting $\theta'(x,v)=\eta(x,v)$ for $(x,v)$ in $\mathcal R_1$ with $(x,v) \ne (x_1,v_1)$ and $\theta'(x_1,v_1)=0$. $\theta'(x,v)$ is defined analogously in $\mathcal R_2$. $\theta''(x,v)$ is obtained from $\theta'(x,v)$ by adding a particle in the first empty point of $\mathcal R_1$ (according to a fixed but arbitrary order) if $\eta'(1,-1)=1$, otherwise $\theta''=\theta'$ in $\mathcal R_1$. $\theta''$ is defined analogously in $\mathcal R_2$. Finally $\theta'''$ is obtained from $\theta''$ in the following way. With probability $1-\ga p $ we let $\theta'''=\theta''$ while with probability $\ga p $ we do the following: we choose with uniform probability $(y_1,v_1)\in \mathcal R_1$ and $(y_2,v_2)\in \mathcal R_2$ and exchange $\theta''(y_1,v_1)$ with $\theta''(y_2,v_2)$. To be well defined we have tacitly supposed that $ \ga p \le 1$, actually $ \ga p \ll 1$ in the simulations.
Heuristically $\ga p$ is the rate at which particles jump directly from a reservoir to the other. Without the channel these exchanges would eventually make the densities of the two reservoirs equal to each other.
[**Magnetization variables.**]{}
To exploit the symmetries in the system it is convenient to introduce spin variables. We set in the CC-CA: $$\label{j1.13}
\si(x) = \eta(x,1)+\eta(x,-1)-1$$ both in the channel and in the reservoirs. (possibly adding $t$ when the variables are computed at time $t$). We call ${\displaystyle}{S_{\rm ch}= \sum_{x=1}^L\si(x)}$ the total spin in the channel, thus $$\label{2.4aA}
S_{\rm ch} = N_{\rm ch}-L,\qquad N_{\rm ch}:=\sum_{x=1}^L\eta(x)$$ Recalling , we define analogously to $$\label{j1.20.3.1}
S_{\mathcal R_1}= N_{\mathcal R_1} - \frac R2,\quad S_{\mathcal R_2}=N_{\mathcal R_2}- \frac R2$$ We define also the magnetization density in the two reservoirs $$\label{2.5aa}
m^{CC}_{-} =\frac { S_{\mathcal R_1}}{R/2}
,\qquad m^{CC}_{+} =\frac { S_{\mathcal R_2}}{R/2}$$ In the OS-CA the magnetization density in the “reservoirs” is $$\label{2.5bb}m_{\pm}= 2\rho_{\pm} -1$$
[**Currents.**]{} For the OS-CA we define $j_{\mathcal R_1\to {\rm ch}}(t)$ and $j_{{\rm ch} \to\mathcal R_2}(t)$ as the number of particles which go from $\mathcal R_1$ to the channel minus those which go from the channel to $\mathcal R_1$ in the time step $t\to t+1$ and respectively, the number of particles which go from the channel to $\mathcal R_2$ minus those which go from $\mathcal R_2$ to the channel in the time step $t\to t+1$. Thus $$\begin{aligned}
&&j_{\mathcal R_1\to {\rm ch}}(t)= \eta'''(1,1;t) - \eta'(1,-1;t){\nonumber}\\ \label{2.7aa} \\&&{\nonumber}j_{{\rm ch} \to\mathcal R_2}(t)= \eta'(L,1;t)-\eta'''(L,-1;t)
$$ with $\eta'$, $\eta''$ and $\eta'''$ the occupation numbers after the three updates which lead from $t$ to $t+1$.
In the CC-CA the currents $j^{CC}_{\mathcal R_1\to {\rm ch}}(t)$ and $j^{CC}_{{\rm ch} \to\mathcal R_2}(t) $ are defined by the same expression as in with the new $\eta$’s. The current between the reservoirs is defined as the number of particles which go from $\mathcal R_2$ to $\mathcal R_1$ minus those which go from $\mathcal R_1$ to $\mathcal R_2$ in the time step $t\to t+1$, thus: $$\begin{aligned}
\label{2.9ab}
j_{\mathcal R_2\to \mathcal R_1}(t)
= -\sum_{(x,v)\in \mathcal R_2} [\theta'''(x,v;t)-\theta''(x,v;t)]\end{aligned}$$
[**Conservation laws.**]{} In the OS-CA we have $$\label{2.9aA}
N_{\rm ch}(t+1)- N_{\rm ch}(t)= j_{\mathcal R_1\to {\rm ch}}(t)-j_{{\rm ch} \to\mathcal R_2}(t)$$ In the CC-CA the analogue of holds as well: $$\label{2.10aA}
N_{\mathcal R_1} (t+1)- N_{\mathcal R_1} (t)= j^{CC}_{\mathcal R_2\to \mathcal R_1}(t)-
j^{CC}_{\mathcal R_1\to {\rm ch}}(t)$$ with analogous formula for $N_{\mathcal R_2}$. As a consequence, in the CC-CA, the total number of particles $N_{\mathcal R_1} +N_{\mathcal R_1} + N_{\rm ch}$ is conserved as well as the total spin $S_{\mathcal R_1} +S_{\mathcal R_1} + S_{\rm ch}$.
[**Initial conditions.** ]{} In the OS-CA we impose $\rho_+> \rho_-$, $\rho_++\rho_-=1$. Observe that this implies $m_+>0$ and $m_++m_-=0$. Analogously in the CC-CA we initially impose that $$\label{j1.3.1}
N_{\mathcal R_1} + N_{\mathcal R_2} = R, \quad N_{\mathcal R_2} > \frac R2$$ The initial state in the channel will be specified in the sequel.
[**Parameters of the simulations.**]{} We conclude the section by recalling the values of the parameters that will be used in the simulations: $$\label{j1.3.0}
\ga^{-1}=30, \;C=1.25, \;\beta=2.5,\;L=600,\;
R= 10^5,\quad R=:\ga^{-1}a,\;L =: \ga^{-1} \ell$$
The OS-CA {#sec.j1.4}
=========
In this section we present the simulations obtained by running the OS-CA, recall that this CA has been defined in terms of two fixed densities $\rho_+$ and $\rho_-$, $\rho_++\rho_-=1$, which in the magnetization variables, see , amounts to fix $m_+>0$, $m_-=-m_+$. As already mentioned the OS-CA simulates the typical Fourier law experiments therefore the physically most relevant quantity is the stationary current $j(m_+)$: $j=j(m_+)$ plays the role of the equation of state in a non equilibrium context (due to the presence of the reservoirs) and defines the “non equilibrium thermodynamics” of the system.
Thus our first task is to consider the currents in the CA, since the instantaneous currents defined in are strongly fluctuating, we take averages: $$\begin{aligned}
\label{j1.20.2.1}
&&j^{T}_{\mathcal R_1\to {\rm ch}}= \frac 1{T}\sum_{t=0}^{T-1}
j_{ \mathcal R_1\to {\rm ch}}(t)\end{aligned}$$ In general with $f^T$ we will denote the average of $f$, thus $j^{T}_{{\rm ch} \to\mathcal R_2}$ is the averaged current from the channel to $\mathcal R_2$.
Strictly speaking stationarity is reached as $T\to \infty$, existence of the limit should follow (almost everywhere) from the Birkhoff theorem. Of course in the simulations we cannot take such a limit and the value of $T$ is chosen empirically in such a way that $j^{T}_{{\rm ch} \to\mathcal R_2}
\approx j^{T}_{\mathcal R_1\to {\rm ch}}$ looks independent of $T$. The initial condition in the channel is with all phase points empty, we have checked that with other conditions the final current does not change appreciably. The stationarity condition $j^{T}_{{\rm ch} \to\mathcal R_2}(m_+)
- j^{T}_{\mathcal R_1\to {\rm ch}}(m_+)\approx 0$ is also satisfied, typical values are $10^{-8}$ while the currents have order $10^{-5}$. $10^{-8}$ is also considerably smaller than the a-priori bound $$|j^{T}_{\mathcal R_1\to {\rm ch}}- j^T_{ {\rm ch}\to \mathcal R_2}| \le \frac{2L}{T} = 1.2 \cdot 10^{-7}, \text{when $T= 10^{10}$}$$
The black dots in Fig. \[fig:2\] are the values of $j^{T}_{{\rm ch} \to\mathcal R_2}(m_+)$ in the simulations done with $T= 3 \cdot 10^9$ for $m_+ \in (0,m')$ and $m_+>m'''$ and $T= 10^{10}$ elsewhere. The continuous line in Fig. \[fig:2\], denoted by $j(m_+)$, is a continuous interpolation of $j^{T}_{{\rm ch} \to\mathcal R_2}(m_+)$ which we presume to be a good approximation of simulations done with the other values of $m_+$, it is therefore the “experimental” value for the non equilibrium equation of state $j=j(m_+)$. The main features in Fig. \[fig:2\] (where $m'= 0.500$, $m''= 0.825$, $m'''=0.912$ and $m^{iv}= 0.985$)) are:
- For $m_+\in (m^{iv},1]$ the current $j(m_+)$ is negative in agreement with the Fourier law, while for $m_+<m^{iv}$ the current is positive going from smaller to larger values of the magnetization (i.e. from $m_-$ to $m_+$).
- $j(m_+)$ is first increasing till $m'$, then decreasing till $m''$, again increasing till $m'''$ and finally decreasing then after.
- In Fig. \[fig:3\] we plot $j^{t}_{\mathcal R_1\to {\rm ch}}(m_+)t$ and $j^t_{ {\rm ch}\to \mathcal R_2}(m_+)t$, $t \le T$ with $m_+ \in [m',m''']$. We see significant fluctuations around the linear slope $j^{T}_{\mathcal R_1\to {\rm ch}}(m_+) t$, while for $m_+ \notin [m',m''']$ the fluctuations are “negligible”.
The most striking feature in the simulations is undoubtedly the fact that the current is positive when $m_+<m^{iv}$ so that it flows along the gradient going from the reservoir with smaller magnetization to the one with larger magnetization. If we dropped the interaction among particles in the channel, namely put $\eps_{x,\ga}\equiv 0$, then the current would flow according to the Fourier law opposite to the gradient, namely from $\mathcal R_2$ to $\mathcal R_1$.
[**A heuristic argument.**]{} Let us now imagine to have two channels connected to $\mathcal R_1$ and $\mathcal R_2$, channel 1 is the channel considered so far while channel 2 is some other channel where the Fourier law is satisfied (for instance the OS-CA with no bias, $\eps_{x,\ga}\equiv 0$, or some simpler connection as the one discussed later). When $m_+<m^{iv}$, in channel 1 there is a current $j(m_+)$ going from $\mathcal R_1$ to $\mathcal R_2$, while in channel 2 the current is $j_2 = \kappa m_+$, $\kappa >0$, going from $\mathcal R_2$ to $\mathcal R_1$ (recall $m_-=-m_+$). Thus in a time $t$ the reservoir $\mathcal R_1$ will loose a magnetization $j(m_+)t$ through channel 1 and gain a magnetization $\kappa m_+t$ through channel 2; the opposite happens to $\mathcal R_2$. This will go forever because the reservoirs in the OS-CA are not changed by what comes and goes; if instead the reservoirs were realized by large but finite systems (as in CC-CA) then after a time which depends on the size of the reservoirs and the difference $j(m_+)- \kappa m_+$ the magnetization in the reservoirs would change and stationarity would be lost. However if we choose channel 2 so that $\kappa m_+ = j(m_+)$ there is a perfect balance so that what $\mathcal R_1$ gives to $\mathcal R_2$ through channel 1 comes back from channel 2. We may thus hope that even if the reservoirs are finite (yet sufficiently large) this is again approximately true and that there is a non zero current which looks stationary for long times.
The simplest choice for channel 2 leads to the CC-CA of Section \[sec.j1\] where channel 2 is made by just allowing direct exchanges between the two reservoirs. Then, as we shall see later, the average current in the CC-CA from $\mathcal R_2$ to $\mathcal R_1$ is equal to $\ga p m_+$, hence the conjecture that for such a particular value of $\ga p$ there is a non zero stationary current in the circuit which is close to $j(m_+)$.
To check this we have defined for each $m_+<m^{iv}$ in Fig. \[fig:2\] $\ga p= j(m_+)/m_+$ as a function of $m_+$, see Fig. \[fig:4\].
We have then run the CC-CA with such values of $\ga p$, putting $m_{+}(0) = m_+$ in $\mathcal R_2$, $m_{-}(0) = -m_+$ in $\mathcal R_1$ and choosing the initial state in the channel equal to the configuration in the OS-CA simulation at the final time $T$. For all the values of $m_+$ considered in Fig. \[fig:2\] we have run the CC-CA for a same time $T= 3\cdot 10^9$ and computed the averaged currents $j^{T,CC}_{\mathcal R_1\to {\rm ch}}$, $j^{T,CC}_{{\rm ch} \to\mathcal R_2}$ and $j^{T,CC}_{\mathcal R_2\to \mathcal R_1}$ defined as in . Recalling we have also defined the averaged magnetization $m^{T,CC}_{\pm}$ in the two reservoirs writing $
j^{T,CC}_{\mathcal R_2\to \mathcal R_1}(m_+), j^{T,CC}_{\mathcal R_1\to {\rm ch}}(m_+)$, $j^{T,CC}_{{\rm ch} \to\mathcal R_2}(m_+), m^{T,CC}_{\pm}(m_+)
$ when we want to underline that the values are obtained starting from $m_+$.
The previous heuristic argument suggests that the three currents above are all close to each other and thus approximately equal to $j^{T}_{{\rm ch} \to\mathcal R_2}(m_+)$ and moreover that $m^{T,CC}_{\pm}(m_+)\approx
\pm m_+$. In the next section we will see what the simulations say.
Self sustained currents {#sec.j1.1}
=======================
Fig. \[fig:5\] is obtained by running the CC-CA in the setup described at the end of the previous section. It reports the values of the differences $10^5[j^{T,CC}_{\mathcal R_2\to \mathcal R_1}(m_+)-j^{T,CC}_{{\rm ch} \to\mathcal R_2}(m_+)]$ and $m_+^{T,CC}(m_+)-m_+$ as a function of $m_+$, recall from Fig. \[fig:2\] that the typical values of the current have order $10^{-5}$.
3 pt
We have also reported for each $m_+$ in Fig. \[fig:2\] the values of the pair $(\ga p,j^{T,CC}_{\mathcal R_2\to \mathcal R_1})$, see Fig. \[fig:6\] left, the continuous line is obtained by interpolating between such values. Analogously in Fig. \[fig:6\] right the dots are the values of $(\ga p,m_+^{T,CC})$ and the continuous line is obtained by interpolation. The continuous lines are multi-valued functions denoted respectively by $j^{CC}(\ga p)$ and $m_+^{CC}(\ga p)$, we presume they are a good approximation of what would be obtained by following the same procedure for other values of $m_+$ in Fig. \[fig:2\].
Let us point out the main features of our simulations.
3 pt
- Fig \[fig:5\] shows that the simulations are in good agreement with the conjectures stated at the end of the previous section except in the interval $m_+ \in (m'',m''')$. The values of $ j^{T,CC}_{\mathcal R_2\to \mathcal R_1}(m_+)$ and $m_+^{T,CC}(m_+)$ when $m_+ \in (m'',m''')$, are however approximately the same as those obtained for different values of $m_+$, see the black circles in Fig. \[fig:6\]).
- The values of $\ga p$ are all in the interval $(0,q_c)$, $q_c= 11.25\times10^{-5}$, and $j^{CC}(\ga p)$ is positive for all such values of $\ga p$. We have also done simulations with $\ga p > q_c$ with several choices of the initial condition and we have always seen zero current (not reported here).
- $j^{CC}(\ga p)$ is multi-valued, it has two distinct branches (separated from each other), the upper one in the interval $(0,q'')$, $q''=5.26\times10^{-5}$, the lower one in the interval $(q',q_c)$; $q''>q'$, $q'=1.98\times10^{-5}$. In the interval $(q',q'')$ there are two positive currents different from each other.
- $m_+^{CC}(\ga p)$ has the analogous structure, being two valued in $(q',q'')$. Both branches are decreasing, $m_+^{CC}(\ga p)\to m^{iv} = 0.985$, as $\ga p \to 0$, and to $0$ as $\ga p \to q_c$.
- There is a gap in the range of $m_+^{CC}(\ga p)$, namely the interval $(m',m'')$.
[**Conclusions.**]{} The simulations in Fig. \[fig:5\] show good agreement with the conjectures of Section \[sec.j1.4\] except when $m_+\in (m'',m''')$. Thus, with such exception, we may say that the stationary state found in the OS-CA evolution persists in the CC-CA provided that $\ga p=j(m_+)/m_{+}$.
There is no mystery about the current between the two reservoirs being $\ga p(m_{+}-m_-)/2\approx \ga p m_+$ because we can prove (see Appendix \[aappA\]) that $$\label{C.0}
E\Big[ \{j^{T,CC}_{\mathcal R_2\to \mathcal R_1} - \ga p \frac 12
[ m_+^{T,CC}-m_-^{T,CC}]\}^2 \Big] \le \frac{\ga p}T +16 \frac {(\ga p)^2}R
+ \; \text{\rm corrections}$$ Since $\ga p \approx 10^{-5}$, $R = 10^{5}$ and $T \approx 10^{9}$, the corrections have order $10^{-19}$, see .
Fig. \[fig:6\] can be obtained from Fig. \[fig:2\]: in fact according to the above statements $j^{CC}(\ga p)$ is (approximately) equal to $j(m_+)$ with $j(m_+)= \ga p m_+$. Since this may have multiple roots, $j^{CC}(\ga p)$ will be correspondingly multi-valued. However the roots with $m_+\in (m'',m''')$ are absent in the simulations (see the black circles in Fig. \[fig:6\]) but their values are the same as those obtained with other values of $m_+$. Same if we look at $m^{CC}(\ga p)$ and compare with Fig. \[fig:2\].
As a conclusion we have a consistent explanation of what seen in the OS-CA and the CC-CA, but we still need to explain (i) what happens when $m_+\in (m'',m''')$; (ii) why the typical values of $j(m_+)$ have order $10^{-5}$ which is much smaller than $1/L \approx 10^{-3}$ which is what expected from Fourier law experiments; (iii) why the true reservoir current has the behavior shown in Fig. \[fig:2\].
We can gain a theoretical insight on what is going on by looking at what happens in the mesoscopic limit $\ga \to 0$ which we study in the next section.
The mesoscopic limit {#sec.j1.3}
====================
This is defined by letting $\ga \to 0$ with $$\label{j1.12}
L= \ga^{-1} \ell,\quad R = \ga^{-1}a,\quad \ell,a >0\;{\rm fixed}$$ In the channel space and time are scaled diffusively, thus $x \to r =\ga x$ and $t \to \tau = \ga^{2}t$. In mesoscopic units the channel after the limit $\ga\to 0$ becomes the real interval $[0,\ell]$. We will prove existence of the limit (for the relevant quantities) under the assumption of a strong form of propagation of chaos, the details are given in an appendix.
We denote by $E_\ga$ the expectation in the CA processes (randomness coming from the initial datum and from the updating rules of the CA’s).
[**Assumptions**]{}. We suppose that
1. In both CA the limit below (denoted in the same way for both CA) exists and is smooth $$\label{5.3a}
\lim_{\ga \to 0} \lim_{\ga x\to r, \ga^{2}t\to \tau} E_\ga[ \eta(x,v,t)] =\frac{m(r,\tau)+1}2, \quad r\in[0,\ell], v\in\{-1,1\}, \tau\ge 0$$
2. In the CC-CA $$\label{j1.15.2}
\lim_{\ga \to 0, \ga^{2}t\to \tau} m_{\pm,\ga}(\ga^2 t)=
m_{\pm}(\tau)$$ where, recalling , we have set $m_{\pm ,\ga}(\ga^2 t):= E_\ga[m^{CC}_{\pm} (t)]$
3. In both CA for all $r,r_1,r_2 \in (0,\ell)$, $r_1\ne r_2$, $v\in\{-1,1\}$ and $\tau\ge 0$ $$\begin{aligned}
{\nonumber}&& \lim_{\ga \to 0} \lim_{\ga x\to r, \ga^{2}t\to \tau} | E_\ga[ \eta(x,v,t)\eta(x,-v,t)]
- E_\ga[ \eta(x,v,t)]E_\ga[\eta(x,-v,t)] | =0 \\\label{j1.15.3}
\\&&
\lim_{\ga \to 0} \lim_{\ga x\to r_1, \ga y\to r_2, \ga^{2}t\to \tau} | E_\ga[ \eta(x,t)\eta(y,t)]
- E_\ga[ \eta(x,t)]E_\ga[\eta(y,t)] | =0
{\nonumber}\end{aligned}$$
4. In the CC-CA for all $\tau\ge 0$ $$\label{j1.15.44}
\lim_{\ga \to 0, \ga^{2}t\to \tau} R^{-1} E_\ga\Big[ \big|N_{\mathcal R_i}(t)
-E_\ga[N_{\mathcal R_i}(t)]\big|\Big]=0,\;\; i=1,2$$
In Appendix \[A1\] we will prove the following two Theorems.
\[Mesoscopic limit\] \[thm5.1\] Under the above assumptions, in both CA, the limit magnetization $m(r,t)$ satisfies: $$\begin{aligned}
\label{j1.16}
&&\frac{\partial}{\partial t}m(r,t) = -
\frac{\partial}{\partial r} I(r,t),\quad r \in (0,\ell)
\\&& \hskip-.3cm
I(r,t)= -\frac 12\big\{\frac{\partial m(r,t)}{\partial r}
- 2C[1-
m(r,t)^2]\int_{r}^{r+1}[m(r+\xi,t)
- m(r-\xi,t)] d\xi\big\}{\nonumber}\end{aligned}$$ with $m(r+\xi,t)= m_+(t)$ if $r+\xi\ge \ell$ and $m(r-\xi,t)= m_-(t)$ if $r-\xi\le 0$ in the CC-CA; same expression holds in the OS-CA but with $m_\pm(t)$ replaced by $m_\pm$. Moreover $$\label{5.7}
m(0,t)=m_{-},\quad m(\ell,t)=m_{+},\qquad \text{in the OS-CA}$$ while in the CC-CA $$\label{j1.17}
m(0,t)=m_{-}(t),\quad m(\ell,t)=m_{+}(t) $$ $$\label{j1.18}
\frac{d}{dt}m_{+}(t) = \frac 1a \Big(2I(\ell, t)+ p [m_{-}(t)-m_{+}(t)]\Big)$$ $$\label{j1.19}
\frac{d}{d t}m_{-}(t) = \frac 1a \Big( -2I(0,t)+ p [m_{+}(t)-m_{-}(t)]\Big)$$
A proof which avoids our assumptions of propagation of chaos has been obtained in [@GL] for a lattice gas with Kac potential and Kawasaki dynamics in a torus. In magnetization variables the system becomes the Ising model with Kac potential and the limit equation is . In [@LOP] it has been studied the macroscopic scaling limit of this system with space scaled by $\ga^{-\alpha}$ and time by $\ga^{-2\alpha}$, $\alpha>1$ ( $\alpha =1$ is the mesoscopic limit considered above).
\[Currents\] \[thm5.1b\] Denote by $j_{x,x+1}(t)$ the number of particles which in the time step $t, t+1$ cross the bond $(x,x+1)$, $x\in\{1,..,L-1\}$ (counting as positive those which jump from $x$ to $x+1$ and as negative those from $x+1$ to $x$). Then, under the above assumptions, in both CA, for all $r\in(0,\ell)$ and $\tau>0$ $$\label{B.12}
\lim_{\ga\to 0: \ga x\to r} \ga \sum_{t=0}^{T-1} E_\ga[j_{x,x+1}(t)]= \int_0^\tau I(r,s)ds,\qquad T=[\ga^{-2}\tau]$$ $$\begin{aligned}
\label{j1.19.1}
&&\lim_{\ga\to 0}\ga \sum_{t=0}^{T-1} E_\ga[ j_{\mathcal R_1\to {\rm ch}}(t)]= -
\int_{0}^{\tau} I(0,s)ds,{\nonumber}\\&&
\lim_{\ga\to 0}\ga \sum_{t=0}^{T-1} E_\ga[j_{{\rm ch}\to \mathcal R_2}(t)]=
\int_{0}^{\tau} I(\ell,s)ds
\end{aligned}$$ where $I(r,s)$ is given in .
In the CC-CA the current between reservoirs converges by to: $$\begin{aligned}
\label{j1.19.1.1}
&&\lim_{\ga\to 0}\ga \sum_{t=0}^{T-1} E_\ga[ j_{\mathcal R_2\to \mathcal R_1}(t)]=
\int_{0}^{\tau} p m_+(s)ds\end{aligned}$$
In the simulations we have plotted the quantity $ j^{T}_{{\rm ch}\to \mathcal R_2}$. This is related by to the mesoscopic current $I$ by $$\begin{aligned}
\label{5.12a}
&&E_\ga[ j^{T}_{\mathcal R_1\to {\rm ch}}]=\frac{\ga}{T} \sum_{t=0}^{T-1} E_\ga[ j_{\mathcal R_1\to {\rm ch}}(t)] \approx - \frac {\ga}{\tau}
\int_{0}^{\tau} I(0,s)ds ,{\nonumber}\\&&
E_\ga[ j^{T}_{{\rm ch}\to \mathcal R_2}]=\frac{\ga}{T} \sum_{t=0}^{T-1} E_\ga[j_{{\rm ch}\to \mathcal R_2}(t)] \approx \frac {\ga}{\tau}
\int_{0}^{\tau} I(\ell,s)ds
{\nonumber}\\&&
E_\ga[ j^T_{\mathcal R_2\to \mathcal R_1}]=
\frac{\ga}{T} \sum_{t=0}^{T-1} E_\ga[ j_{\mathcal R_2\to \mathcal R_1}(t)]
\approx \frac {\ga}{\tau}
\int_{0}^{\tau} p m_+(s)ds
\end{aligned}$$ so that the experimental values of the three currents scale all as $\ga$ when $\ga\to 0$.
We next show that there is a natural interpretation of the solutions of the system – in terms of statistical mechanics, which then allows to relate what seen in the simulations to phase transitions and metastable–unstable magnetization values.
[**Free energy functional and thermodynamic potentials.**]{} The evolution equation in $[0,\ell]$ with periodic boundary conditions is the gradient flow relative to a non local free energy functional $F(m)$, in fact $$\begin{aligned}
\label{j1.23}
&&
I(r)= - \chi \frac {\partial}{\partial r}\frac {\delta F(m)}{\delta m(r)},\quad \chi =\frac\beta 2(1-m^2), \quad \beta=2C
\\&& F(m) = \int \Big(-\frac {m^2}{2} - \frac {S}{\beta}\Big) + \frac 14\int\int
J (r,r') [m(r)-m(r')]^2
{\nonumber}\\&&S(m)= -\frac{1-m} {2}\log \frac{1-m} {2} - \frac{1+m} {2}\log \frac{1+m} {2}
{\nonumber}\\&& J(r,r')=1-|r-r'|,\qquad \text {for } |r-r'|\le 1\quad \text{ and }=0 \quad \text {elsewhere}
{\nonumber}\end{aligned}$$ $F(m)$ is “the mesoscopic free energy functional”, the Ginzburg-Landau functional is a local approximation of $F(m)$ where the non local term becomes a gradient squared. The corresponding gradient flow evolution is the Cahn-Hilliard equation, which can then be viewed as a local approximation of .
The important point for us is that $F(m)$ specifies the thermodynamics of the system. In fact $$\begin{aligned}
\label{j1.24}
&&
f_\beta(m)= -\frac {m^2}{2} - \frac {S(m)}{\beta}\end{aligned}$$ is the van der Waals mean field free energy; its convex envelope $f^{**}_\beta(m)$ is the thermodynamic free energy. $f^{**}_\beta(s)$ is obtained by minimizing $F(m)/\ell$ under the constraint $\int m(r) = \ell s$ and then taking the limit $\ell \to \infty$, see for instance [@presutti], Ch. 6.
The equilibrium magnetization density when there is a magnetic field $h$ is the solution of the mean field equation $$\begin{aligned}
\label{j1.25}
&&
m = \tanh\{\beta (m+h)\}\end{aligned}$$ When $\beta>1$ there is $h_c(\beta)>0$ so that for any $|h|<h_c(\beta)$, $f_\beta(m,h)=f_\beta(m)-hm $ is a double well function of $m$. The local minima are $m_+(h)$ and $m_-(h)$ and their graph is the hysteresis cycle, see Fig. \[fig:1\]. In particular at $h=-h_c(\beta)$, $m_+(h)= m^*$ $$\begin{aligned}
\label{j1.26}
&&
m^*>0 : \beta [1-(m^*)^2] =1\end{aligned}$$ so that the magnetization in $(m^*,m_\beta)$ and in $(-m_\beta,- m^*)$ is metastable. At $h=0$ the double well is symmetric and the local minima are global minima, they are attained at $m=\pm m_\beta$, $m_\beta $ the positive solution of with $h=0$. $\pm m_\beta$ are the equilibrium magnetization at the phase transition with $h=0$ and $\beta>1$. $m_+(h)$ and $m_-(h)$ are the unique equilibrium magnetization at $h> 0$ and respectively $h<0$.
The adiabatic limit. {#sec.j1.5}
====================
Some of the characteristic parameters of the simulations are related to the thermodynamics associated to the mesoscopic equations, see the end of Section \[sec.j1.3\]. Indeed in fig \[fig:2\] which refers to simulations with the OS-CA, the value $ 0.985$ is very close to $m_\beta$ so that the simulation shows that the current is negative when $m_+$ is stable, namely $m_+ > m_\beta$ and positive when $m_+ < m_\beta$ (metastable or unstable). Correspondingly when there is a current in the CC-CA then $m_+ < m_\beta$, see fig. \[fig:6\] right. The above validates the considerations in the Introduction about the relation between the appearance of a current in the circuit and the occurrence of phase transitions.
Also the metastable region $(m^*,m_\beta)$ has a role in the simulations as the interval $(m'',m''')$ is a subset of $(m^*,m_\beta)$ (because $m''=0.825$ and $m'''=0.912$ while $m^*\approx 0.775$ and $m_\beta \approx 0.985$); thus the gap phenomenon (i.e. that some values of the magnetization in $\mathcal R_2$ are never seen for all $\ga p$) occurs only inside the metastable region. We turn now to the heuristic argument at the end of Section \[sec.j1.4\] by observing that it becomes rigorous in the context of the mesoscopic equations. In fact if $m$ is a stationary solution of the mesoscopic equation for the OS-CA when the reservoirs magnetizations are $m_\pm$, and the corresponding current $I$ is positive, then $u=m$ in the channel and $u=m_\pm$ in $\mathcal R_i$ $i=1,2$ is a stationary solution of the CC-CA mesoscopic equations with $p= \frac I{\frac 12 (m_+-m_-)}$, recall that the ratio between the mesoscopic current and the current in the CA scales as $\ga^{-1}$. Thus for sufficiently small $\ga$ we may expect to see what conjectured at the end of Section \[sec.j1.4\].
We can also give an explanation of the gap phenomenon (i.e. that for all $\ga p$ the magnetization in the reservoirs $\mathcal R_2$ is never in the interval $(m'',m''')$) by assuming that the evolution of the OS-CA is well approximated by the mesoscopic equations in the adiabatic limit that we are going to define. We first observe that the OS-CA can be regarded as the “infinite reservoirs limit” of the CC-CA, in fact in the limit $R\to\infty$ the updating rules of the CC-CA become those of the OS-CA. This is true also at the mesoscopic level: when $a\to\infty$ the magnetizations $m_\pm(t)$ converge to their initial value $m_\pm(0)$ and the evolution becomes that of the OS-CA. The above is true when we let $a\to \infty$ keeping the time finite, more interesting behaviour is seen if we scale time proportionally to $a$, which is the so called adiabatic scaling limit. Suppose (in agreement with the simulations in fig. \[fig:2\]) that for each value of $m_+$ (and with $m_-=-m_+$) there is a unique stationary solution of the mesoscopic equations for the OS-CA, $I_{\rm{stat}}(m_+)$ being the corresponding current. We then say that the CC-CA mesoscopic equations have a “good adiabatic behavior” if in the adiabatic limit the magnetizations $m_\pm(t)$ satisfy the equations $$\begin{aligned}
\label{j4.1}
\frac{d m_{+}(t)}{d t} = 2\Big( I_{\rm{stat}}(m_{+}(t)) - pm_{+}(t))
\Big) \qquad m_-(t)=-m_+(t)
\end{aligned}$$ Suppose now that $ I_{\rm{stat}}$ is positive with a graph like $j(m_+)$, see fig. \[fig:2\]. Then the stationary solutions of $p m_+= I_{\rm{stat}}(m_+)$ with $m_+\in(m'',m''')$ are linearly unstable because $I_{\rm{stat}}(m_+)$ is decreasing while $p m_+$ is increasing. Thus a small perturbation will lead the magnetization away from the stationary value $p m_+= I_{\rm{stat}}(m_+)$, $m_+\in(m'',m''')$, and presumably it will converge to one of the two other solutions of $p m_+= I_{\rm{stat}}(m_+)$. This may therefore explain why in the simulations we do not see the magnetization $m_+\in(m'',m''')$ and instead find another solution of $p m_+= I_{\rm{stat}}(m_+)$.
We can check experimentally whether the CC-CA has a good adiabatic behavior by doing simulations with non stationary initial data. In fig. \[fig:adia\] we report the experimental values and those obtained by solving numerically the adiabatic equations.
We do not have an analytic proof of good adiabatic behavior which instead can be rigorously proved for another particle model. This is the simple symmetric exclusion process in an interval with boundary processes at the endpoints which simulate reservoirs with densities $\rho_{\pm}(t)$ dependent on time. In [@DO] it is proved that in a scaling limit where $\rho_{\pm}(t)$ are “slowly varying” the current in the system becomes at each time $t$ the same as the stationary current when the densities at the endpoints are kept fixed at the values $\rho_{\pm}(t)$.
Summarizing, we have a reasonable explanation of the simulations in the CC-CA once we accept the behavior of the current $j(m_+)$ in the OS-CA as given in Fig. \[fig:2\]. To explain the latter we need to go deeper in the analysis of the simulations discussing the magnetization profile in the channel, which will be the argument of the remaining sections.
The instanton and the Stefan problem {#sec6}
====================================
We have a good understanding of what happens when $m_+\in (m_\beta,1]$. In Fig. \[fig:j3.2\] we plot the time evolution of the magnetization pattern when $m_+ = 1$, but a similar picture is observed for the other values of $m\in (m_\beta,1]$. The simulation shows convergence as time increases to a profile which is therefore stationary (in the times of the simulation) and it agrees with what found studying the mesoscopic equations. The existence of stationary solutions $m_{\rm st}(r;\ell;m_{\pm})$ of with boundary conditions when $m_+>m_\beta$ has been proved in [@DPT] for $\ell$ large enough. It is also shown that $$\label{j3.5}
\lim_{\ell \to \infty}m_{\rm st}(r\ell;\ell;m_{\pm})=m_{\rm st}(r;m_{\pm}),\quad r\in (0,1)$$ where the limit $m_{\rm st}(r;m_{\pm})$ is antisymmetric around $r= 1/2$ and satisfies the equation $$\label{j3.6}
-\frac 12 [1-\beta(1-m_{\rm st}^2)]\frac{dm_{\rm st}}{dr} =I_{\rm st}(m_+),\quad r\in[\frac 12,1]$$ where $I_{\rm st}(m_+)$ is determined by requiring that $m_{\rm st}(1/2)=m_\beta$ and $m_{\rm st}(1)=m_+$. For $m_+=1$, $\beta=2.5$ and $m_\beta=0.985$ from we get $I_{\rm st}(1)\simeq -7.2\times 10^{-3}$. To compare with the simulations we have to divide by $L=\ga^{-1}\ell= 600$ getting $-1.2 \times 10^{-5}$ the current in the simulations of the cellular automata OS -CA is instead $\simeq -2.2\times 10^{-5}$. The discrepancy is possibly due to $\ell$ not being large enough. In [@DPT] it is also proved that $$\label{j3.6.1}
\lim_{\ell \to \infty}m_{\rm st}(\frac 12 + x;\ell)=\bar m (x),\quad x\in \mathbb R$$ where $\bar m (x)$ is the instanton solution of $$\label{j3.6.2}
\bar m (x) = \tanh\{J*\bar m (x)\}$$ namely the antisymmetric function solution of which converges to $m_\beta$ as $x \to \infty$. See for instance [@presutti] for existence and properties of the instanton.
In Fig. \[fig:j3.2\] it is also plotted the time evolution of the magnetization pattern when starting away from the stationary one. The approach to the latter occurs on the time scale $L^2$.
$$\label{j3.9}
2m_\beta\frac{dr_t}{dt} = I(r_t^-,t) - I(r_t^+,t)$$
.5cm
We do not have a proof that $m(r,t)\to m_{\rm st}(r;m_{\pm})$ as $t\to \infty$ ($m_{\rm st}(r;m_{\pm})$ as in ). However if the pattern looks like the one in Fig. \[fig:j3.2\], i.e. essentially linear away from $\pm m_\beta$, then the current (being proportionally to the slope) when $r> r_t > 1/2$ is larger (in absolute value) than the one when $r< r_t$. Thus the magnetization increases and therefore $r_t$ moves to the left.
has been derived in [@LOP] from the spin dynamics on a torus when the initial profile $m_0(r)$ has values in $(m^*,1)$ for all $r$ or when it has values in $(-1,-m^*)$. The result does not apply in the case of the Stefan problem where there are both positive and negative values of the magnetization: the derivation of the Stefan problem for Ising spins with Kawasaki dynamics and Kac potential is still an open problem.
Boundary layers, the bump {#sec7}
=========================
When $m_+<m_\beta$ we get a completely different picture. Compare in fact the simulations in Fig. \[fig:j3.2\] and Fig. \[fig:j3.3\] where the initial state is the same but $m_+$ is stable in the former ($m_+=1$) and metastable ($m_+=0.93$) in the latter. In both cases, after a transient, we see a profile with a sharp (instanton-like) transition from $-m_\beta$ to $+m_\beta$ and then approximately linear profiles which connect $m_-$ to $-m_\beta$ and $m_\beta$ to $m_+$. But, in the stable case the instanton-like region moves towards the center, while in the metastable case it moves towards 0 which is eventually reached. The same \[heuristic\] argument which explained in the case of Fig. \[fig:j3.2\] the motion of the instanton towards the center, now explains its motion away from the center: since $m_+<m_\beta$ the slope of the pattern from the endpoint to the instanton is negative in the case of Fig. \[fig:j3.3\] (as it connects $m_-$ to $-m_\beta$ and $m_\beta$ to $m_+$); consequently the current in the interval from $m_-$ to $-m_\beta$ is positive and larger than the one from $m_\beta$ to $m_+$ (as the instanton in Fig. \[fig:j3.3\] is closer to 0 than to $\ell$), thus the total magnetization increases and the instanton moves further towards 0.
The transition region in the stable case is approximated by an instanton which is a stationary solution of the evolution equations on the whole line. Analogously, when $m_+<m_\beta$ we speculate that the transition region is approximated by a bump which is again a stationary solution $m(r)$, $r\ge 0$, of on the half line with zero current and given boundary condition at 0, say $\mu$, namely: $$\begin{aligned}
\label{j3.11}
&& m(r) = \tanh\Big\{ \beta[ J*m(r)+h]\Big\},\quad r \ge 0
\\&& h= -J*m(0) + \frac 1 \beta\tanh^{-1} \mu, \qquad m(r)=\mu \,\,\text{for $r<0$} \label{j3.11bis}
\end{aligned}$$ Indeed it can be easily seen that a stationary solution of with zero current is necessarily a solution of . The Gibbsian formula (in the mesoscopic limit) would give with $h=0$, thus the problem is not in the framework of the equilibrium theory. This is reflected by the appearance of an auxiliary magnetic field which has to be determined consistently with the magnetization pattern (as in the FitzHugh Nagumo models of the introduction where however by a mean field assumption the magnetization was simply a real number).
Observe that if $m(r)$ solves with boundary condition $\mu$ then $-m(r)$ solves with boundary condition $-\mu$, this symmetry will play an important role in the sequel. Besides the trivial solution $m(r)\equiv \mu$, existence of other solutions of is an open problem. The simulations indicate the existence of increasing solutions, we thus define:
[**Definition.**]{} [*The bump $B_{\mu}(r)$, $\mu \in (-m_\beta, m^{*})$, is a non constant solution of which is monotone non decreasing, We call $b(\mu)$ its asymptotic value:*]{} $$\label{j3.12}
\lim_{r \to \infty} B_{\mu}(r)=: b(\mu)$$
Analogously we call $B^-_\mu$, $m\in (-m^*,m_\beta)$ a non constant solution which is monotonic non increasing and denote by $b^-_\mu$ its asymptotic value. The existence $B_\mu$ implies the existence of $B^-_\mu$, in fact by simmetry $B^-_\mu=-B_{-\mu}$. Thus what we will say for $B_\mu$ extends to $B^-_\mu$ and in the sequel we will consider only $B_\mu$.
As mentioned above the existence of bumps is an open problem, the simulations indicate that bumps do indeed exist. The relation between bump and instanton can be understood in the following way. Call $\bar x (\mu)$ the value of $r$ such that $\bar m(r)=\mu$. Replace the boundary condition $m(r)=\mu$, $r<0$, in the definition of the bump by $m(r)= \bar m(r+\bar x (\mu))$, $r<0$. Then the solution of would be $m(r)= \bar m(r +\bar x (\mu))$, $r>0$, with $h=0$, the asymptotic value at $r=+\infty$ being $m_\beta$. Replacing $m(r)=\mu$ by $m(r)= \bar m(r+\bar x (\mu))$ for $r<0$ is a small error if $\mu$ is close to $-m_\beta$ (because the instanton converges exponentially to its asymptotic values). One may then hope to prove in such a case the existence of the bump using perturbative techniques as in [@DOP]–[@DOP00]. This has been done successfully in [@DOP00] for the equation $$ m(r) = \tanh\{ \beta[ J^{\rm neum}*m(r)+h]\},\quad r \ge 0$$ where $J^{\rm neum}$ is defined with Neumann conditions; $h$ above is fixed and sufficiently small.
We have numerical evidence of the existence of bumps. We have simulated by looking at its discrete version with $\ga^{-1}=120$, $\ell=5$ and Neumann conditions at the right boundary. We have solved such an equation by iteration: we start with $m\equiv 1$, compute $h$ via with such $m$ and then define the first iterate $m_1$ as $m_1= \tanh{\beta(J*m+h)}$. We then repeat the procedure till we find a fixed point. This is indeed reached (approximately) after a few iterations (in fact, three iterations already suffice to obtain good numerical convergence), see Fig. \[fig:j4.1\].
The numerical values of $b(\mu)$ are reported in Fig. \[fig:j4.2-3\], the main features are:
- the values of $b(\mu)$ are all in the metastable region,
- $b(m_+)=m_+$ if $m_+\in(m^*,m_\beta)$, i.e. in the plus metastable region (left panel)
- $b(m_-)>b(m_+)$ for $m_-\in(-m_\beta, 0)$ and $m_+=-m_-$ (right panel).
3 pt
[**Conjecture**]{} [*The bump $B_\mu$ exists for all $\mu < m^*$, when $\mu \in [m^*,m_\beta)$ there is no bump and we call $b(\mu)=\mu$. When $m_+ < m_\beta$ for all $\ell$ large enough there is a stationary solution $m_{\rm st}(r;\ell;m_{\pm})$ of , such that $$\label{j3.13}
\lim_{\ell \to \infty}m_{\rm st}(r\ell;\ell;m_{\pm})=m_{\rm st}(r;m_{\pm}),\quad r\in (0,1)$$ $$\label{j3.14}
m_{\rm st}(0;m_{\pm})= b(m_-),\quad m_{\rm st}(1;m_{\pm})= b(m_+)$$ and*]{} $$\label{j3.15}
-\frac 12 [1-\beta(1-m_{\rm st}^2)]\frac{dm_{\rm st}}{dr} =I_{\rm st}(m_+),\quad r\in[0,1]$$
[**Remark.**]{} [*Under the above Conjecture the channel has a positive current if*]{} $$\label{j3.16}
b(\mu)>-\mu, \; \mu \in (-m_\beta,-m^*);\quad
b(\mu)>b(-\mu) , \; \mu \in (-m^*,0)$$
As shown in the right panel of Fig. \[fig:j4.2-3\] there is clear numerical evidence of the validity of .\
The equation with boundary conditions can be easily solved analytically thus determining $ I_{\rm st}(m_+)$. By using the numerical values obtained for $b(m_-)$ and $b(m_+)$ we get the graph shown with empty circles in Fig. \[fig:10b\], where however $ I_{\rm st}$ is divided by $L$ in order to compare it with the experimental value $j(m_+)$ (black circle in Fig. \[fig:10b\]) as given in Fig. \[fig:2\]. The agreement is good except in the interval $m_+\in (m',m''')$, such a discrepancy will be discussed in the next section.
![[We plot $I_{\rm st}/L$ (empty circles) and $j:=j^{T}_{{\rm ch} \to\mathcal R_2}$ (black circles) as functions of $m_+$ .]{}[]{data-label="fig:10b"}](10b){width="70.00000%"}
Stability of the bump {#sec.j9}
=====================
The numerical analysis of suggests the following:
- there exists a bump solution $B_\mu$ for all $\mu\in(-m_\beta,m^*)$,
- when $\mu\in (-m_\beta,-m^*)$ there are two solutions: $m(x)\equiv \mu$ and $m(x)=B_\mu(x)$,
- when $\mu\in (-m^*,0)$ there are three solutions: $m(x)\equiv \mu$, $m(x)=B_\mu(x)$ and $m(x)=-B_{-\mu}(x)$,
An alternative way to study the existence of the bump is by running the OS-CA with boundary conditions $\mu$ on the left and Neumann on the right. We take the same parameters $\ga^{-1}=120$ and $\ell=5$ used for the numerical analysis of the solutions of and start with an initial condition where all sites in the channel are occupied. Referring to Fig. \[fig:2\], when $\mu\in (-m_\beta,m')$ we see, after a transient, a steady pattern close to $B_\mu$. When $\mu\in (m^{\text{iv}},m_\beta)$ we see, after a transient, a steady pattern close to $m(x)\equiv \mu$. When $\mu\in (m',m^{\text{iv}})$ the final pattern is close to $-B_{-\mu}$, see Fig. \[fig:j4.4\].
![[$\beta=2.5$, $\gamma^{-1}=120$, $\ell=5$ and $\mu=0.7\in (m^{\text{iv}},m'')$. The black dashed line represents $B_\mu$, the black thin line represents the asymptotic pattern of the CA which is close to $-B_{-\mu}$. ]{}[]{data-label="fig:j4.4"}](j44){width="70.00000%"}
Observe that the OS-CA does not select the bump solution when $\mu\in (m',m^{\text{iv}})$ which is approximately the region where there is discrepancy between the theoretical and the experimental curves in Fig \[fig:10b\]. We conjecture that this is due to $\ga$ being not small enough so we are far from the mesoscopic regime and stochastic fluctuations are relevant. Stochastic fluctuations may then determine tunnelling from the bump to patterns where there is a bump on the left and a minus bump on the right with an instanton in between them and patterns where the two bumps are both up. Indeed we have numerical evidence of all that, in the times of the simulations we see in fact the magnetization patterns oscillate as described above, see Fig.\[fig:j4.7-8\].\
3 pt
Conclusions {#sec:concl}
===========
We have presented two sets of simulations: the first one, see Fig. \[fig:6\], shows that in the CC-CA there is a non zero current $j^{CC}(\ga p)$ (provided the rate $\ga p$ of exchanges between reservoirs is in some non zero interval); the second set of simulations, see Fig. \[fig:2\], refers to the OS-CA at magnetization $m_+ = -m_->0$ and shows that when $m_+ \in (0,m_\beta)$ then the current $j(m_+)$ goes “in the wrong direction”, namely from the reservoir with $m_-$ to that with $m_+$. We have a heuristic proof that what seen in Fig. \[fig:6\] follows from the behavior of the channel in the OS-CA, as shown in Fig. \[fig:2\]; the proof relies on the validity of the mesoscopic and the adiabatic limits.
In the case of Fig. \[fig:2\] the current is negative when $m_+>m_\beta$ and positive when $m_+<m_\beta$, in the former case the magnetization pattern in the channel shows the coexistence of the plus and minus phases while in the latter case only one phase appears (the statement in both cases refers to what happens in most of the volume). When the current is negative the values of the magnetization in the plus phase are larger than $m_\beta$ and smaller than $-m_\beta$ in the negative one. Instead when the current is positive we are in the one phase regime and the values of the magnetization are metastable (thus there is a state with positive current which in the bulk takes positive metastable values and another state also with positive current which in the bulk takes negative metastable values). When the current is negative the plus and minus phases are connected via an instanton-like profile around the center of the channel, when the current is negative the unstable values of the magnetization are localized in a small region close to the endpoints. We thus have a boundary layer which leads quite abruptly from the imposed values of the magnetization at the boundaries to some metastable value after which the magnetization pattern is smooth and the current flows opposite to the magnetization gradient in agreement with the Fourier law.
The strange phenomenon of the current going “in the wrong direction” depends on the fact that the magnetization-jump in the boundary layer is more pronounced if it starts from lower values of the magnetization. Such a property, see , follows from the solution (the bump) of a non local equation describing the boundary layer, but its solution is obtained only numerically and we do not have a mathematical proof or even a heuristic explanation of why should hold.
We expect that also in the Cahn-Hilliard equation the graph of $j(m_+)$ has a qualitatively similar shape as in Fig. \[fig:2\], but we miss a proof.
We imagine that our results extend to more general systems with Kac potentials and maybe to physical systems where a van der Waals type of phase transition is present. In such cases a metastable interval is well defined and the relevant density (or magnetization) patterns in the bulk of the channel should have metastable values. Also for short range interactions, as in the n.n. Ising model with ferromagnetic interactions there are metastable values but the metastable region depends on the size of the system and shrinks to 0 as the volume diverges. Take the 2D Ising model in a squared box of side $L$: in the periodic case for $\beta$ large it is proved that if $\pm m_\beta$ are the equilibrium magnetizations then for the canonical Gibbs measure with average magnetization $m \in (-m_\beta, -m_\beta + c L^{-2/3})$ and $c$ small enough the phenomenon of phase separation is absent. Consider the Kawasaki dynamics at such values of $\beta$ with periodic conditions on the horizontal sides of the box and exchanges of the spins in the vertical ones with infinite reservoirs at magnetization $m_-$ and $m_+$ on the left and right. If what we have observed extends to this 2D Ising model we should see in the bulk magnetization patterns in the metastable phase, hence with values in an interval of size $L^{-2/3}$. The current should therefore scale as $L^{-1}L^{-2/3}= L^{-5/3}$ and if the boundary layer goes like in our case then the current would go from the small to the large values of the reservoirs magnetization.
Estimates on the current between reservoirs {#aappA}
===========================================
Recalling we have $$\label{C.1}
j_{\mathcal R_2\to \mathcal R_1}(t)
=j_t := \zeta_t \sum_{i_+,i_-}\mathbf 1_{\xi_t=(i_+,i_-)}[\theta''_t(i_+)- \theta''_t(i_-)]$$ where $ \zeta_t$ and $\xi_t$ are random variables independent of the process till time $t$ and of $\theta''_t$, they are also independent of each other. $ \zeta_t$ takes value 1 with probability $\ga p$ and value 0 with probability $1-\ga p$; the values of $\xi_t$ are pairs $(i_+,i_-)$, $i_+\in \mathcal R_2$, $i_-\in \mathcal R_1$ and $P(\xi_t=(i_+,i_-))=\frac 1{R^2}$. The sum $ \sum_{i_+,i_-}$ is over $i_+\in \mathcal R_2$ and $i_-\in \mathcal R_1$. We first estimate the expected value of $ j_{\mathcal R_2\to \mathcal R_1}(t)$: $$\label{C.1a}
E_\ga[ j_{\mathcal R_2\to \mathcal R_1}(t)]=
E_\ga[\frac{N''_{\mathcal R_2}(t)-N''_{\mathcal R_1}(t)}{R}]\ga p$$ where $$\label{C.1b}
N''_{\mathcal R_i}(t)=\sum_{i\in\mathcal R_i}\theta''_t(i),\quad i=1,2$$ Since $|N''_{\mathcal R_i}(t)-N_{\mathcal R_i}(t)|\le 2$ for all $t$ we have $$\label{C.1c}
\Big| E_\ga[ j_{\mathcal R_2\to \mathcal R_1}(t)]-\ga p
E_\ga[\frac{N_{\mathcal R_2}(t)-N_{\mathcal R_1}(t)}{R}] \Big|\le \ga p\frac 4R$$
We will next prove . Since $\theta''$ has values 0, 1 we have from $$\label{C.2}
E[ j_t ]\le \ga p$$
By the left hand site of , can be written as $$\label{C.3}
A_T:=E\Big[\{ \frac 1T \sum_{t=0}^{T-1} [j_t- \ga p R^{-1}(N_{+,t}-N_{-,t}) ] \}^2\Big]$$ where $$\label{C.4}
N_{+,t}= \sum_{i_+}\eta_t(i_+)=N_{\mathcal R_2}(t),
\quad N_{-,t}= \sum_{i_-}\eta_t(i_-)=N_{\mathcal R_1} (t)$$ Define $N''_{\pm,t}$ as in but with $\theta''_t$ instead of $\eta_t$ and $A''_T$ as in but with $N''_{\pm,t}$.
$$\label{C.5}
A_{T} \le A''_T + \frac {16} R (\ga p)^2 + \frac{16}{R^2} (\ga p)^2$$
[**Proof.**]{} Call $$\label{C.6}
a_{t} =j_t - \ga p R^{-1}(N''_{+,t}-N''_{-,t})$$ $$\label{C.7}
b_{t} = \ga p R^{-1}\{(N''_{+,t}-N''_{-,t}) - (N_{+,t}-N_{-,t})\}$$ Then $$\label{C.8}
A_{T} = E\Big[\frac 1{T^2} \sum_{s,t} (a_{t}-b_{t})(a_{s}-b_{s})\Big]$$ Hence $$\label{C.9}
A_{T} \le A''_{T} +2 E\Big[\frac 1{T^2} \sum_{s,t}|a_{t}||b_{s}|\Big]
+ E\Big[\frac 1{T^2} \sum_{s,t}|b_{s}| |b_t|\Big]$$ $|b_t|\le \ga p \frac 4R$ because $|N''_{+,t} - N''_{-,t}| \le R$ and $|N''_{\pm,t} - N_{\pm,t}| \le 2$. By and $|N''_{+,t} - N''_{-,t}| \le R$ we get $E_\ga[|a_t|] \le 2 \ga p$, therefore $$\label{C.10}
A_{T} \le A''_{T} +2 \ga p \frac 8 R \ga p + [\ga p \frac 4R]^2$$
Let $s<t$ and $a_t$ as in then $$\label{C.11}
E\Big[ a_s a_t\Big] =0$$
[**Proof.**]{} By the independence properties of $\zeta_t$ and $\xi_t$: $$\label{C.12}
E [ a_s j_t ] = E \Big[ a_s \ga p \sum_{i_+,i_-} R^{-2}[\theta''_t(i_+)- \theta''_t(i_-)]\Big] = E \Big[ a_s \ga p R^{-1}[N''_{+,t}- N''_{-,t}]\Big]$$
As a consequence $$\label{C.13}
A''_T = \frac 1{T^2} \sum_{t=0}^{T-1} E[a_t^2
]$$ We expand the square in $E[a_t^2
]$, the first term is $$ E\Big[ \zeta_t \sum_{i_+,i_-}\sum_{i'_+,i'_-}\mathbf 1_{\xi_t=(i_+,i_-)}
\mathbf 1_{\xi_t=(i'_+,i'_-)}[\theta''_t(i_+)- \theta''_t(i_-)][\theta''_t(i'_+)- \theta''_t(i'_-)]\Big]$$ Due to the characteristic functions $i_{\pm}=i'_{\pm}$ so that the above is bounded by $\ga p$. The double product in the expansion of $E[a_t^2
]$ is bounded by $2(\ga p)^2$ and the third term by $(\ga p)^2$, so that $$\label{C.14}
A''_T \le \frac 1T \{\ga p + 3(\ga p)^2\}$$ Going back to we get $$\label{C.15}
A_{T} \le \frac 1T \{\ga p + 3(\ga p)^2\} +16 \frac {(\ga p)^2}R + 16[ \frac{\ga p}R]^2$$ which concludes the proof of .
Proof of Theorems \[thm5.1\] and \[thm5.1b\] {#A1}
============================================
[**Proof of .** ]{} Here we prove that $m(r,t)$ satisfies both in the CC-CA and in the OS-CA.
Let $u(r,t)=m(r,t)+1$ then $m$ satisfies if and only if $u$ satisfies $$\begin{aligned}
\label{B.1a}
&&\frac{\partial}{\partial t}u(r,t) = \frac 12
\frac{\partial^2 u}{\partial r^2} - C \frac{\partial }{\partial r} \Big\{ [u(2-
u]\int_{r}^{r+1}[u(r+\xi,t)
- u(r-\xi,t)] d\xi\Big\}
\end{aligned}$$ with $u(r+\xi,t)= u_+(t)=m_+(t)+1$ if $r+\xi\ge \ell$ and $u(r-\xi,t)= u_-(t)=m_-(t)+1$ if $r-\xi\le 0$. In the OS-CA $m_\pm(t)\equiv m_\pm$.
By $$\begin{aligned}
{\nonumber}&& \lim_{\ga \to 0} \lim_{\ga x\to r, \ga^{2}t\to \tau} E_\ga[ \eta(x,v,t)] =\frac 12 u(x,t), \qquad v\in\{-1,1\}
\\&& \lim_{\ga \to 0} \lim_{\ga x\to r, \ga^{2}t\to \tau} u_\ga(x,t)=u(x,t),\qquad u_\ga(x,t)=E_\ga[ \eta(x,t)] \\\label{B.1b}
{\nonumber}\end{aligned}$$ So that we need to prove that the limit of $u_\ga$ satisfies .
By assumpotion $u(r,t)$ is smooth so that it is enough to prove weak convergence namely that for any smooth test function $f(r,t)$ with compact support in $(0,\ell)\times (0,\infty)$, $$\begin{aligned}
\label{B.2b}
&&\hskip-1cm \int u(r,t)\frac{\partial f(r,t)}{\partial t} dr dt= - \frac 12 \int u(r,t)\frac{\partial^2 f(r,t)}{\partial r^2} dr dt {\nonumber}\\&&\hskip.1cm - \int
\frac{\partial f(r,t)}{\partial r} C \Big\{ [u(2-
u]\int_{r}^{r+1}[u(r+\xi,t)
- u(r-\xi,t)] \Big\}dr dt\end{aligned}$$ By an integration by parts $$\begin{aligned}
&&\int u(r,t)\frac{\partial f(r,t)}{\partial t} dr dt=-\lim_{\ga \to 0} \ga^3\sum_{x,t} f(\ga x,\ga^2t)
\ga^{-2}\{u_\ga(x;t+1)- u_\ga(x;t)\}
\end{aligned}$$ We will next consider $u_\ga(x;t+1)- u_\ga(x;t)$. Recalling that $j_{x,x+1}(t)$ is the number of particles which in the time step $t, t+1$ cross the bond $(x,x+1)$, $x\in\{1,..,L-1\}$ (counting as positive those which jump from $x$ to $x+1$ and as negative those from $x+1$ to $x$), we have $$u_\ga(x;t+1)- u_\ga(x;t)=E_\ga[j_{x-1,x}(t)]-E_\ga[j_{x,x+1}(t)]$$ We then have denoting by $\nabla_\ga$ the discrete derivative ($\nabla_\ga \varphi(x)=\varphi(x+1)-\varphi(x)$), $$\label{B.5}
\int u(r,t)\frac{\partial f(r,t)}{\partial t} dr dt=-\lim_{\ga \to 0} \ga^3\sum_{x,t} \ga^{-1} \nabla_\ga f(\ga x,\ga^2t)
\ga^{-1}E_\ga[j_{x,x+1}(t)]$$
\[lemmaA3\] $$\label{B.4}
E_\ga[j_{x,x+1}(t)]= \frac 12[u_\ga(x;t)- u_\ga(x+1;t)+E_\ga\Big[\chi_{x,\ga;t}
\eps_{x,\ga;t}+\chi_{x+1,\ga;s}
\eps_{x+1,\ga;t}\Big]$$ where $\eps_{x,\ga;t}$ is $\eps_{x,\ga}$ computed at time $t$ and $$\chi_{x,\ga;t} = \eta (x,1;t)\Big(1- \eta (x,-1;t)\Big)+
\eta (x,-1;t)\Big(1- \eta (x,1;t)\Big)$$
[**Proof.**]{} Observe that the expected number of particles that goes from $x$ to $x+1$ is $$E_\ga\Big[\eta(x,1;t)\eta(x,-1,t)+\chi_{x,\ga;t}(\frac 12+\eps_{x,\ga;t})\Big]=\frac 12 u_\ga(x,t)+ E_\ga\Big[\chi_{x,\ga;t}\eps_{x,\ga;t}\Big]$$ The expected number of particles that goes from $x+1$ to $x$ is $$\begin{aligned}
&&E_\ga\big[\eta(x+1,1;t)\eta(x+1,-1,t)+\chi_{x+1,\ga;t}(\frac 12-\eps_{x+1,\ga;t})\big]=\frac 12 u_\ga(x+1,t)
\\&& \hskip7cm - E_\ga\big[\chi_{x+1,\ga;t}\eps_{x+1,\ga;t}\big]
\end{aligned}$$ so that we get .
We insert in and, denoting by $\Delta_\ga$ the discrete laplacian, we get $$\begin{aligned}
{\nonumber}&&
\ga^3\sum_{x,t} \ga^{-1} \nabla_\ga f(\ga x,\ga^2t)
\ga^{-1}j_\ga(x,x+1,t)= \frac 12 \ga^3\sum_{x,t}\ga^{-2} \Delta_\ga f(\ga x,\ga^2t)
u_\ga(x,t)\\&&\hskip1cm + \ga^3\sum_{x,t} \ga^{-1} 2 f'(\ga x,\ga^2t)E_\ga[\chi_{x,\ga;t}]
E_\ga[\ga^{-1}\eps_{x,\ga;t}] +R_\ga {\nonumber}\\
\label{B.6}
\end{aligned}$$ where $2 f'(\ga x,\ga^2t)= [\nabla_\ga f(\ga x,\ga^2t)+ \nabla_\ga f(\ga (x-1),\ga^2t)]$ and $$\begin{aligned}
R_\ga :=2 \ga^3\sum_{x,t} \ga^{-1} 2 f'(\ga x,\ga^2t) E_\ga[\chi_{x,\ga;t}\Big(\ga^{-1}\eps_{x,\ga;t}-
E_\ga[\ga^{-1}\eps_{x,\ga;t}]\Big)]
\end{aligned}$$ By and $$\begin{aligned}
{\nonumber}&& \lim_{\ga \to 0} \lim_{\ga x\to r, \ga^{2}t\to \tau} E_\ga[\chi_{x,\ga;t}]=\frac 12 u(r,t)[2-u(r,t)]\\ \label{B.8}
\\&& \lim_{\ga \to 0} \lim_{\ga x\to r, \ga^{2}t\to \tau} E_\ga[\ga^{-1}\eps_{x,\ga;t}] = \int_{r}^{r+1} C[u(r+\xi,t)
- u(r-\xi,t)]d\xi {\nonumber}\end{aligned}$$ We postpone the proof of $$\label{A1.6}
\lim_{\ga \to 0} \ga^3\sum_{x=2}^{\ga^{-1}\ell-1} \sum_{t=1}^{\ga^{-2}T} E_\ga\Big[\big|\ga^{-1}\eps_{x,\ga;t}-
E_\ga[\ga^{-1}\eps_{x,\ga;t}]\big| \Big] = 0$$ where $(0,\ell)\times (0,T)$ contains the support of $f(r,t)$.
Observe that , , and yield concluding the proof of
By Cauchy-Schwartz it is enough to prove that $$\label{A1.6.1}
\lim_{\ga \to 0} \ga^3\sum_{x=2}^{\ga^{-1}\ell-1} \sum_{t=1}^{\ga^{-2}T} E_\ga\Big[\big |\ga^{-1}\eps_{x,\ga;t}-
E_\ga[\ga^{-1}\eps_{x,\ga;t}]\big|^2 \Big] = 0$$ We thus need to compute the limit of $$\label{A1.6.2}
\ga^5\sum_{r,r',r''\in \ga \mathbb Z} \sum_{\tau\in \ga^2 \mathbb Z}
g_\ga(r,r',r'',\tau)$$ where $\ga^{-1}r \in [2,\ga^{-1}\ell-1]$, $|r'-r| \le 1$, $|r''-r| \le 1$, $\ga^{-2}\tau \in [1,\ga^{-2}T]$ and $$\label{A1.6.3}
g_\ga(r,r',r'',\tau) = C^2E_\ga[\tilde \eta_{\ga^{-2}\tau} (\ga^{-1} (r'-r))
\tilde \eta_{\ga^{-2}\tau} (\ga^{-1} (r''-r)) ]$$ where $\tilde \eta_t(x) = \eta (x,t)-E_\ga[\eta (x,t)]$ if $x\in [1,L]$, otherwise it is $=\frac {2 N_{\mathcal R_i}}{R}- E_\ga[\frac {2 N_{\mathcal R_i}}{R}]$ where $i=2$ if $x>L$ and $i=1$ if $x<1$ otherwise in the OS-CA is equal to $m_\pm$ respectively. By and , vanishes as $\ga\to 0$. [**Proof of .**]{} We call $$\label{B.13}
I_{x,\ga}^T=\ga \sum_{t=0}^{T-1} E_\ga[j_{x,x+1}(t)]$$
There are $c$ and $c'$ so that for all $r'<r''$ in $(0,\ell)$ $$\label{B.9.2}
\Big|\frac{1}{x''-x'}\sum_{y=x'}^{x''}I_{y,\ga}^T\Big|\le c,\qquad x'=[\ga^{-1}r'],\quad x''=[\ga^{-1}r'']$$ $$\label{B.9.3}
\Big|I_{x'',\ga}^T-I_{x',\ga}^T\Big|\le c' |r''-r'|$$
[**Proof.**]{} By , using that $|\chi_{x,\ga;t}|\le 2$ and $|\eps_{x,\ga;t}| \le 2C\ga$ for all $x$ and $t$ and after telescopic cancellations we get $$\begin{aligned}
\Big|\frac{1}{x''-x'}\sum_{y=x'}^{x''}I_{y,\ga}^T\Big|\le\big| \ga \sum_{s=0}^{T-1}\frac{1}{x''-x'} E_\ga[\frac 12(\eta(x',s)-\eta(x''+1,s))]\big|+ 8 C \ga^2T
\end{aligned}$$ The right hand side converges to $\frac 1{r''-r'}\int_0^\tau \frac 12 [m(r'.s)-m(r'',s)] ds+8C^2\tau$ which, by the smoothness of $m$, proves .
We have that $$\Big| \ga \sum_{t=0}^{T-1} j_{x',x'+1}(t)]-\ga \sum_{t=0}^{T-1} j_{x'',x''+1}(t)\Big|\le c' \ga |x''-x'|$$ because the particles which contribute to the left hand site are: (1) those which reach for the first time $x'+1$ jumping from $x'$ and at the final time are in $[x'+1,x'']$; (2) those which reach for the first time $x''$ jumping from $x''+1$ and at the final time are in $[x'+1,x'']$; (3) those initially in $[x'+1,x'']$ and which leave this interval for the last time jumping to $x''+1$; (4) those initially in $[x'+1,x'']$ and which leave this interval for the last time jumping to $x'$.
The family $\{I_{x,\ga}^T\}$ thought as functions of $r=\ga x$ are equibounded and equicontinuous in any compact of $(0,\ell)$, thus they converge pointwise by subsequences. We will then prove by identifying the limit. By continuity it will be enough to prove $$\label{A.9.2}
\lim_{\ga\to 0}\frac{1}{x''-x'}\sum_{y=x'}^{x''} I_{x,\ga}^T=\frac{1}{r''-r'} \int_{r'}^{r''}dr\int_0^\tau I(r,s)ds$$ By $$\begin{aligned}
\frac{1}{x''-x'}\sum_{y=x'}^{x''} I_{x,\ga}^T&=&
\ga \sum_{s=0}^{T-1}\Big\{\frac{1}{x''-x'} E_\ga[\frac 12(\eta(x',s)-\eta(x''+1,s))]
\\&+& \frac{1}{x''-x'}\sum_{y=x'}^{x''}E_\ga[\chi_{x,\ga;s}\eps_{x,\ga;s}+\chi_{x+1,\ga;s}\eps_{x+1,\ga;s}]\Big\}
\end{aligned}$$ The first term converges to $$\label{A.9.4}
\frac{1}{r''-r'} \frac 12\int_0^\tau [m(r',s)-m(r'',s)]ds= \frac{1}{r''-r'} \frac 12\int_0^\tau ds\int_{r'}^{r''}dr\frac{\partial m(r,s)}{\partial s}$$ By and the second one converges to $$-\frac{1}{r''-r'} \int_0^\tau \int_{r'}^{r''} C [1-
m^2]\int_{r}^{r+1}[m(r+\xi,s)
- m(r-\xi,s)] d\xi dr ds$$
[**Proof of .**]{} As the two are similar, we just prove the second equality in . The same proof as the one for shows that $$\label{B.16}
\Big|I_{x,\ga}^T-I_{{\rm ch}\to \mathcal R_2,\ga}^T\Big|\le c' |\ell-r|,\qquad x=[\ga^{-1}r]$$ where $$I_{{\rm ch}\to \mathcal R_2,\ga}^T= \ga \sum_{t=0}^{T-1} E_\ga[j_{{\rm ch}\to \mathcal R_2}(t)]$$ Let $\tilde I$ be a limit point of $I_{{\rm ch}\to \mathcal R_2,\ga}^T$ as $\ga\to 0$ then $$\Big|\int_0^\tau I(r,s)ds -\tilde I\Big|\le c' |\ell-r|$$ Using the expression for $I(r,t)$ and the continuity of $m$, we get in the limit $r\to \ell$ that $\tilde I=\int_0^\tau I(\ell,s)ds$.
[**Proof of** ]{}. As the proofs are similar, we just prove the second equality in for the CC-CA. Suppose by contradiction that there is $t>0$ such that $m(\ell,t)\ne m_{+}(t)$ and for the sake of definiteness $m(\ell,t)< m_{+}(t)$. Then there is $\delta>0$ and an interval $[t',t'']$ so that for $s\in [t',t'']$, $m_{+}(t)> m(\ell,t) +\delta$. Recalling the proof of Lemma \[lemmaA3\] $$\begin{aligned}
E_\ga[j_{{\rm ch}\to \mathcal R_2}(s)]&=&E_\ga[ \frac{N_{\mathcal R_2}(s)}{R}]
- \frac 12 u_\ga(L,s)- E_\ga\Big[\chi_{L,\ga;s}\eps_{L,\ga;s}\Big] \\&\ge &
E_\ga[ \frac{N_{\mathcal R_2}(s)}{R}]-\frac 12 u_\ga(L,s) -c\ga
\end{aligned}$$ $c$ a suitable constant, $c\ga$ bounding the term with $\eps_{x,\ga}$. Then, recalling , , and using the assumptions in Theorem \[thm5.1\] we get $$\begin{aligned}
\liminf_{\ga \to 0}\ga^2\sum_{s\in \mathbb Z \cap \ga^{-2}[t',t'']} E_\ga[j_{{\rm ch}\to \mathcal R_2}(s)]&\ge &\frac 12 \int_{t'}^{t''}[m_+(s)- m(\ell,s)]ds\ge \frac{\delta}2 [t''-t']
\end{aligned}$$ which contradicts .
[**The dynamics of the reservoirs.**]{} We just prove . Let $\tau_0\ge 0$, $\tau>0$, $t_0=[\ga^{-2}\tau_0]$, $T=[\ga^{-2}\tau]$, then $$\begin{aligned}
N_{\mathcal R_2}(t_0+T)-N_{\mathcal R_2}(t_0)=\sum_{t=t_0}^{t_0+T-1}\Big[
j_{{\rm ch}\to \mathcal R_2}(t)-j_{\mathcal R_2\to \mathcal R_1}(t)\big]
\end{aligned}$$ We take the expectation and we use to get $$\begin{aligned}
{\nonumber}&&\Big|E_\ga[N_{\mathcal R_2}(t_0+T)-N_{\mathcal R_2}(t_0)] -\sum_{t=t_0}^{t_0+T-1}\Big[E_\ga[j_{{\rm ch}\to \mathcal R_2}(t)]-
E_\ga[\frac{N_{\mathcal R_2}(t)-N_{\mathcal R_1}(t)}{R}]\ga p
\Big]\Big|
\\&&\hskip3cm\le \frac{4\ga p}RT\label{B.18}
\end{aligned}$$ We then get $$\begin{aligned}
{\nonumber}\frac a2 [m_+(\tau_0+\tau)-m_+(\tau_0)]=\int_{\tau_0}^{\tau_0+\tau} I(\ell,s)ds-p\int_{\tau_0}^{\tau_0+\tau} \frac 12[m_+(s)-m_-(s)]ds
\\ \label{B.19}
\end{aligned}$$ which is obtained from by multiplying by $\ga$ and taking the limit $\ga\to 0$ after using that (1) $R=a\ga^{-1}$, (2) by $$\begin{aligned}
\lim_{\ga \to 0}E_\ga[\frac{N_{\mathcal R_2}(t)-N_{\mathcal R_1}(t)}R]=\frac{m_+(\tau)-m_-(\tau)}2, \quad t =[\ga^{-2}\tau]
\end{aligned}$$ (3) by $$\begin{aligned}
\lim_{\ga\to 0}\ga \sum_{t=t_0}^{t_0+T-1} E_\ga[j_{{\rm ch}\to \mathcal R_2}(t)]=\int_{\tau_0}^{\tau_0+\tau} I(\ell,s)ds
\end{aligned}$$ Then is obtained from by dividing by $\tau$ and taking the limit $\tau\to 0$.
**Acknowledgements** The authors acknowledge very useful discussions with Dima Ioffe.
[3]{}
P. Bruno, [*Comment on “Quantum Time Crystals”*]{}, Phys. Rev. Lett. [**110**]{}, 118901 (2013).
L. A. Bunimovich, [*Mushrooms and other billiards with divided phase space*]{}, Chaos **11**, 802–808 (2001).
M. Colangeli, A. De Masi, E. Presutti, [*Latent heat and the Fourier law*]{}, Physics Letters A **380**, 1710–1713 (2016).
P. Dai Pra , M. Fischer, D. Regoli, [*A Curie-Weiss Model with Dissipation*]{}, Journal of Statistical Physics **152**, 37–53 (2013).
A. De Masi, E. Presutti, D. Tsagkarogiannis, [*Fourier law, phase transitions and the stationary Stefan problem*]{}, Archive for Rational Mechanics and Analysis **201**, 681–725 (2011).
A. De Masi, S. Olla, [*Quasi-static hydrodynamic limit*]{}, Journal of Statistical Physics [**161**]{}, 1037–1058 (2015).
A. De Masi, E. Olivieri, E. Presutti, [*Spectral properties of integral operators in problems of interface dynamics and metastability*]{}, Markov Processes and Related Fields [**4**]{}, 27–112, (1998).
A. De Masi, E. Olivieri, E. Presutti, [*Critical droplet for a non local mean field equation*]{}, Markov Processes and Related Fields [**6**]{}, 439–471 (2000).
G. B. Giacomin, C. Poquet, [*Noise, interaction, nonlinear dynamics and origin of rhythmic behaviors*]{}, Brazilian Journal of Probability and Statistics **29**, 460–493 (2015).
G.B. Giacomin, J. L. Lebowitz, [*Phase segregation dynamics in particle systems with long range interactions. I. Macroscopic limits*]{}, Journal of Statistical Physics [**87**]{}, 37–61 (1997).
Y. Kuramoto, [*Chemical oscillations, waves and turbulence*]{}, Springer Series in Synergetics, Springer - Verlag Heidelberg (1984).
J. L. Lebowitz, E. Orlandi, E. Presutti, [*A particle model for spinodal decomposition*]{}, Journal of Statistical Physics [**63**]{}, 933–974 (1991).
J. L. Lebowitz, O. Penrose, [*Rigorous Treatment of the Van Der Waals-Maxwell Theory of the Liquid-Vapor Transition*]{}, Journal of Mathematical Physics [**7**]{}, 98–113 (1966).
E. Presutti, [*Scaling limits in statistical mechanics and microstructures in continuum mechanics*]{} Springer, Theoretical and Mathematical Physics, Springer-Verlag Berlin Heidelberg (2009).
P. A. Tass, [*Phase resetting in Medicine and Biology: Stochastic Modelling and Data Analysis*]{}, Springer Series in Synergetics, Springer-Verlag Berlin Heidelberg (1999).
F. Wilczek, [*Quantum time crystals.*]{} Phys. Rev. Lett. [**109**]{}, 160401 (2012).
A. Shapere, F. Wilczek, Classical time crystals, Phys. Rev. Lett. [**109**]{}, 160402 (2012).
J. Zhang, P. W. Hess, A. Kyprianidis, P. Becker, A. Lee, J. Smith, G. Pagano, I.-D. Potirniche, A. C. Potter, A. Vishwanath, N. Y. Yao, C. Monroe, [*Observation of a Discrete Time Crystal*]{}. arXiv:1609.08684 (2016).
[^1]: Università degli Studi dell’Aquila, Via Vetoio, 67100 L’Aquila, Italy.\
E-mail: [email protected]
[^2]: Università degli Studi dell’Aquila, Via Vetoio, 67100 L’Aquila, Italy.\
E-mail: [email protected]
[^3]: Gran Sasso Science Institute, Viale F. Crispi 7, 00167 L’ Aquila, Italy.\
E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'K. Sysoliatina[^1], A. Just, I. Koutsouridou, E.K. Grebel, G. Kordopatis, M. Steinmetz, O. Bienaymé, B.K. Gibson, J. Navarro, W. Reid,'
- 'G. Seabroke'
bibliography:
- 'ld.bib'
date: 'Printed: '
title: Local disc model in view of Gaia DR1 and RAVE data
---
[We test the performance of the semi-analytic self-consistent Just-Jahrei[ß]{} disc model (JJ model) with the astrometric data from the Tycho-Gaia Astrometric Solution (TGAS) sub-catalogue of the first Gaia data release (Gaia DR1), as well as the radial velocities from the fifth data release of the Radial Velocity Experiment survey (RAVE DR5).]{} [We used a sample of 19,746 thin-disc stars from the TGAS$\times$RAVE cross-match selected in a local solar cylinder of 300 pc radius and 1 kpc height below the Galactic plane. Based on the JJ model, we simulated this sample via the forward modelling technique. First, we converted the predicted vertical density laws of the thin-disc populations into a mock sample. For this we used of the Modules and Experiments in Stellar Astrophysics (MESA) Isochrones and Stellar Tracks (MIST), a star formation rate (SFR) that decreased after a peak at 10 Gyr ago, and a three-slope broken power-law initial mass function (IMF). Then the obtained mock populations were reddened with a 3D dust map and were subjected to the selection criteria corresponding to the RAVE and TGAS observational limitations as well as to additional cuts applied to the data sample. We calculated the quantities of interest separately at different heights above the Galactic plane, taking into account the distance error effects in horizontal and vertical directions into account separately.]{} [The simulated vertical number density profile agrees well with the data. An underestimation of the stellar numbers begins at $\sim$800 pc from the Galactic plane, which is expected as the possible influence of populations from $|z|>1$ kpc is ignored during the modelling. The lower main sequence (LMS) is found to be thinner and under-populated by 3.6% relative to the observations. The corresponding deficits for the upper main sequence (UMS) and red giant branch (RGB) are 6% and 34.7%, respectively. However, the intrinsic uncertainty related to the choice of stellar isochrones is $\sim$10% in the total stellar number. The vertical velocity distribution function f($|W|$) simulated for the whole cylinder agrees to within 1$\sigma$ with the data. This marginal agreement arises because the dynamically cold populations at heights $<200$ pc from the Galactic plane are underestimated. We find that the model gives a fully realistic representation of the vertical gradient in stellar populations when studying the Hess diagrams for different horizontal slices. We also checked and confirm the consistency of our results with the newly available second Gaia data release (DR2).]{} [Based on these results and considering the uncertainties in the data selection as well as the sensitivity of the simulations to the sample selection function, we conclude that the fiducial JJ model confidently reproduces the vertical trends in the thin-disc stellar population properties. Thus, it can serve as a starting point for the future extension of the JJ model to other Galactocentric distances.]{}
Introduction {#intro}
============
In the past two decades, the amount of data collected on the Milky Way’s stellar content has increased by several orders of magnitude. A typical present-day large-scale survey contains measurements for millions of objects, and with the European astrometric mission Gaia [@gaia16; @gaia16a], this number has already increased by a factor of ten. The core of the best dataset available so far for the Galactic studies includes photometry from the 2 Micron All Sky Survey (2MASS, ) and Gaia G-band [@jordi10; @carrasco16; @leeuwen17], proper motions of the PPM-Extended catalogue (PPMX, ) and the fifth US Naval Observatory CCD Astrograph Catalog (UCAC5, ), as well as astrometric parameters of the Tycho-Gaia Astrometric Solution (TGAS, [@michalik15; @lindegren16]). Extensive information on the stellar chemical abundances is now available from spectroscopic surveys such as GALactic Archaeology with HERMES (GALAH, ), the Large sky Area Multi-Object Spectroscopic Telescope (LAMOST) General Survey , the Sloan Digital Sky Survey III Apache Point Observatory Galactic Evolution Experiment (SDSS-III/APOGEE, ), the Radial Velocity Experiment (RAVE, ), and the Gaia-ESO survey . Thus, we are entering an era in which observational data available for Galactic stellar populations will be not only various and abundant, but also precise enough to enable the detailed study of the Galactic components and to unravel their evolution. To achieve these goals, developed and robust methods are required for a comprehensive analysis of observational data.
The most sophisticated current machinery for modelling the Galaxy is the Besançon Galaxy model (BGM, [@robin3; @robin12; @robin17; @czekaj14]), which accounts for the stellar, gaseous, and non-baryonic content of the Galaxy. It includes the non-axisymmetry of the disc in the form of the bar and the spiral arms and the disc warp. Its first version, presented in , is implemented in the tool *Galaxia*[^2] , which allows constructing synthetic catalogues of the Milky Way and accounting for the selection effects of the different surveys. Another example of a semi-analytic model is the TRIdimensional modeL of thE GALaxy (TRILEGAL, ), which has been designed specifically for star count simulations.
Each Galactic model has its own focus. In previous papers we developed and presented the semi-analytic Just-Jahrei[ß]{} model (JJ model). Its main purpose is to enable a detailed study of the thin-disc vertical structure (, hereafter ). The JJ model describes an axisymmetric thin disc in a steady state consisting of a set of isothermal stellar populations moving in the total gravitational potential. Thick and gaseous discs as well as a dark matter (DM) component are added to the general local mass budget in order to treat the pair of density-potential fully self-consistently. The main input functions describing the thin-disc evolution are the star formation rate (SFR) and age-velocity dispersion relation (AVR), as well as the age-metallicity relation (AMR) and the initial mass function (IMF). In the parameters of the JJ model were calibrated with the local kinematics of the main-sequence (MS) stars from the *Hipparcos* catalogue . The model was later tested with the SDSS star counts towards the north Galactic pole, and the parameters of the thick disc were derived (, ). The model IMF was constrained with the sample of *Hipparcos* data combined with the Catalog of Nearby Stars (, ). As discussed in , the median model-to-data deviations over the Hess diagrams constructed with the SDSS apparent magnitudes and colours towards the north Galactic pole are 5.6% for the JJ model, but 26% for TRILEGAL and 20-53% for the old Besançon model .
In the long term, the development of the JJ model is focused on its extension to Galactocentric distances beyond solar. A special tool for testing chemical evolution scenarios of the Galaxy, the code Chempy[^3], has recently been developed . Constraining Chempy parameters with high-quality spectroscopic data and combining it with the radius-dependent SFR and AVR (first steps to the radial extension of these functions are presented in and ) will be the road to build a global Milky Way disc model.
In this paper we test the JJ model locally with the high-quality astrometric data from the TGAS sub-catalogue of the first Gaia data release (DR1) and the radial velocities from the fifth data release of the RAVE survey (DR5, ). As more than one billion stars with high-precision astrometry of the second data release (DR2) of Gaia have recently become available (published on 25 April, 2018; [@gaia18]), the impact of Gaia DR2 on our results is also discussed.
In the framework of the JJ model, we construct the thin-disc stellar populations in the local solar cylinder, including the full stellar evolution. For the sample of mock stellar populations we predict parallaxes as observed in the TGAS$\times$RAVE cross-match, although for the purposes of illustration, the results are presented as a function of a simple distance estimate calculated both in the data and model as the inverse observed parallax. We also account for the reddening with a realistic three-dimensional (3D) extinction model based on the map from and simulate the complicated selection effects.
This paper has the following structure. In Section \[data\] we describe the TGAS$\times$RAVE data sample and discuss our selection criteria together with the selection functions of the TGAS and RAVE catalogues. Section \[model\] explains the basics of the JJ model and the details of our modelling method. Section \[results\] presents the results of this study. In Section \[discus\] we discuss the intrinsic uncertainty of the modelling associated with the choice of stellar library, and also the usefulness of the applied reddening map and the robustness of our results in view of more accurate and abundant Gaia DR2 data. Finally, we conclude in Section \[final\].
Data
====
TGAS$\times$RAVE thin-disc sample {#cuts}
---------------------------------
To study kinematics and spatial distribution of the thin-disc stellar populations, the full 6D information in the dynamic phase space has to be known for individual stars. The most recent and accurate five astrometric parameters, that is, positions, proper motions, and parallaxes, are provided in the TGAS catalogue. The astrometric solution was found by combining Gaia measurements of the first 14 months of the mission and information on positions from the *Tycho-2* Catalogue . Typical errors for the parallaxes and proper motions are $\sim$0.3 mas and $\sim$0.3 mas yr$^{-1}$, respectively. TGAS covers the whole sky and contains $\sim$2 million stars. However, this data sample alone cannot serve for our purposes as it lacks the radial velocities, which prevents deriving 3D stellar space velocity vectors. This motivated us to complement TGAS with the information from RAVE DR5 , where not only accurate radial velocities are provided (typical errors are 1-2 km s$^{-1}$), but also chemical abundances of six elements as well as photometry from other surveys such as 2MASS and The AAVSO Photometric All-Sky Survey (APASS, ). The total cross-match between TGAS and RAVE contains 257,288 stars[^4]. For these stars we calculated the estimate of heliocentric distance by inverting the TGAS parallaxes, $\tilde{d}=1/\varpi$. Here and elsewhere in this paper notations with tilde are used to emphasise that a given quantity is related to the observed distance derived through this simple inversion. We also transformed it to the distance from the Galactic plane $\tilde{z}$. Using the TGAS proper motions and parallaxes as well as the RAVE radial velocities, we calculated for each star a 3D velocity in Cartesian coordinates $(U,V,W)$. The components of the peculiar velocity of the Sun were chosen as $U_\odot=11.1$ , $W_\odot=7.25$ , $V_\odot=4.47$ . Currently , only the vertical motion of stars is included in the JJ model, such that only the $W$ component of the spatial velocities is used in practice. We also calculated the observational errors $(\Delta U,\Delta V, \Delta W)$ with the help of the TGAS error covariance matrix.
The present-day version of the JJ model contains a detailed recipe for constructing the individual populations of the Galactic thin disc while the other components are added in a simple way. Thus, to perform a detailed model-to-data comparison, we needed to construct a clean thin-disc sample from the data. To do so, we applied the following selection criteria to the TGAS$\times$RAVE cross-match:
1. \[geomcut\]*Geometry cut.* We selected stars with $|b|>20^\circ$ as the RAVE survey avoids the Galactic plane and the number of observed stars drops quickly at low Galactic latitudes. As later we account for the incompleteness of the selected sample, we considered only a special region on the sky where the selection function of the TGAS catalogue is defined [@bovy17a]. We further selected only stars below the Galactic plane, $\tilde{z}<0$, as after all the cuts listed in this section the fraction of stars left above the midplane is only 6.6% of the total final sample, which can be safely neglected in order to reduce the modelled volume and thus speed up the calculations. We also restricted ourselves to the local cylinder by applying $\tilde{d}\cos{b}<300$ pc and $|\tilde{z}|<1$ kpc. Hereafter we drop modulus in our notation of height and recall that our model is plane-symmetric and we work in the region below the midplane. The number of stars left is 51,234.
2. \[parcut\]*Parallax cut.* About $1.3\%$ of the stars in the TGAS$\times$RAVE cross-match have negative parallaxes because of the impact of large observational errors on astrometric solutions for faint and/or distant stars. As a straightforward conversion $1/\varpi$ is not applicable in this case, we further included only stars with positive parallaxes, $\varpi>0$. Next we selected stars with a relative parallax error smaller than 30%, $\sigma_{\varpi}/{\varpi}<0.3$. After this cut, 50,491 stars remained.
3. \[photocut\]*Photometric cut.* We selected stars with known (B-V) APASS colour that belonged to the range of visual magnitudes and colours prescribed by the observation windows of the RAVE and Gaia surveys: , and . The remaining sample contains 45,478 stars.
4. \[qualcut\]*Quality cut.* We set a lower limit to the signal-to noise ratio (S/N) of the RAVE spectra, $S/N>30$. We also selected only stars with $algo\_conv \neq 1$, which corresponds to the robust stellar parameters derived from the RAVE spectra. Additionally, we ensured that only stars with reliable values of TGAS astrometric parameters entered our sample. For this we applied the cuts to astrometric excess noise, $\epsilon_i<1$, and its significance, $D[\epsilon_i]>2$. At this stage, 34,501 stars were left.
5. \[chemcut\]*Abundance cut.* To select the stars that can be identified as thin-disc members in chemical abundance plane, we first applied the cut $4000<T_{\mathrm{eff}}/K<7000$ as only for this range of effective temperatures the chemical abundances were determined in RAVE . Then the cuts $\mathrm{[Fe/H]}>-0.6$, $\mathrm{[Mg/Fe]}<0.2$ were added. As discussed in , this is a reasonable criterion for the separation of the thin and thick discs in the RAVE data.
After applying all the selection criteria, we had a sample of 19,746 stars (Fig. \[data\_sample\], blue). The numbers of stars given for each stage of the data cleaning should not be interpreted straightforwardly in terms of strictness of the applied criteria. Because of the strong correlations between some of the criteria (e.g. cuts \[geomcut\] and \[parcut\]), these values are very sensitive to the order of applying the cuts. Alternatively, all the criteria might be sorted according to the origin of the quantities and the fractions of stars might be estimated that were removed from the initial sample due to the problems with TGAS or RAVE or even entries from TGAS and RAVE together. The corresponding fractions removed from the sample after a simple geometric pre-selection (300-pc radius, $-1000 \, \mathrm{pc}<\tilde{z}<0 \,
\mathrm{pc}$) were 38%, 14.5%, and 22%, respectively.
Data incompleteness {#data_incompl}
-------------------
By reducing the TGAS$\times$RAVE cross-match with criteria \[geomcut\]-\[chemcut\], we selected a clean thin-disc sample in the local solar cylinder with high quality of the measured quantities. This final sample is not representative for the direct study of the vertical disc structure, however: the TGAS and RAVE catalogues are both incomplete, and the applied cuts exacerbate this incompleteness. To account for this, we thoroughly examined the sources of possible biases in our sample and constructed its selection function for further practical use.
The final thin-disc sample clearly contains imprints of the selection functions from both of its parent catalogues. The selection function of the RAVE DR5 depends in general on $I$-band photometry, $J-K_s$ colour, and position on the sky, $S_{\mathrm{RAVE}}(\alpha,\delta,I,J-K_s)$. However, the dependence on colour is relevant only for the low latitudes, $5^\circ<|b|<25^\circ$ and is not of primary importance. The selection function of RAVE DR5, which was derived recently in and is used in this paper, ignores the colour dependence, $S_{\mathrm{RAVE}}(\alpha,\delta,I)$. The completeness factor of the RAVE data varies strongly with the line of sight and visual $I$-band magnitude. The TGAS selection function was investigated in and was found to be quite homogeneous over a large part of the sky. It is defined in terms of 2MASS magnitudes and colours, . Because of the properties of the Gaia scanning law, the number of observations per star in Gaia DR1 is lowest near the ecliptic, that is, the measured quantities in the near-ecliptic regions are characterised by high uncertainties and are expected to be most biased by systematic errors. For this reason, near-ecliptic ‘bad’ regions were excluded from the analysis in , such that the TGAS selection function is not defined for the whole sky. We considered only the region where the TGAS selection function is known (cut \[geomcut\]).
Our selection criteria listed in Section \[cuts\] also add to the total incompleteness of the final sample. The first cut specifies the geometry of the modelled volume and is of no interest here. The parallax cut was modelled as described in Section \[model\_parallax\_cut\]. We also lost stars when applying the high-quality criteria (cut \[qualcut\]) together with the selection of stars with available APASS colour (part of cut \[photocut\]). Finally, not all of the stars in the volume have measurements of Fe and Mg, even within the safe effective temperature range given. As a result, by applying cut \[chemcut\], we also removed a part of the upper main sequence (UMS) and lost thin-disc stars. To quantify the fraction of the thin-disc stars missing in our final sample, we constructed two additional data sets. The first contains stars with known chemical abundances that are classified as thin-disc populations but have low-quality or incomplete records (if any condition from cut \[qualcut\] is not fulfilled or the APASS (B-V) colour is missing). The second data set contains stars that are unclassified as a result of missing chemical abundances, regardless of the quality of their spectra and astrometric solution or available photometry. At this point, we can derive a reliable estimate of the missing star fraction. To determine the total number of the thin-disc stars in some vertical bin that also pass our parallax cut, we summed all the three samples in the following manner: $$\label{td_fr1}
N_{tot} = N_{f} + N_{\sigma} + N_{x} \cdot \frac{ \sum_{\tilde{z}_{min}}^{\tilde{z}_{max}} \rho_d}{\sum_{\tilde{z}_{min}}^{\tilde{z}_{max}} (\rho_d+\rho_t)},$$ where $N_f$, $N_{\sigma}$, and $N_x$ are the number of stars in the final, low-quality, and unclassified samples. The densities $\rho_d$ and $\rho_t$ are the local vertical density profiles of the thin and thick discs as inferred in and , respectively. The two limits $\tilde{z}_{min}$ and $\tilde{z}_{max}$ are the boundaries of the corresponding vertical bin. In order to simplify the expression, we dropped the dependence on $\tilde{z}$ of all quantities in this equation. With this estimate we calculated the total fraction of the missing stars $F_{tot}$ as a function of height: $$\begin{aligned}
\label{td_fr2}
F_\sigma &=& N_{\sigma}/N_{tot}; \ \ \ F_x = N_x/N_{tot} \\ \nonumber
F_{tot} &=& (N_{\sigma}+N_x)/N_{tot}\end{aligned}$$ Here $F_\sigma$ and $F_x$ are the fractions of stars that are missing in our final clean thin-disc sample because of the high noise and the absence of chemical abundances, respectively. The fractions given by Eq. \[td\_fr2\] are shown in the top panel of Fig. \[missing\_stars\]. All three curves are smoothed with a window of 10 pc width. After summation over all vertical bins, we find that $\sim$37% of the total expected number $N_{tot}$ are missing in our final sample. We also examined the fraction of missing stars in the Hess diagram. The bottom panel of Fig. \[missing\_stars\] with the $F_{tot}$ ratio calculated in colour-magnitude bins for the whole cylinder shows that the unclassified stars mostly influence the UMS; the LMS and red giant branch (RGB) regions are affected to a somewhat smaller extent. As it is not fully clear which of the unclassified stars really belong to the thin disc (the ratio of the thin- and thick-disc densities presented in Eq. \[td\_fr1\] was used only to estimate the number of stars, it gives us no clue which population an individual star belongs to), we did not attempt to use the derived weights of the Hess diagram as an additional selection function during the modelling procedure. We rather kept it as a key for understanding the nature of the discrepancies between the data and model when they arose (see Section \[discus\]).
Now, treating all the described selection effects as independent, we defined a completeness factor for the TGAS$\times$RAVE thin-disc sample as follows: $$\begin{aligned}
S = S_{Q}(\tilde{z}) \times S_{\varpi}(\tilde{z}) &\times& S_{\mathrm{RAVE}}(\alpha,\delta,I) \times \nonumber\\
\label{Sfunc} &\times& S_{\mathrm{TGAS}}(\alpha,\delta,J,J-K_s). \end{aligned}$$ Here $S_{\varpi}(\tilde{z})$ corresponds to the selection effect arising from the parallax cut (practical realisation described in Section \[model\_parallax\_cut\]). $S_{Q}(\tilde{z}) = 1 - F_{tot}(\tilde{z})$ is the additional completeness factor related to the missing stars problem. All the factors of the selection function except for $S_{Q}$ were applied during the creation of the stellar populations and calculation of such quantities of interest as Hess diagrams or velocity distribution functions, whereas $S_{Q}$ was taken into account at the post-processing of the results (Section \[model\]).
Model and simulations {#model}
=====================
Basics of the JJ model {#basics}
----------------------
Our self-consistent chemodynamical model of the thin disc has been presented and discussed in a series of papers. We refer to our original for the model details and also for and for its later improvement. Here we briefly review the model basics that are relevant for the further discussions.
The thin disc is constructed from a set of 480 isothermal monoage subpopulations with ages in the range of $\tau=0...12$ Gyr with an equal step of $\Delta \tau=25$ Myr. The dynamical heating of the subpopulations is given by the AVR function, which describes an increase in vertical velocity dispersion with stellar age as a result of the response to small perturbations in the gravitational potential. We model the AVR as a power law with the vertical velocity dispersion starting at $\sim$5 km s$^{-1}$ for the newly born stars and increasing up to 25 km s$^{-1}$ for the oldest subpopulation. Another input function of interest is the SFR, which is implemented as a two-parametric analytic function. When combined with the AVR and IMF, it is a sensitive tool to predict star counts and age distributions. A local chemical enrichment in terms of a monotonously increasing AMR is added to the model in order to reproduce the observed metallicity distributions. In addition to the thin disc, other Galactic components are included in the total mass budget. The gas is modelled with the scaled thin-disc AVR, its scale height was optimised to $h_g=100$ pc in order to reproduce the observed surface density of the gas. The thick disc is included as a single-birth population with an age of 12 Gyr. The spherical isothermal DM halo is added in a simple thin-disc approximation. The thin-disc subpopulations are assumed to be in dynamic equilibrium in the total gravitational potential generated by the stellar, gaseous, and DM components of the Galaxy. By solving Poisson’s equation iteratively, we obtain a self-consistent vertical potential and density laws of the thin-disc subpopulations. Originally, the AVR parameters were calibrated against the kinematics of MS stars from the *Hipparcos* catalogue combined at the faint end with the Fourth Catalogue of Nearby Stars (CNS4, ). To constrain the AMR, the local metallicity distribution from the Copenhagen F and G star sample (GCS1, ) was used . The thin-disc vertical profile is essentially non-exponential close to the plane, characterised by half-thickness $h_d=400$ pc. However, kinematic data alone do not help to distinguish SFR and IMF. Further comparison of the model to the Sloan Digital Sky Survey (SDSS) star counts towards the north Galactic pole allowed us to distinguish the SFR and IMF and constrain both of them . Our best SFR has a peak at $\tau \approx 10$ Gyr and is characterised by the mean and present-day star formation rate of 3.75 $\mathrm{M_\odot \ pc^{-2} \ Gyr^{-1}}$ and 1.4 $\mathrm{M_\odot \ pc^{-2} \ Gyr^{-1}}$, correspondingly. At this stage, the thick-disc parameters were also improved. The local surface density was set to 5.4 $\mathrm{M_\odot \ pc^{-2}}$ (18% of the thin-disc surface density), the scale height was fixed to $h_t = 800$ pc, and the vertical velocity dispersion was $\sigma_t=45.4$ km s$^{-1}$.
We note that although numerous recent studies indicate that the thick disc might be more complex than we adopt here (e.g. ), the topic is still under debate. It is generally recognised that the thick disc is an old, alpha-rich, and metal-poor population formed on a short timescale. The span of the suggested timescale varies significantly in different studies (4-5 Gyr in , but less than 2 Gyr with as low as 0.1 Gyr in the two-infall chemical models in ). Some authors reported an increase in vertical velocity dispersion with the age of thick-disc populations. On the other hand, within the framework of the Jeans analysis developed in , the alpha-rich metal-poor thick disc behaves as a well-mixed kinematically homogeneous population. As the local data alone are not enough to distinguish the thin and thick discs in a robust way, we did not attempt to re-define them in this work. For a quick test against Gaia DR2 (see end of Section \[discus\]), we modelled the thick disc with a metallicity spread. Further possible improvements are postponed to the stage when the model is fully extended, which will also include the detailed chemical evolution. This will allow us to study different Galactic populations in the abundance plane.
Thus, at the moment we arrived at a monotonous AVR and AMR, a decreasing SFR, and a three-slope broken power-law IMF for the thin disc and a single-age, scale height, and vertical velocity dispersion thick disc. Following the convention established in -, we address this best set of parameters as the fiducial model A and use it for the modelling throughout this paper.
Modelling procedure {#modelling}
-------------------
Before going into the details, we sketch a general overview of the modelling procedure summarising all steps discussed below in Sections \[table\]-\[verted\].
We start with discretisation of the age-metallicity space and assigning an isochrone to each age-metallicity pair (Section \[table\]). After this, we create a cylindrical grid (Section \[grid\]) and populate each of its space volumes with the predicted types of stars (characterised by age, metallicity, and mass from the isochrone) and take into account the parallax cut (Section \[model\_parallax\_cut\]), reddening, the TGAS$\times$RAVE selection function (Section \[incompl\]), and the corresponding abundance and photometric cuts. Then we account for the vertical effect of the distance error (Section \[verted\]) and investigate the properties of the mock sample as a function of distance from the midplane (Section \[results\]).
Creation of the stellar assemblies {#table}
----------------------------------
We started our modelling with thin-disc mono-age subpopulations. Each age corresponds to a unique value of metallicity as prescribed by the AMR law. We added a Gaussian scatter in metallicity $\sigma_{\mathrm{[Fe/H]}}=0.15$ dex. Each age $\tau_i$ was then associated with seven metallicities representing a Gaussian with a mean AMR$(\tau)$ and a dispersion $\sigma_{\mathrm{[Fe/H]}}$. Each metallicity subpopulation is given an appropriate weight $w_k$, such that $\sum_{k=1}^7 w_k = 1$. In a similar scatter in metallicity was modelled to reproduce the observed metallicity distributions of Geneva-Copenhagen F and G stars, which can be interpreted either in terms of observational errors or in terms of real physical scatter in metallicity of mono-age populations. The added dispersion $\sigma_{\mathrm{[Fe/H]}}$ here corresponds to a physical scatter in metallicity as the age-metallicity grid was then used to select isochrones. This complexity roughly accounts for the effect of radial migration, which is not explicitly introduced in our model.
At this point, it is useful to define a new notation in order to avoid confusion in the future. From here on, we refer with ‘subpopulation’ to the thin-disc mono-age isothermal subpopulations. For the additional splitting in metallicity-mass parameter space we introduce the term ‘stellar assembly’. Thus, the described splitting of the subpopulations in metallicity results in $480 \cdot 7 = 3360$ stellar assemblies.
The next step is to include the full stellar evolution. We used the Modules and Experiments in Stellar Astrophysics (MESA, [@paxton11; @paxton13; @paxton15]) Isochrones and Stellar Tracks (MIST, [@dotter16; @choi16])[^5]. We took a set of 67 isochrone tables that safely cover the whole range of the modelled metallicities, $\mathrm{[Fe/H]}=-0.6...0.46$. The ages of isochrones are given in logarithmic range $\log{(\tau/\mathrm{yr})}= 7...10.08$ with a step of 0.02. Each of the 3360 stellar assemblies constructed earlier was associated with the isochrone that had the closest age and metallicity to the modelled ones. The MIST isochrones cover a range of initial masses $0.1...300 \ M_{\odot}$ from which we selected the range $0.1...100 \ M_{\odot}$ as prescribed by the model IMF. The resulting number of the stellar assemblies characterised now by the same ages, metallicities, and initial stellar masses is $\sim 5 \cdot 10^6$. For each stellar assembly $j$, a surface number density $N^\Sigma_j$ was calculated by weighting the SFR with the IMF and accounting for the weights $w_k$. Isochrones also provide us with the present-day stellar masses, stellar parameters $\log{L}$, $\log{T_{\mathrm{eff}}}$, $\log{g}$, and absolute magnitudes in the standard $UBVRI$ system, 2MASS $JHK_s$, and Gaia $G$ band. At this step we have a table of the mock stellar assemblies with their known surface densities in the solar neighbourhood. The remaining modelling procedure may be briefly described as checking which of these types of stars were actually observed in the TGAS$\times$RAVE cross-match and then were selected by us for the thin-disc sample as well as for further distributing them in the local cylinder.
The volume number density is calculated by: $$N^V_{j}(z) = \frac{N^{\Sigma}_j}{2 h(\tau_j)}\exp{\big\{- \Phi(z)/\sigma_W^2(\tau_j)\big\}},
\label{N_in_volume}$$ where $N^{\Sigma}_j$ is the surface number density of a given stellar assembly, $h(\tau_j)$ and $\sigma_W(\tau_j)$ are the half-thickness and the vertical velocity dispersion defined by the AMR for the corresponding age $\tau_j$, respectively, and $\Phi(z)$ is the total vertical potential at a height $z$, as predicted by the fiducial model A. As an example outcome of the local model, we calculated the volume densities $N_{V}$ at $z=0$ kpc for all modelled stellar assemblies and constructed a Hertzsprung-Russel (HR) diagram for the thin disc in the solar neighbourhood (Fig. \[N\_in\_volume\_plot\]).
Before proceeding to the next steps of the modelling, we pre-selected stellar assemblies with the effective temperatures belonging to the RAVE range (cut \[chemcut\]). An additional pre-selection was based on the expectations for the range of absolute G magnitudes in the sample, which we set to $-2...10$ mag (see bottom panel of Fig.\[missing\_stars\]). The reduced subset of the stellar assemblies used further during the simulations falls into the blue frame on the bottom right plot in Fig. \[N\_in\_volume\_plot\].
Sample geometry {#grid}
---------------
To model the geometry of the thin-disc sample, we created a 3D grid in cylindrical coordinates $(r,\theta,z)$. To adequately account for the parallax cut (see below), we set the initial radius of the cylinder to $r_{in}=500$ pc. This is larger than we used for the data (cut \[geomcut\] in Section \[cuts\]). The modelled cylinder extends vertically up to $z_{max}=1$ kpc below the midplane. The region above the Galactic plane was not included. We binned this cylindrical volume (1) in height $z$ with a step of $\Delta z = 5$ pc, (2) in angle $\theta$, which was measured from the cylinder axis, with a step of $\Delta \theta = 3^\circ$, and (3) in radial direction $r$ by binning the space in 20 intervals in logarithmic scale. This gives $4.8\cdot10^5$ space volumes, but only those of them that correspond to the directions of observations were used in the actual modelling.
Parallax cut {#model_parallax_cut}
------------
The distances $r$ and $z$ defining the model grid are not directly comparable to the distance estimates derived for the data sample through the inversion of observed parallaxes. As a work-around, we translated the model true distances into the space of observed parallaxes. This allowed us to include in the model the same parallax cut as was applied to the data, and also to account for the distance error effects.
We split the procedure into two independent steps. First, we applied Eq. \[parpdf\] at fixed z and translated the model distances to the observed parallaxes. This allowed us to reduce the number of the stellar assemblies left for the modelling at a given height by applying a parallax cut identical to the one we introduced in Section \[cuts\]. Then we reduced the radius of the modelled cylinder to 300 pc using the distances calculated from the mock observed parallaxes. At this stage, we treated different z-slices as independent, that is to say, the modelled stellar assemblies remained at their initial height. The fact that a given star is observed at $\tilde{z}$ different from its true distance from the midplane z was taken into account at the second stage during a post-processing of the quantities calculated in the different z-slices. This strategy was chosen in order to simplify the modelling process and speed up the calculations, as it allowed us to include the effect of the distance error in vertical direction at the post-processing stage of the simulations. We now consider the first step with the parallax cut more closely; the second step is described in Section \[verted\].
We started with the usual assumption that the observed parallax $\varpi$ is distributed normally about the true parallax $p=1/d$ with a standard deviation $\sigma_{\varpi}$ depending on the stellar brightness, exposure time, and number of observations per star. This is a good approximation of the real shape of the TGAS parallax error distribution which is known to deviate slightly from normality only beyond $\sim2\sigma$ . Under this assumption, the normalised probability of the observed parallax $\varpi$ is given by the Gaussian PDF: $$\begin{aligned}
\label{parpdf}
P(\varpi|\sigma_{\varpi},p) &=& \frac{1}{A} \exp{\Big[ - \frac{(\varpi - p)^2}{2 \sigma_{\varpi}^2} \Big]} \\ \nonumber
\mathrm{with} \ A &=& \sqrt{2 \pi} \sigma_{\varpi}. \end{aligned}$$ An example PDF of the observed parallax is shown in the inset of Fig. \[epar\_plot\].
At each fixed height z, we have 2400 volume spaces, as implied by the model grid. We assumed that the parallax error PDF as derived from the TGAS$\times$RAVE cross-match (Fig. \[epar\_plot\], black histogram) does not depend on parallax or direction of observation (see coloured PDFs in Fig. \[epar\_plot\]). After this, we can easily assign parallax errors to our stellar assemblies when modelling the stellar content in each volume space. Then the observed parallaxes can be derived from the Gaussian distribution in accordance with Eq. \[parpdf\]. We performed this separately for each radial bin in a given z-slice. In summary, at each fixed z, we selected a pie of 20 space volumes corresponding to the same angle $\theta$ and different distances from the cylinder axis, and in each of them, we reduced the table of stellar assemblies in the following way:
- A random true heliocentric distance $d$ within the given space volume was assigned to each stellar assembly and was then converted into the true parallax $p=1/d$.
- Parallax errors $\sigma_{\varpi}$ were drawn randomly from the parallax error PDF.
- The corresponding observed parallaxes $\varpi$ were derived as random values from the Gaussian PDF $P(\varpi|\sigma_{\varpi},p)$.
- The observed parallaxes $\varpi$ were converted into observed distances $\tilde{d}=1/\varpi$.
- Cuts identical to those used for the data sample were applied: $\varpi>0$, $\sigma_{\varpi}/\varpi<0.3$.
- Only stellar assemblies with $\tilde{d} \cos{b}<300$ pc were selected. This allowed us to account for the fact that the parallax errors cause some stars to lie outside the modelled cylinder, but may yet be observed within it (alternatively, stars that actually belong to the cylinder of 300 pc radius may appear at larger distance and be excluded from the sample).
- We also removed the stellar assemblies located at $|b|<20^\circ$ to reproduce the spatial geometry of the modelled sample.
The resulting set of tables of the stellar assemblies were then subjected to the TGAS$\times$RAVE incompleteness factor.
TGAS$\times$RAVE selection function {#incompl}
-----------------------------------
After the parallax cut was applied, in each space volume at a given $z$ we calculated the visual magnitudes in the $I$ and $J$ bands as well as colour $(J-K_s)$ for all the stellar assemblies that were left. Apparent or absolute magnitudes in other photometric bands such as $B,V$, or $G$ were also added when necessary. Visual magnitudes were calculated with the true model distances $d$.
The recent progress in mapping the dust content of the Milky Way allowed us to use a fully realistic 3D extinction model instead of the 2D map of or the simple analytic extinction model from . We implemented the 3D dust map from [^6], which is a combination of 3D extinction models from , , and .
The photometric cuts , , and were then applied to the stellar assemblies, in full analogy with the corresponding data selection criterion (cut \[photocut\], Section \[cuts\]). The surface number densities of the remaining stellar assemblies were weighted with the completeness factor $S_{\mathrm{RAVE}}\times S_{\mathrm{TGAS}}$. For this we binned the stellar assemblies in $I$ and $J$ magnitudes with the corresponding steps of $\Delta I=0.5$ mag and $\Delta J=0.1$ mag as well as in colour $(J-K_s)$ with a $\Delta (J-K_s)=0.14$ mag step. After this, we calculated the resulting stellar number in a given grid cell by converting the surface stellar densities $N^{\Sigma}_j$ into $N^V_j$ in accordance with Eq. \[N\_in\_volume\] and multiplying the volume density by the volume of the cell. With the known number of stars as well as their multi-band photometry, ages, and metallicities we can predict the vertical density laws for different populations, calculate vertical kinematics with the AVR, and study the modelled sample in the colour-magnitude plane.
Vertical distance error {#verted}
-----------------------
![*Top.* Mean observed heliocentric distance $\tilde{d}$ vs. height $\tilde{z}$ as derived from the data (blue curve). The thin-disc sample was binned in $\tilde{z}$, and at each height, the mean heliocentric distance was calculated. The observed dependence was fitted with a polynom of the fifth order to obtain a smooth function. The distance error as a function of height was then derived with the derived mean distance (magenta curve). *Bottom.* Diagram showing the vertical effect of the distance error. The normalised probability for a true height $z$ to be associated with the observed distance from the midplane $\tilde{z}$ given the thin-disc sample geometry and the typical distance errors from the top panel is colour-coded. The horizontal and vertical projections are example PDFs $P(\tilde{z}|\sigma_z,z)$ and $P(z|\sigma_z,\tilde{z})$ for the true and observed heights of 0.5 kpc. The skewness of the two distributions results in two effects: (1) an individual star is more likely to be observed at a height larger than its true distance from the midplane, and (2) the observations at some $\tilde{z}$ are more likely to be affected by smaller heights. The magenta lines in the projection subplots show distribution maxima.[]{data-label="verted_plot"}](DZ.pdf)
At the stage of post-processing the results, when the quantities of interest were calculated for different $z$, we also accounted for the effects of the distance errors in the vertical direction. To do so, we introduced a probability $P(\tilde{z}|\sigma_z,z)$, which defines a likelihood for a quantity $Q$ calculated at its true $z$ to be observed at another height $\tilde{z}$ given a corresponding vertical error $\sigma_z$. In order to derive the $P(\tilde{z}|\sigma_z,z)$ expression, we rewrote Eq. \[parpdf\] in terms of heliocentric distance: $$\begin{aligned}
P(\tilde{d}|\sigma_d,d) &=& \frac{1}{B} \exp{\Big[ - \frac{d^4}{\sigma_d^2} (1/\tilde{d} - 1/d)^2 \Big]} \\ \nonumber
\mathrm{with} \ B &=& \int_{0}^{\infty} P(\tilde{d}|\sigma_d,d) d \tilde{d},
\label{distpdf}\end{aligned}$$ where $\tilde{d}$ and $d$ are the observed and true heliocentric distances and $\sigma_d$ is the distance error. At this point, we need to know how both distances as well as the distance error correspond to the true and observed heights, that is, the functions of interest are $\tilde{d}(\tilde{z})$, $d(z)$ and $\sigma_d(z)$. The first can be easily derived from the thin-disc sample itself (blue curve in the top panel of Fig. \[verted\_plot\]). Its behaviour is ruled by the sample geometry. Owing to the cut $|b|>20^\circ$, close to the midplane the radius of the cylinder grows linearly with $\tilde{z}$ until it reaches 300 pc, so that the mean distance $\tilde{d}$ increases almost linearly with height. At large $\tilde{z}$, the mean distance asymptotically approaches $\tilde{z}$. The function $d(z)$ can be calculated within the model framework, but for our accuracy, it is sufficient to use an approximation $\tilde{d}(\tilde{z}) \approx d(z)$. Then the distance error is estimated as $\sigma_d(z)=1/d^2(z) \cdot \sigma_{\varpi}$ with a constant $\sigma_{\varpi}=0.3$ mas, which is a typical parallax error of the TGAS catalogue (magenta curve in the top panel of Fig. \[verted\_plot\]).
The vertical effect of the distance error is illustrated in the bottom panel of Fig. \[verted\_plot\]. The colour-coding represents the normalised probability . The two projections of this diagram are of special interest. The vertical projection, calculated for a true height $z=500$ pc, gives a skewed PDF that describes a likelihood for a star located at $z$ to be observed at other heights $\tilde{z}$, $P(\tilde{z}|\sigma_z,z)$. The horizontal projection shown for the same but observed height $\tilde{z}=500$ pc describes the impact from the different vertical bins on a given $\tilde{z}$, which is expressed with the PDF $P(z|\sigma_z,\tilde{z})$. While the maximum of the probability density $P(\tilde{z}|\sigma_z,z)$ belongs to a slightly larger height than the corresponding true $z$, the shift is opposite in case of $P(z|\sigma_z,\tilde{z})$: a star observed at a given $\tilde{z}$ most probably belongs in reality to a smaller height. It is also clear that the vertical effect of the distance error is negligible close to the midplane, where the distance errors are small, and has the largest impact on the location of the most distant stars. We also have to consider heights larger than $z_{max}=1$ kpc in order to allow the stars to not just leave the cylinder, but also to enter it from the larger heights (in full analogy with the radial direction, when we first consider a cylinder of a larger radius and then cut it to 300 pc after allowing stars to migrate horizontally due to the distance error effect). However, as the stellar density decreases approximately exponentially with increasing height, we expect the effect of not including $z>z_{max}$ to be very small except in the most distant parts of the cylinder, so we did not model it.
Finally, we write $$P(\tilde{z}|\sigma_z,z) = P(\tilde{d}(\tilde{z})|\sigma_d(z),d(z)).
\label{Pz}$$ With Eq. \[Pz\] we calculated the quantities of interest as functions of observed height $\tilde{z}$: $$Q_{\tilde{z}} = S_{Q} (\tilde{z}) \cdot \sum_{z=0}^{z_{max}} Q_{z} P(\tilde{z}|\sigma_z,z)
\label{Qz1}$$ Here an additional and final correction was included in the form of the $S_{Q}(\tilde{z})$ factor discussed in Section \[data\_incompl\]. Alternatively, when we wish to derive the model predictions in the horizontal slice of a thickness of $|\tilde{z}_2-\tilde{z}_1|$, we write $$Q_{\tilde{z_1} < \tilde{z} < \tilde{z_2}} = \sum_{\tilde{z}=\tilde{z}_1}^{\tilde{z}=\tilde{z}_2} Q_{\tilde{z}}.
\label{Qz2}$$ Eq. \[Qz1\] and \[Qz2\] are used to account for the vertical effects of the distance error for the modelling of the stellar density laws and vertical kinematics. An alternative procedure for the Hess diagrams is presented in Section \[hess\_results\].
Results
=======
Vertical density law {#Nz_results}
--------------------
The first and most straightforward test that can be conducted with the JJ model is a comparison of the observed and predicted vertical density profiles because the primary focus of the model is the vertical structure of the stellar populations in the disc. By following the steps described in Sections \[table\]-\[verted\], we calculated the vertical stellar density law for all stars in the cylinder, $n(\tilde{z}) = N^V_{\tilde{z}}/V_{\tilde{z}}$ with $V_{\tilde{z}}$ being the volume of the corresponding vertical bin. The top panel of Fig. \[Nz\_plot\] shows the stellar number density as calculated from the data (blue line) and as predicted by the model (violet and red curves). The final model prediction is plotted in red, while the violet curve is given for comparison to show the number density law as predicted without the impact of the vertical distance error taken into account. The procedure of weighting the model output with the PDF $P(\tilde{z}|\sigma_z,z)$ results in migration of stars to larger observed heights $\tilde{z}$, which in turn leads to a better consistency with the data far from the midplane. On the other hand, the effect is very small close to the Galactic plane, predictably in view of the negligible distance error in close vicinity to the Sun (Fig. \[verted\_plot\], top panel). The observed and the modelled density profiles, agree well at all heights, which also implies also a good agreement in the local region close to the Galactic plane where the model is well-calibrated and is therefore expected to produce robust predictions. By analogy to the vertical profiles, we derived a distribution of stars over the observed heliocentric distances that allows us to quantify the model-to-data agreement in spherical volumes as well. The error in the predicted stellar number within a 25 pc local sphere (considered for the model calibration in ) is $\sim$6%. However, we note that the actual number of observed stars in this volume is smaller than ten, such that the statistics is strongly influenced by Poisson noise.
When investigated in terms of the cumulative number of stars as a function of observed height, the model and data demonstrate 7.8% disagreement at $\tilde{z}=1$ kpc (Fig. \[Nz\_plot\], middle panel). The ratio of the observed to modelled stellar number in individual vertical bins as shown in the bottom panel of Fig. \[Nz\_plot\] helps to understand the systematic trends in the model-to-data differences. The ratio is close to unity over a wide range of heights, which means that the robustness of the modelled vertical profile and our treatment of the stars that are missing in the volume as a result of the lack of chemical abundances and low quality of other parameters are reliable (Fig. \[missing\_stars\]). It is also clear that starting from $\sim$600 pc, the Poisson noise begins to dominate star count statistics in the bins. From a height of approximately 800 pc, the number of stars is systematically under-represented in the model. Consistently, we expect this region to be biased in our simulation as we did not model stars at heights larger than 1 kpc and thus did not account for their possible presence in our sample owing to the large distance errors.
It is also interesting to perform a comparison in the space of observed parallaxes in order to validate our two-step treatment of the parallax uncertainties and the vertical distance errors. Following our standard steps, we calculated the parallax distributions at individual model heights z and then weighted the results with the corresponding probabilities $P(\tilde{z}|\sigma_z,z)$ and completeness factors $S_Q(\tilde{z})$. The final distribution is shown in Fig. \[Par\_plot\]. Both data and model are normalised to the total number of observed and predicted stars, such that the shapes of the distributions can be clearly compared. We find a very good agreement between the two curves which demonstrates the reliability of our forward-modelling approach.
Vertical kinematics {#fW_results}
-------------------
It is illustrative to test the model performance in terms of the vertical kinematics that was originally used to calibrate the model parameters . Again, following the routine described in Sections \[table\]-\[verted\], we evaluated for each observed height $\tilde{z}$ the velocity distribution function $f(|W|)$ as a superposition of Gaussians with the standard deviations given by the AVR. We considered the absolute W-velocities in a range of $0...60$ km s$^{-1}$ with a step of $\Delta |W| = 2$ km s$^{-1}$. The width of the step was selected such that it allowed tracing the shape of the velocity distribution function but at the same time was larger than the typical observational error in the velocity bin. Following Eq. \[Qz2\], we modelled $f(|W|)$ in six horizontal slices probing the dynamical heating of the disc as a function of distance from the midplane. The observed and predicted normalised velocity distribution functions calculated for the different ranges in $\tilde{z}$ as well as for the whole cylinder are shown in Fig. \[fWz\_plot\]. Green points represent the data with the error bars calculated as standard deviation of the Poisson distribution. The red curves are the model predictions. In general, the observed and modelled vertical kinematics of the sample agree very well. The total predicted $f(|W|)$ (Fig. \[fWz\_plot\], right plot) is consistent with the data within 1$\sigma$ at almost all $|W|$. This is also true for the velocity distribution functions compared to the data at different heights (Fig. \[fWz\_plot\], three left columns). A noticeable discrepancy appears in the two lowest bins, for 0 pc $<\tilde{z}<$ 100 pc and 100 pc $<\tilde{z}<$ 200 pc where the fraction of the dynamically cold populations is underestimated in our model. In general, a trend with height is clearly evident: with the increase of the distance to the midplane, the shape of $f(|W|)$ becomes less peaked at small $|W|$. This is a natural and straightforward result as more dynamically heated populations reach larger distances from the Galactic plane because their velocities are higher.
We also compared $f(|W|)$ in four magnitude bins, for which we selected different parts of the MS (CMD in Fig. \[fWMv\_plot\]). The bin ranges were the same as used in our previous work (compare to Fig. 3 in ). The CMD shows the whole TGAS$\times$RAVE thin-disc sample colour-coded with the RAVE metallicity. A metallicity gradient across the MS is visible such that each of the defined magnitude bins contains a mixture of stars with different metallicities and ages. To model $f(|W|)$ for the selected magnitude bins, we added the same colour-magnitude cuts to the calculation procedure and removed the stars that did not fall under the given criteria after adding the reddening, at the stage of applying the TGAS$\times$RAVE selection function (Section \[incompl\]). The remaining calculation procedure was the same. The velocity distribution functions for the MS also show a good consistency with the data (Fig. \[fWMv\_plot\], left), although the model slightly underestimates the role of the dynamically cold populations for the bin $3.5 \, \mathrm{mag}<M_V<4.5 \, \mathrm{mag}$.
Hess diagrams {#hess_results}
-------------
In order to achieve deeper insight into the space distribution of the thin-disc populations, we also investigated the modelled cylindric volume in the 2D colour-magnitude space, that is, we built Hess diagrams. We used the Gaia $G$ and 2MASS $K_s$ filters and constructed the Hess diagrams in $(G-K_s,M_G)$ within the range of colours and absolute magnitudes of $-0.5...3.5$ mag and $9.5..1.5$ mag, respectively, with corresponding steps of $\Delta (G-K_s)=0.05$ mag and $\Delta M_G=0.14$ mag. The typical uncertainty of the visual G-band magnitude is $\sim$0.003 mag and for the colour $(G-K_s)$, it is larger by about a factor of ten, $\sim$0.025 mag. Both values are well within our resolution on the Hess diagram axes, thus the choice of $G$ and $K_s$ photometry allows us to construct high-quality Hess diagrams without the need to complicate the simulations by adding photometric errors.
The calculation procedure in this case differs from the one described previously at the stage of accounting for the vertical distance error. For each modelled height $z$, we visualised the stellar content of this thin horizontal slice on the colour-absolute magnitude axes. The absolute magnitude is itself a logarithmic function of the distance related to $z$ through $d(z)$ (Fig. \[verted\_plot\], top panel). This additional distance dependence is the reason why we cannot model the resulting Hess diagram by appropriate weighting of the predictions calculated at the true heights in accordance to Eqs. \[Qz1\] and \[Qz2\], as we did for the other quantities. In order to account for the impact of the populations lying at some true height $z_1$ on the Hess diagram modelled for the height $z_2$, we took into account that the populations from $z_1$, when observed at $z_2$, has different absolute magnitudes shifted by $\Delta M = 5 \log{(z_1/z_2)}$. Thus, the vertical error effect adds a spread on the vertical axis on the Hess diagram.
To approach the problem, we rewrote Eq. \[Pz\] in terms of the absolute magnitudes: $$\begin{aligned}
&& P(\tilde{M}|\sigma_M,M) = \\ \nonumber
&& = \frac{1}{C} \exp{\Big[ - \frac{1}{2}\Big(\frac{5}{\ln{10} \sigma_M} \Big)^2 \Big(10^{0.2(\tilde{M}-M)} - 1 \Big)^2 \Big]} \\ \nonumber
&& \mathrm{with} \ C = \int_{-\infty}^{\infty} P(\tilde{M}|\sigma_M,M) d \tilde{M}.
\label{magpdf}\end{aligned}$$ Here $P(\tilde{M}|\sigma_M,M)$ gives the probability of a star with the true absolute magnitude $M$ to be associated with another magnitude $\tilde{M}$ given a magnitude error $\sigma_{M}$. The limits of integration $M_{min}$ and $M_{max}$ formally correspond to $\mp \infty$, but we replace them with the values of the physical meaning. We put $M_{min}=-8$ mag and $M_{max}=17$ mag as these are the maximum and minimum values of the absolute $G$-magnitudes as predicted by our set of the MIST isochrones. The error in the absolute magnitude is a function of distance, that is, of the height above or below the midplane: $$\sigma_M = \frac{5}{\ln{10} d(z)} \sigma_d(z)
\label{sigM}$$ Now, we can smooth the Hess diagrams predicted for the different true heights $z$ by taking the corresponding magnitude errors $\sigma_{M}(z)$ from Eq. \[sigM\]. We refer to the value in some row of the Hess diagram calculated for a given $z$ as $H_M$. Then the smoothing procedure can be expressed as $$\tilde{H}_{M} = \sum_{M=M_1}^{M=M_2} H_{M} P(\tilde{M}|\sigma_M,M).
\label{HM1}$$ The limits $M_1$ and $M_2$ are related to the range of the heights of interest $\tilde{z}_1 < \tilde{z} < \tilde{z}_2$: $$\begin{aligned}
M_1 = M + 5 \log{(d(z)/d(\tilde{z}_1))} \\ \nonumber
M_2 = M + 5 \log{(d(z)/d(\tilde{z}_2))}
\label{HM2}\end{aligned}$$ The Hess diagram $\tilde{H}$ smoothed in such a way shows where in the absolute magnitude-colour plane the populations from the height $z$ will appear when the range of observed heights $|\tilde{z}_2-\tilde{z}_1|$ is considered. The resulting Hess diagram for the given range of heights can be evaluated as a sum over all $\tilde{H}$: $$H_{\tilde{z_1} < \tilde{z} < \tilde{z_2}} = \sum_i S_{Q}(z_i) \tilde{H}_i.
\label{HM3}$$
We constructed Hess diagrams for the same six vertical bins as were used in Section \[fW\_results\]. Observed and modelled Hess diagrams were additionally smoothed with a window $0.1\times0.28$ mag$^2$ (2$\times$2 bins in $(G-K_s,M_G)$). The results are plotted in Fig. \[hess\_plot\]. The two left columns show the Hess diagrams as derived from the data and predicted by the model given for increasing $\tilde{z}$ (top to bottom). A pronounced vertical trend appears: close to the Galactic plane, the LMS stars dominate; at the middle heights, the contribution of the the UMS becomes important and RGB stars appear as well; and starting from $\tilde{z} \approx 400$ pc, the stellar populations in Hess diagrams are represented by UMS and RGB alone. All these changes are very confidently traced in the model. The right column in Fig. \[hess\_plot\] shows three plots for the whole modelled cylinder (top to bottom) corresponding to the observed and predicted Hess diagrams, as well as their relative ratio calculated with the suppression of noise in the almost empty bins. The relative ratio plot demonstrates a general good agreement between data and model as it has already been inferred from investigating the vertical number density profiles, although a comparison in 2D space reveals the sources of model-to-data differences in more details. The typical model-to-data deviations over the Hess diagram are found to be $\pm$20%, with the red and blue regions indicating the problematic areas (see Section \[discus\]).
We further investigated the UMS, LMS, and RGB populations separately. We defined them with the white lines in the middle plot of the right panel of Fig. \[hess\_plot\]. The border between the lower and upper parts of the MS approximately corresponds to the stellar masses of 1.5 $M_{\odot}$. We plot two projections of the total Hess diagrams onto the magnitude and colour axis. The resulting colour distributions and luminosity functions are shown in the top and middle panels of Fig. \[3pop\_plot\]. In addition to the general difference in the number of stars, we see more clearly that the modelled LMS is more peaked than the observed one, while the UMS and RGB regions are less pronounced in the model. The same information can be obtained from the individual vertical density laws plotted for the three populations (Fig. \[3pop\_plot\], bottom panel). The LMS, UMS, and RGB regions are under-populated in the model by 3.6%, 6%, and 34.7%, respectively.
Discussion {#discus}
==========
In this section we discuss several potential problems of our modelling procedure.
When we analyzed the Hess diagrams (Fig. \[hess\_plot\]), we found a few tenth of stars in the data that were clearly outliers (see the reddest and bluest $G-K_s$ values). These might be misidentified stars as well as contamination of the metal-poor thick-disc and halo stars and objects with unrealistic metallicities, magnitudes, or colours. One of the problematic red areas is associated with the LMS region: the observed MS is noticeably wider than the predicted one. This might be caused by the impact of binarity, which is ignored in our modelling. Additionally, metal-rich stars are present in the data sample ($\mathrm{[Fe/H]}>0.2$), and their number is underestimated with the model AMR (see Fig.16 in , the metallicity prescribed to the youngest stellar population is $+0.02$ dex). Moreover, we predict an underestimated stellar density in the RGB region. This we attribute to the simplicity of accounting for the $S_Q$ completeness factor, which is included in the modelling as a function of height above the plane (Fig. \[missing\_stars\], top panel), although it also shows a variation with colour and absolute magnitude (Fig. \[missing\_stars\], bottom panel). By ignoring this dependence of $S_Q$ on magnitude and colour, we may over- or underestimate stellar numbers in the regions of the Hess diagram where the values of incompleteness in Fig. \[missing\_stars\] deviate considerably from the average. This effect can also be responsible for a blue region near the UMS at $G-K_s \approx 0.8$ mag (compare to the bottom panel of ). On the other hand, when considered together with a blue region in the LMS (see colour range $1.8 \lesssim (G-K_s)/\mathrm{mag} \lesssim 2.8$ at the relative ratio plot in Fig. \[hess\_plot\]), it points to a small colour shift between the modelled and observed Hess diagrams of $\sim$0.1 mag in $G-K_s$. Several reasons may be responsible for this: an underestimated reddening, a systematic shift of isochrones, or a lack of metal-poor thin-disc populations in the data due to an incorrect separation of thin- and thick-disc stars; the interplay of all of these factors is also possible. Taking into account the locality of our volume, however, it is improbable that this shift is related to the underestimated reddening. We return to this question in view of the new results obtained on the basis of Gaia DR2 (see below). Furthermore, we discuss the sensitivity of our results to the choice of dust map and stellar library.
A possible unreliability of the dust map can have a twofold impact on our results. First, the predicted Hess diagram may appear shifted in the colour-magnitude plane relative to the observed one. Second, as the data selection function strongly depends on apparent magnitude and colour and the predicted number of stars may change significantly over the colour-magnitude bins (this is particularly related to the UMS and RGB regions), under- or overestimation of extinction and reddening may add up to the uncertainty in star counts. The 3D dust map used in our simulations is mainly based on the high-angular resolution extinction model from constructed by the statistical technique from photometric data of Pan-STARRS and 2MASS. The safe range of distances covered by this map lies approximately in a range from 300 pc to 4-5 kpc, although the values of the minimum and maximum reliable distance may significantly vary over the sky. In most of our modelled sky region, the minimum reliable distance is $\sim$200 pc and extends to 300 pc. This implies that reddening may not be modelled reliably in those regions of the cylinder that are in close vicinity to the Sun. However, we do not expect the roughness of the extinction map in the nearby areas to have a significant effect on the modelling process because with the cut of $|b|>20^\circ$, we avoid the most problematic near-plane regions. To quantify the impact of the extinction model on our results, we tested an additional 3D reddening map from [^7] when we used Gaia DR2 data (see below).
With the full stellar evolution now included in the modelling, our predictions are sensitive to the choice of the stellar evolution package. To quantify the corresponding uncertainty, we tested an alternative set of isochrones generated by the PAdova and TRieste Stellar Evolution Code (PARSEC, , ) for the same range of metallicities and ages as described in Section \[table\]. The total number of predicted stars increased by $\sim$10%, but no appreciable difference of the model behaviour was discovered apart from this.
As the Gaia DR2 astrometry and completeness are significantly improved with respect to TGAS and the Gaia radial velocities have an even better precision than those of RAVE , we performed a sanity check complementary to the local test using the Gaia DR2 astrometric parameters and radial velocities. To mimic the test presented in this paper, we used the Gaia$\times$RAVE cross-match with the Gaia radial velocities [@soubiran18] and selected the local subset belonging to the same spatial volume and area on the sky as the TGAS$\times$RAVE sample described in Section \[data\]. We did not attempt to repeat the full forward-modelling described in Section \[modelling\] by accounting for parallax errors; instead, we used geometric distances from derived in the framework of Bayesian approach on the basis of Gaia DR2 parallaxes. In this test we did not apply metallicity scatter to the AMR eiter, which allowed us to reduce the number of stellar assemblies and speed up the calculations. These simplifications do not influence the star counts much: (1) even if the distance errors reach $\sim$50% for the individual stars at 1 kpc, the overall stellar density at this heliocentric distance is quite low in our geometry, and the corresponding uncertainty introduced in the number of stars in the volume is not significant; and (2) the metallicity grid affects the modelled number of stars only indirectly through the sample selection function, which is sensitive to the apparent magnitudes and colours. In order to avoid additional incompleteness, no cuts on RAVE Fe or Mg abundances were used. Correspondingly, the thick disc was added to the model as a mono-age population with a metallicity of $-0.7$ dex and a Gaussian dispersion of 0.2 dex. As some stars of the RAVE DR5 are missing in the Gaia$\times$RAVE cross-match (because of unknown Gaia radial velocities or match problems; this makes up $\sim$ 11% of the full RAVE DR5), we built a modified RAVE selection function that incorporates this additional incompleteness and used it for our test. The selected Gaia$\times$RAVE local sample contains 63,408 stars. We built absolute and apparent Hess diagrams using the updated Gaia DR2 $G$-band and $G_{BP}-G_{RP}$ colour from PARSEC isochrones. Because the metallicity scatter and distance errors were not modelled, all features of the predicted Hess diagram look more pronounced than those seen in the data (). Regardless of the width of the observed and modelled MS, we expect their centres to coincide; however, the right panel of Fig. \[hess\_abs\_md\] shows that this is not the case, which implies a small colour shift of about $\sim$0.05 mag in $G_{BP}-G_{RP}$ between the data and the model. As the thick disc is included at this stage, the shift is most likely related to a systematic shift of the isochrones. We also observe an excess of stars in the UMS region which indicates that a better understanding of the Gaia selection function is necessary. Additionally, there is a deficit of stars in the modelled LMS in the colour range of $1.7 \lesssim (G_{BP}-G_{RP})/\mathrm{mag} \lesssim 1.9$. As we did not include any special colour cuts and the applied selection function is essentially independent of colour, this feature must arise from the stellar library and needs to be further investigated in future work. With this setup we reproduced star counts with an accuracy of $\sim$3%. We report the presence of non-reproduced dynamically cold populations at $|z|<100$ pc even more confidently than in case of the TGAS$\times$RAVE test (Fig. \[fw\_local\], also see the upper left panel of Fig. \[fWz\_plot\]). Finally, we compared the outcome of the sanity check star counts for the and dust maps. We find only a difference between the two runs of $\sim$1%. Thus, the outcome of our tests gives us a good insight into potential problems of the model and can be viewed as a first step to the future and more comprehensive work with Gaia DR2.
![Normalised $|W|$-velocity distribution function at $|z|<100$ pc. The data sample is based on Gaia DR2.[]{data-label="fw_local"}](fw_local.pdf)
Conclusions {#final}
===========
We used the forward-modelling technique to test the semi-analytic JJ model. We selected a clean sample of thin-disc stars from the TGAS$\times$RAVE cross-match providing precise astrometry as well as radial velocities and chemical abundances. We investigated the large local volume of the solar cylinder with radius of 300 pc extending to 1 kpc below the Galactic plane. The full stellar evolution in the form of MIST stellar isochrones was implemented in the model. For each modelled height $z$, we accounted for the parallax errors and reddening, and reproduced the incompleteness of the sample. The results were assigned with the additional weights to model the effect of the distance error in the vertical direction.
We found that the model predictions and data agree well with each other. A deviation of $\sim 1 \sigma$ is found between the modelled and observed velocity distribution functions $f(|W|)$ close to the Galactic plane, indicating that the role of the cold young stellar populations may be underestimated in our local disc model. When complemented with the velocity distribution functions for the different magnitude bins of the MS stars, this implies that the source of the discrepancy in vertical kinematics is related to the lower part of the UMS region. A complementary test based on the RAVE and new Gaia DR2 data confirmed the results presented in Section \[results\], and the reported underestimation of dynamically cold populations is even more prominent when tested with Gaia DR2 astrometry and radial velocities.
With this realistic performance test, we demonstrated the robustness of the local JJ model when compared to the data up to 1 kpc away from the Galactic plane. However, non-negligible deviations from the data are identified in case of the vertical kinematics of the disc, which suggests that the model should be adapted to match the new Gaia data. The next step will include an extension of the model to other Galactocentric distances: the code Chempy will be used to build the chemical evolution model of the disc, consistent with the adopted SFR and AVR varying with Galactic radius. For this purpose, Gaia DR2 and its next releases as well as high-resolution spectroscopic surveys such as APOGEE, Gaia-ESO, and GALAH will be of high benefit.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported by Sonderforschungsbereich SFB 881 ‘The Milky Way System’ (subprojects A6 and A5) of the German Research Foundation (DFG).
Funding for RAVE has been provided by the Australian Astronomical Observatory, the Leibniz-Institut für Astrophysik Potsdam (AIP), the Australian National University, the Australian Research Council, the French National Research Agency, the German Research Foundation (SPP 1177 and SFB 881), the European Research Council (ERC-StG 240271 Galactica), the Istituto Nazionale di Astrofisica at Padova, The Johns Hopkins University, the National Science Foundation of the USA (AST-0908326), the W. M. Keck foundation, the Macquarie University, the Netherlands Research School for Astronomy, the Natural Sciences and Engineering Research Council of Canada, the Slovenian Research Agency, the Swiss National Science Foundation, the Science & Technology Facilities Council of the UK, Opticon, the Strasbourg Observatory, and the Universities of Groningen, Heidelberg, and Sydney. The RAVE web site is at <http://www.rave-survey.org>.
This work has made use of data from the European Space Agency (ESA) mission [*Gaia*]{} (<https://www.cosmos.esa.int/gaia>), processed by the [*Gaia*]{} Data Processing and Analysis Consortium (DPAC, <https://www.cosmos.esa.int/web/gaia/dpac/consortium>). Funding for the DPAC has been provided by national institutions, in particular the institutions participating in the [*Gaia*]{} Multilateral Agreement.
[^1]: Fellow of the International Max Planck Research School for Astronomy and Cosmic Physics at the University of Heidelberg (IMPRS-HD).
[^2]: <http://galaxia.sourceforge.net>
[^3]: <https://github.com/jan-rybizki/Chempy>
[^4]: The cross-match between catalogues was performed by the RAVE Collaboration, see <https://www.rave-survey.org/downloads>
[^5]: <http://waps.cfa.harvard.edu/MIST>
[^6]: <https://github.com/jobovy/mwdust>
[^7]: <http://stilism.obspm.fr>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $G$ and $H$ be finite graphs without loops or multiple edges. If for any two-coloring of the edges of a complete graph $K_n$, there is a copy of $G$ in the first color, red, or a copy of $H$ in the second color, blue, we will say $K_n\rightarrow (G,H)$. The Ramsey number $r(G, H)$ is defined as the smallest positive integer $n$ such that $K_{n} \rightarrow (G, H)$. A two-coloring of $K_{r(G, H)-1}$ such that $K_{r(G, H)-1} \not \rightarrow (G,H)$ is called a critical coloring. A Ramsey critical $r(G, H)$ graph is a graph induced by the first color of a critical coloring. In this paper, when $n \geq 15$, we show that there exist exactly sixty eight non-isomorphic Ramsey critical $r(C_n, K_6)$ graphs.'
author:
- |
C. J. Jayawardene\
Department of Mathematics\
University of Colombo\
Sri Lanka\
email: c\[email protected]\
\
W. C. W. Navaratna and J.N. Senadheera\
Department of Mathematics\
The Open University of Sri Lanka\
Sri Lanka\
email: [email protected], [email protected]\
title: 'All Ramsey $(C_n,K_6)$ critical graphs for large $n$'
---
Keywords: Graph theory, Ramsey theory, Ramsey critical graphs\
Mathematics Subject Classification: 05C55, 05C38, 05D10\
Introduction
============
For any two graphs $G$ and $H$, the *Ramsey number* $r(G,H)$ is the smallest positive integer $n$ such that $K_n\rightarrow (G,H)$. The *classical Ramsey number* $r(s, t)$ is defined as $r(K_s,K_t)$ and the *diagonal Ramsey number* is defined as $r(K_n,K_n)$. These numbers have been studied extensively in the last five decades. The difficulty in exact determination of *diagonal Ramsey number* $r(n, n)$, swifts expeditiously from the apparent $r(3,3)=6$ to the unmanageable $r(5,5)$. Currently, the best known lower and upper bounds for $r(5,5)$ are 43 and 48 (see [@Ra]). Many interesting variations of the basic problem of finding classical Ramsey numbers have emerged. One such variation is the calculation of the number of *Ramsey critical $(G,H)$ graphs*, for any pair of graphs $(G,H)$. In this paper, we show that there are exactly 68 Ramsey critical $r(C_n, K_6)$ graphs, for all $n$ exceeding fourteen.
Notation
========
The *complete graph* on $n$ vertices, a *cycle* on $n$ vertices and a *Star* on $n+1$ (see [@JaSa]) vertices are denoted by $K_n$, $C_n$ and $K_{1,n}$ respectively. Given a graph $G$ and a vertex $v \in V(G)$, we define the *neighbourhood of $v$ in $G$*, $\Gamma(v)$, as the set of vertices adjacent to $v$ in $G$. The *degree* of a vertex $v$, $d(v)$, is defined as the cardinality of $ $ $\Gamma(v)$, $ $ i.e. $d(v)=| \Gamma(v)|$. The minimum degree of a graph $G(V,E)$ denoted by $\delta(G)$ is defined as $ $ $ \min\{ d(v) | v\in V \}$.
Given a graph $G$, we say $I \subseteq V(G)$ is an *independent set*, if no pair of vertices of $I$ is adjacent to each other in $G$. Equivalently, $I$ forms a clique in $G^c$. Given a graph $G=(V,E)$, we define *the independence number*, $\alpha(G)$, as the size of the largest independent set. Thus, $\alpha(G)=\max \{ \left| I\right| : I$ is an independent set of $G \}$. In the special case of $H=K_m$, alternatively $r(G,K_m)$ can be viewed as the smallest positive integer $n$ such that every graph of order $n$ either contains $G$ as a subgraph or else satisfies $\alpha(G) \geq m$. For a non-empty subset $S$ of $V$, *the induced subgraph of $S$ in $G$* denoted by $G[S]$ is defined as the subgraph obtained by deleting all the vertices of $S^c$ from $G$. For two disjoint subgraphs $H$ and $K$ of $G$, we denote the set of edges between $H$ and $K$ by $E(H,K)$.
Lemmas used to generate Ramsey critical $(C_n,K_6)$ graphs for $n \geq 15$
==========================================================================
In an attempt to prove Bondy and Erd[ö]{}s conjecture $r(C_n,K_m) = (n-1)(m-1)+1$, for all $(n,m)\neq(3,3)$ satisfying $n \geq m \geq 3$ under certain restrictions, Schiermeyer has proved that $r(C_n,K_6) = 5(n - 1)+1$, for $n \geq 6$ (see [@Ra; @Sc]). Characterizing all Ramsey critical $(C_n,K_6)$ graphs boils down to finding all (red/blue) colorings of $K_{r(C_n,K_6)-1}$ such that there is no red $C_n$ or a blue $K_6$. This is achieved by finding all $C_n$-free graphs on $K_{r(C_n,K_6)-1}$ vertices such that $\alpha(G)<6$. We first prove that any $C_n$ -free graph (where $n \geq 15$) of order $5(n-1)$ with $\alpha(G) \leq 5$ contains a $5K_{n-1}$. To prove this, we use seven lemmas of which the first three are already proven results. For ease of reference, we reiterate Lemma \[l1a\] from [@JaRo1], Lemma \[l1b\] from [@Ja] and Lemma \[l2\] by Bollob[á]{}s et al from [@BoJaSh].
\[l1a\] ([@JaRo1], Lemma 2; [@JaRo1]). A $C_n$- free graph $G$ of order $N$ with independent number less than or equal to $m$ has minimal degree greater than or equal to $N-r(C_n,K_{m})$.
\[l1b\] ([@Ja], Lemma 8). A $C_n$- free graph (where $n \geq 7$) of order $4(n-1)$ with no independent set of 5 vertices contains a $4K_{n-1}$.
([@BoJaSh], Lemma 5) \[l2\] Suppose $G$ contains the cycle $(u_1, u_2, ... , u_{n-1}, u_1)$ of length $n - 1$ but no cycle of length $n$. Let $Y = V(G) \setminus \{u_1, u_2, ... , u_{n-1}\}$. Then,
**(a)** No vertex $x \in Y$ is adjacent to two consecutive vertices on the cycle.
**(b)** If $x \in Y$ is adjacent to $u_i$ and $u_j$ then $u_{i+1}u_{j+1} \notin E(G)$.
**(c)** If $x \in Y$ is adjacent to $u_i$ and $u_j$ then no vertex $x' \in Y$ is adjacent to both $u_{i+1}$ and $u_{j+2}$.
**(d)** Suppose $\alpha(G) = m - 1$ where $m \leq \frac {n + 2}{2}$ and $\{x_1, x_2, ... ,x_{m -1} \} \subseteq Y$ is an $(m - 1)$-element independent set. Then, no member of this set is adjacent to $m -2$ or more vertices on the cycle (We have taken the liberty of making a slight correction to the inequality $m \leq \frac {n + 2}{2}$ of the original [@BoJaSh], Lemma 5(d)).
The next lemma plays a pivotal role in proving the main results of this paper.
\[l3\] A $C_n$ -free graph (where $n \geq 15$) of order $5(n-1)$ with no independent set of 6 vertices contains a $5K_{n-1}$.
[**Proof.**]{} We shall assume that in each of the three cases $n = 15$, $n = 16$ and $n \geq 17$ we consider, $G$ as a graph on $5(n-1)$ vertices satisfying $C_n \not \subseteq G$ and $\alpha(G) \leq 5$. Since $r(C_{n-1}, K_6)=5n-9 \leq 5(n-1)$ (see [@BoJaSh; @Ra]), there exists a cycle $C=(u_1, u_2, ... , u_{n-1}, u_1)$ of length $n - 1$ in $G$. In consistent with the notation of [@BoJaSh], define $H$ as the induced subgraph of $G$ not containing the vertices of the cycle $C$. Then, $|V(C)|=n-1$ and $|V(H)|=4(n-1)$.
Suppose there exists an independent set $Y=\{y_1, y_2,y_3, y_4, y_5 \}$ of size 5 in $H$, so that $\alpha(G)=5$. From Lemma \[l2\] (as $5\leq \frac{n+2}{2}$), it follows that no vertex of $Y$ is adjacent to four or more vertices of the $C_{n-1}$. Thus, $\left| E(Y,V(C)) \right| \le 15$. For ease of reference, we define such a graph structure as a **Standard Configuration ($n$)**.
**Case 1: $n \geq 17$**
Now, $\left| E(Y,V(C)) \right| \le 15 < n-1$. Thus, there exists a vertex $x\in V(C)$ adjacent to no vertex of $Y$. This gives, an independent set $Y \cup \{x\}$ of size 6, a contradiction.
**Case 2: $n = 16$**
In this case as $n-1=15$, in order to avoid an independent set of size 6, each vertex of $V(C)$ must be adjacent to at least one vertex of $Y$. Thus, we get that for each $1\leq i\leq 5$, $\left|\Gamma (y_i) \cap V(C) \right|=3$ and for each $1\leq j < j' \leq 5$, $\Gamma (y_j) \cap \Gamma (y_{j'}) \cap V(C)=\phi$.
By Lemma \[l1a\], as $\delta(G) \geq 14$, $\left|\Gamma (y_i) \cap V(H \setminus Y)\right| \geq 11$ for $i=1,2$. Since $r(P_3,K_6)=11$ and $\alpha(G) < 6$, each of $G[\Gamma (y_1) \cap V(H \setminus Y)]$ and $G[\Gamma (y_2) \cap V(H \setminus Y)]$ contains a copy of $P_3$. Thus, $P_3 \subseteq \Gamma (y_1) \cap V(H \setminus Y)$, where the $P_3$ is induced by $\{x,y,z\}$ such that $(x,y), (y,z) \in E(G)$ and $P_3 \subseteq \Gamma (y_2) \cap V(H \setminus Y)$, where this $P_3$ is induced by $\{p,q,r\}$ such that $(p,q), (q,r) \in E(G)$.
Suppose that $x$ is not adjacent to any vertex of $\{y_2,y_3,y_4,y_5\}$ and $p$ is not adjacent to any vertex of $\{y_1,y_3,y_4,y_5\}$. Re-order the vertices of the cycle such that $y_1\in Y$ is adjacent to $u_1$. In this ordering, let $y_1$ be also adjacent to $u_t$ where $2 \leq t \leq 15$.
Figure 1. Configuration for $n=16$
By Lemma \[l2\](a), $t \neq 2$. In order to avoid an independent set of size 6, induced by $\{x,u_t,y_2,y_3,y_4,y_5\}$, we get that $(x,u_t) \in E(G)$. However, $t \neq 3$, in order to avoid a $C_{16}$ comprising $(u_1,y_1,x, u_3,...,u_{15},u_1)$. Also, $t \neq 4$ in order to avoid a $C_{16}$ comprising $(u_1,y_1,y,x,u_4,...,u_{15},u_1)$ and $t \neq 5$ in order to avoid a $C_{16}$ comprising $(u_1,y_1,z,y,x, u_5,...,u_{15},u_1)$.
Thus, any pair of vertices adjacent to $y_1$ in $C$ cannot be separated by a path of length 1, 2, 3 or 4 along $C$. Thus, $\Gamma (y_1) \cap C =\{u_1,u_6,u_{11}\}$. In this scenario, we use the prerogative that $(y_2,u_2)\in E(G)$. Then, by the previous argument $\Gamma (y_2) \cap C =\{u_2,u_7,u_{12}\}$. But by Lemma \[l2\](b), $(u_2,u_7)\notin E(G)$. Henceforth, we will get that $\{u_2,u_7,y_1,y_3,y_4,y_6\}$ is an independent set of size 6, a contradiction.
This implies that there is a vertex of $X=\{x,y,z\}$ adjacent to some vertex of $\{y_2,y_3,y_4,y_5\}$ or there is a vertex of $\{p,q,r\}$ adjacent to some vertex of $\{y_1,y_3,y_4,y_5\}$. Therefore, without loss of generality, we may assume that $y_1$ is adjacent to $X=\{x,y,z\} \subseteq V(H \setminus Y)$ and $y_1$ is adjacent to $X'=\{x',y',z'\} \subseteq V(C)$ where $X'$ induces a $P_3$ and $y_2$ is adjacent to $x$. Next since $C$ has 15 points without loss of generality, $\{x',y',z'\}=\{u_1,u_3,u_5\}$ or $\{x',y',z'\}=\{u_1,u_3,u_6\}$ or $\{x',y',z'\}=\{u_1,u_3,u_7\}$ or $\{x',y',z'\}=\{u_1,u_3,u_8\}$ or $\{x',y',z'\}=\{u_1,u_3,u_9\}$ or $\{x',y',z'\}=\{u_1,u_4,u_7\}$ or $\{x',y',z'\}=\{u_1,u_4,u_8\}$ or $\{x',y',z'\}=\{u_1,u_4,u_9\}$ or $\{x',y',z'\}=\{u_1,u_4,u_{10}\}$ or $\{x',y',z'\}=\{u_1,u_5,u_9\}$ or $\{x',y',z'\}=\{u_1,u_5,u_{10}\}$ or $\{x',y',z'\}=\{u_1,u_6,u_{11}\}$. Moreover, as $y_1$ and $y_2$ are connected by paths of lengths 2, 3 and 4 in $H$, no pair of vertices selected from $\Gamma (y_1) \cap V(C)$ and $\Gamma (y_2) \cap V(C)$ can be separated by a path of length 3, 4 or 5 along the cycle $C$. Using this we argue that when $\{x',y',z'\}=\{u_1,u_3,u_5\}$, $\Gamma (y_2) \cap V(C) = \phi$, when $\{x',y',z'\}=\{u_1,u_3,u_6\}$, $\Gamma (y_2) \cap V(C) = \phi$ and when $\{x',y',z'\}=\{u_1,u_3,u_7\}$, $\Gamma (y_2) \cap V(C) \subseteq \{u_9\}$. Similarly, when $\{x',y',z'\}=\{u_1,u_3,u_8\}$, $\Gamma (y_2) \cap V(C) \subseteq \{u_2,u_9,u_{10}\}$, when $\{x',y',z'\}=\{u_1,u_3,u_9\}$, $\Gamma (y_2) \cap V(C) \subseteq \{u_2,u_{10}\}$, when $\{x',y',z'\}=\{u_1,u_4,u_7\}$, $\Gamma (y_2) \cap V(C) = \phi$ and when $\{x',y',z'\}=\{u_1,u_4,u_8\}$, $\Gamma (y_2) \cap V(C) \subseteq \{u_2,u_{10}\}$. When $\{x',y',z'\}=\{u_1,u_4,u_9\}$, $\Gamma (y_2) \cap V(C) \subseteq \{u_2,u_3,u_{10}\}$, when $\{x',y',z'\}=\{u_1,u_4,u_{10}\}$, $\Gamma (y_2) \cap V(C) \subseteq \{u_2,u_3\}$, when $\{x',y',z'\}=\{u_1,u_5,u_9\}$, $\Gamma (y_2) \cap V(C) \subseteq \{u_3,u_7\}$ and when $\{x',y',z'\}=\{u_1,u_5,u_{10}\}$, $\Gamma (y_2) \cap V(C) \subseteq \{u_{3}\}$ and when $\{x',y',z'\}=\{u_1,u_6,u_{11}\}$, $\Gamma (y_2) \cap V(C) =\phi$. Since none of these give a viable configuration, we get a contradiction.
**Case 3: $n = 15$**
To deal with the case $n = 15$, we first prove three Lemmas. Lemma $\ref{l31}$, deals with the possible scenarios generated by the Standard Configuration ($15$). Lemmas $\ref{l32}$ and $\ref{l33}$ deal with showing that none of the scenarios generated by Lemma $\ref{l31}$ give viable configurations.
\[l31\] In the Standard Configuration $(n=15)$, one of the following three scenarios $(a)$, $(b)$ and $(c)$ will occur:
**(a)** $y_1 \in Y$ is a vertex of the subgraph $K_4$ (see Figure 2(a)) in $H$.
**(b)** $y_1,y_2\in Y$ are vertices of the subgraph $K$ (see Figure 2(b1)) or subgraph $K'$ (see Figure 2(b2)) in $H$.
**(c)** $y_1,y_2\in Y$ are vertices of the subgraph $L$ (see Figure 2(c)) in $H$.
*scenario (a)* *scenario (b)* *scenario (c)*
Figure 2: (a),(b1), (b2) and (c). The first three scenarios, *scenario* (a), *scenario* (b) and *scenario* (c) of the Standard Configuration $(n=15)$.
[**Proof.**]{} As in the case of $n=16$, we get that without loss of generality $1\leq i\leq 4$, $\left|\Gamma (y_i) \cap V(C)\right| =3$ and for each $1\leq j < j' \leq 4$, $\Gamma (y_j) \cap \Gamma (y_{j'}) \cap V(C)=\phi$. Also $\left|\Gamma (y_5) \cap V(C)\right| \in \{ 2,3\}$. In particular, if $\left|\Gamma (y_5) \cap V(C) \right|=2$ then, for each $1\leq j < j' \leq 5$, $\Gamma (y_j) \cap \Gamma (y_{j'}) \cap V(C)=\phi$ and if $\left|\Gamma (y_5) \cap V(C)\right| =3$ then, for each $1\leq j < j' \leq 5$, $$\left|\Gamma (y_j) \cap \Gamma (y_{j'}) \cap V(C)\right|=
\begin{cases}
$ $ 1 & \text{if } j,j' \in \{4,5\}, \\
\hspace{20pt} & \\
$ $ 0 & \text{otherwise. } \\
\end{cases}$$
By Lemma \[l1a\], as $\delta(G) \geq 13$, we get that $\left|\Gamma (y_i) \cap V(H \setminus Y)\right| \geq 10$. Suppose that there is some $1\leq i\leq 3$ (say $i=1$) such that $\left|\Gamma (y_i) \cap V(H \setminus Y)\right|\geq 11$. Then as $r(P_3,K_6)=11$ we get scenario $(a)$. Next, assume that for all $1\leq i\leq 3$, $\left|\Gamma (y_i) \cap V(H \setminus Y)\right|=10$. By the classification of the Ramsey critical $(P_3,K_6)$ graphs, we get that for all $1\leq i\leq 3$, $G[\Gamma (y_i) \cap V(H \setminus Y)] \supseteq 5K_2$.
This gives two possibilities. The first possibility is $\left|{\cup}_{i=1}^3 \Gamma (y_i) \cap V(H \setminus Y)\right|=10$. In this case, as for all $1\leq i\leq 3$, $5K_2 \subseteq G[\Gamma (y_i) \cap V(H \setminus Y)]$ we get scenario $(c)$. The second possibility if $\left|{\cup}_{i=1}^3 \Gamma (y_i) \cup V(H \setminus Y)\right|\geq 11$. Without loss of generality, we may assume that $\left| \Gamma (y_1) \cup \Gamma (y_2) \cup V(H \setminus Y)\right|\geq 11$. Let $x_{11}$ be any vertex of $\Gamma (y_2) \cap (\Gamma (y_1))^c \cup V(H \setminus Y)$. Since $r(P_3,K_6)=11$, we get that $G[\Gamma (y_1) \cup V(H \setminus Y) \cup \{x_{11}\}]$ contains a subgraph $P$ isomorphic to a $P_3$. If $P$ is contained in $G[\Gamma (y_1) \cup V(H \setminus Y)]$ we get scenario $(a)$. Otherwise, $x_{11} \in P$. However, $x_{11}$ is an element of $5K_2 \subseteq G[\Gamma (y_2) \cap V(H \setminus Y)]$ and therefore, $x_{11}$ is adjacent to some other vertex say $w$ in $G[\Gamma (y_2) \cap V(H \setminus Y)]$. Depending on whether or not $w$ belongs to $V(P)$, we get scenarios $(b2)$ or $(b1)$ respectively. Hence the Lemma.
\[l32\] In the Standard Configuration $(n=15)$, $y_1\in Y$ can not be a vertex of a $K_4$ in $H$ (see Figure 2(a)).
[**Proof.**]{} As indicated in Figure 2(a), let $x \in \Gamma (y_1) \cap V(H \setminus Y)$. Then we get two possibilities depending on whether or not $x$ is adjacent to a vertex of $\{y_2,y_3,y_4,y_5\}$. In the first possibility, $x$ is adjacent to some vertex of $Y$ (say $y_2$). Then, as $y_1$ and $y_2$ are connected by paths of lengths 2, 3 and 4 in $H$, no pair of vertices selected from of $\Gamma (y_1) \cap V(C)$ and $\Gamma (y_2) \cap V(C)$ can be separated by a path of length 3, 4 or 5 along the cycle $C$. However, as argued in $n=16$, we get that $\left|\Gamma (y_2) \cap V(C)\right| \leq 2$ and $\left|\Gamma (y_2) \cap V(C)\right| = 2$ only when $\{x',y',z'\}=\{u_1,u_3,u_{8}\}$, $\{x',y',z'\}=\{u_1,u_4,u_{9}\}$ or $\{x',y',z'\}=\{u_1,u_5,u_{9}\}$. This gives a contradiction. In the second possibility, re order the vertices of the cycle such that $y_1\in X$ is adjacent to $u_1$. In this ordering, suppose further that $y_1$ is also adjacent to $u_t$ where $2 \leq t \leq 14$. By the argument used in $n=16$, we get that any pair of vertices adjacent to $y_1$ in $C$ cannot be separated by a path of length 1, 2, 3 or 4 along $C$. However, this again leads to a contradiction.
\[l33\] In the Standard Configuration $(n=15)$, the vertices $y_1,y_2\in Y$ can not be vertices of the subgraph $K$, $K'$ or $L$ in $H$ (see Figure 2).
[**Proof.**]{} In the case $y_1,y_2 \in K$, since $y_1$ and $y_2$ are connected by paths of lengths 3, 4 and 5 in $H$, no pair of vertices selected from of $\Gamma (y_1) \cap V(C)$ and $\Gamma (y_2) \cap V(C)$ can be separated by a path of length 4, 5 or 6 along the cycle $C$. The cardinality of the possible vertex sets of $\Gamma (y_i) \cap V(C)$ ($i=1,2$), subject to this condition, are presented in Table 1.
[ ||p[1.89cm]{}||p[4.55cm]{}||p[6.8cm]{}|| ]{}
$\{x',y',z'\}$ equals
&
$\Gamma (y_2) \cap V(C)$ is contained in
&
Cardinality of $\Gamma (y_2) \cap V(C)$
\
$\{u_1,u_3,u_5\}$ & $\{u_2,u_4\}$ & $\left|\Gamma (y_2) \cap V(C)\right| \leq 2$\
$\{u_1,u_3,u_6\}$ & $\{u_4\}$ & $\left|\Gamma (y_2) \cap V(C)\right| \leq 1$\
$\{u_1,u_3,u_7\}$ & $\{u_4,u_{14}\}$ & $\left|\Gamma (y_2) \cap V(C)\right| \leq 2$\
$\{u_1,u_3,u_8\}$ & $ $ $ $ $\phi$ & $\left|\Gamma (y_2) \cap V(C)\right| = 0$\
$\{u_1,u_3,u_9\}$ & $\{u_{2}\}$ & $\left|\Gamma (y_2) \cap V(C)\right| \leq 1$\
$\{u_1,u_4,u_7\}$ & $ $ $ $ $\phi$ & $\left|\Gamma (y_2) \cap V(C)\right| =0$\
$\{u_1,u_4,u_8\}$ & $ $ $ $ $\phi$ & $\left|\Gamma (y_2) \cap V(C)\right| =0$\
$\{u_1,u_4,u_9\}$ & $\{u_2\}$ & $\left|\Gamma (y_2) \cap V(C)\right| \leq 1$\
$\{u_1,u_5,u_9\}$ & $\{u_2,u_8,u_{12}\}$ & $\left|\Gamma (y_2) \cap V(C)\right| \leq 3$\
$\{u_1,u_5,u_{10}\}$ & $\{u_3,u_8,u_{12}\}$ & $\left|\Gamma (y_2) \cap V(C)\right| \leq 3$\
Table 1: Cardinality of $\Gamma (y_2) \cap V(C)$: Graph $K$.
Because $\Gamma (y_2) \cap V(C)=3$, we are only left to deal with the last two possibilities of Table 1 for $\Gamma (y_2) \cap V(C)$. In both possibilities, $\Gamma (y_1) \cap V(C)$ will induce a $C_3$ by Lemma \[l2\](b). In the first possibility, $\Gamma (y_1) \cap V(C)=\{u_1,u_5,u_9\}$ gives rise to the 15-cycle given by $(u_1,u_5,...,u_8,y_2,y,x,y_1, u_9,...,u_{14},u_1)$, a contradiction. In the second possibility, $\Gamma (y_1) \cap V(C)=\{u_1,u_5,u_{10}\}$ gives rise to the 15-cycle given by $ (u_1,u_5,u_6,...,u_8,y_2,z,y,x,y_1,u_{10},...,u_{14},u_1)$, a contradiction.
In the case $y_1,y_2 \in K'$, since $y_1$ and $y_2$ are connected by paths of lengths 2, 3 or 4 in $H$, no pair of vertices selected from of $\Gamma (y_1) \cap V(C)$ and $\Gamma (y_2) \cap V(C)$ can be separated by a path of length 3, 4 or 5 along the cycle $C$. The cardinality of the possible vertex set of $\Gamma (y_1) \cap V(C)$ is presented in Table 2 and each of these leads to a contradiction as $\Gamma (y_2) \cap V(C)<3$
[ ||p[1.9cm]{}||p[2.7cm]{}||p[4.7cm]{}|| ]{}
$\{x',y',z'\}$ equals
&
$\Gamma (y_2) \cap V(C)$ is contained in
&
Cardinality of $\Gamma (y_2) \cap V(C)$
\
$\{u_1,u_3,u_5\}$ & $ $ $ $ $\phi$ & $\left|\Gamma (y_2) \cap V(C)\right| =0$\
$\{u_1,u_3,u_6\}$ & $ $ $ $ $\phi$ & $\left|\Gamma (y_2) \cap V(C)\right| =0$\
$\{u_1,u_3,u_7\}$ & $\{u_9\}$ & $\left|\Gamma (y_2) \cap V(C)\right| \leq 1$\
$\{u_1,u_3,u_8\}$ & $\{u_2,u_9\}$ & $\left|\Gamma (y_2) \cap V(C)\right| \leq 2$\
$\{u_1,u_3,u_9\}$ & $\{u_2\}$ & $\left|\Gamma (y_2) \cap V(C)\right| \leq 1$\
$\{u_1,u_4,u_7\}$ & $ $ $ $ $\phi$ & $\left|\Gamma (y_2) \cap V(C)\right| =0$\
$\{u_1,u_4,u_8\}$ & $\{u_2\}$ & $\left|\Gamma (y_2) \cap V(C)\right| \leq 1$\
$\{u_1,u_4,u_9\}$ & $\{u_2,u_3\}$ & $\left|\Gamma (y_2) \cap V(C)\right| \leq 2$\
$\{u_1,u_5,u_9\}$ & $\{u_3,u_7\}$ & $\left|\Gamma (y_2) \cap V(C)\right| \leq 2$\
$\{u_1,u_5,u_{10}\}$ & $\{u_3\}$ & $\left|\Gamma (y_2) \cap V(C)\right| \leq 1$\
Table 2: Cardinality of $\Gamma (y_2) \cap V(C)$: Graph $K'$.
In the case $y_1,y_2 \in K'$, since $y_1$ and $y_2$ are connected by paths of lengths 2, 3, 4 and 5 in $H$, no pair of vertices selected from of $\Gamma (y_1) \cap V(C)$ and $\Gamma (y_2) \cap V(C)$ can be separated by paths of length 3, 4, 5 or 6 along the cycle $C$. Thus, Table 2 will give us the required contradiction. Similarly, in the case $y_1,y_2 \in L$, since $y_1$ and $y_2$ are connected by paths of lengths 2, 3, 4, 5 and 6 in $H$, no pair of vertices selected from of $\Gamma (y_1) \cap V(C)$ and $\Gamma (y_2) \cap V(C)$ can be separated by paths of length 3, 4, 5, 6 or 7 along the cycle $C$. As before, for all possibilities $\left|\Gamma (y_2) \cap V(C)\right| <3$, a contradiction. Thus, lemmas 5, 6 and 7 imply that $H$ cannot have an independent set of size 5.
Having proved that $H$ cannot have an independent set of size 5 in all three cases $n=15,16$ and 17, we next continue with the main proof. Since, $H$ satisfies all conditions of Lemma \[l1b\], $H$ contains a $4K_{n-1}$.
Next we show that $V(C_{n-1})$ induced a $K_{n-1}$. Suppose that there exists two vertices of $V(C)$, say $v$ and $w$, such that $(v,w) \not \in E(G)$. In order to avoid a $C_n$ both $v$ and $w$ will have to be adjacent to at most one vertex of each of the four copies of $K_{n-1}$ in $H$. Moreover, any vertex of any copy of $K_{n-1}$ in $H$ will have to be adjacent to at most one vertex of another copy of a $K_{n-1}$ in $H$. Thus, each copy of a $K_{n-1}$ will have at most 5 vertices adjacent to some vertex outside that of $K_{n-1}$, in $V(H) \cup \{v,w\}$. Since $(n-1)-5 \geq 1$, we can select $x_1$ in the first $K_{n-1}$, $x_2$ in the second $K_{n-1}$, $x_3$ in the third $K_{n-1}$ and $x_4$ in the fourth $K_{n-1}$ such that $\{x_1, x_2, x_3, x_4\}$ is an independent set of size four and no vertex of $\{x_1, x_2, x_3, x_4\}$ is adjacent to any vertex of $\{v, w\}$. Hence $\{x_1, x_2, x_3, x_4, v, w\}$ is an independent set of size 6, a contradiction. Therefore, we get that any two pair of vertices of $V(C)$ are connected by an edge. Hence, $G[V(C_{n-1})]=K_{n-1}$ as required. This $K_{n-1}$ along with the $4K_{n-1}$ contained in $H$ gives the required result.
All Ramsey $(C_n,K_6)$ critical graphs for $n \geq 15$
=======================================================
We have already observed that any Ramsey $(C_n,K_6)$ critical graph will consist of a red graph containing $5K_{n-1}$, with respect to the red/blue coloring. Let $\{V_i : $ $ i \in \{1,2,...,5\} \}$ be the vertex set of the five $K_{n-1}$ graphs. We notice that there are two types of Ramsey $(C_n,K_6)$ critical graphs. The first type (Type1) of Ramsey $(C_n,K_6)$ critical graphs will satisfy the condition that at most one vertex of each $V_i$ is adjacent to any other vertex in $V_i^c$. The second type (Type2) of Ramsey $(C_n,K_6)$ critical graphs will satisfy the condition that there exists a $V_k$ for some $1 \leq k \leq 5$ such that at least two vertices of $V_k$ have neighbors in $V_k^c$. Moreover, it is worth noting that a Type1 critical graph is completely determined by the structure of the external edges between $V_i$’s and not by the $\binom{n}{2}$ edges inside each of the five $V_i$’s. This fact is taken into consideration when representing the Ramsey $(C_n,K_6)$ critical graphs.
Each subgraph of $K_5$ generates a unique Ramsey $(C_n,K_6)$ critical graph of Type1. Thus, as illustrated in the following figure, there are 34 critical graphs ($R_i$,$ $ $1 \leq i \leq 34$) of Type1 generated by the 34 subgraphs of $K_5$.
Figure 3. Ramsey $(C_n,K_6)$ critical graphs of Type1, $R_i$ ($1 \leq i \leq 34$)
First note that each and every Type2 critical graph is obtained by an appropriate vertex splitting of some Type1 critical graph. As illustrated in the Figure 4, there are exactly 34 Type2 critical graphs (labeled $S_i$ where $1 \leq i \leq 34$) generated by 18 critical graphs of Type1, since exactly sixteen Type1 critical graphs do not generate Type2 critical graphs.
Figure 4. Ramsey $(C_n,K_6)$ critical graphs of Type2 ($S_i$, $1 \leq i \leq 34$)
Henceforth, we conclude that there are exactly 68 Ramsey $(C_n,K_6)$ critical graphs out of which 34 are categorized as Type1 critical graphs (labeled $R_i$, $1 \leq i \leq 34$ ) and the balance 34 are categorized as Type2 critical graphs (labeled $S_i$, $1 \leq i \leq 34$).
[99]{}
J. A. Bondy and P. Erd[ö]{}s, Ramsey number for cycles in graphs, [*Journal of Combinatorial Theory Series B*]{}, [**14**]{}, (1973), 46-54.
E.T. Baskoro, H. Surahmat and H.J. Broersma, The Ramsey numbers of fans versus $K_4$, [*Bulletin of the Institute of Combinatorial Applications*]{}, [**43**]{}, (2005), 96-102.
B. Bollob[á]{}s, C. J. Jayawardene, Yang Jian Sheng, Huang Yi Ru, C. C. Rousseau and Zhang Ke Min, On a conjecture involving cycle-complete graph Ramsey numbers, [*The Australasian Journal*]{}, [**22**]{}, (2000), 63-71.
J. Hook, The Classification of Critical Graphs and Star-Critical Ramsey Numbers, [*Ph.D. Thesis, Lehigh University*]{}, (2010).
J. Hook and G. Isaak, Star-critical Ramsey numbers, [*Discrete Applied Mathematics*]{}, [**159**]{}, (2011), 328-334.
C.J. Jayawardene, The Star-critical Ramsey Number for any Cycle vs. a $K_5$, arXiv:1901.04802.
C. J. Jayawardene and C. C. Rousseau, An upper bound for the Ramsey number of a quadrilateral versus a complete graph on seven vertices, [*Congressus Numerantium*]{}, (1998), 175-188.
C. J. Jayawardene and C. C. Rousseau, The Ramsey number for a quadrilateral vs. a complete graph on six vertices, [*Congressus Numerantium*]{}, (1997), 97-108.
C.J. Jayawardene and B.L. Samarasekara, Size multipartite Ramsey numbers for $K_4 -e$ verses all graphs up to 4 vertices, [*Annals of Pure and Applied Mathematics*]{} [**13(1)**]{} (2017), 9-26.
S.P. Radziszowski, Small Ramsey numbers, [*Electronic Journal of Combinatorics*]{}, [**14**]{}, (2014), DS1.
I. Schiermeyer, All cycle-Complete Graph Ramsey Number $r(C_m,K_6)$, [*Journal of Graph Theory*]{}, [**44**]{} (2003), 251-260.
Wu Yali, Sun Yongqi and S.P. Radziszowski, Wheel and star-critical Ramsey numbers for quadrilaterals, [*Discrete Applied Mathematics*]{}, [**185**]{}, (2015), 260-271.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper we prove that, in the deep limit, the stochastic gradient descent on a ResNet type deep neural network, where each layer shares the same weight matrix, converges to the stochastic gradient descent for a Neural ODE and that the corresponding value/loss functions converge. Our result gives, in the context of minimization by stochastic gradient descent, a theoretical foundation for considering Neural ODEs as the deep limit of ResNets. Our proof is based on certain decay estimates for associated Fokker-Planck equations.\
2010 [Primary: 68T05, 65L20, Secondary: 34A45, 35Q84, 62F10, 60H10]{}
[*Keywords and phrases: Machine Learning, Deep Neural Network, ResNet, Neural ODE, Ordinary Differential Equation, Stochastic Gradient Descent, Partial Differential Equations, Fokker-Planck.*]{}
address:
- |
Benny Avelin\
Department of Mathematics, Uppsala University\
S-751 06 Uppsala, Sweden
- |
Kaj Nyström\
Department of Mathematics, Uppsala University\
S-751 06 Uppsala, Sweden
author:
- 'Benny Avelin and Kaj Nystr[ö]{}m'
bibliography:
- 'references.bib'
title: |
Neural ODEs as the Deep Limit of\
ResNets with constant weights
---
dotsep[5]{}
Introduction {#sec:intro}
============
ResNet and Neural ODEs {#resnet-and-neural-odes .unnumbered}
----------------------
Ever since the very popular ResNet paper [@He] was published several authors have made the observation that the ResNet architecture is structurally similar to the Euler discretization of an ODE ([@EHL; @LCW]). However, the original ‘ResNet block’ considered in [@He] is defined as $$\begin{aligned}
\label{resnet1}
h_{t+1} = \sigma (h_t + K_t^{(1)} \sigma (K_t^{(2)} h_t + b_t^{(2)} )+b_t^{(1)}), \quad t=0,\ldots, N-1\end{aligned}$$ where $h_t,b_t^{(1)},b_t^{(2)} \in \R^d$ and $K_t^{(1)}, K_t^{(2)} : \R^d \to \R^d$ are linear operators, and $h_N$ represents the output of the network. The activation function $\sigma(x) = \max\{0,x\}$ is applied component-wise on vectors in $\R^d$. $N$ is the number of layers used in the construction. For standard neural networks the operators $K_t^{(1)},K_t^{(2)}$ are matrices, but for convolutional neural networks (CNNs) the operators are discrete convolution operators[^1].
It is not immediately obvious that \[resnet1\] can be viewed as the Euler discretization of an ODE as the activation function $\sigma$ is applied after the addition of $h_t$ and $K_t^{(1)} \sigma (K_t^{(2)} y_t + b_t^{(2)})+b_t^{(1)}$. However, removing[^2] the outermost activation $\sigma$ and instead consider the difference equation $$\begin{aligned}
\label{resnet2}
h_{t+1} = h_t + K_t^{(1)} \sigma (K_t^{(2)} h_t + b_t^{(2)} )+b_t^{(1)}, \quad t=0,\ldots, N-1\end{aligned}$$ we see that this is the Euler discretization (with time-step 1) of the ODE $$\begin{aligned}
\label{neuralODE---}
\dot{h}_t = f(h_t,\theta_t), \quad t=[0,N],\end{aligned}$$ where $$\begin{aligned}
f(\cdot,\theta):\mathbb R^d\to \mathbb R^d,\ f(x,\theta):= K^{(1)} \sigma (K^{(2)} x + b^{(2)} )+b^{(1)},\end{aligned}$$ and where $\theta$ is the short notation for $\theta = (K^{(1)}, K^{(2)},b^{(1)},b^{(2)})$. Note that in \[neuralODE—\] the time-step is 1 and hence the time horizon will be the number of layers $N$. This timescale is not optimal in the sense that if the ODE is stable then the system will attract to zero as $N \to \infty$.
In this paper we consider (general) autonomous ODEs as in \[neuralODE—\], with time independent $\theta$ and a time horizon of $1$, i.e.$$\begin{aligned}
\label{neuralODE--}
\dot{h}_t = f(h_t,\theta), \quad t=[0,1].\end{aligned}$$ This type of model is called a Neural ODE (NODE) [@CRBD] and has recently garnered a lot of attention as it has been proven to work very well in practice. Naturally we also consider the Euler discretization of \[neuralODE–\], which can be written as the difference equation $$\begin{aligned}
\label{resnet3}
h_{t_{i+1}} = h_{t_i} + \frac{1}{N} f(h_{t_i},\theta),\ i=0,...,N-1,\ t_i=i/N.\end{aligned}$$ The models in \[neuralODE–,resnet3\] map $\R^d$ into $\R^d$ for $t=1$. We call the models in \[neuralODE–,resnet3\], (general) ResNet type deep neural networks defined on the interval $[0,1]$. Furthermore, if the input data is in $\R^d$ and the output data in $\R^l$, we have to complement the model with a function mapping $\mathbb R^d\to \mathbb R^l$ resulting in the final output.
Empirical risk minimization {#empirical-risk-minimization .unnumbered}
---------------------------
Assume that we are given data distributed according to a probability measure $\mu$ where $(x,y) \sim \mu$, and $x \in \R^d$, $y \in \R^l$. Let $L : \R^l \times \R^l \to \R$ be a non-negative convex function. The learning problem for a model $h=h(\cdot,\theta):\R^d \times \R^m \to \R^l$, where $\theta\in \R^m$ indicates that the function $h$ is parameterized by weights $\theta$, can be formulated as $$\begin{aligned}
\label{minima}
\min_{\theta}\mathcal{R}(\theta), \ \mathcal{R}(\theta):=\E_{(x,y) \sim \mu} \left [ L(h(x,\theta),y) \right ],\end{aligned}$$ where $\mathcal{R}$ is often referred to as the *risk* or the *risk function*. In practice, the risk we have to minimize is the empirical risk, and it is a well-established fact that for neural networks the minimization problem in \[minima\] is, in general, a non-convex minimization problem [@SafS; @BGMSS; @SJL; @DuLee]. As such many search algorithms may get trapped at, or converge to, local minima which are not global minima [@SafS]. Currently, a variety of different methods are used in deep learning when training the model, i.e. when trying to find an approximate solution to the problem in \[minima\], we refer to [@VBGS] for an overview of various methods. One of the most popular methods, and perhaps the standard way of approaching the problem in \[minima\], is back-propagation using stochastic gradient descent, see [@HH] for a more recent outline of the method. While much emphasis is put on back-propagation in the deep learning community, from a theoretical perspective it does not matter if we use a forward or a backward mode of auto-differentiation.
Continuous approximation of SGD {#continuous-approximation-of-sgd .unnumbered}
-------------------------------
In [@LTE; @LTE2] it is proved that the stochastic gradient descent can be approximated by a continuous time process $$\begin{aligned}
\label{SGDfirst}
d\theta_t = - \grad_\theta \mathcal{R} dt + \sqrt{\Sigma} dW_t,\end{aligned}$$ where $\Sigma$ is a covariance matrix and $dW_t$ is a standard $m$-dimensional Wiener process defined on a probability space. The idea of approximating stochastic gradient descent with a continuous time process has been noted by several authors, see [@COOSC; @CS; @CLTZ; @GL; @MGIVW; @MHB]. A special case of what we prove in this paper, see \[thm:modified\_risk\] below, is that the stochastic gradient descent \[SGDfirst\] used to minimize the risk for the ResNet model in \[resnet3\] converges to the stochastic gradient descent used to minimize the risk for the Neural ODE model in \[neuralODE–\]. This convergence is proved in the sense of expectation with respect to the random initialization of the weights in the stochastic gradient descent. Furthermore, we prove that the corresponding discrepancy errors decay as $N^{-1}$ where $N$ is the number of layers or discretization steps.
Novelty and significance {#novelty-and-significance .unnumbered}
------------------------
It is fair to say that in general there are very few papers making more fundamental and theoretical contributions to the understanding of deep learning and more specifically ResNet like neural networks. However, in the latter case there is a strain of recent and interesting contributions. In [@TVG] the authors allow the parameters of the model to be layer- and time-dependent resulting in non-autonomous ODEs with corresponding Euler discretization: $$\begin{aligned}
\label{neuralODE}
\dot{h}_t = f(h_t,\theta_t),\quad h_{t+1} = h_t + \frac{1}{N} f(h_t,\theta_t).\end{aligned}$$ In particular, based on more restrictive assumptions on $f$, more restrictive compared to what we use, it is proved in [@TVG] that as the number of layers tends to infinity in \[neuralODE\], the risk associated to \[neuralODE\] and defined as in \[minima\], converges in the sense of gamma convergence to the risk associated to the corresponding (continuous) ODE in \[neuralODE\]: we refer to Theorem 2.1 in [@TVG] and to [@Maso] for an introduction to gamma convergence. The authors obtain that the minima for finite layer networks converge to minima of the continuous limit, infinite layer, network. To prove that the limit exists and has nice properties they introduce a regularization which penalizes the norm of the difference between the weights in subsequent layers. We emphasize that in [@TVG] the authors only consider the convergence of minimizers and not the convergence of the actual optimization procedure. In [@EHL] the authors study the limit problem directly and reformulates the problem as an optimal control problem for an ODE acting on measures. However, the complexity of such networks can be quite substantial due to the time-dependency of the weights, and it is unclear what would be the best way to construct a penalization such that the limit has nice properties.
As we mentioned before, in [@CRBD] the authors consider the autonomous ODE in \[neuralODE–\], i.e. they make the assumption that all layers share the same weights, and they develop two things. Firstly, they develop an adjoint equation that allows them to approximately compute the gradient with a depth independent memory cost. Secondly, they show through numerical examples that the approach works surprisingly well for some problems.
In general, the upshot of the ODE approach is the increased regularity, since trajectories are continuous and do not intersect, for autonomous ODEs, they are reversible, see [@DDW]. However, the increased regularity comes with a cost as Neural ODEs can have difficulties solving certain classification problems, see \[sec:numerical\].
Our main contribution is that we establish, in the context of minimization by stochastic gradient descent, a theoretical foundation for considering Neural ODEs as the deep limit ($N \to \infty$) of ResNets.
Overview of the paper {#overview-of-the-paper .unnumbered}
---------------------
The rest of the paper is organized as follows. In \[sec:main\_result\] we introduce the necessary formalism and notation and state our results: \[thm:penalization,thm:modified\_risk,thm:modified\_risk\_msq\]. In \[sec:resnet\_error\] we estimate, for $\theta$ fixed, the discretization error arising as a consequence of the Euler scheme, and we prove some technical estimates. In \[sec:SDE\_FP\] we collect and develop the results concerning stochastic differential equations and Fokker-Planck equations that are needed in the proofs of \[thm:penalization,thm:modified\_risk,thm:modified\_risk\_msq\]. \[sec:fp\_estimates\] is devoted to the Fokker-Planck equations for the probability densities associated to the stochastic gradient descent for the Euler scheme and the continuous ODE, respectively. We establish some slightly delicate decay estimates for these densities, of Gaussian nature, assuming that the initialization density has compact support: see \[lemma1\] below. In \[sec:proof\_of\_main\_thm\] we prove \[thm:penalization,thm:modified\_risk,thm:modified\_risk\_msq\]. In \[sec:numerical\] we discuss a number of numerical experiments. These experiments indicate that in practice the rate of convergence is highly problem dependent, and that it can be considerably faster than indicated by our theoretical bounds. Finally, \[sec:conclusion\] is devoted to a few concluding remarks.
Statement of main results {#sec:main_result}
=========================
Our main results concern (general) ResNet type deep neural networks defined on the interval $[0,1]$. To outline our setup we consider $$\begin{aligned}
f_\theta:\R^d\to\R^d\end{aligned}$$ where $\theta\in \R^m$ is a potentially high-dimensional vector of parameters acting as a degree of freedom. Given $N\in\mathbb Z_+$ we consider $[0,1]$ as divided into $N$ intervals each of length $N^{-1}$, and we define $x_i^{(N)}(x,\theta)$, $i = 0,\ldots, N$, recursively as $$\begin{aligned}
\label{diffeq}
x^{(N)}_{i+1}(x,\theta)&= x^{(N)}_i(x,\theta) + \frac{1}{N} f_\theta(x^{(N)}_i(x,\theta)),\ i = 0,\ldots, N-1,\notag\\
x^{(N)}_0(x,\theta) &= x.\end{aligned}$$ We define $x^{(N)}(t,x,\theta)=x^{(N)}_{i}(x,\theta)$ whenever $t\in [i/N,(i+1)/N)$. We will not indicate the dependency on $x$, $\theta$ when it is unambiguous.
We are particularly interested in the case when $f_\theta(x)$ is a general vector valued (deep) neural network having parameters $\theta$ but in the context of ResNets a specification for $f_\theta$ is, as discussed in the introduction, $$\begin{aligned}
\label{nnspec}
f_\theta(x)=K^{(1)} \sigma (K^{(2)} x + b^{(2)} )+b^{(1)}\end{aligned}$$ where $\theta=(K^{(1)},K^{(2)},b^{(1)},b^{(2)})$ are parameters and $\sigma$ is a globally Lipschitz activation function, for instance sigmoid, tanh, ReLU, Leaky ReLU and Swish, [@RZL]. However, our arguments rely only on certain regularity and growth properties of $f_\theta$. We will formulate our results using the following classes of functions.
\[def:F\] Let $g:\R_+ \to \R_+$, $g(0) \geq 1$, be a non-decreasing function. We say that the function $f_\theta(x):\R^m \times \R^d \to \R^d$ is in regularity class $\mathcal{A}(g)$ if $$\label{eq3++a+l}
\begin{aligned}
\|f_\theta(x)-f_{\theta'}(x)\|&\leq \max\{g(\|\theta\|),g(\|\theta'\|)\}\|\theta-\theta'\|\|x\|,\\
\|f_\theta(x)-f_\theta(x')\|&\leq g(\|\theta\|)\|x-x'\|,\\
\|\grad_{\theta}f_\theta(x)-\grad_{\theta}f_\theta(x')\|&\leq g(\|\theta\|)\max\{\|x\|,\|x'\|\}\|x-x'\|,\\
\|\grad_{x}f_\theta(x)-\grad_{x}f_\theta(x')\|&\leq g(\|\theta\|)\|x-x'\|,
\end{aligned}$$ whenever $x,x'\in\mathbb R^d$, $\theta$, $\theta'\in\mathbb R^m$.
Some remarks are in order concerning \[def:F\]. Firstly, it contains the prototype neural network in \[nnspec\], with $g(s) = s+1$. Secondly, it is essentially closed under compositions, see \[lem:AgComposition\] below. Therefore, finite layer neural networks satisfy \[def:F\]. We defer the proof of the following lemma to \[sec:resnet\_error\].
\[lem:AgComposition\] Let $f_\theta,g_\theta \in \mathcal{A}(g)$. Then $F_{\theta} = f_{\theta_2} \circ g_{\theta_1} \in \mathcal{A}(2g^3)$ for $\theta = (\theta_1,\theta_2)$.
Certain results in this paper require us to control the second derivatives of the risk. We therefore also introduce the following class of functions.
\[def:F+\] Let $g:\R_+ \to \R_+$, $g(0) \geq 1$, be a non-decreasing function. We say that the function $f_\theta(x):\R^m \times \R^d \to \R^d$ is in regularity class $\mathcal{A}_+(g)$ if $f_\theta\in\mathcal{A}(g)$ and if there exists a polynomial $P:\R \to \R$ such that $$ \begin{aligned}
\|\grad_\theta^2 f_\theta(x)\| + \|\grad_\theta \grad_x f_\theta(x)\| &\leq g(\|\theta\|)P(\|x\|), \\
\|\grad^2_x f\| &\leq g(\|\theta\|),
\end{aligned}$$ whenever $x\in\mathbb R^d$, $\theta \in\mathbb R^m$.
The following lemma follows from \[lem:AgComposition\] and an elementary calculation using \[def:F+\].
Let $f_\theta,g_\theta \in \mathcal{A}_+(g)$, then $F_{\theta} = f_{\theta_2} \circ g_{\theta_1} \in \mathcal{A}_+(2g^3)$ for $\theta = (\theta_1,\theta_2)$.
Given a probability measure $\mu$, $(x,y) \sim \mu$, $x \in \R^d$, $y \in \R^d$, and with $x^{(N)}(\cdot,\theta)$ defined as in \[diffeq\], we consider the penalized risk $$\begin{aligned}
\mathcal{R}^{(N)}(\theta) := \E_{(x,y) \sim \mu} \left [ \|y-x^{(N)}(1,\theta)\|^2\right ]+\gamma H(\theta)\end{aligned}$$ where $\gamma\in \R_+$ is a hyper-parameter and $H=H(\theta)$ is a non-negative and convex regularization. The finite layer model in \[diffeq\] is, as described in \[sec:intro\], the forward Euler discretization of the autonomous system of ordinary differential equations $$\begin{aligned}
\label{ODE}
\dot{x}(t)&=f_\theta(x(t)),\ t\in (0,1],\ x(0)=x,\end{aligned}$$ where $x(t)=x(t,\theta) \in\mathbb R^d$. Given data from the distribution $\mu$ and with $x(\cdot,\theta)$ solving the system of Neural ODEs in \[ODE\], we consider the penalized risk $$\begin{aligned}
\mathcal{R}(\theta) := \E_{(x,y) \sim \mu} \left [ \|y-x(1,\theta)\|^2\right ]+\gamma H(\theta).\end{aligned}$$ Throughout this paper we will assume that **moments of all orders are finite** for the probability measure $\mu$. By construction the input data, $x$, as well as $x(t,\theta)$ are vectors in $\mathbb R^d$. In the case when the output data is $\mathbb R^l$, we need to modify the $\mathcal{R}^{(N)}(\theta)$ and $\mathcal{R}(\theta)$ by performing final transformations of $x^{(N)}(1,\theta)$ and $x(1,\theta)$ achieving outputs in $\mathbb R^l$. These modifications are trivial to incorporate and throughout the paper we will in our derivations therefore simply assume that $l=d$.
We will base our analysis on the following continuous in time approximations of the stochastic gradient descent, see [@LTE; @LTE2], $$\begin{aligned}
\label{sgdsys}
d\theta^{(N)}_t &= - \nabla \mathcal{R}^{(N)}(\theta^{(N)}_t)dt + \Sigma dW_t,\notag \\
d\theta_t &= - \nabla \mathcal{R}(\theta_t)dt + \Sigma dW_t,\end{aligned}$$ for $t \in [0,T]$. Throughout the paper we will assume, for simplicity, that the constant covariance matrix $\Sigma$ has full rank something which, in reality, may not be the case, see [@CS]. We want to understand in what sense $\theta^{(N)}_t$ approximates $\theta_t$ as $N \to \infty$. To answer this, we first need to make sure that $\theta^{(N)}_t$ and $\theta_t$ exist. Note that $\nabla \mathcal{R}^{(N)}(\theta)$ and $\nabla \mathcal{R}(\theta)$ can, as functions of $\theta$, grow exponentially: simply consider the ODE $\dot{x} = \theta x$ which has $e^{\theta t}$ as a solution. This creates problems as the setting does not fit the standard theory of SDEs, see [@GS], a theory which most commonly requires that the growth of the drift term is at most linear. However, if the drift terms are *confining potentials*, i.e. $$-\nabla \mathcal{R}(\theta) \cdot \theta \leq c(1+\|\theta\|^2),\ -\nabla \mathcal{R}^{(N)}(\theta) \cdot \theta \leq c(1+\|\theta\|^2)$$ then we have existence and uniqueness for the SDEs in \[sgdsys\], see \[sec:SDE\_FP\]. In particular, if we have a bound on the growth of $\grad_\theta x^{(N)}(\cdot,\theta)$, $\grad_\theta x(\cdot,\theta)$ then, as we will see, we can choose $H$ to have sufficiently strong convexity to ensure the existence of a large constant $c$ such that the drift terms are confining potentials in the sense stated.
If we choose $H$ so that $\mathcal{R}^{(N)}$ and $\mathcal{R}$ are convex outside some large ball then $\mathcal{R}^{(N)}$ and $\mathcal{R}$ can be seen as bounded perturbations of strictly convex potentials, see the proof of \[thm:penalization\]. Using this we can use the log-Sobolev inequality and hyper-contractivity properties of certain semi-groups, see \[sec:SDE\_FP\], to obtain very good tail bounds for the densities of $\theta_t^{(N)}$ and $\theta_t$. Actually these tail bounds are good enough for us to prove that the expected risks are bounded, expectation is over trajectories, and that $\theta^{(N)}_t \to \theta_t$ in probability. However, these bounds do not seem to be strong enough to give us quantitative convergence estimates for the difference $|\mathbb{E} [\mathcal{R}(\theta_T)]-\mathbb{E} [\mathcal{R}^{(N)}(\theta^{(N)}_T)]|$. The main reason for this is that even though we have good tail bounds for the densities of $\theta^{(N)}_t$ and $\theta_t$ we do not have good estimates for $\theta^{(N)}_t - \theta_t$ in terms of $N$. The following is our first theorem.
\[thm:penalization\] Let $g:\R_+ \to \R_+$, $g(0) \geq 1$, be a non-decreasing function and assume that $f_\theta\in \mathcal{A}_+(g)$. Given $d$, $m$, there exists a regularizing function $H=H(g)$ such that if we consider the corresponding penalized risks $\mathcal{R}^{(N)}$ and $\mathcal{R}$, defined using $H$ and with $\gamma=1$, then $\mathcal{R}^{(N)}$ and $\mathcal{R}$ are bounded perturbations of strictly convex functions. Furthermore, given $T>0$ and a compactly supported initial distribution $p_0$ for $\theta_0$, we have $$\begin{aligned}
\sup_{t\in[0,T]} \|\theta_t-\theta^{(N)}_t\| \to 0 \quad \text{in probability as $N \to \infty$}
\end{aligned}$$ and $$\begin{aligned}
\mathbb{E} [\mathcal{R}(\theta_T)] < \infty, \quad \mathbb{E} [\mathcal{R}^{(N)}(\theta^{(N)}_T)] < \infty.
\end{aligned}$$
\[thm:penalization\] remains true but with a different rate of convergence, if we replace $\gamma=1$ with $0<\gamma < 1$.
There are a number of ways to introduce more restrictive assumptions on the risk in order to strengthen the convergence $\theta^{(N)}_t \to \theta_t$ and in order to obtain quantitative bounds for the difference $|\mathbb{E} [\mathcal{R}(\theta_T)]-\mathbb{E} [\mathcal{R}^{(N)}(\theta^{(N)}_T)]|$. Our approach is to truncate the loss function. This can be done in several ways, but a very natural choice is to simply restrict the hypothesis space by not allowing weights with too large norm. Specifically, we let $\Lambda>0$ be a large degree of freedom, and we consider $$\begin{aligned}
\label{eq:modified_risk-}
{\widetilde}{\mathcal{R}}^{(N)}(\theta)&:= \E_{(x,y) \sim \mu} \left [ \|y-x^{(N)}(1,T_{\Lambda}(\theta))\|^2\right ]+\gamma H(\theta),\notag\\
{\widetilde}{\mathcal{R}}(\theta)&:= \E_{(x,y) \sim \mu} \left [ \|y-x(1,T_{\Lambda}(\theta))\|^2\right ]+\gamma H(\theta),\end{aligned}$$ instead of $\mathcal{R}^{(N)}$ and $\mathcal{R}$, where $T_{\Lambda}(\theta)$ is a smooth function such that $T_{\Lambda}(\theta)= \theta$ if $\|\theta\| \leq {\Lambda}$ and $T_{\Lambda}(\theta) = \frac{\theta}{\|\theta\|} 2{\Lambda}$ if $\|\theta\| \geq 2 {\Lambda}$. It should be noted that, the choice of considering only the quadratic loss function is purely for simplicity and our technique works with minor modification for any other convex loss.
Having truncated the loss functions we run continuous forms of SGDs $$\begin{aligned}
d\theta^{(N)}_t &= - \nabla {\widetilde}{\mathcal{R}}^{(N)}(\theta^{(N)}_t)dt + \Sigma dW_t, & \theta_0^{(N)} \sim p_0\\
d\theta_t &= - \nabla {\widetilde}{\mathcal{R}}(\theta_t)dt + \Sigma dW_t, & \theta_0 \sim p_0\end{aligned}$$ to minimize the modified risks. Using this setup, assuming also that $H(\theta)\approx \|\theta\|^2$ when $\|\theta\|$ is large, the modified risks in \[eq:modified\_risk-\] will satisfy quadratic growth estimates at infinity, and the modified risk will be globally Lipschitz. As a consequence all the tools from classical stochastic differential equations are at our disposal, see \[sec:SDE\_FP\]. This allows us to prove that $\theta^{(N)}_t\to \theta_t$ in the sense of mean square convergence. However, still the classical SDE theory does not easily seem to allow us to conclude that ${\widetilde}{\mathcal{R}}^{(N)}(\theta_t^{(N)})$ converges in any meaningful way to ${\widetilde}{\mathcal{R}}(\theta_t)$. To overcome this difficulty we develop a PDE based approach to obtain further estimates for the densities of $\theta^{(N)}_t$ and $\theta_t$, and their differences. In particular, we prove the following theorem.
\[thm:modified\_risk\] Let $g:\R_+ \to \R_+$, $g(0) \geq 1$, be a non-decreasing function and assume that $f_\theta\in \mathcal{A}(g)$. Let $\Lambda\gg 1$ and $T>0$ be fixed and assume that the initial weight randomization distribution $p_0$ has compact support in $B(0,R_0)$, $R_0\geq 1$. Assume also that $$\lambda^{-1}\|\theta\|^2\leq H(\theta)\leq \lambda\|\theta\|^2$$ on $\mathbb R^m\setminus B(0,\rho_0)$, $\rho_0\geq 1$ and for some $\lambda\geq 1$. Then for given $\gamma > 0$, $\Lambda > 0$ and positive definite $\Sigma$, there exists a positive and finite constant $c$, depending on the function $g$ as well as $d$, $m$, $\Lambda$, $\gamma$, $\Sigma$, $T$, $R_0$, $\lambda$ and $\rho_0$, such that $$\begin{aligned}
\label{thm2.4a}
\sup_{t\in[0,T]}\bigl \|\E[\theta_t-\theta^{(N)}_t]\bigr \|&\leq c N^{-1}\|p_0\|_2,\\
\label{thm2.4b}
\sup_{t\in[0,T]}\bigl |\E[{\widetilde}{\mathcal{R}}(\theta_t)-{\widetilde}{\mathcal{R}}^{(N)}(\theta^{(N)}_t) ]\bigr |&\leq c N^{-1}\|p_0\|_2.
\end{aligned}$$ Furthermore, if ${\widetilde}R_0\geq 2R_0$, then $$\begin{aligned}
\sup_{t\in[0,T]}\bigl \|
\E[\theta_t | \theta_t \in B_{{\widetilde}R_0}]-\E[\theta^{(N)}_t | \theta^{(N)}_t \in B_{{\widetilde}R_0}] \bigr \|&\leq c N^{-1}e^{-{\widetilde}R_0^2/T}\|p_0\|_2,\notag\\
\sup_{t\in[0,T]}\bigl |
\E[{\widetilde}\Risk(\theta_t) | \theta_t \in B_{{\widetilde}R_0}]-\E[{\widetilde}\Risk^{(N)}(\theta^{(N)}_t) | \theta^{(N)}_t \in B_{{\widetilde}R_0}] \bigr |&\leq c N^{-1}e^{-{\widetilde}R_0^2/T}\|p_0\|_2.\notag
\end{aligned}$$
The requirement that the regularization term $H$ grows quadratically is crucial for our quantitative estimates, and as such our result only covers $L^2$ regularization (weight decay) and not sparse type regularizations like $L^1$.
To prove \[thm:modified\_risk\] we develop certain estimates for $p^{(N)}(\theta,t)$, $p(\theta,t)$, i.e. for the to $\theta^{(N)}(t)$, $\theta(t)$, associate probability densities, by exploring the associated Fokker-Planck equations: see \[eq2a\]. In fact, we prove several estimates which give that $p^{(N)}(\theta,T)\to p(\theta,T)$ in a very strong sense, stronger than initially is recognized from the statement of \[thm:modified\_risk\]. Particular consequences of our proofs are the estimates $$\begin{aligned}
\label{esta0}
\int\limits_{\mathbb R^m}e^{\gamma H(\theta)/4}(p(\theta,T)+p^{(N)}(\theta,T))\, d\theta&\leq c \|p_0\|_2,\\
\int\limits_{B(0, 2^{k+1}{\widetilde}R_0)\setminus B(0, 2^{k}{\widetilde}R_0)}e^{\gamma H(\theta)/4}(p(\theta,T)+p^{(N)}(\theta,T))\, d\theta&\leq c e^{-2^k{\widetilde}R_0^2/T} \|p_0\|_2, \notag\end{aligned}$$ and $$\begin{aligned}
\label{esta1}
\int\limits_{\mathbb R^m}e^{\gamma H(\theta)/4}|p(\theta,T)-p^{(N)}(\theta,T)|\, d\theta&\leq c N^{-1}\|p_0\|_2,\\
\int\limits_{B(0, 2^{k+1}{\widetilde}R_0)\setminus B(0, 2^{k}{\widetilde}R_0)}e^{\gamma H(\theta)/4}|p(\theta,T)-p^{(N)}(\theta,T)|\, d\theta&\leq c e^{-2^k{\widetilde}R_0^2/T} N^{-1}\|p_0\|_2, \notag\end{aligned}$$ whenever ${\widetilde}R_0\geq 2R_0$, $k\in\mathbb N$ and where $c= c(g,d,m, \Lambda, \gamma, \Sigma, T,R_0, \lambda, \rho_0)$. In particular, these estimates indicate that $p(\theta,T)$, $p^{(N)}(\theta,T)$ and $|p(\theta,T)-p^{(N)}(\theta,T)|$ have Gaussian tails away from the (compact) support of $p_0$.
The estimates \[esta0,esta1\] can be interpreted from a probabilistic point of view. The bound \[esta0\] for $p$ is equivalent to $\E[e^{\gamma H(\theta_T)/4}] \leq c \|p_0\|_2$, which implies that all moments of $H(\theta_T)$ are finite. Secondly we can interpret \[esta1\] as the total variation distance between $p$ and $p^{(N)}$ being of order $N^{-1}$ with respect to the measure $e^{\gamma H(\theta)/4} d\theta$.
A direct consequence of \[esta1\] is the following corollary which states that $\theta_t^{(N)}$ is a weak order 1 approximation of $\theta_t$.
Let $\theta_t,\theta_t^{(N)}$ be as in \[thm:modified\_risk\]. Let $\varphi: \R^m \to \R$ be a continuous function satisfying the growth condition $$\begin{aligned}
|\varphi(x)| \leq P(x), \quad x \in \R^m,
\end{aligned}$$ for some polynomial $P:\R^m \to \R$ of order $k$. Then there exists a constant $c$ depending on $g$ as well as $d$, $m$, $\Lambda$, $\gamma$, $\Sigma$, $T$, $R_0$, $\lambda, \rho_0$ and $P$ such that $$\begin{aligned}
\sup_{t\in[0,T]}\bigl |\E[\varphi(\theta_t)-\varphi(\theta^{(N)}_t)]\bigr |&\leq c N^{-1}\|p_0\|_2.
\end{aligned}$$
Using our estimates we can also derive the following theorem from the standard theory.
\[thm:modified\_risk\_msq\] Let $g:\R_+ \to \R_+$, $g(0) \geq 1$, be a non-decreasing function and assume that $f_\theta\in \mathcal{A}(g)$. Let $\Lambda\gg 1$ and $T>0$ be fixed and assume that $\E[\|\theta_0\|^2]< \infty$. Then $$\begin{aligned}
\E [\sup_{t\in[0,T]} \|\theta(t)-\theta^{(N)}(t)\|^2 ] \to 0.
\end{aligned}$$
ResNet and the Neural ODE: error analysis {#sec:resnet_error}
=========================================
In this section we will estimate the error formed by the discretization of the ODE. This will be done on the level of the trajectory as well as on the level of the first and second derivatives with respect to $\theta$. Note that by construction, see \[diffeq,ODE\] we have $$\begin{aligned}
x^{(N)}(t,x,\theta)&= x+\int_0^tf_\theta(x^{(N)}(\tau,x,\theta))\, d\tau,\notag\\
x(t,x,\theta)&= x+\int_0^t f_\theta(x(\tau,x,\theta))\, d\tau,\end{aligned}$$ for all $t\in [0,1]$. We assume consistently that the paths $x^{(N)}(t,x,\theta)$ and $x(t,x,\theta)$ start at $x$ for $t=0$, and are driven by the parameters $\theta$. We will in the following, for simplicity, use the notation $x^{(N)}(t):=x^{(N)}(t,x,\theta)$, $x_i^{(N)}(t)=x^{(N)}(i/N)$, $x(t):=x(t,x,\theta)$. Recall that ${\widetilde}{\mathcal{R}}^{(N)}$ and ${\widetilde}{\mathcal{R}}$ are introduced in \[eq:modified\_risk-\] using the cut-off parameter $\Lambda$. Let $\hat{\mathcal{R}}^{(N)}(\theta):={\widetilde}{\mathcal{R}}^{(N)}(\theta)-\gamma H(\theta)$ and $\hat{\mathcal{R}}(\theta):={\widetilde}{\mathcal{R}}(\theta)-\gamma H(\theta)$. In this section we prove estimates on the discrepancy error between the trajectories of $x(t)$ and the discrete trajectories $x_i^{(N)}$. We begin by bounding the difference.
\[lemmadiff1\]Let $g:\R_+ \to \R_+$, $g(0) \geq 1$, be a non-decreasing function and assume that $f_\theta\in \mathcal{A}(g)$ as in \[def:F\]. Then there exists a function ${\widetilde}g:\mathbb R_+\to\mathbb R_+$ such that $$\label{eq:xdiff}
\left \|x \left ({i}/{N}\right ) - x_i^{(N)} \right \| \leq \frac {{\widetilde}g(g(\|\theta\|))}N \|x\|$$ holds for all $i \in \{0,\ldots,N\}$, and such that $$\label{eq:xbdd}
\big \|x \left ({i}/{N}\right ) \big \| + \big \|x_i^{(N)} \big \| \leq {{\widetilde}g(g(\|\theta\|))} \|x\|,$$ holds for all $i \in \{0,\ldots,N\}$.
To start the proof we first write $$\begin{gathered}
\label{eq:discContDiff}
\left \|x({i}/{N}) - x_i^{(N)} \right \|= \left \| x (({i-1})/{N}) + \int_{\frac{i-1}{N}}^{\frac{i}{N}} \dot{x}(t) dt - x_{i-1}^{(N)} - (x_{i}^{(N)}-x_{i-1}^{(N)}) \right \| \\
\leq \left \| x( ({i-1})/{N}) - x_{i-1}^{(N)}\right \| + \left \| \int_{\frac{i-1}{N}}^{\frac{i}{N}} \dot{x}(t) dt - (x_i^{(N)} - x_{i-1}^{(N)}) \right \|.
\end{gathered}$$ To bound the second term in \[eq:discContDiff\] we note that $$\begin{gathered}
\label{eq:discContDiff2}
\left \| \int_{\frac{i-1}{N}}^{\frac{i}{N}} \dot{x}(t) dt - (x_i^{(N)} - x_{i-1}^{(N)}) \right \| = \left \| \int_{\frac{i-1}{N}}^{\frac{i}{N}} f_\theta(x(t)) -
f_\theta( x_{i-1}^{(N)}) dt \right \| \\
\leq g(\|\theta\|) \int_{\frac{i-1}{N}}^{\frac{i}{N}} \left \|x(t) - x_{i-1}^{(N)} \right \|dt,
\end{gathered}$$ by \[def:F\]. By the triangle inequality, the integral on the right-hand side of \[eq:discContDiff2\] is bounded by $$\begin{gathered}
\label{eq:discContDiff2int}
\int_{\frac{i-1}{N}}^{\frac{i}{N}} \left \| x(t) - x_{i-1}^{(N)} \right \| dt\leq \frac{1}{N} \left \| x_{i-1}^{(N)} - x(({i-1})/{N}) \right \| \\
+ \int_{\frac{i-1}{N}}^{\frac{i}{N}} \left \| x(t) - x(({i-1})/{N}) \right \|dt.
\end{gathered}$$ By the definition of $x(t)$ and \[def:F\] we have that the second term in \[eq:discContDiff2int\] is bounded as $$\begin{gathered}
\label{eq:second}
\int_{\frac{i-1}{N}}^{\frac{i}{N}}\left \|x(t) - x(({i-1})/{N}) \right \| dt \leq \int_{\frac{i-1}{N}}^{\frac{i}{N}} \int_{\frac{i-1}{N}}^{t}\|f_\theta(x(r))\|\, dr dt \\
\leq g(\|\theta\|)\int_{\frac{i-1}{N}}^{\frac{i}{N}} \int_{\frac{i-1}{N}}^{t}\|x(r)\|\, dr dt+\frac{1}{N^2}\|x\|
\leq \frac {2g(\|\theta\|)}{N^2}\|x\|_{L^{\infty}([0,1])}
\end{gathered}$$ whenever $t\in [(i-1)/N,i/N]$. Assembling \[eq:discContDiff,eq:discContDiff2,eq:discContDiff2int,eq:second\] we arrive at $$\begin{aligned}
\label{eq:Aibound}
A_i \leq \left (1+\frac{g(\|\theta\|)}{N} \right ) A_{i-1} + 2\frac{g^2(\|\theta\|)}{N^2} \|x\|_{L^\infty([0,1])}
\end{aligned}$$ where $A_i = \left \|x({i}/{N}) - x_i^{(N)} \right \|$ for $i = 0,1,\ldots,N$, and $A_0 = 0$. \[eq:Aibound\] can be rewritten as $A_i \leq C_0 A_{i-1} + C_1$. Iterating this inequality we see that $$\begin{gathered}
\label{eq:AiIteration}
A_k \leq C_0 A_{k-1} + C_1 \leq C_0 (C_0 A_{k-1} + C_1) + C_1 = C_0^2 A_{k-2} + C_0C_1 + C_1 \\
\leq C_0^{k} A_0 + C_1 \sum_{j=0}^{k-1} C_0^j = C_1 \sum_{j=0}^{k-1} C_0^j \leq C_1 \frac{C_0^N - 1}{C_0-1},
\end{gathered}$$ for any $k \in \{1,\ldots, N\}$. Elementary calculations give us $$\begin{aligned}
\label{eq:diffxbound}
\left \|x({i}/{N}) - x_i^{(N)} \right \| \leq & \frac {2g^2(\|\theta\|)}{N}e^{g(\|\theta\|)}\|x\|_{L^\infty([0,1])}.
\end{aligned}$$ To finalize the proof of \[eq:xdiff\] we need to establish a bound for $\|x\|_{L^\infty([0,1])}$ in terms of the initial value $\|x\|$. By the definition of $x(t)$ and \[def:F\] we have for any $t_1 \in [0,1]$ such that $t_1+t \leq 1$ $$\begin{aligned}
\|x(\cdot)\|_{L^{\infty}([t_1,t_1+t])} &= \sup_{s \in [t_1,t_1+t]} \left \|\int_{t_1}^{t_1+s} \dot{x}(r) dr + x({t_1}) \right \| \\
&= \sup_{s \in [t_1,t_1+t]} \left \|\int_{t_1}^{t_1+s} f_\theta(x(r)) dr + x({t_1}) \right \|\\
&\leq g(\|\theta\|)t \|x(\cdot)\|_{L^{\infty}([t_1,t_1+t])} +t\|f_\theta(x(t_1))\|+\|x({t_1})\|\\
&\leq g(\|\theta\|)t \|x(\cdot)\|_{L^{\infty}([t_1,t_1+t])} + (g(\|\theta\|)t+1)\|x({t_1})\|+\|x\|.
\end{aligned}$$ Choosing $t$ small enough i.e. $t = \frac{1}{2 g(\|\theta\|)}$ we deduce $$\begin{aligned}
\|x\|_{L^{\infty}([t_1,t_1+t])} &\leq 4(\|x_{t_1}\|+\|x\|).
\end{aligned}$$ Iterating the above inequality $N =2 \lceil g(\|\theta\|) \rceil$ times we obtain $$\begin{aligned}
\label{bound2}
\|x\|_{L^{\infty}([0,1])} &\leq 4^{2 g(\|\theta\|)}\|x\|.
\end{aligned}$$ Combining \[eq:diffxbound,bound2\] proves \[eq:xdiff\].
It remains to prove \[eq:xbdd\] and to start the proof note that the first term on the left in \[eq:xbdd\] is already bounded by \[bound2\]. Thus, we only need to establish that $\|x_i^{(N)}\|$ is bounded. We first note, using the definition of $x_i^{(N)}$ and \[def:F\], that $$\begin{aligned}
\left \| x_{i+1}^{(N)} - x_{i}^{(N)}\right \|\leq \frac{2g(\|\theta\|)}{N} \left ( \|x_i^{(N)} \|+\|x \|\right ).
\end{aligned}$$ By the triangle inequality we get by rearrangement that $$\begin{aligned}
\| x_{i+1}^{(N)} \| \leq \left ( 1 + \frac{2g(\|\theta\|)}{N} \right ) \|x_{i}^{(N)}\| + \frac{2g(\|\theta\|)}{N}\|x \|.
\end{aligned}$$ This is again an estimate of the form $A_i \leq C_0 A_{i-1} + C_1$, where $A_0 = \|x\|$. Iterating this we obtain as in \[eq:AiIteration\] that $$\begin{aligned}
A_i \leq C_0^N A_0 + C_1 \frac{C_0^N - 1}{C_0-1},
\end{aligned}$$ and by elementary calculations as in \[eq:diffxbound\] we can conclude that $$\begin{aligned}
\|x_{i}^{(N)}\| \leq e^{2 g(\|\theta\|)} \|x\|, \quad i=0,1,\ldots,N.
\end{aligned}$$ This proves the final estimate stated in \[eq:xbdd\] and finishes the proof.
We next upgrade the previous lemma to the level of the gradient of the trajectories with respect to $\theta$.
\[lemmadiff20\] Let $g:\R_+ \to \R_+$, $g(0) \geq 1$, be a non-decreasing function and assume that $f_\theta\in \mathcal{A}(g)$ as in \[def:F\]. Then there exists a function ${\widetilde}g:\mathbb R_+\to\mathbb R_+$ such that $$\label{eq:xdiffgrad}
\left \|\grad_\theta x \left ({i}/{N}\right ) - \grad_\theta x_i^{(N)} \right \| \leq \frac {{\widetilde}g(g(\|\theta\|))}N \|x\|$$ holds for all $i \in \{0,\ldots,N\}$, and such that $$\label{eq:xbddgrad}
\big \|\grad_\theta x \left ({i}/{N}\right ) \big \| + \big \|\grad_\theta x_i^{(N)} \big \| \leq {{\widetilde}g(g(\|\theta\|))} \|x\|,$$ holds for all $i \in \{0,\ldots,N\}$.
We begin by proving \[eq:xdiffgrad\] under the assumption that \[eq:xbddgrad\] holds. First note that $$\begin{gathered}
\label{eq:5.1}
\| \grad_\theta x(i/N)-\grad_\theta x^{(N)}_i\|
\leq \|\grad_\theta x_{i-1}^{(N)} - \grad_\theta x((i-1)/N)\| \\
+\|\grad_\theta x_i^{(N)} - \grad_\theta x_{i-1}^{(N)} - \grad_\theta \int_{\frac{i-1}{N}}^{\frac{i}{N}} \dot{x}(r)\, dr \|.
\end{gathered}$$ As in the proof of \[lemmadiff1\] we want to build iterative inequalities, and our first goal is to bound the second term in \[eq:5.1\]. First note that $$\begin{gathered}
\bigg \|\grad_\theta x_i^{(N)} - \grad_\theta x_{i-1}^{(N)} - \grad_\theta \int_{\frac{i-1}{N}}^{\frac{i}{N}} \dot{x}(r)\, dr \bigg \| \notag\\
= \left \|\frac{1}{N}\grad_\theta f_\theta(x_{i-1}^{(N)}) - \grad_\theta \int_{\frac{i-1}{N}}^{\frac{i}{N}} f_\theta(x(r))\, dr \right \|\notag \\
= \left \|\int_{\frac{i-1}{N}}^{\frac{i}{N}} \grad_\theta(f_\theta(x(r) - f_\theta(x_{i-1}^{(N)}) )\, dr \right \|.
\end{gathered}$$ Now using \[def:F\] multiple times we see that $$\begin{aligned}
\label{eq:5.1B}
\|\grad_\theta (f_\theta(x(r)) - &f_\theta(x_{i-1}^{(N)}) )\|\notag
\\
\leq& \|\grad_x f_\theta(x(r))\grad_\theta x(r)-\grad_x f_\theta(x_{i-1}^{(N)})\grad_\theta x_{i-1}^{(N)}\|\notag
\\
&+g(\|\theta\|)\max\{\|x(r)\|,\|x_{i-1}^{(N)}\|\}\|x(r)-x_{i-1}^{(N)}\|\notag\\
\leq&\|\grad_x f_\theta(x(r))\|\|\grad_\theta x(r)-\grad_\theta x_{i-1}^{(N)}\|\notag\\
&+\|\grad_x f_\theta(x(r))-\grad_x f_\theta(x_{i-1}^{(N)})\|\|\grad_\theta x_{i-1}^{(N)}\|\\
&+g(\|\theta\|)\max\{\|x(r)\|,\|x_{i-1}^{(N)}\|\}\|x(r)-x_{i-1}^{(N)}\|, \notag\\
\leq & g(\|\theta\|)\|\grad_\theta x(r)-\grad_\theta x_{i-1}^{(N)}\|\notag\\
&+g(\|\theta\|)\|x(r)-x_{i-1}^{(N)}\|(\max\{\|x(r)\|,\|x_{i-1}^{(N)}\|\}+\|\grad_\theta x_{i-1}^{(N)}\|). \notag
\end{aligned}$$ We want to bound the terms on the right-hand side in \[eq:5.1B\] and to do this we first note that $$\begin{gathered}
\label{eq:5.gradABdiff21}
\|\grad_\theta x(r) - \grad_\theta x_{i-1}^{(N)}\| \leq \|\grad_\theta x(r) - \grad_\theta x\bigl ({({i-1})/{N}}\bigr )\|\\
+ \|\grad_\theta x\bigl ({({i-1})/{N}}\bigr ) - \grad_\theta x_{i-1}^{(N)}\|.
\end{gathered}$$ The second term appearing in \[eq:5.gradABdiff21\] is what we want to bound in an iterative scheme. Focusing on the first term in \[eq:5.gradABdiff21\], using \[def:F\], we see that $$\begin{aligned}
\label{eq:5.1C}
\notag \|\grad_\theta x(r) - \grad_\theta x\bigl ({({i-1})/{N}}\bigr )\|&=\bigg \|\grad_\theta \int_{\frac{i-1}{N}}^r \dot{x}(s)\, ds \bigg \|\notag\\
&=\bigg \|\grad_\theta \int_{\frac{i-1}{N}}^r f_\theta (x(s))\, ds \bigg \|\notag\\
&\leq \bigg \|\int_{\frac{i-1}{N}}^r \grad_x f_\theta (x(s))\grad_\theta x(s)+\grad_\theta f_\theta (x(s))\, ds \bigg \|\notag\\
&\leq g(\|\theta\|) \int_{\frac{i-1}{N}}^r ( \| x(s)\| + \|\grad_\theta x(s)\| )\, ds.
\end{aligned}$$ Again, since we assume \[eq:xbddgrad\] we can use \[eq:5.1,eq:5.1B,eq:5.1C,eq:5.gradABdiff21,lemmadiff1\] to get $$\begin{aligned}
A_i \leq \left ( 1 + \frac{g}{N} \right )A_{i-1} + \frac{1}{N^2} g_1(g(\| \theta \|))\max\{\|x\|,\|x\|^2\},
\end{aligned}$$ for some non-decreasing function $g_1$, $A_i := \| \grad_\theta x(i/N)-\grad_\theta x^{(N)}_i\|$, and $A_0 := 0$. The above is an iterative inequality of the same type as \[eq:Aibound\], and we get $$\begin{aligned}
\| \grad_\theta x(i/N)-\grad_\theta x^{(N)}_i\| \leq \frac{g_2(g(\|\theta\|))}{N} \max\{\|x\|,\|x\|^2\}, \quad i=0,1,\ldots,N,
\end{aligned}$$ where $g_2$ is another non-decreasing function. This completes the proof of \[eq:xdiffgrad\] under the assumption \[eq:xbddgrad\].
We now prove \[eq:xbddgrad\]. Let $t \in [0,1]$ and $\epsilon > 0$, and note that from \[def:F\] we have $$\begin{gathered}
\sup_{r \in [t,t+\epsilon]} \|\grad_\theta x(r)\| = \sup_{r \in [t,t+\epsilon]} \bigg \| \grad_\theta \bigg ( \int_{t}^{r} f_\theta(x(s)) ds + x(t)\bigg ) \bigg\| \\
\leq g(\|\theta\|) \epsilon \sup_{r \in [t,t+\epsilon]}\bigg ( \| x(r)\|+ \|\grad_\theta x(r)\| \bigg ) + \|\grad_\theta x(t)\|.
\end{gathered}$$ Fix $\epsilon=1/(2g(\|\theta\|))$. Then $$\begin{aligned}
\sup_{r \in [t,t+\epsilon]}\|\grad_\theta x(r)\| \leq \frac 1 2\| x(r)\| + \frac {1}{2g(\|\theta\|)} \|\grad_\theta x(t)\|.
\end{aligned}$$ Iterating this inequality, using also \[eq:xbdd\], we get $$\begin{aligned}
\sup_{r \in [0,1]} \|\grad_\theta x(r)\| \leq {\widetilde}g(\|\theta\|)\|x\|,
\end{aligned}$$ for some non-decreasing function ${\widetilde}g:\mathbb R_+\to\mathbb R_+$. Similar bounds can be deduced for $\grad_\theta x^{(N)}$. This establishes \[eq:xbddgrad\] and completes the proof of the lemma.
In the proof of \[thm:penalization\] we need to control the second derivatives of the risk and thus we need to bound the second derivatives of the trajectories w.r.t. $\theta$. We have collected what is needed in the following lemma.
\[lemmagrowth\] Let $g:\R_+ \to \R_+$, $g(0) \geq 1$, be a non-decreasing function and assume that $f_\theta\in \mathcal{A}_+(g)$. Then there exists a non-decreasing positive function $\hat g:\R_+ \to \R_+$ such that $$\begin{aligned}
\sup_{t \in [0,1]} \|\grad^2_\theta x(t)\| \leq \hat g(\|\theta\|){\widetilde}P(\|x\|),
\end{aligned}$$ for some polynomial ${\widetilde}P$.
We will proceed similarly as in the proof of \[lemmadiff20\]. Starting with $t \in [0,1]$ and $\epsilon > 0$, using \[def:F+\] we have $$\begin{gathered}
\sup_{r \in [t,t+\epsilon]} \|\grad^2_\theta x(r)\| = \sup_{r \in [t,t+\epsilon]} \bigg \| \grad^2_\theta \bigg ( \int_{t}^{r} f_\theta(x(s)) ds + x(t)\bigg ) \bigg\| \\
\leq g^3(\|\theta\|) \epsilon \left ( \sup_{r \in [t,t+\epsilon]} \|\grad^2_\theta x(r)\| + {\widetilde}P(\|x\|) \right )+ \|\grad^2_\theta x(t)\|,
\end{gathered}$$ for a polynomial ${\widetilde}P$. This gives, with $\epsilon=1/(2g^3(\|\theta\|))$ and after absorption, using \[lemmadiff1,lemmadiff20\], that $$\begin{aligned}
\sup_{r \in [t,t+\epsilon]} \|\grad^2_\theta x(r)\| \leq {\widetilde}P(\|x\|) + \frac{1}{2g^3(\theta)} \|\grad^2_\theta x(t)\|.
\end{aligned}$$ Iterating this inequality we deduce that $$\begin{aligned}
\sup_{r \in [0,1]} \|\grad^2_\theta x(r)\| \leq \hat g(\|\theta\|){\widetilde}P(\|x\|)
\end{aligned}$$ for some function $\hat g:\mathbb R_+\to\mathbb R_+$. Similar bounds can be deduced for $\grad^2_\theta x^{(N)}$.
The next lemma of this section uses \[lemmadiff1,lemmadiff20\] to bound the difference of the risks for the discrete and the continuous systems, as well as the difference at the gradient level. Recall that $\hat \Risk, \hat \Risk^{(N)}$ are the un-penalized truncated risks as defined at the beginning of this section.
\[lemmadiff2\] Let $g:\R_+ \to \R_+$, $g(0) \geq 1$, be a non-decreasing function and assume that $f_\theta\in \mathcal{A}(g)$. Then $$\begin{aligned}
|\hat{\mathcal {R}}^{(N)}(\theta)|+|\hat{\mathcal {R}}(\theta)|+\|\nabla \hat{\mathcal {R}}^{(N)}(\theta)\|+\|\nabla \hat{\mathcal {R}}(\theta)\|&\leq c(g,\Lambda),\notag\\
\|\nabla^2 \hat{\mathcal {R}}^{(N)}(\theta)\|+\|\nabla^2 \hat{\mathcal {R}}(\theta)\|&\leq c(g,\Lambda),\notag\\
|\hat{\mathcal {R}}(\theta)-\hat{\mathcal {R}}^{(N)}(\theta)|+\|\nabla (\hat{\mathcal {R}}(\theta)-\hat{\mathcal {R}}^{(N)}(\theta))\|&\leq \frac 1 N c(g,\Lambda).
\end{aligned}$$
We will only prove the estimates $$\begin{aligned}
\label{term1}
\|\nabla \hat{\mathcal {R}}^{(N)}(\theta)\|+\|\nabla \hat{\mathcal {R}}(\theta)\|&\leq c(g,\Lambda),\notag\\
\|\nabla (\hat{\mathcal {R}}(\theta)-\hat{\mathcal {R}}^{(N)}(\theta))\|&\leq \frac 1 N c(g,\Lambda),
\end{aligned}$$ as all the other estimates are proved similarly. To get started we first note that $$\begin{aligned}
\label{term2}
\grad \hat{\mathcal {R}}^{(N)}(\theta)&= 2\E_{(x,y) \sim \mu} \left [ (y-x^{(N)}_1) \grad_\theta x^{(N)}_N\right ],\notag\\
\grad \hat{\mathcal {R}}(\theta)&= 2\E_{(x,y) \sim \mu} \left [ (y-x(1)) \grad_\theta x(1)\right ],
\end{aligned}$$ where now $x^{(N)}_N=x^{(N)}_N(T_\Lambda(\theta))$, $x(1)=x(1,T_\Lambda(\theta))$. As $T_\Lambda(\theta)$ is constant when $\|\theta\|\geq 2\Lambda$, and $\|\nabla_\theta T_\Lambda(\theta)\|$ is bounded, we see that it suffices to simply derive the bounds when $\|\theta\|\leq \Lambda$. In this case $T_\Lambda(\theta)=\theta$. Using \[term2\] we see that to estimate the terms in \[term1\] it suffices to bound $\|x(1)\|,\|x_N^{(N)}\|,\|\grad_\theta x^{(N)}_N\|$, $\|\grad_\theta x(1)\|$, $\| x(1)- x^{(N)}_N\|$, and $\| \grad_\theta x(1)-\grad_\theta x^{(N)}_N\|$ which are all provided by \[lemmadiff1,lemmadiff20\].
We end this section with the proof of \[lem:AgComposition\].
Proof of \[lem:AgComposition\] {#proof-of-lemagcomposition .unnumbered}
------------------------------
The proof is just a matter of verifying each part of \[eq3++a+l\]. We begin with the Lipschitz continuity in $\theta$. From the triangle inequality, and repeatedly applying \[def:F\], we get $$\begin{aligned}
\|F_\theta(x)-F_{\theta'}(x)\|
\leq& g(\|\theta_2\|) \max\{g(\|\theta_1\|),g(\|\theta'_1\|)\}\|\theta_1-\theta'_1\|\|x\| \\
&+\max\{g(\|\theta_2\|),g(\|\theta'_2\|)\}\|\theta_2-\theta'_2\| g(\|\theta'_1\|)\|x\| \\
\leq& 2\max\{g^2(\|\theta\|),g^2(\|\theta'\|)\}\|\theta-\theta'\| \|x\|.\end{aligned}$$ The Lipschitz continuity in $x$ is a simple consequence of \[def:F\] $$\begin{aligned}
\|F_\theta(x)-F_\theta(x')\|
\leq g(\|\theta_2\|)g(\|\theta_1\|)\|x-x'\|
\leq g^2(\|\theta\|)\|x-x'\|.\end{aligned}$$ For the $\theta$ gradient of $F$ the Lipschitz continuity requires a bit more work, and we note that $$\begin{aligned}
\label{eq:1}
\|\grad_{\theta}F_\theta(x)-\grad_{\theta}F_\theta(x')\|^2 =& \|\grad_{\theta_1} (f_{\theta_2} \circ g_{\theta_1}(x))-\grad_{\theta_1} (f_{\theta_2} \circ g_{\theta_1}(x'))\|^2 \notag \\
&+ \|\grad_{\theta_2} (f_{\theta_2} \circ g_{\theta_1}(x))-\grad_{\theta_2} (f_{\theta_2} \circ g_{\theta_1}(x'))\|^2.\end{aligned}$$ From \[def:F\] we see that the first term in \[eq:1\] is bounded as $$\begin{aligned}
\|\grad_{\theta_1} (f_{\theta_2} \circ &g_{\theta_1}(x))-\grad_{\theta_1} (g_{\theta_2} \circ g_{\theta_1}(x'))\| \\
=& \|(\grad_{x} f_{\theta_2} \circ g_{\theta_1}(x)) \grad_{\theta_1} g_{\theta_1}(x)-(\grad_{x} f_{\theta_2} \circ g_{\theta_1}(x')) \grad_{\theta_1} g_{\theta_1}(x')\| \\
\leq& g(\|\theta_2\|)\|g_{\theta_1}(x)-g_{\theta_1}(x')\| g(\|\theta_1\|)\|x\| \\
&+ g(\|\theta_2\|) g(\|\theta_1\|)\max\{\|x\|,\|x'\|\}\|x-x'\| \\
\leq& 2g^2(\|\theta\|) \max\{\|x\|,\|x'\|\}\|x-x'\|.\end{aligned}$$ The second term in \[eq:1\] is slightly simpler to bound using \[def:F\], and we get $$\begin{aligned}
\|\grad_{\theta_2} &(f_{\theta_2} \circ g_{\theta_1}(x))-\grad_{\theta_2} (f_{\theta_2} \circ g_{\theta_1}(x'))\| = \|\grad_{\theta_2} f_{\theta_2} \circ g_{\theta_1}(x)-\grad_{\theta_2} f_{\theta_2} \circ g_{\theta_1}(x')\| \\
\leq& g(\|\theta_2\|) \max\{\|g_{\theta_1}(x)\|,\|g_{\theta_1}(x')\|\} \|g_{\theta_1}(x)-g_{\theta_1}(x')\| \\
\leq& g^3(\|\theta\|)\max\{\|x\|,\|x'\|\} \|x-x'\|.\end{aligned}$$ The last part is to prove the Lipschitz continuity of the $x$ gradient of $F_\theta$. To do this we simply note that $$\begin{aligned}
\|\grad_x (f_{\theta_2} \circ &g_{\theta_1} (x))-\grad_x (f_{\theta_2} \circ g_{\theta_1} (x'))\| \\
=& \|\grad_x f_{\theta_2} \circ g_{\theta_1} (x) \grad_x g_{\theta_1}(x)-\grad_x f_{\theta_2} \circ g_{\theta_1} (x') \grad_x g_{\theta_1}(x)\| \\
\leq& g(\|\theta_2\|) \|g_{\theta_1} (x)-g_{\theta_1} (x')\|g(\|\theta_1\|) + g(\|\theta_2\|)g(\|\theta_1\|)\|x-x'\| \\
\leq& g^3(\|\theta\|)\|x-x'\|.\end{aligned}$$
SDEs and Fokker-Planck equations {#sec:SDE_FP}
================================
In this section we collect and develop the results concerning stochastic differential equations and Fokker-Planck equations that are needed in the proofs of \[thm:penalization,thm:modified\_risk,thm:modified\_risk\_msq\].
Existence, Uniqueness and Stability of SDEs {#subsec:sde}
-------------------------------------------
We here prove results about a type of SDEs including the ones in \[sgdsys\]. In particular, as the risk in \[thm:penalization\] does not satisfy the standard linear growth assumption, see [@GS], we need to verify that the processes in \[thm:penalization\] exist, that they are unique and that they satisfy a stability result w.r.t. parameters.
The notation in this section deviates slightly from the rest of the paper, this is by design, as the results are general in nature. We consider $$\label{eq:SDEdiff}
dx_t = a(t,x_t)dt + b(t,x_t) dW_t$$ which is interpreted as the stochastic integral equation $$\label{eq:SDEint}
x_t = x_{t_0} + \int_{t_0}^t a(t,x_t)dt + \int_{t_0}^{t} b(t,x_t) dW_t$$ and where the second integral should be interpreted in the sense of Itô. We make the following assumptions.
1. \[A1\] $a = a(t,x)$ and $b = b(t,x)$ are jointly $L^2$-measurable in $(t,x) \in [t_0,T] \times \R^{d}.$
2. \[A2\] For each compact set $\mathbf{K} \subset \R^d$ there is a Lipschitz constant $L_\mathbf{K} > 0$ such that for $x,y \in \mathbf{K}, t \in [t_0,T]$ $$\begin{aligned}
\|a(t,x) - a(t,y)\| + \|b(t,x) - b(t,y)\| \leq L_\mathbf{K} \|x-y\|.
\end{aligned}$$
3. \[A3\] There exists a constant $K > 0$ such that for all $(t,x) \in [t_0,T] \times \R^d$ $$\begin{aligned}
a(t,x)\cdot x &\leq K^2(1+\|x\|^2),\ \|b(t,x)\|^2 \leq K^2.
\end{aligned}$$
4. \[A4\] $x_{t_0}$ is independent of the Wiener process $W_t$, $t \geq t_0$, and $\E[x_{t_0}^2] < \infty$.
[@GS Theorem 3, §6] If \[A1,A2\] hold, then the solutions to the stochastic differential equations \[eq:SDEint\] on $[t_0,T]$, corresponding to the same initial value and the same Wiener process, are path-wise unique.
Assuming \[A1,A2,A3,A4\], the stochastic differential equation \[eq:SDEint\] has a path-wise unique strong solution $x_t$ on $[t_0,T]$ with $$\sup_{t_0 \leq t \leq T} \E (\|x_t\|^2) < \infty.$$
As we have been unable to find a reference to this result, we here outline the modifications of [@GS Theorem 3,§6, p.45] needed when dropping the linear growth assumption on $a$.
To begin, let us first remark that we can take $t_0 = 0$ without loss of generality. Consider the truncation operator $\mathcal{T}^\Lambda(x) = \Lambda \frac{x}{\|x\|}$ for $\|x\| > \Lambda$ and $\mathcal{T}^\Lambda(x) = x$ otherwise. Then define the truncated version of \[eq:SDEdiff\] as follows. Let $x^\Lambda_{0} = \mathcal{T}^\Lambda(x_0)$, $a^\Lambda(t,x) = a(t,\mathcal{T}^\Lambda(x))$, $b^\Lambda(t,x) = b(t,\mathcal{T}^\Lambda(x))$, and consider $$\begin{aligned}
\label{eq:SDEtrunc}
dx^\Lambda_t = a^\Lambda(t,x^\Lambda_t)dt + b^\Lambda(t,x^\Lambda_t) dW_t.
\end{aligned}$$ It is easily seen that the truncations $x^\Lambda_0, a^\Lambda, b^\Lambda$ satisfy all the requirements of [@GS Theorem 1,§6, p.40]. This gives us existence and uniqueness for the solution $x^\Lambda_t$ to \[eq:SDEtrunc\].
In [@GS Remark 3,§6, p.48], the authors state that the existence of a solution to \[eq:SDEint\] can be proven with the linear growth assumption replaced with \[A3\]. They claim that it is enough to prove that $$\begin{aligned}
\label{eq:SDE_existence}
\E\left [ \psi(x_0)\|x^\Lambda_t\|^2 \right ] \leq c,
\end{aligned}$$ where $\psi(x) = \frac{1}{1+\|x\|^2}$ and where $c$ is independent of the truncation parameter $\Lambda$.
The claim \[eq:SDE\_existence\] can be proved with Itô’s formula together with Gr[ö]{}nwall’s Lemma, see [@GS]. However, we also need to prove that $$\begin{aligned}
\label{eq:SDE_existence2}
\E\left [ \psi(x_0)\sup_{0 \leq t \leq T}\|x^\Lambda_t\|^2 \right ] \leq c. \quad \text{(independent of $\Lambda$)}
\end{aligned}$$ To do this we first apply Itô’s formula to $\|x^\Lambda_t\|^2$, $$\begin{aligned}
\|x^\Lambda_t\|^2 = \|x_0\|^2 + \int_0^t x^\Lambda_t \cdot a^\Lambda(t,x^\Lambda_t) dt + \int_0^t x^\Lambda_s \cdot b^\Lambda(s,x^\Lambda_s) dW_s.
\end{aligned}$$ Next, taking the supremum, multiplying with $\psi(x_0)$, and finally taking the expectation we get $$\begin{gathered}
\label{eq:Ito_sup}
E\left [ \psi(x_0)\sup_{0 \leq t \leq T}\|x^\Lambda_t\|^2 \right ]
\leq E\left [ \psi(x_0) \|x_0\|^2 \right ]\\
+
E\left [ \psi(x_0)\sup_{0 \leq t \leq T} \int_0^t x^\Lambda_t \cdot a^\Lambda(t,x^\Lambda_t) dt \right ] \\
+ E\left [ \psi(x_0)\sup_{0 \leq t \leq T} \int_0^t x^\Lambda_s \cdot b^\Lambda(s,x^\Lambda_s) dW_s \right ].
\end{gathered}$$ The first term on the right in \[eq:Ito\_sup\] is bounded by $1$. We will now focus on the third term in \[eq:Ito\_sup\]. Using H[ö]{}lder’s inequality together with Doob’s $L^p$ inequality, see for instance [@GS Theorem 1,§3, p.20], and \[A3\], we get $$\begin{gathered}
\label{eq:third}
\E\left [ \psi(x_0) \sup_{0 \leq t \leq T} \left \|\int_0^t x^\Lambda_s \cdot b^\Lambda(s,x^\Lambda_s) dW_s \right \|\right ] \\
\leq \E\left [ \psi(x_0)^2 \sup_{0 \leq t \leq T} \left \|\int_0^t x^\Lambda_s \cdot b^\Lambda(s,x^\Lambda_s) dW_s \right \|^2 \right ] \\
\leq C \int_0^T \E \left [ \psi(x_0) \|x^\Lambda_s\|^2 \right ] ds,
\end{gathered}$$ as $\psi^2(x) \leq \psi(x)$. The boundedness of the right-hand side in \[eq:third\] follows from \[eq:SDE\_existence\]. The second term in \[eq:Ito\_sup\] can be bounded in the same way as in \[eq:third\], due to \[A3\], which concludes the proof of \[eq:SDE\_existence2\]. Note that the constant $c$ in \[eq:SDE\_existence2\] depends only on the structural assumptions of $a$, $b$ and not on $\Lambda$. Using \[eq:SDE\_existence2\] one can now argue as in [@GS Theorem 3, p.45] to conclude that $$\begin{aligned}
\label{eq:stoppingTimeConvergence}
\lim_{\Lambda \to \infty} \P\left ( \sup_{0 \leq t \leq T} \|x_{t}^\Lambda - x_t\| > 0 \right ) \to 0
\end{aligned}$$ with a rate of convergence depending only on the structural assumptions \[A1,A2,A3,A4\]. We have now established all needed modifications and the rest of the existence proof follows as in [@GS Theorem 3, p.45].
\[thm:limitTheorem\_smpl\]
Let $x_{n,t}$, $n = 0,1,2,\ldots$ be solutions of $$\begin{aligned}
x_{n,t} = x_0 + \int_0^t a_n(s,x_{n,s}) ds + \int_0^t b_n(s,x_{n,s})dW_s,
\end{aligned}$$ where the coefficients satisfy \[A1,A4\], with $a$ and $b$ globally Lipschitz continuous (with a constant independent of $n$), as well as $$\begin{aligned}
\|a_n(t,x)\|^2 &\leq K^2(1+\|x\|^2) \\
\|b_n(t,x)\|^2 &\leq K^2(1+\|x\|^2).
\end{aligned}$$ If for each $\Lambda > 0$, and $s \in [0,T]$, $$\begin{aligned}
\lim_{n \to \infty} \sup_{\|x\| \leq \Lambda} \|a_n(s,x)-a_0(s,x)\| + \|b_n(s,x)-b_0(s,x)\| = 0,
\end{aligned}$$ then $$\begin{aligned}
\E \left [\sup_{0 \leq t \leq T} \left \| x_{n,t} - x_{0,t} \right \|^2\right ] \to 0,
\end{aligned}$$ as $n \to \infty$.
\[thm:limitTheorem\] Let $x_{n,t}$, $n = 0,1,2,\ldots$ be solutions of $$\begin{aligned}
x_{n,t} = x_0 + \int_0^t a_n(s,x_{n,s}) ds + \int_0^t b_n(s,x_{n,s})dW_s
\end{aligned}$$ where $a_n$, $b_n$ satisfy \[A1,A2,A3,A4\] with constants independent of $n$. If for each $\Lambda > 0$, and $s \in [0,T]$, $$\begin{aligned}
\lim_{n \to \infty} \sup_{\|x\| \leq \Lambda} \|a_n(s,x)-a_0(s,x)\| + \|b_n(s,x)-b_0(s,x)\| = 0,
\end{aligned}$$ then $$\begin{aligned}
\sup_{0 \leq t \leq T} \left \| x_{n,t} - x_{0,t} \right \| \to 0
\end{aligned}$$ in probability as $n \to \infty$.
We will follow [@GS Theorem 3, §7] and modify accordingly. Let $a_n^\Lambda(t,x) = a_n(t,\mathcal{T}^\Lambda(x))$, $b_n^{\Lambda}(t,x) = b_n(t,\mathcal{T}^\Lambda(x))$, $x_0^\Lambda = \mathcal{T}^\Lambda(x_0)$ and consider the solution $x_{n,t}^\Lambda$ to $$\begin{aligned}
x_{n,t}^{\Lambda} = x_0^{\Lambda} + \int_0^t a_n^{\Lambda}(s,x_{n,s}^{\Lambda}) ds + \int_0^t b_n^{\Lambda}(s,x_{n,s}^{\Lambda})dW_s.
\end{aligned}$$ Following the proof of [@GS Theorem 3, §6] we first note that $$\begin{aligned}
\label{eq:diffeqlarge}
P(\sup_{0 \leq t \leq T} \|x_{n,t}^\Lambda - x_{n,t}\| > 0) \leq P(\sup_{0 \leq t \leq T} \|x_{n,t}^\Lambda\| > \Lambda).
\end{aligned}$$ I.e. as long as the truncated process $x_{n,t}^\Lambda$ stays within a ball of radius $\Lambda$ then we have $x_{n,t}^\Lambda = x_{n,t}$ a.s.: for a detailed proof see [@GS Theorem 3, §7]. Using \[eq:diffeqlarge\] and the triangle inequality we see that the following holds for any $\epsilon > 0$ $$\begin{gathered}
\label{eq:sdediff}
\P(\sup_{0 \leq t \leq T} \|x_{n,t} - x_{0,t}\| > \epsilon) \leq \P(\sup_{0 \leq t \leq T} \|x_{n,t}^\Lambda - x_{0,t}\| > \epsilon) \\
+ \P(\sup_{0 \leq t \leq T} \|x_{n,t}^\Lambda\| > \Lambda) + \P(\sup_{0 \leq t \leq T} \|x_{0,t}^\Lambda\| > \Lambda).
\end{gathered}$$ The first term in \[eq:sdediff\] goes to zero as $n \to \infty$ because our truncated equations satisfy the requirements of \[thm:limitTheorem\_smpl\]. The two other summands converge to 0 as $\Lambda \to \infty$ uniformly in $n$ because the rate in \[eq:stoppingTimeConvergence\] only depends on the constants in \[A1,A2,A3,A4\] and the initial data, see the proof of \[eq:stoppingTimeConvergence\]. Thus, we conclude $$\begin{aligned}
\lim_{n \to \infty} \P(\sup_{0 \leq t \leq T} \|x_{n,t} - x_{0,t}\| > 0) = 0
\end{aligned}$$ which completes the proof.
Fokker-Planck equations for stochastic minimization {#sub:FP_min}
---------------------------------------------------
In general the density of a diffusion process, for us a solution to a stochastic differential equation, satisfies a partial differential equation usually called the Fokker-Planck equation or Kolmogorov’s first (forward) equation. In this section we will explore this connection for the SDEs as in \[sgdsys\] of the form $$\begin{aligned}
d \theta_t = -\grad V(\theta_t)dt + \sqrt{2} dW_t, \quad \text{in } \R^m,\end{aligned}$$ with initial datum $\theta_0 \sim p_0$ and for $V \in C^2(\R^m)$. Furthermore, we have for simplicity assumed $\Sigma=\sqrt{2}I_m$: the generalization to general constant full rank $\Sigma$ is straightforward. If $-\grad V$ and $\theta_0$ satisfy \[A1,A2,A3,A4\], then the associated density $p_t$ for $\theta_t$ satisfies the following Cauchy problem for the Fokker-Planck equation, see for instance [@GS], $$\begin{aligned}
\label{eq:appFP}
p_t = {\operatorname{div}}(\nabla p + p \nabla V), \quad p(x,0) = p_0, \quad \text{in $\R^m$}.\end{aligned}$$ The equation \[eq:appFP\] is sometimes called the **linear Fokker-Planck equation**, and for an interesting overview see for instance [@VBook2 Chapter 8]. The Cauchy problem in \[eq:appFP\] has unique ‘probability’ solutions in the sense of measures, i.e. non-negative solutions that define a probability density, see [@BRS; @S]. Using the identity $$\begin{aligned}
{\operatorname{div}}(\nabla p+p\nabla V)={\operatorname{div}}(e^{-V}\nabla(e^{V} p))\end{aligned}$$ we see that we can formally rewrite \[eq:appFP\] as $$\begin{aligned}
\label{eq:appFPtilde}
\partial_t {\widetilde}p = e^{V} {\operatorname{div}}(e^{-V} \grad {\widetilde}p) = \Delta {\widetilde}p - \langle \grad V , \grad {\widetilde}p \rangle,\end{aligned}$$ for ${\widetilde}p = e^V p$, and where $\langle \cdot , \cdot \rangle$ denotes the dot–product in $\R^m$. Introducing the operator $L (\cdot) = -\Delta (\cdot) + \langle \nabla V, \nabla (\cdot) \rangle$ we note that $L$ is self-adjoint in $L^2_\mu = L^2_\mu(\R^m)$ for functions in $W_\mu^{1,2} = W_\mu^{1,2}(\R^m)$, where $d\mu := e^{-V} dx$. $L^2_\mu$ and $W_\mu^{1,2}$ are the standard $L^2$ and Sobolev spaces but defined with respect to $\mu$. Indeed, integrating by parts we have $$\begin{aligned}
\label{eq:self_adj}
\langle Lh, g \rangle_{L^2_\mu} = \langle \nabla h, \nabla g \rangle_{L^2_\mu},\end{aligned}$$ where $\langle \cdot, \cdot \rangle_{L^2_\mu}$ denotes the $L^2$ inner product w.r.t. the measure $\mu$. Assuming that $V$ satisfies a **logarithmic-Sobolev inequality**, see \[subsec:logsob\_hyper\], it follows, see [@Gross Theorem 6], that $e^{tL}$ defines a **strong semi-group** which is **hyper-contractive**. Furthermore, the semi-group is positivity preserving. This establishes the existence and uniqueness of solutions to the problem in \[eq:appFP\] for initial data in $L^2_\mu$.
The question is now, how do the solutions to \[eq:appFP,eq:appFPtilde\] relate to each other? To answer this question, first note if ${\widetilde}p_0 = e^V p_0 \in L^2_\mu$, $p_0 \geq 0$, then for ${\widetilde}p = C_0^{-1} e^{tL} {\widetilde}p_0 \geq 0$, with $C_0 = \int {\widetilde}p_0 d\mu$, we have from \[eq:self\_adj\] that $$\begin{aligned}
\partial_t \int {\widetilde}p d\mu = \int L {\widetilde}p \cdot 1 d\mu = \langle \grad {\widetilde}p, 0 \rangle_{L^2_\mu} = 0 \implies \int {\widetilde}p d\mu = 1, \quad \forall t \geq 0.\end{aligned}$$ Thus we have established that if the initial data ${\widetilde}p_0$ is a probability density w.r.t. $\mu$ then the solution ${\widetilde}p$ is also a probability density for all $t$ w.r.t. $\mu$. This implies that $\hat p = C_1^{-1} e^{-V} {\widetilde}p$ with $C_1 = \int e^{-V} dx$ solves \[eq:appFP\] and that it is a probability density w.r.t. the Lebesgue measure. The uniqueness of ‘probability’ solutions to \[eq:appFP\], again see [@BRS; @S], finally gives us that $\hat p = p$.
Logarithmic Sobolev inequalities and hyper-contractivity {#subsec:logsob_hyper}
--------------------------------------------------------
The consequence of the previous subsection is that we can switch between solutions to \[eq:appFP\] and solutions related to the semi-group $e^{tL}$ provided that the Gibbs measure $e^{-V}dx$ satisfies a **logarithmic Sobolev inequality**. In this section we discuss when the logarithmic Sobolev inequality holds. We begin with the following lemma which is a direct consequence of the Bakry-Émery theorem, see for instance [@VBook Thm 21.2].
\[thm:conv\_logsob\] Assume that $V(x) \in C^2(\R^m)$ and $\grad^2 V \geq K I_m$ then $d\mu = e^{-V(x)} dx$ satisfies $$\begin{aligned}
\label{eq:logsob}
\int u^2 \log(u^2) d\mu - \|u\|^2_{L^2_\mu}\log(\|u\|_{L^2_\mu}) \leq \frac{2}{K} \|\grad u\|^2_{L^2_\mu}.
\end{aligned}$$
The inequality in \[thm:conv\_logsob\] is the so called logarithmic Sobolev inequality with constant $2/K$. Often the potential $V$ is not strictly convex, but satisfies some form of ‘convexity at infinity’. For such potentials the logarithmic Sobolev inequality carries over, as is made rigorous by the Holley-Stroock perturbation lemma, see for instance [@R Prop 3.1.18]. This lemma states that if we perturb the potential with a function of bounded oscillation, then the logarithmic Sobolev inequality is preserved at the expense of a larger constant.
\[lem:perturb\] Assume that the probability measure $\mu$ on $\R^m$ satisfies the logarithmic Sobolev inequality \[eq:logsob\] with constant $c$. Assume that $W: \R^m \to \R$ is a bounded and measurable function. Then the modified probability measure $$\begin{aligned}
d \nu = Z^{-1}e^{-W(x)} d\mu, \qquad Z := \int e^{-W(x)} d \mu
\end{aligned}$$ satisfies the log-Sobolev inequality with constant $c e^{\osc W}$, where $\osc W = \sup W - \inf W$.
Let us now consider hyper-contractivity of the semi-group $e^{tL}$. We define the $p$ to $q$ norm of the semi-group $e^{tL}$ as $$\begin{aligned}
\|e^{tL}\|_{p \to q} = \sup \{\|e^{tL} f\|_{L^q_\mu}, f \in L^2_\mu \cap L^p_\mu, \|f\|_{L^p_\mu} \leq 1\}.\end{aligned}$$ It then follows from a theorem of Stroock, see [@FFGKRS Section 4, Theorem 4.1, Gross] or [@Stroock], that there exists, for each $p$ and $q$, where $q \geq p$, a time $t_{p\to q}$ such that if $t \geq t_{p\to q}$ then $$\begin{aligned}
\label{eq:hyper-contract}
\|e^{t L}\|_{p \to q} \leq 1.\end{aligned}$$ The estimate in \[eq:hyper-contract\] is the hyper-contractivity of the semi-group $e^{tL}$.
Estimates for Fokker-Planck equations {#sec:fp_estimates}
=====================================
In the forthcoming section we prove our main results, and we will for simplicity only give the proof in the case $\Sigma=\sqrt{2}I_m$: the generalization to general constant full rank $\Sigma$ being straightforward. Consequently, we in this section consistently assume $\Sigma=\sqrt{2}I_m$.
To prove \[thm:modified\_risk\] we will first prove perturbation estimates for the linear Fokker-Planck equation \[eq:appFPtilde\] and then utilize this to prove convergence of the processes, $\theta^{(N)}(t)$, and the value, $\E[{\widetilde}\Risk(\theta^{(N)}(t))]$. However, we first discuss what is a priori known about solutions to \[eq:appFP,eq:appFPtilde\] in the sense of regularity and integrability.
We begin with the Fokker-Planck equations, see \[eq:appFP\], for the processes $\theta^{(N)}(t)$, $\theta(t)$, $$\begin{aligned}
\label{eq2a}
\partial_tp^{(N)}&= \mbox{div}_\theta(\nabla_\theta p^{(N)}+p^{(N)}\nabla_\theta {\widetilde}{\mathcal{R}}^{(N)}),\ p^{(N)}(\theta,0)=p_0(\theta),\notag\\
\partial_tp&= \mbox{div}_\theta(\nabla_\theta p+p\nabla_\theta {\widetilde}{\mathcal{R}}),\quad\quad\quad\quad\quad p(\theta,0)=p_0(\theta),\end{aligned}$$ for $t\in (0,T]$ and where $T$ denotes the time frame used in the stochastic gradient descent. Recall that ${\widetilde}{\mathcal{R}}^{(N)}$ and ${\widetilde}{\mathcal{R}}$ are introduced in \[eq:modified\_risk-\], and that $p^{(N)}$, $p$, are the probability densities to $\theta^{(N)}(t)$, $\theta(t)$, respectively. Recall from \[sec:SDE\_FP\] the equations in \[eq2a\] have unique ‘probability‘ solutions in the sense of measures. Furthermore, since ${\widetilde}{\mathcal{R}}^{(N)}$, ${\widetilde}{\mathcal{R}}$ are smooth, it follows from Schauder estimates, see [@Lieb], and a bootstrap argument that $p^{(N)},p$ are smooth functions of $(\theta,t)$. This implies that uniqueness holds in the point-wise sense for probability solutions. Let us now consider the integrability properties of solutions to \[eq2a\]. First note that \[lemmadiff2\] implies that the coefficients in \[eq2a\] satisfies all the regularity and integrability assumptions of [@LeBLi Proposition 1, Remark 7], i.e.$$\begin{aligned}
\notag \nabla {\widetilde}{\mathcal{R}} &\in (W^{1,1}_{\text{loc}}(\mathbb R^m)^m, \\
\max\{0,\Delta {\widetilde}{\mathcal{R}}\} &\in L^\infty(\mathbb R^m), \\
\notag (1+\|\theta \|)^{-1}\nabla {\widetilde}{\mathcal{R}} &\in (L^\infty(\mathbb R^m))^m,\end{aligned}$$ and the same holds true for ${\widetilde}\Risk^{(N)}$. Thus, there exists a unique solution $$\begin{aligned}
\label{eq4}
p\in L^\infty([0,T], L^2(\mathbb R^m)\cap L^\infty(\mathbb R^m)),\ \nabla p\in L^2([0,T], L^2(\mathbb R^m))\end{aligned}$$ to the problem for $p$ in \[eq2a\], in the sense that $$\begin{gathered}
\label{eq2+}
-\int\limits_0^T\int\limits_{\mathbb R^m}p\partial_t\phi\, d\theta dt-\int\limits_{\mathbb R^m}p_0\phi(\theta ,0)\, d\theta\\
+\int\limits_0^T\int\limits_{\mathbb R^m}p\nabla {\widetilde}{\mathcal{R}}\cdot\nabla \phi\, d\theta dt
+\int\limits_0^T\int\limits_{\mathbb R^m}\nabla p\cdot\nabla\phi\, d\theta dt=0,\end{gathered}$$ for **test-functions** $\phi\in C^\infty(\mathbb R^m\times(0,T])\cap C(\mathbb R^m\times[0,T])$ with $\mbox{supp }\phi(\cdot,t)\subset K$ for all $t\in [0,T]$ and for some compact set $K\subset\mathbb R^m$. Solutions of \[eq2+\] are called **weak solutions** and more info about weak solutions can be found in [@EvansBook]. Now, as $p$, $\grad p$ are both in $L^2$ we are allowed to use $p$ as a test-function in \[eq2+\], resulting in the Caccioppoli estimate $$\begin{aligned}
\label{eq2ha-}
\frac d{dt} \int\limits_{\mathbb R^m}p^2(\theta ,t)\, d\theta +2\int\limits_{\mathbb R^m}\|\nabla p\|^2\, d\theta =\int\limits_{\mathbb R^m}p^2\Delta {\widetilde}{\mathcal{R}}\, d\theta \leq {\rho} \int\limits_{\mathbb R^m}p^2(\theta ,t)\, d\theta ,\end{aligned}$$ if $\Delta {\widetilde}{\mathcal{R}}\leq \rho$. By Gr[ö]{}nwall’s lemma we get from the above Caccioppoli estimate $$\begin{aligned}
\label{eq2ha}
\int\limits_{\mathbb R^m}p^2(\theta ,t)\, d\theta +2\int\limits_0^T\int\limits_{\mathbb R^m}\|\nabla p(\theta ,t)\|^2\, d\theta dt\leq e^{\rho T}\int\limits_{\mathbb R^m}p_0^2\, d\theta ,\end{aligned}$$ for $0\leq t\leq T$. \[eq4,eq2+,eq2ha-,eq2ha\] also hold true with $p$, ${\widetilde}{\mathcal{R}}$, replaced by $p^{(N)}$, ${\widetilde}{\mathcal{R}}^{(N)}$. However, this type of estimate is not strong enough for our purposes.
Another approach is to switch to the semi-group $e^{tL}$ related to the linear Fokker-Planck equation \[eq:appFPtilde\] as in \[sub:FP\_min\], i.e. we consider weak solutions in spaces defined with respect to the measure $d\mu:=e^{-{\widetilde}{\mathcal{R}}}d\theta$. As can be easily seen, the definition of ${\widetilde}\Risk$ implies that ${\widetilde}\Risk$ is, due to the truncation, a bounded perturbation of a strictly convex potential. Hence, the relations stated in \[sub:FP\_min\] hold. Therefore, we have a unique $p$ for which ${\widetilde}p = e^{{\widetilde}\Risk} p$ satisfies the weak formulation of \[eq:appFPtilde\] $$\begin{aligned}
\label{eq2}
- \int_0^T \int_{\R^m} {\widetilde}p \partial_t \phi d\mu dt - \int_{\R^m} e^{{\widetilde}\Risk} p_0 \phi(\theta,0) d\mu + \int_0^T \int_{\R^m} \grad {\widetilde}p \cdot \grad \phi d\mu dt = 0.\end{aligned}$$ Here ${\widetilde}p \in L^2_\mu$ and the test-functions satisfy $\phi \in C^\infty(\mathbb R^m\times(0,T])\cap C(\mathbb R^m\times[0,T])$ and $\phi(\cdot,t) \in L^2_\mu$ for all $t \in [0,T]$. In addition, the initial data is assumed to satisfy ${\widetilde}p_0 := e^{{\widetilde}\Risk} p_0 \in L^2_\mu$. In this case we deduce the a priori estimate $$\begin{aligned}
\label{eq2hap}
&\int\limits_{\mathbb R^m}|e^{{\widetilde}{\mathcal{R}}}p(\theta,t)|^2\, d\mu +2\int\limits_0^T\int\limits_{\mathbb R^m}\|\nabla (e^{{\widetilde}{\mathcal{R}}}p(\theta,t))\|^2\, d\mu dt\leq \int\limits_{\mathbb R^m}|e^{{\widetilde}{\mathcal{R}}} p_0|^2\, d\mu,\end{aligned}$$ for $0\leq t\leq T$. Again, \[eq2,eq2hap\] also hold true with $p$, ${\widetilde}{\mathcal{R}}$, replaced by $p^{(N)}$, ${\widetilde}{\mathcal{R}}^{(N)}$. In particular, for our problems at hand we get stronger a priori estimates compared to \[eq2ha\] by using the latter approach instead of the one in [@LeBLi]. The purpose of the section is to prove the following lemma.
\[lemma1\] Let $T>0$, $t\in [0,T]$ and consider ${\widetilde}{\mathcal{R}}^{(N)}$, $p^{(N)}$, ${\widetilde}{\mathcal{R}}$, $p$. Assume $p_0\in L^2(\mathbb R^m)\cap L^\infty(\mathbb R^m)$ and let $\alpha\in\mathbb R$, $\psi\in C^\infty(\mathbb R^m)$, $0\leq \psi\leq 1$. Then there exists a constant $c=c(m)$, $1\leq c<\infty$, such that the following holds $$\begin{aligned}
\label{set 1}
\int\limits_{\mathbb R^m}(p^{(N)})^2(\theta,t)e^{2\alpha\psi(\theta)}\, e^{{\widetilde}{\mathcal{R}}^{(N)}}d\theta&\leq e^{c\alpha^2\|\nabla \psi\|_\infty^2t}\int\limits_{\mathbb R^m}p^2_0(\theta)e^{2\alpha\psi(\theta)}\, e^{{\widetilde}{\mathcal{R}}^{(N)}}d\theta,\notag\\
\int\limits_{\mathbb R^m}p^2(\theta,t)e^{2\alpha\psi(\theta)}\, e^{{\widetilde}{\mathcal{R}}}d\theta&\leq e^{c\alpha^2\|\nabla \psi\|_\infty^2t}\int\limits_{\mathbb R^m}p^2_0(\theta)e^{2\alpha\psi(\theta)}\, e^{{\widetilde}{\mathcal{R}}}d\theta.
\end{aligned}$$ Furthermore, let $E$, $F$ be compact sets of $\mathbb R^m$ such that $d(E,F)>0$, $d(E,F)$ is the distance between $E$ and $F$, and assume that $\mbox{supp } p_0\subset E$. Then $$\begin{aligned}
\label{set 2}
\int\limits_{F}(p^{(N)})^2(\theta,t)\, e^{{\widetilde}{\mathcal{R}}^{(N)}}d\theta&\leq ce^{-c(d(E,F))^2/t}\int\limits_{E}p^2_0(\theta)\, e^{{\widetilde}{\mathcal{R}}^{(N)}}d\theta,\notag\\
\int\limits_{F}p^2(\theta,t)\, e^{{\widetilde}{\mathcal{R}}}d\theta&\leq ce^{-c(d(E,F))^2/t}\int\limits_{E}p^2_0(\theta)\, e^{{\widetilde}{\mathcal{R}}}d\theta.
\end{aligned}$$
We will only supply the proof of the estimates for $p$ as the proofs in the case of $p^{(N)}$ are analogous. Recall that $p$ is smooth. Let $\alpha$ and $\psi$ be as in the statement of the lemma and let $\varphi(\theta)=e^{\alpha\psi(\theta)}$. We let ${\widetilde}p=e^{{\widetilde}{\mathcal{R}}}p$, and we introduce the measure $d\mu=e^{-{\widetilde}{\mathcal{R}}}d\theta $. Using this notation, and using the test-function $p \varphi^2 \in W^{1,2}_\mu$ in \[eq2\] together with an integration by parts, we see that $$\begin{gathered}
\label{eq5g-}
\frac 1 2\frac d{dt}\int\limits_{\mathbb R^m}{\widetilde}p^2\varphi^2\, d\mu=-\int\limits_{\mathbb R^m}\nabla {\widetilde}p\cdot\nabla ({\widetilde}p\varphi^2)\, d\mu\\
=-\int\limits_{\mathbb R^m}\|\nabla {\widetilde}p\|^2\varphi^2\, d\mu-2\alpha\int\limits_{\mathbb R^m}{\widetilde}p\varphi^2 \nabla {\widetilde}p\cdot\nabla \psi\, d\mu.\end{gathered}$$ Using \[eq5g-\] we first deduce that $$\begin{gathered}
\label{eq5gg}
\quad\quad\frac 1 2\frac d{dt}\int\limits_{\mathbb R^m}{\widetilde}p^2\varphi^2\, d\mu\leq-\int\limits_{\mathbb R^m}\|\nabla {\widetilde}p\|^2\varphi^2\, d\mu+c\alpha \|\nabla \psi\|_\infty
\int\limits_{\R^m}{\widetilde}p\varphi^2\|\nabla {\widetilde}p\|\, d\mu,\end{gathered}$$ and then by Young’s inequality, Cauchy-Schwartz and re-absorption we get $$\begin{aligned}
\label{eq5gg+}
\frac d{dt}\int\limits_{\mathbb R^m}{\widetilde}p^2\varphi^2\, d\mu&\leq c\alpha^2 \|\nabla \psi\|_\infty^2
\int\limits_{\R^m}{\widetilde}p^2\varphi^2\, d\mu.\end{aligned}$$ Let $\gamma:=c\alpha^2\|\nabla \psi\|_\infty^2$ and note that \[eq5gg+\] together with Gr[ö]{}nwall’s lemma give $$\begin{aligned}
\label{eq5+}
\int\limits_{\mathbb R^m}{\widetilde}p^2 e^{2\alpha\psi}\, d\mu&\leq e^{\gamma t}\int\limits_{\R^m}{\widetilde}p^2_0 e^{2\alpha\psi}\, d\mu,\end{aligned}$$ where $c=c(m)>0$. In particular going back to $p$, \[eq5+\] becomes, $$\begin{aligned}
\label{final}
\int\limits_{\mathbb R^m}p^2 e^{2\alpha\psi}\, e^{{\widetilde}{\mathcal{R}}}d\theta&\leq e^{c\alpha^2\|\nabla \psi\|_\infty^2t}\int\limits_{\R^m}p^2_0 e^{2\alpha\psi}\, e^{{\widetilde}{\mathcal{R}}}d\theta,
\end{aligned}$$ and this concludes the proof of \[set 1\] for $p$.
To prove the estimate in \[set 2\] for $p$, let $E$, $F$ be compact sets of $\mathbb R^m$ such that $d:=d(E,F)>0$, $d(E,F)$ is the distance between $E$ and $F$, and assume that $\mbox{supp } p_0\subset E$. Let $\psi=1$ on $F$ and $\psi=0$ on $E$, $0\leq\psi\leq 1$. One can construct $\psi$ so that $\|\nabla \psi\|_\infty\leq cd^{-1}$, $c=c(m)>0$. Using \[final\] we see that $$\begin{aligned}
\int_{F}p^2(\theta,t)\, e^{{\widetilde}{\mathcal{R}}}d\theta&\leq e^{c\alpha^2d^{-2}t-2\alpha}\int\limits_{E}p^2_0(\theta)\, e^{{\widetilde}{\mathcal{R}}}d\theta\end{aligned}$$ Letting $\alpha=\beta d^2/t$ with $\beta$ so that $(c\beta-2)=-1$ we see that $$c\alpha^2d^{-2}t-2\alpha=(c\beta-2)\beta d^2/t=-{\widetilde}c d^2/t$$ for some ${\widetilde}c={\widetilde}c(m)>0$. This concludes the proof of \[set 2\] and the proof of the lemma.
Proof of the main results: \[thm:penalization,thm:modified\_risk,thm:modified\_risk\_msq\] {#sec:proof_of_main_thm}
==========================================================================================
In this section we prove \[thm:penalization,thm:modified\_risk,thm:modified\_risk\_msq\], and we will for simplicity only give the proof in the case $\Sigma=\sqrt{2}I_m$.
Proof of \[thm:penalization\]
-----------------------------
Let us first prove $$\begin{aligned}
\label{eq:penal_first}
\sup_{t \in [0,T]} \|\theta_t - \theta_t^{(N)}\| \to 0, \quad \text{in probability as $N \to \infty$}.\end{aligned}$$ We will prove \[eq:penal\_first\] under the assumption that $f_\theta \in \mathcal{A}(g)$. The first step is to show that there exists a penalization $H(\theta)$ such that $$\begin{aligned}
\label{esti}
-\grad \Risk(\theta) \cdot \theta \leq c(1+\|\theta\|^2),\end{aligned}$$ for some constant $c$. Let $\hat {\mathcal {R}}(\theta):=\mathcal {R}(\theta)-\gamma H(\theta)$, $\hat {\mathcal {R}}^{(N)}(\theta):={\mathcal {R}}^{(N)}(\theta)-\gamma H(\theta)$, denote the un-penalized risk. As in \[term2\] we have $$\begin{aligned}
\grad \hat {\mathcal {R}}(\theta)&= 2\E_{(x,y) \sim \mu} \left [ (y-x(1)) \grad_\theta x(1)\right ].\end{aligned}$$ A bound on the gradient of the risk follows from \[lemmadiff1\], $$\begin{aligned}
\label{eq:riskGradBdd}
\|\grad \hat {\mathcal {R}}(\theta)\| \leq 2\E_{(x,y) \sim \mu} \left [ (y-x(1)) \grad_\theta x(1)\right ] \leq {\widetilde}g(g(\|\theta\|))^2 \E[\|x\|^2],\end{aligned}$$ and the same bound holds for $\grad \hat {\mathcal {R}}^{(N)}$. Let $\phi_R$ be a cutoff function supported in $B(0,2R)$ such that $\phi_R=1$ on $B(0,R)$ and $\|\grad \phi_R\| \leq \frac{c}{R}$. Let $G(\theta)$ be a convex function satisfying $\grad G(\theta) = c{\widetilde}g(g(\|\theta\|))^2 \theta$. With $\theta_R$ and $G$ as above we obtain from \[eq:riskGradBdd\] $$\begin{aligned}
-\grad \left ( \hat {\mathcal {R}}^{(N)}(\theta) + (1-\phi_R(\theta))G(\theta) \right ) \cdot \theta \leq c(1+\|\theta\|^2),\end{aligned}$$ with a constant $c(R,\mu) \geq 1$. Letting $H(\theta):=(1-\phi_R(\theta))G(\theta)$ proves \[esti\]. Collecting what we proved in \[subsec:sde\], and applying \[thm:limitTheorem,lemmadiff2\], we get \[eq:penal\_first\].
We next prove the second part of \[thm:penalization\], i.e.$$\begin{aligned}
\label{eq:penal_second}
\mathbb{E} [\mathcal{R}(\theta_T)] < \infty, \quad \mathbb{E} [\mathcal{R}^{(N)}(\theta^{(N)}_T)] < \infty.\end{aligned}$$ We will prove \[eq:penal\_second\] under the more restrictive assumption $f_\theta \in \mathcal{A}_+(g)$. We begin by showing that there is a rotationally symmetric penalization $H$ such that $\hat {\mathcal {R}} + H$ is strictly convex on $\mathbb R^m\setminus B(0,R)$ for $R\geq 1$ large. To do this we will bound the growth of the second derivatives of $\hat {\mathcal {R}}$. We have $$\begin{aligned}
\grad^2 \hat {\mathcal {R}}(\theta) &= 2\E (\grad_\theta x_1 \otimes \grad_\theta x_1 + (y-x_1)\grad^2_\theta x_1).\end{aligned}$$ Using \[lemmadiff20,lemmagrowth\] we derive the bound $$\begin{aligned}
\|\grad^2 \hat {\mathcal {R}}(\theta)\| &\leq {\widetilde}g(\|\theta\|)^2\E[\|x\|^2] + \E[(\|x\|+\|y\|){\widetilde}P(\|x\|)]\hat g(\|\theta\|).\end{aligned}$$ Hence, if we penalize using a function ${\widetilde}H(\theta)$ satisfying $$\begin{aligned}
\grad^2 {\widetilde}H(\theta) \geq C(g(\|\theta\|)^2+\hat g(\|\theta\|)+1) I_m,\end{aligned}$$ for a large enough $C(\mu) \geq 1$, then $\hat {\mathcal {R}}(\theta) + {\widetilde}H(\theta)$ is strictly convex outside $B(0,R)$. In fact, consider a cutoff function $\phi_R$ as in the first part of the proof, and define $V = (1-\phi_R)\hat {\mathcal {R}} + {\widetilde}H, W = \phi_R \hat {\mathcal {R}} - \phi_R {\widetilde}H$. Then, for a possibly slightly larger constant $C=C(R,\mu)$, we see that $V$ is strictly $K$-convex and that $W$ is bounded. Thus, $\hat {\mathcal {R}} + (1-\phi_R) {\widetilde}H$ is the sum of a strictly convex potential $V$ and a bounded perturbation $W$. Define the penalization as $H = (1-\phi_R){\widetilde}H$, then $\hat {\mathcal {R}} + H$ is a confining potential that satisfies \[A1,A2,A3,A4\]. From \[thm:conv\_logsob,lem:perturb\] we see that $d\mu=c e^{-(\hat \Risk+H)}d\theta$ satisfies a logarithmic Sobolev inequality and thus we can conclude the validity of the estimate in \[eq:hyper-contract\].
Recall from \[subsec:logsob\_hyper\] that for $V = \hat {\mathcal {R}} + H$, we have ${\widetilde}p = e^{tL} c^{-1} e^V p_0 = c^{-1} e^V p$ where $p$ is the density for the stochastic process $$\begin{aligned}
d\theta_t = - \grad V(\theta_t)dt + \sqrt{2} dW_t.\end{aligned}$$ The hyper-contractivity, i.e. \[eq:hyper-contract\] shows that ${\widetilde}p(\cdot, t) \in L^r_\mu$ for $t \geq t_r$, $r \geq 2$. This implies that $$\begin{aligned}
\int {\widetilde}p^r d\mu = \int e^{r V(\theta)} p^r(\theta) e^{-V(\theta)} d\theta = \int c^{1-r} e^{(r-1) V(\theta)} p^r(\theta) d\theta \leq 1.\end{aligned}$$ Using H[ö]{}lder’s inequality and the above we get $$\begin{aligned}
\label{eq:hyper_bound}
\notag \E[V(\theta_t)] &= \int V(\theta) p(\theta,t) d\theta
= \int c^{\frac{r-1}{r}} e^{-\frac{r-1}{r} V} V(\theta) c^{-\frac{r-1}{r}} e^{\frac{r-1}{r} V}p(\theta,t) d\theta \\
\notag &\leq
\left (\int c^{1-r} e^{(r-1) V} p^r d\theta \right )^{1/r} \left (\int c e^{-\frac{r-1}{r}\frac{r}{r-1} V} V^{\frac{r}{r-1}} d\theta \right )^{\frac{r-1}{r}} \\
&\leq \left (\int c e^{-V} V^{\frac{r}{r-1}} d\theta \right )^{\frac{r-1}{r}}.\end{aligned}$$ The first term on the second line in \[eq:hyper\_bound\] is bounded by $1$. To bound the last term in \[eq:hyper\_bound\], note that if $V > \frac{r}{r-1}$ then the expression $e^{-V}V^{\frac{r}{r-1}}$ decreases as $V$ increases. We can thus use the estimate $$\begin{aligned}
e^{-V}V^{\frac{r}{r-1}} \lesssim e^{-K\|\theta\|^2}(K\|\theta\|^2)^{\frac{r}{r-1}}\end{aligned}$$ for $V > \frac{r}{r-1}$. The term on the right-hand side in the last display is integrable and thus by the Lebesgue dominated convergence theorem we see that $\E[V(\theta_t)] < \infty$. Specifically we could use $r = 2$ and thus get \[eq:penal\_second\], completing the proof of the theorem.
If $t \to \infty$ then $p_\infty = c e^{-V}$ and from \[eq:hyper\_bound\] we get $$\begin{aligned}
\lim_{t \to \infty}\E[V(\theta_t)] \leq \lim_{r \to \infty} \left (\int c e^{-V} V^{\frac{r}{r-1}} d\theta \right )^{\frac{r-1}{r}} = \int c e^{-V} V d\theta = \E[V(\theta_\infty)].
\end{aligned}$$
Proof of \[thm:modified\_risk\]
-------------------------------
To start the proof of \[thm:modified\_risk\] we let $\delta^{(N)}(x,t):=p(\theta,t)-p^{(N)}(\theta,t)$ and we note, as $\theta^{(N)}$ and $\theta$ share the same initial density, that $\delta^{(N)}$ satisfies the initial value problem $$\begin{aligned}
\label{eq2b}
\partial_t\delta^{(N)}&= \mbox{div}(\nabla \delta^{(N)}+\delta^{(N)}\nabla {\widetilde}{\mathcal{R}})-\mbox{div}(p^{(N)}\nabla ({\widetilde}{\mathcal{R}}-{\widetilde}{\mathcal{R}}^{(N)})),\notag\\
\delta^{(N)}(\theta,0)&= 0.\end{aligned}$$ Rewriting the equation \[eq2b\] as in \[eq:appFPtilde\], we get $$\begin{aligned}
\partial_t(e^{{\widetilde}{\mathcal{R}}}\delta^{(N)})e^{-{\widetilde}{\mathcal{R}}}&= \mbox{div}(e^{-{\widetilde}{\mathcal{R}}}\nabla(e^{{\widetilde}{\mathcal{R}}}\delta^{(N)}))-\mbox{div}(e^{-{\widetilde}{\mathcal{R}}}e^{{\widetilde}{\mathcal{R}}}p^{(N)}\nabla ({\widetilde}{\mathcal{R}}-{\widetilde}{\mathcal{R}}^{(N)})).\end{aligned}$$ Now, we would like to test the above equation with $e^{{\widetilde}\Risk} \delta^{(N)}$ but we do not a priori know that this test-function is in $L^2_\mu$, for $d\mu = e^{-{\widetilde}\Risk} d\theta$. Instead, we use the test-function $\phi^2(\theta)e^{{\widetilde}{\mathcal{R}}}\delta^{(N)}$, where $\phi\in C_0^\infty(\mathbb R^m)$, $0\leq \phi\leq 1$. From the above display using integration by parts we get $$\begin{gathered}
\frac{1}{2}\int\limits_{\mathbb R^m}|e^{{\widetilde}{\mathcal{R}}}\delta^{(N)}|^2(\theta,T)\phi\, d\mu+ \int\limits_0^T\int\limits_{\mathbb R^m}\|\nabla (e^{{\widetilde}{\mathcal{R}}}\delta^{(N)})\|^2\phi^2\, d\mu dt \\
\begin{aligned}
\leq&2\int\limits_0^T\int\limits_{\mathbb R^m} e^{{\widetilde}{\mathcal{R}}} p^{(N)}
\|\nabla ({\widetilde}{\mathcal{R}}-{\widetilde}{\mathcal{R}}^{(N)})\| |e^{{\widetilde}{\mathcal{R}}} \delta^{(N)} | \|\nabla \phi\|\phi\, d\mu dt\\
&+2\int\limits_0^T\int\limits_{\mathbb R^m}\|\nabla(e^{{\widetilde}{\mathcal{R}}}\delta^{(N)})\| \|\nabla\phi\| |e^{{\widetilde}{\mathcal{R}}} \delta^{(N)}|\phi\, d\mu dt\\
&+\int\limits_0^T\int\limits_{\mathbb R^m}e^{{\widetilde}{\mathcal{R}}} p^{(N)}\|\nabla ({\widetilde}{\mathcal{R}}-{\widetilde}{\mathcal{R}}^{(N)})\|\|\nabla(e^{{\widetilde}{\mathcal{R}}}\delta^{(N)})\|
\phi^2\, d\mu dt,
\end{aligned}\end{gathered}$$ where $d\mu = e^{-{\widetilde}\Risk} d\theta$. Using Youngs inequality and Cauchy-Schwarz to reabsorb terms on the left-hand side we conclude that $$\begin{aligned}
\label{eq:tildeRdeltaN}
\notag\int\limits_{\mathbb R^m}|e^{{\widetilde}{\mathcal{R}}}\delta^{(N)}|^2(\theta,T)\phi^2\, d\mu\leq&c\int\limits_0^T\int\limits_{\mathbb R^m}(e^{{\widetilde}{\mathcal{R}}} p^{(N)})^2
\|\nabla ({\widetilde}{\mathcal{R}}-{\widetilde}{\mathcal{R}}^{(N)})\|^2\phi^2\, d\mu dt\notag\\
&+c\int\limits_0^T\int\limits_{\mathbb R^m}|e^{{\widetilde}{\mathcal{R}}} \delta^{(N)}|^2\|\nabla\phi\|^2\, d\mu dt.\end{aligned}$$ Furthermore, $$\begin{gathered}
\label{eq:tildeRdeltaN2}
\int\limits_0^T\int\limits_{\mathbb R^m}|e^{{\widetilde}{\mathcal{R}}} \delta^{(N)}|^2\|\nabla\phi\|^2\, d\mu dt\leq 2\int\limits_0^T\int\limits_{\mathbb R^m}p^2\|\nabla\phi\|^2e^{{\widetilde}{\mathcal{R}}}\, d\theta dt\\
+2\int\limits_0^T\int\limits_{\mathbb R^m}(p^{(N)})^2\|\nabla\phi\|^2e^{{\widetilde}{\mathcal{R}}-{\widetilde}{\mathcal{R}}^{(N)}}e^{{\widetilde}{\mathcal{R}}^{(N)}}\, d\theta dt.\end{gathered}$$ To get rid of the localization that $\phi$ provides we let $\phi=\phi_R\in C_0^\infty(B(0,2R))$ be such that $\phi_R=1$ on $B(0,R)$ and $\|\grad \phi\| < \frac{C}{R}$. We will consider the limit as $R \to \infty$ in \[eq:tildeRdeltaN,eq:tildeRdeltaN2\]. Specifically, using the estimate for $p$ stated in \[set 2\], \[lemma1\], we see that $$\begin{gathered}
\label{eq:plimit}
\int\limits_0^T \int\limits_{\R^m} p^2 \|\grad \phi\|^2 e^{{\widetilde}\Risk} d\theta dt \leq
\frac{c}{R^2} \int\limits_0^T \int\limits_{B(0,2R) \setminus B(0,R)} p^2 e^{{\widetilde}\Risk} d\theta dt \\
\leq \frac{c}{R^2} T e^{-cR^2/T} \int_{\supp(p_0)} p_0^2 e^{{\widetilde}\Risk} d\theta \to 0\end{gathered}$$ as $R\to \infty$ since $d(\supp p_0, B(0,2R) \setminus B(0,R)) \approx R$ for large values of $R$. Similarly, $$\begin{gathered}
\label{eq:pnbound}
\int\limits_0^T\int\limits_{\mathbb R^m}(p^{(N)})^2\|\nabla\phi\|^2 e^{{\widetilde}{\mathcal{R}}-{\widetilde}{\mathcal{R}}^{(N)}}e^{{\widetilde}{\mathcal{R}}^{(N)}}\, d\theta dt \\
\leq
\frac{c}{R^2} T e^{-cR^2/T} \max_{R\leq\|\theta\|\leq 2R}e^{{\widetilde}{\mathcal{R}}(\theta)-{\widetilde}{\mathcal{R}}^{(N)}(\theta)} \int_{\supp(p_0)} p_0^2 e^{{\widetilde}\Risk} d\theta.\end{gathered}$$ Now, recall the definitions of ${\widetilde}{\mathcal{R}}^{(N)}$ and ${\widetilde}{\mathcal{R}}$ introduced in \[eq:modified\_risk-\]. Using \[lemmadiff2\] we see that $$\begin{aligned}
\label{eq:modified_risk-c}
|{\widetilde}{\mathcal{R}}(\theta)-{\widetilde}{\mathcal{R}}^{(N)}(\theta)|\leq&\frac {c(g,\Lambda)}N.\end{aligned}$$ Thus applying \[eq:modified\_risk-c\] in \[eq:pnbound\] we get $$\begin{aligned}
\label{eq:pnlimit}
\int\limits_0^T\int\limits_{\mathbb R^m}(p^{(N)})^2\|\nabla\phi\|^2e^{{\widetilde}{\mathcal{R}}-{\widetilde}{\mathcal{R}}^{(N)}}e^{{\widetilde}{\mathcal{R}}^{(N)}}\, d\theta dt\to 0\end{aligned}$$ as $R\to\infty$. In particular collecting \[eq:tildeRdeltaN,eq:tildeRdeltaN2,eq:plimit,eq:pnlimit\], and letting $R \to \infty$ for $\phi=\phi_R$ we can conclude that $$\begin{aligned}
\label{eq:finaltest}
\int\limits_{\mathbb R^m}|\delta^{(N)}|^2(\theta,T)e^{{\widetilde}{\mathcal{R}}}\, d\theta\leq&c\int\limits_0^T\int\limits_{\mathbb R^m}(p^{(N)})^2
\|\nabla ({\widetilde}{\mathcal{R}}-{\widetilde}{\mathcal{R}}^{(N)})\|^2e^{{\widetilde}{\mathcal{R}}}\, d\theta dt.\end{aligned}$$ Now, to bound the right-hand side of \[eq:finaltest\] we use \[lemmadiff2\] again to see that $$\begin{aligned}
\label{eq:modified_risk-d}
\|\nabla ({\widetilde}{\mathcal{R}}-{\widetilde}{\mathcal{R}}^{(N)})\|\leq&\frac {c(g,\Lambda)}N.\end{aligned}$$ Hence, using \[eq:modified\_risk-c\] and \[eq:modified\_risk-d\], we arrive at $$\begin{gathered}
\label{eq5g}
\int\limits_0^T\int\limits_{\mathbb R^m}(p^{(N)})^2
\|\nabla ({\widetilde}{\mathcal{R}}-{\widetilde}{\mathcal{R}}^{(N)})\|^2e^{{\widetilde}{\mathcal{R}}}\, d\theta dt\\
\leq \frac {c(g,\Lambda)}{N^2} e^{c(g,\Lambda)/N}
\int\limits_0^T\int\limits_{\mathbb R^m}(p^{(N)})^2e^{{\widetilde}{\mathcal{R}}^{(N)}}\, d\theta dt.\end{gathered}$$ We can apply the estimate for $p^{(N)}$ stated in \[set 1\], see \[lemma1\], and we deduce that $$\begin{aligned}
\label{eq5g+}
\int\limits_0^T\int\limits_{\mathbb R^m}( p^{(N)})^2e^{{\widetilde}{\mathcal{R}}^{(N)}}\, d\theta dt\leq& c(T)\int\limits_{B(0,R_0)}p_0^2e^{{\widetilde}{\mathcal{R}}^{(N)}}\, d\theta.\end{aligned}$$ Now collecting \[eq:finaltest,eq5g,eq5g+\] we get $$\begin{gathered}
\int\limits_0^T\int\limits_{\mathbb R^m}(p^{(N)})^2
\|\nabla ({\widetilde}{\mathcal{R}}-{\widetilde}{\mathcal{R}}^{(N)})\|^2e^{{\widetilde}{\mathcal{R}}}\, d\theta dt\\
\leq \frac {c(g,\Lambda, T)}{N^2} e^{c(g,\Lambda,R_0)/N}\int\limits_{B(0,R_0)}p_0^2\, d\theta,\end{gathered}$$ and finally $$\begin{aligned}
\label{eq5g+++}
\int\limits_{\mathbb R^m}|\delta^{(N)}|^2(\theta,T)e^{{\widetilde}{\mathcal{R}}}\, d\theta\leq&\frac {c(g,\Lambda, T)}{N^2} e^{c(g,\Lambda,R_0)/N}\int\limits_{B(0,R_0)}p_0^2\, d\theta.\end{aligned}$$ Using Cauchy-Schwartz and \[eq5g+++\], we get $$\begin{gathered}
\label{eq5ga}
\int\limits_{\mathbb R^m}e^{{\widetilde}{\mathcal{R}}/4}| \delta^{(N)}|(\theta,T)\, d\theta \\
\leq \left [\int_{\mathbb R^m}e^{{\widetilde}{\mathcal{R}}}| \delta^{(N)}|^2(\theta,T)\, d\theta \right ] ^{\frac{1}{2}}\left [\int_{\mathbb R^m}e^{-{\widetilde}{\mathcal{R}}/2}\, d\theta \right ]^{\frac{1}{2}} \\
\leq c(m,\gamma,g,\Lambda,R_0, T) N^{-1}\|p_0\|_2.\end{gathered}$$ We will now use the estimate in \[eq5ga\] to conclude the proof of \[thm:modified\_risk\]. Indeed, by \[eq5ga\] $$\begin{aligned}
\label{eq:final_first}
\bigl \|\E[\theta(T)-\theta^{(N)}(T)]\bigr \|&=\left \|\int_{\mathbb R^m}\theta (p(\theta,T)- p^{(N)}(\theta,T))\, d\theta\right \|\notag\\
&\leq \int\limits_{\mathbb R^m}\|\theta\| |\delta^{(N)}(\theta)|\, d\theta\notag\\
&=\int\limits_{\mathbb R^m}\|\theta\| e^{- {\widetilde}{\mathcal{R}}/4} |\delta^{(N)}(\theta)|e^{{\widetilde}{\mathcal{R}}/4}\, d\theta\notag\\
&\leq \sup_{\|\theta\|} \|\theta\| e^{- {\widetilde}{\mathcal{R}}/4} c(m,\gamma,g,\Lambda, R_0, T) N^{-1}\|p_0\|_2.\end{aligned}$$ Now by the definition of ${\widetilde}\Risk$ as the truncated risk + a quadratic penalization we see that ${\widetilde}\Risk \approx \|\theta\|^2$ for large $\|\theta\|$ and thus $\sup_{\|\theta\|} \|\theta\| e^{- {\widetilde}{\mathcal{R}}/4} \leq c(g,\Lambda,\lambda,\rho_0)$. This together with \[eq:final\_first\] concludes the proof of \[thm2.4a\] in \[thm:modified\_risk\]. To prove the other estimate, \[thm2.4b\], we note that $$\begin{gathered}
\label{eq:valuebound}
\bigl |\E [{\widetilde}{\mathcal{R}}(\theta(T))-{\widetilde}{\mathcal{R}}^{(N)}(\theta^{(N)}(T)) ]\bigr |\leq \int\limits_{\mathbb R^m}{\widetilde}{\mathcal{R}}(\theta)|p(\theta,T)-p^{(N)}(\theta,T)|\, d\theta\\
+\int\limits_{\mathbb R^m}p^{(N)}(\theta,T)|{\widetilde}{\mathcal{R}}(\theta)-{\widetilde}{\mathcal{R}}^{(N)}(\theta)|\, d\theta.\end{gathered}$$ Again using \[eq5ga\] we see that $$\begin{aligned}
\label{esta}
\int\limits_{\mathbb R^m}{\widetilde}{\mathcal{R}}(\theta)|p(\theta,T)-p^{(N)}(\theta,T)|\, d\theta\leq c(m,\gamma,g,\Lambda, \lambda,\rho_0, R_0, T) N^{-1}\|p_0\|_2.\end{aligned}$$ Furthermore, using \[eq:modified\_risk-c\], we see that $$\begin{aligned}
\label{eq:pRRN}
\int\limits_{\mathbb R^m}p^{(N)}(\theta,T)|{\widetilde}{\mathcal{R}}(\theta)-{\widetilde}{\mathcal{R}}^{(N)}(\theta)|\, d\theta\leq \frac{c}{N}.\end{aligned}$$ Collecting \[eq:valuebound,esta,eq:pRRN\] completes the proof of \[thm2.4b\] in \[thm:modified\_risk\]. To prove the second set of estimates we simply have to repeat the argument and use the estimates stated in \[set 2\], \[lemma1\].
Proof of \[thm:modified\_risk\_msq\]
------------------------------------
The result follows immediately from the fact that based on the truncation procedure $T_\Lambda$, and \[lemmadiff1,lemmadiff2\], we can conclude that the processes $\theta$ and $\theta^{(N)}$ both satisfy the assumptions of \[thm:limitTheorem\_smpl\], and this proves our theorem.
Numerical experiments {#sec:numerical}
=====================
To investigate how the optimized networks depend on the number of residual connections, i.e. layers, we have conducted a number of numerical simulations. The idea is simply to optimize \[resnet3\] for different values of $N$. To do so we consider a few different topologies and problems. We have conducted our experiments on the following datasets:
1. Cifar10 [@KH]
2. Starlike Annuli + Starlike Annuli with augmentation.
Cifar10
-------
Cifar10 is a well-known dataset consisting of images of 10 classes, \[airplanes, cars, birds, cats, deer, dogs, frogs, horses, ships, and trucks\]. The task is to take a 32x32 image as input and as output the class of the image. A lot of research papers have proposed different topologies for Cifar10 and currently the accuracy of the best performing model is $96.53\%$ [@Graham]. However, the model closest to our experiments is the Residual Network (ResNet) [@He] based on which they obtained an accuracy of $\approx 93.4\%$. For comparison, the average human performance is $93.91\pm 1.52 \%$ [@H]. From our perspective, the interesting point about this dataset is that it is the simplest standard dataset where convolutional networks vastly outperform other models, like for instance fully connected neural networks. For this dataset we have designed a simple network that performs fairly well compared to state of the art mentioned above.
### **Network topology** {#network-topology .unnumbered}
To continue let us describe the network topology used in our experiments, see \[figCNN\] for a schematic. Since networks of the type in \[resnet3\] need to have the same input-dimension and output-dimension, channels, we first apply a convolutional layer with $3 \times 3$ kernels and $256$ output channels with a stride of $2$. This is then the input to the iterated block, of type \[resnet3\], that has $256$ input channels and $256$ output channels. The block is applied $N$ times. After this we apply a batch normalization layer, [@IoffeSzeg], as well as a $3 \times 3$ kernel layer with $64$ channels with a stride of $4$ and ReLU activation. This is then followed by another batch normalization layer, a 4x4 average pooling layer and finally a softmax layer. The network in the main iterated block has the type \[nnspec\]. Specifically, it consists of a $3 \times 3$ kernel layer with $256$ output channels and ReLU activation. This feeds into another $3 \times 3$ kernel layer with again $256$ output channels, with no activation. Although this network topology is quite simple the layers are fairly wide, and we end up with a total number of about 1M parameters.
### **Experimental setup** {#experimental-setup .unnumbered}
The experimental setup is as follows. The Cifar10 dataset consists of 50k images in the training set and 10k images in the test set. In order to get a better estimate on the generalization performance, we performed 10-fold cross validation on the training set, for $N = [1,2,3,4,5,9,15,20,30,100]$. We use a weight decay of $10^{-6}$, a momentum of $0.9$, and we use the weight initialization from [@Glorot]. The models are trained with a batch size of $128$ on a single GPU (RTX 2080ti) for $N < 10$ and a batch size $512$ split over $4$ GPUs (RTX 2080ti) for $N \geq 10$. We start with a learning rate of $0.1$, divide it by $10$ at $80$, $120$, $160$ and by $2$ at $180$ epochs, where an epoch consists of 45k/batchsize iterations. Training is terminated at $200$ epochs. We use the simple data augmentation depicted in [@Lee] when training: i.e. $3$ pixels are padded on each side, and a $32\times32$ crop is randomly sampled from the padded image or its horizontal flip. For testing, we only evaluate the single view of the original $32\times32$ image. All images are normalized by the pixel mean.
### **Results** {#results .unnumbered}
The results are depicted in \[figCNNResult\], and we note that our best model gives us $90.5\%$ which is good for such a simple topology. It is interesting to note that the cross validation accuracy is increasing with layer count, and we saw a similar increase in the training accuracy. As the parameter count stays the same with increasing depth this means that for this problem deeper is better. Also, note that the accuracy flattens out fairly quickly. This can be interpreted as to indicate that there is limited value in using too many layers for this type of image classification problems. It may be worthwhile to implement a limit on the number of layers in the methods developed in [@CRBD]. Another observation is that the speed of convergence for the validation accuracy is close to $1/N$, albeit slightly faster. We believe that the convergence rate for the risk (or related metric) will in practice almost always be faster than our theoretical bound of $1/N$, see \[thm:modified\_risk\].
![CNN Topology for Cifar10[]{data-label="figCNN"}](NODE_CIFAR10.png){width="50.00000%"}
![Standard Box-plot of 10 fold cross validation result on Cifar10 with different number of residual connections using the topology in \[figCNN\].[]{data-label="figCNNResult"}](CIFAR10_ACC.pdf){width="\textwidth"}
Starlike Annuli dataset
-----------------------
An interesting observation that was made in [@DDW], is that certain simple functions cannot be represented as a neural ODE of the type defined in \[neuralODE–\]. In [@DDW] the authors give the following example. Let $0 < r_1 < r_2 < r_3$ and let $g:\R^d \to \R$ be such that $$\begin{aligned}
\label{eq:neuralODEBad}
\begin{cases}
g(x) = -1, & \text{if $\|x\|\leq r_1$} \\
g(x) = 1, & \text{if $r_2 \leq \|x\|\leq r_3$}.
\end{cases}\end{aligned}$$ To attempt to represent $g$ using a neural ODE $h_t$ as in \[neuralODE—\] we compose the output of the neural ODE $h_1(x)$ with a linear map $\mathcal{L}:\R^d \to \R$. In [@DDW] they prove that $h_1(x)$ is a homeomorphism. In order for $\mathcal{L}(h_1(x)) = g(x)$ to hold, the neural ODE $h_1(x)$, which is a homeomporhism, have to map the ball $B_{r_1}$ to one side of a hyper-plane and the outer annuli to the other side, which is impossible. For the numerical simulations in [@DDW] the authors consider $g$ as defining a binary classification problem. The data-points are sampled uniformly at random from the two classes defined by $g$ with a 50/50 class balance. When solving this binary classification problem using the neural ODE model, the authors in [@DDW] noted that the corresponding ODEs are numerically hard to solve and require many costly function evaluations. They propose a solution to this problem by simply embedding the dataset in a bigger space, for instance $\R^d \to \R^{d+1}$. This solves the problem of representability, in the sense above, and the corresponding classification problem becomes easy to solve using Neural ODEs. The process of embedding the dataset in a bigger space is called augmentation.
### **Dataset** {#dataset .unnumbered}
Our purpose is to compare the convergence rate of the risk for non-representable problems vs representable problems, in the sense above. As such, a comparison between \[eq:neuralODEBad\] and its augmentation makes sense. However, the augmented version of \[eq:neuralODEBad\] is very easy as it is essentially a conical solution, which is easy to represent with a neural network using ReLU activation functions. We therefore consider a related dataset that is harder than \[eq:neuralODEBad\] even in the augmented case, we call this dataset the Starlike Annuli. To describe our construction, see \[figStar\], we let $r_1 < r_2 < r_3$ and define, using polar coordinates, $$\label{eq:starlikeAnnuli}
g(r,\theta) =
\begin{cases}
-1, & \text{ if } r \leq r_1 (2+\cos(5\theta)) \\
1, & \text{ if } r_2 (2+\cos(5\theta)) \leq r \leq r_3 (2+\cos(5\theta)). \\
\end{cases}$$ Here the difference $r_3-r_1$ denotes the class separation. The data is generated by sampling uniformly at random from each of the two class regions. The augmented version is the dataset embedded in 3 dimensions. For our experiments we chose $r_1 = 1$, $r_2 = 1.5$, and $r_3=3$.
### **Network topology** {#network-topology-1 .unnumbered}
The topology of our network is \[nnspec\] and iterated as in \[resnet3\], i.e. $$\begin{aligned}
f_\theta(x) = K^{(1)} \sigma (K^{(2)} x + b^{(2)}) + b^{(1)}\end{aligned}$$ where $K^{(2)} \in \R^{m} \times \R^{2+n}$, $b^{(2)} \in \R^m$, $K^{(1)} \in \R^{2+n} \times \R^m$, $b^{(1)} \in \R^{2+n}$ and $\sigma$ is the ReLU activation function. Following [@DDW], $n \geq 0$ is the number of augmented dimensions, $m > 0$ is the size of the inner layer. In our experiments we chose $m=16$ and $n=0,1$. At the end of the network we have a softmax layer. We are using a quadratic penalization, as in the theoretical result \[thm:modified\_risk\], with constant $\gamma = 0.001, \lambda=1$.
### **Experimental setup** {#experimental-setup-1 .unnumbered}
The experimental setup is as follows. The Starlike Annuli dataset consists of 50k points drawn uniformly at random from \[eq:starlikeAnnuli\] with a 50/50 class balance. As in the Cifar10 experiment, we estimate generalization by performing 10-fold cross validation on the training set itself, for $N = [1,2,5,10,20,100,200]$. We use a weight decay of $0$ and a momentum of $0.9$, and we use the weight initialization in [@Glorot]. These models are trained with a batch size of $128$ on a single GPU (RTX 2080ti). We start with a learning rate of $0.1$, divide it by $10$ at $80$, $120$, $160$ and by $2$ at $180$ epochs where an epoch consists of 45k/batchsize. Training is terminated at $200$ epochs. We ran one experiment with zero augmented dimensions and one experiment with one augmented dimension.
### **Results** {#results-1 .unnumbered}
The results are depicted in \[figAnnulus,figAnnulusAug\]. We chose to display the Cross Entropy instead of accuracy for visual clarity. In the non-augmented case, results displayed in \[figAnnulus\], it is interesting to note a few things. First of all the Cross-Entropy (Risk) is increasing with the layer count, indicating that deeper is worse for this problem. This is reasonable to expect as \[eq:starlikeAnnuli\] is of the same type as \[eq:neuralODEBad\] for the limit problem (NODE). Thus, it is non-representable, in the sense stated previously, within the NODE function class, [@DDW]. Secondly the convergence rate of the risk is indeed very close to the theoretical bound $1/N$.
On the other hand, the one augmented dimension problem, see \[figAnnulusAug\], tells a different story. For $N \geq 2$ we see no noticeable difference in performance, and we seem to gain nothing from the added depth. We interpret this as to indicate that the convergence rate is very fast and that we essentially reach the NODE performance using only a small value of $N$.
Conclusion
----------
It is interesting to observe that for some problems the performance either increase or stay the same as depth increase, see \[figCNNResult,figAnnulusAug\], while for others the performance decrease significantly, see \[figAnnulus\].
What is the reason for this? Looking at this from the context of regularity, we note that the deeper the network is, the closer we are to the NODE, \[thm:modified\_risk\]. This forces the trajectories to become smoother and smoother, and we see that depth is a regularizer for the learning problem. On one hand, the idea of regularization is supported by the observations above, especially \[figAnnulus\]. On the other hand, for the experiments regarding Cifar10 we saw a similar increase in training accuracy for increasing depth, as we saw in the test accuracy, \[figCNNResult\]. This contradicts the idea of regularization. Putting these observations together we see that regularization alone cannot explain our numerical simulations. We believe that there is likely a complicated interplay between the effect that the depth has on the ‘loss landscape’, [@LiLoss], and it’s regularizing effects.
![Annuli dataset[]{data-label="figStar"}](Starlike_ring.pdf){width="70.00000%"}
![10 fold cross validation result on the starlike annulus dataset[]{data-label="figAnnulus"}](NODE_ANNULUS_NO_AUG.pdf){width="\textwidth"}
![10 fold cross validation result on the annulus dataset with 1 augmented dimension[]{data-label="figAnnulusAug"}](NODE_ANNULUS_AUG_1.pdf){width="\textwidth"}
Concluding remarks {#sec:conclusion}
==================
In this paper we have established several convergence results, \[thm:penalization,thm:modified\_risk,thm:modified\_risk\_msq\], which together give a theoretical foundation for considering Neural ODEs (NODEs) as the deep limit of ResNets. The rate of convergence proved is essentially sharp as it is the order of the discretization error for the ODE. Our results can be seen as a complement to the result in [@TVG], where the authors consider a related but different problem. However, in [@TVG] the authors only consider the convergence of minimizers and not the convergence of the actual optimization procedure. They conjecture that minimizers also converge with the rate $1/N$, which is what we prove for the stochastic gradient descent. It should also be noted that we in this paper assume that the gradient of the NODE can be computed exactly. This is in contrast to [@CRBD] where they used an adjoint method to compute an approximate gradient. On the other hand, [@CRBD] contains no theoretical justifications. The error produced when applying the adjoint method to compute the gradient is not known and it is hard to tell how it would affect the convergence considered in \[thm:modified\_risk\].
\[sec:SDE\_FP\] is of independent interest, and while the results of the section may be ‘folklore’ we have provided this section as it rigorously establishes the existence and uniqueness of stochastic gradient descent as well as the connection to the corresponding Fokker-Planck and linear Fokker-Planck equation. Using what is established in \[sec:SDE\_FP\] we can also make the observation that, instead of considering the rate of convergence to minima of $V$ for $$\begin{aligned}
d \theta_t = -\grad V(\theta_t) dt + \sqrt{2} dW_t,\end{aligned}$$ we can consider the convergence of the density $p_t$ of $\theta_t$ $$\begin{aligned}
p_t \to p_\infty(\theta) = c e^{-V}, \quad t \to \infty.\end{aligned}$$ Using \[eq:appFPtilde\] we see that the corresponding solution ${\widetilde}p = e^V p$ satisfies $$\begin{aligned}
\partial_t ({\widetilde}p_t-1) = L({\widetilde}p_t - 1).\end{aligned}$$ According to the Poincaré inequality for measures that satisfy logarithmic Sobolev inequalities, see [@VBook2], $$\begin{aligned}
\frac{1}{2} \partial_t \|{\widetilde}p_t-1 \|^2_{L_\mu^2} &= \langle \partial_t ({\widetilde}p_t-1), ({\widetilde}p_t-1) \rangle_{L_\mu^2} \\
&= -\langle L({\widetilde}p_t - 1), ({\widetilde}p_t - 1) \rangle_{L_\mu^2} = -\|\grad ({\widetilde}p_t - 1)\|^2_{L_\mu^2} \\
&\leq - c \|{\widetilde}p_t - 1\|^2_{L_\mu^2},\end{aligned}$$ using \[eq:self\_adj\]. Now, by Gr[ö]{}nwall’s lemma we see that $$\begin{aligned}
\|{\widetilde}p_t-1 \|^2_{L_\mu^2} \leq e^{-ct} \|{\widetilde}p_0-1 \|^2_{L_\mu^2}.\end{aligned}$$ The conclusion is that $p_t \to p_\infty$ exponentially fast as $t \to \infty$. So instead of minimizing the possibly non-convex function $V$ we see that stochastic gradient descent is actually minimizing a convex problem in the sense of distributions. This could explain why stochastic gradient descent performs so well in practice.
In all our proofs, and in fact even for the existence of strong solutions to the stochastic gradient descent, \[sgdsys\], we need a penalization term. In particular, this is the case for neural networks with more than one layer, in this case we cannot simply guarantee from the neural network that a condition such as \[A3\] holds. Instead, we have to rely on bounds concerning the absolute value of the drift, which grows polynomially, the order of the polynomial being the number of layers. For the Neural ODE it is clear that for the problem in \[eq:neuralODEBad\] one would need $\|\theta\| \to \infty$ in order to produce a zero risk, and hence the penalization is necessary. However, the practical implications of such a penalization is essentially non-existent. This is because we can choose a penalization that is zero on a ball of arbitrary size, for instance $B(0,1.8 \cdot 10^{308})$, which is the largest value for an IEEE **754** 64–bit double.
[**Acknowledgment**]{} The authors were partially supported by a grant from the G[ö]{}ran Gustafsson Foundation for Research in Natural Sciences and Medicine.
[^1]: Discrete convolutional operators can be phrased as matrices, see for instance [@Z].
[^2]: This was also done by the authors of ResNet in their improved follow-up paper [@He2]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have performed shell-model calculations for the even- and odd-mass $N=82$ isotones, focusing attention on low-energy states. The single-particle energies and effective two-body interaction have been both determined within the framework of the time-dependent degenerate linked-diagram perturbation theory, starting from a low-momentum interaction derived from the CD-Bonn nucleon-nucleon potential. In this way, no phenomenological input enters our effective Hamiltonian, whose reliability is evidenced by the good agreement between theory and experiment.'
author:
- 'L. Coraggio'
- 'A. Covello'
- 'A. Gargano'
- 'N. Itaco'
- 'T. T. S. Kuo'
bibliography:
- 'biblio.bib'
title: 'Shell-model study of the $N=82$ isotonic chain with a realistic effective hamiltonian'
---
Introduction
============
The $N=82$ region has long been the subject of shell-model studies owing to the strong doubly magic character of $^{132}$Sn. In this context, several calculations have been performed employing purely phenomenological two-body effective interactions [@Sarkar01]. Clearly, a main issue in the study of this region is that it provides the opportunity to investigate the effect of adding protons to a doubly-magic core over a large number of nuclei (from mass number A=133 to A=152). Comprehensive studies of this kind were conducted some twenty years ago in Refs. [@Andreozzi90] and [@Wildenthal90]. In the former a simple pairing Hamiltonian and empirical single-particle (SP) energies were used, while in the latter the two-body matrix elements and single-particle energies were determined simultaneously by a least squares fit to the ground-state and excited-state energies drawn from all experimentally studied $N=82$ nuclei.
Starting in the mid 1990s, however, shell-model calculations employing effective interactions derived from realistic nucleon-nucleon ($NN$) potentials have been performed in the $N=82$ region, which have generally yielded very good results [@Andreozzi97; @Covello97; @Holt97; @Suhonen98; @Covello02]. Some of these studies [@Covello97; @Holt97; @Suhonen98] have focused attention on the evolution of the low-energy properties of the $N=82$ isotones as a function of $A$, showing that the observed behavior is well reproduced by the theory. In Refs. [@Holt97; @Suhonen98] the shell-model results were also compared with those of a quasiparticle random-phase approximation calculation leading to the conclusion that the low-lying states are equally well described by these two approaches.
A main merit of the above mentioned realistic shell-model calculations is, of course, that no adjustable parameter appears in the matrix elements of the two-body effective interaction. In all of them, however, the single-particle energies have been taken from experiment. Actually, the calculational techniques used in these studies are based on the time-dependent degenerate linked-diagram perturbation theory [@Kuo90], which provides an effective shell-model hamiltonian, $H_{\rm eff}$, containing both one- and two-body components. Usually, however, the one-body components, which represent the theoretical SP energies, are subtracted from $H_{\rm eff}$ and replaced by those obtained from the experimental spectra of nuclei with one valence nucleon [@Shurpin83]. This procedure has indeed led to a very good description of nuclear properties in different mass regions [@Coraggio09].
Recently, we have performed fully realistic shell-model calculations employing both theoretical SP energies and two-body interactions in studies of $p-$ and $sd-$shell nuclei [@Coraggio05c; @Coraggio07b]. In these studies we have renormalized the high-momentum repulsive components of the bare $NN$ potential $V_{NN}$ by way of the so-called $V_{\rm low-k}$ approach [@Bogner01; @Bogner02], which provides a smooth potential preserving exactly the onshell properties of the original $V_{NN}$ up to a cutoff $\Lambda$. The effective hamiltonian has then been derived within the framework of the time-dependent degenerate linked-diagram perturbation theory. In both papers [@Coraggio05c; @Coraggio07b] a cutoff momentum $\Lambda=2.1$ fm$^{-1}$ was employed. This corresponds to the laboratory energy $E_{lab} \simeq 350$ MeV, which is the inelastic threshold of the $NN$ scattering. The results of these studies compare well with experiment and other shell-model calculations [@Cohen65; @Richter06].
On these grounds, we have found it interesting to carry out a similar shell-model study of the $N=82$ chain, focusing attention on the low-lying states of both even- and odd-mass isotones. In this case, the SP energies and two-body effective interaction are determined from a $V_{\rm low-k}$ derived from the high-precision CD-Bonn potential [@Machleidt01b] with a cutoff momentum $\Lambda=2.6$ fm$^{-1}$. This value of $\Lambda$, which is somewhat larger than that used in our studies of the light nuclei, is needed to obtain a reasonable description of the experimental SP spectrum of the medium-heavy nucleus $^{133}$Sb. This is mainly due to the fact that, at variance with the model spaces for light nuclei, in this case the 50-82 major shell contains the intruder $0h_{11/2}$ state, whose relative energy turns out to be strongly sensitive to the value of $\Lambda$. More precisely, for small values of $\Lambda$ this state lies far away from the other SP levels of the 50-82 shell, while the correct structure of this shell is restored when increasing the value of $\Lambda$ [@Coraggio08b].
The dependence of the results on $\Lambda$ would obviously vanish when complementing the $V_{\rm low-k}$ effective two-body interaction with different three- and higher-body components for each value of $\Lambda$. However, it is at present a very hard task, from the computational point of view, to take into account two- and three-body forces on an equal footing in a third-order perturbative calculation for medium- and heavy-mass nuclei. Therefore, the value of the cutoff $\Lambda$ may be considered as a parameter that we have chosen so as to reproduce satisfactorily the experimental SP spectrum. For the sake of completeness, it should be mentioned that the trend of our calculated SP energies suggests that a larger value of the cutoff could even improve the agreement with experiment. However, this would amount to include higher-momentum components, which would rapidly deteriorate the convergence properties of the perturbative expansion used to derive the effective shell-model hamiltonian. In the following section we show that our choice of $\Lambda$ is a reasonable compromise to assure both good perturbative properties and quality of the results.
The paper is organized as follows. In Sec. II we give a brief outline of our calculations, focusing attention on the convergence properties of the effective hamiltonian. Sec. III is devoted to the presentation and discussion of our results, while some concluding remarks are given in Sec. IV.
Outline of calculations
=======================
Our goal is to derive an effective hamiltonian for shell-model calculations in the proton $2s1d0g0h$ shell, which is the standard model space to describe the spectroscopic properties of $N=82$ isotones. Within the framework of the shell model, an auxiliary one-body potential $U$ is introduced in order to break up the hamiltonian for a system of $A$ nucleons as the sum of a one-body term $H_0$, which describes the independent motion of the nucleons, and a residual interaction $H_1$:
$$\begin{aligned}
H & = & \sum_{i=1}^{A} \frac{p_i^2}{2m} + \sum_{i<j=1}^{A} V_{ij}^{NN} = T +
V = \nonumber \\
~ & = & (T+U)+(V-U) = H_{0}+H_{1}~~.
\label{smham}\end{aligned}$$
Once $H_0$ has been introduced, a reduced model space is defined in terms of a finite subset of $H_0$’s eigenvectors. The diagonalization of the many-body hamiltonian (\[smham\]) in an infinite Hilbert space, that it is obviously unfeasible, is then reduced to the solution of an eigenvalue problem for an effective hamiltonian $H_{\rm eff}$ in a finite space.
In this paper, we derive $H_{\rm eff}$ by way of the time-dependent perturbation theory [@Kuo90]. Namely, $H_{\rm eff}$ is expressed through the Kuo-Lee-Ratcliff (KLR) folded-diagram expansion in terms of the vertex function $\hat{Q}$-box, which is composed of irreducible valence-linked diagrams [@Kuo71; @Kuo81]. We include in the $\hat{Q}$-box one- and two-body Goldstone diagrams through third order in $H_1$. The folded-diagram series is summed up to all orders using the Lee-Suzuki iteration method [@Suzuki80].
The hamiltonian $H_{\rm eff}$ contains one-body contributions, whose collection is the so-called $\hat{S}$-box [@Shurpin83]. As mentioned in the Introduction, in realistic shell-model calculations it is customary to use a subtraction procedure, so that only the two-body terms of $H_{\rm eff}$, which make up the effective interaction $V_{\rm eff}$, are retained while the SP energies are taken from experiment. In this work, we have adopted a different approach employing SP energies obtained from the $\hat{S}$-box calculation. In this regard, it is worth pointing out that, owing to the presence of the $-U$ term in $H_1$, $U$-insertion diagrams arise in the $\hat{Q}$-box. In our calculation we use the harmonic oscillator (HO) potential, $U=\frac{1}{2} m \omega ^2 r^2$, and take into account all $U$-insertion diagrams up to third order. The oscillator parameter is $\hbar \omega = 7.88$ MeV, according to the expression [@Blomqvist68] $\hbar \omega= 45 A^{-1/3} -25 A^{-2/3}$ for $A=132$.
Let us now outline the $V_{\rm low-k}$ approach [@Bogner01; @Bogner02] to the renormalization of $V_{NN}$. The repulsive core contained in $V_{NN}$ is smoothed by integrating out the high-momentum modes of $V_{NN}$ down to a certain cutoff $\Lambda$. This integration is carried out with the requirement that the deuteron binding energy and phase shifts of $V_{NN}$ are preserved by $V_{\rm
low-k}$, which is achieved by the following $T$-matrix equivalence approach. We start from the half-on-shell $T$ matrix for $V_{NN}$ $$\begin{aligned}
T(k',k,k^2) = V_{NN}(k',k) +
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \nonumber \\
\mathcal{P} \int _0 ^{\infty} q^2 dq V_{NN} (k',q)
\frac{1}{k^2-q^2} T(q,k,k^2 ) ~~, ~~~~~~~~~~~~~~~~~~~~~\end{aligned}$$
where $\mathcal{P}$ denotes the principal value and $k,~k'$, and $q$ stand for the relative momenta. The effective low-momentum $T$ matrix is then defined by $$\begin{aligned}
T_{\rm low-k } (p',p,p^2) = V_{\rm low-k }(p',p) +
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \nonumber \\
\mathcal{P} \int _0 ^{\Lambda} q^2 dq V_{\rm low-k }(p',q)
\frac{1}{p^2-q^2} T_{\rm low-k} (q,p,p^2) ~~,~~~~~~~~~~~\end{aligned}$$
where the intermediate state momentum $q$ is integrated from 0 to the momentum space cutoff $\Lambda$ and $(p',p) \leq \Lambda$. The above $T$ matrices are required to satisfy the condition $$T(p',p,p^2)= T_{\rm low-k }(p',p,p^2) \, ; ~~ (p',p) \leq \Lambda \,.$$
The above equations define the effective low-momentum interaction $V_{\rm low-k}$, and it has been shown [@Bogner02] that their solution is provided by the KLR folded-diagram expansion [@Kuo71; @Kuo90] mentioned before. In addition to the preservation of the half-on-shell $T$ matrix, which implies preservation of the phase shifts, this $V_{\rm low-k}$ preserves the deuteron binding energy, since eigenvalues are preserved by the KLR effective interaction. For any value of $\Lambda$, the $V_{\rm low-k}$ can be calculated very accurately using iteration methods. Our calculation of $V_{\rm low-k}$ is performed by employing the method proposed in [@Andreozzi96], which is based on the Lee-Suzuki similarity transformation [@Suzuki80].
![Diagonal $J^{\pi}=0^+$ TBME of $V_{\rm eff}$ as a function of $N_{\rm max}$ (see text for details).[]{data-label="Veffconv"}](N82_Fig01.epsi)
As mentioned in the Introduction, our $V_{\rm low-k}$ has been derived from the CD-Bonn $NN$ potential with a cutoff momentum $\Lambda=2.6$ fm$^{-1}$, and for protons the Coulomb force has been explicitly added to $V_{\rm low-k}$.
A discussion of the convergence properties of $H_{\rm eff}$ is now in order. For the sake of clarity, we first consider the two-body matrix elements (TBME) of $V_{\rm eff}$ and then the SP energies.
------------------------------------------------------------ ----------- --------------
Configuration 3rd order Padé $[2|1]$
$\langle (0g_{7/2})^2 | V_{Qbox} | (0g_{7/2})^2 \rangle$ -0.8624 -0.8757
$\langle (0g_{7/2})^2 | V_{Qbox} | (1d_{5/2})^2 \rangle$ -1.0022 -1.0196
$\langle (0g_{7/2})^2 | V_{Qbox} | (1d_{3/2})^2 \rangle$ -0.6948 -0.6999
$\langle (0g_{7/2})^2 | V_{Qbox} | (2s_{1/2})^2 \rangle$ -0.4642 -0.4662
$\langle (0g_{7/2})^2 | V_{Qbox} | (0h_{11/2})^2 \rangle$ 2.0532 2.058
$\langle (1d_{5/2})^2 | V_{Qbox} | (1d_{5/2})^2 \rangle$ -0.7069 -0.7138
$\langle (1d_{5/2})^2 | V_{Qbox} | (1d_{3/2})^2 \rangle$ -1.7611 -1.8191
$\langle (1d_{5/2})^2 | V_{Qbox} | (2s_{1/2})^2 \rangle$ -0.6807 -0.7555
$\langle (1d_{5/2})^2 | V_{Qbox} | (0h_{11/2})^2 \rangle$ 1.2028 1.2081
$\langle (1d_{3/2})^2 | V_{Qbox} | (1d_{3/2})^2 \rangle$ 0.0838 0.0447
$\langle (1d_{3/2})^2 | V_{Qbox} | (2s_{1/2})^2 \rangle$ -0.4817 -0.4857
$\langle (1d_{3/2})^2 | V_{Qbox} | (0h_{11/2})^2 \rangle$ 1.0457 1.0459
$\langle (2s_{1/2})^2 | V_{Qbox} | (2s_{1/2})^2 \rangle$ -0.882 -0.4686
$\langle (2s_{1/2})^2 | V_{Qbox} | (0h_{11/2})^2 \rangle$ 0.658 0.6586
$\langle (0h_{11/2})^2 | V_{Qbox} | (0h_{11/2})^2 \rangle$ -0.8025 -0.8391
------------------------------------------------------------ ----------- --------------
: $\hat{Q}$-box $J^{\pi}=0^+$ two-body matrix elements at third order in $V_{\rm low-k}$ (in MeV) compared with those obtained by calculating the Padé approximant $[2|1]$.
\[table2b\]
In Fig. \[Veffconv\], we report the diagonal $J^{\pi}=0^+$ TBME of $V_{\rm eff}$ as a function of the maximum allowed excitation energy of the intermediate states expressed in terms of the oscillator quanta $N_{\rm max}$. We have chosen the $J^{\pi}=0^+$ matrix elements because they are the largest ones in the entire set of $V_{\rm eff}$ TBME and play a key role in determining the relative spectra and ground-state (g.s.) properties of the even $N=82$ isotones. From Fig. \[Veffconv\] it is clear that our results have practically achieved convergence at $N_{\rm max}=15$.
As regards the order-by-order convergence, an estimate of the value to which the perturbative series should converge may be obtained by using Padé approximants. In Table \[table2b\], we compare all the $\hat{Q}$-box $J^{\pi}=0^+$ TBME calculated at third order with those given by the $[2|1]$ Padé approximant [@Baker70]
$$[2|1] = V_{Qbox}^0 + V_{Qbox}^1 +
\frac{V_{Qbox}^2}{1-V_{Qbox}^{3}/V_{Qbox}^2}~~,$$
$V_{Qbox}^n$ being the $n$th-order contribution to the $J^{\pi}=0^+$ TBME in the linked-diagram expansion. From Table \[table2b\] we see that our third-order results are in good agreement with those from the $[2|1]$ Padé approximant, the differences being all in the order of few tens of keV with the exception of the diagonal matrix element for the $(2s_{1/2})^2$ configuration. This indicates a weak dependence of our results on higher-order $\hat{Q}$-box perturbative terms.
![(Color online) Theoretical relative SP energies as a function of $N_{\rm max}$ (see text for details). The experimental spectrum of $^{133}$Sb is also reported.[]{data-label="Sboxconv"}](N82_Fig02.epsi)
In Fig. \[Sboxconv\] we report the relative SP energies calculated at third order in $H_1$ as a function of $N_{\rm max}$. In the same figure, the results obtained by calculating the Padé approximant $[2|1]$ are also reported (red lines). Once again, the agreement between third-order results and those from this approximant points to a good perturbative behavior of the calculated relative SP energies. We also see that an $N_{\rm max}=15$ calculation yields satisfactorily converged values for the relative SP energies.
A different situation occurs for the absolute SP energies. Actually, the calculated energy of the $1d_{5/2}$ level relative to $^{132}$Sn is -4.468, -5.225, and -5.954 MeV for $N_{\rm max}$=11, 13, and 15, respectively. This shows that there is no sign of convergence of the absolute SP energies as a function of the number of intermediate states. Moreover, the Padé approximant $[2|1]$ of the $1d_{5/2}$ energy for $N_{\rm max}=15$ is equal to -7.262, showing that also the order-by-order convergence of the absolute SP energies is unsatisfactory. This poor perturbative behavior may be traced to the chosen value of the cutoff momentum, $\Lambda=2.6$ fm$^{-1}$, which is substantially larger than the standard one, $\Lambda \simeq 2.1$ fm$^{-1}$. However, the inaccuracy of the absolute SP energies affects only the g.s. energies of the nuclei considered. In fact, the calculated spectroscopic properties (energy spectra, electromagnetic transition rates) are certainly reliable, based on the good convergence of the relative SP energies and TBME.
Results and discussion
======================
We have performed calculations, using the Oslo shell-model code [@EngelandSMC], for the even-mass $N=82$ isotones up to $^{154}$Hf, which is the last known nucleus belonging to this chain, and for the odd-mass ones up to $^{149}$Ho, the last odd-mass nucleus with a well-established experimental low-lying energy spectrum. This is the longest isotonic chain approaching the proton drip line, and therefore provides an interesting laboratory to study the evolution of nuclear structure when adding pairs of identical particles.
------------- ---------------- ----------------
$nlj$ Calc. Expt.
$0g_{7/2}$ 0.000 (-6.999) 0.000 (-9.663)
$1d_{5/2}$ 1.045 0.962
$1d_{3/2}$ 2.200 2.440
$2s_{1/2}$ 2.006
$0h_{11/2}$ 3.645 2.793
------------- ---------------- ----------------
: Calculated SP relative energies (in MeV), compared with the experimental spectrum of $^{133}$Sb [@nndc]. The values in parenthesis are the absolute SP energies with respect to the doubly closed $^{132}$Sn.
\[spetab\]
In Table \[spetab\] our calculated SP energies are reported and compared with the experimental spectrum of $^{133}$Sb. We see that the latter is on the whole reasonably well reproduced by the theory. However, while the calculated positions of the $1d_{5/2}$ and $1d_{3/2}$ levels come very close to the experimental ones, the energy of the $0h_{11/2}$ level is overestimated by about 0.85 MeV. As we shall discuss later in this section, the energy of the $2s_{1/2}$ level, for which there is no experimental information, appears to be somewhat underestimated.
![Experimental and calculated $^{134}$Te spectra.[]{data-label="134te"}](N82_Fig03.epsi)
A strong test for our effective hamiltonian is given by the calculation of the energy spectrum of $^{134}$Te, since the theory of the effective interaction is tailored for systems with two valence nucleons. From Fig. \[134te\], where the experimental [@nndc] and calculated $^{134}$Te spectra are reported up to 3.5 MeV excitation energy, we see that a very good agreement is indeed obtained.
![(Color online) Experimental and calculated ground-state energies per valence proton for $N=82$ isotones from $A=134$ to 154. $Z_{\rm val}$ is the number of valence protons.[]{data-label="gndsten"}](N82_Fig04.epsi)
In Fig. \[gndsten\], we show the calculated and experimental [@Audi03] g.s. energies (relative to the $^{132}$Sn core) per valence proton as a function of the number of valence particles $Z_{\rm val}$ of even-mass isotopes. We see that the experimental and theoretical curves are practically straight lines having the same slope, while being about 2.4 MeV apart. This discrepancy is essentially the same as that existing between the theoretical and experimental g.s. energies of $^{133}$Sb (see Table \[spetab\]). This confirms the reliability of our SP spacings and TBME, since the pattern of the theoretical curve depends only on these quantities.
![(Color online) Experimental and calculated excitation energies of the yrast $J^{\pi}=2^+$ states for $N=82$ isotones.[]{data-label="J2p"}](N82_Fig05.epsi)
![(Color online) Experimental and calculated excitation energies of the yrast $J^{\pi}=4^+$ states for $N=82$ isotones.[]{data-label="J4p"}](N82_Fig06.epsi)
![(Color online) Experimental and calculated excitation energies of the yrast $J^{\pi}=6^+$ states for $N=82$ isotones.[]{data-label="J6p"}](N82_Fig07.epsi)
In Figs. \[J2p\], \[J4p\], and \[J6p\] we report, as a function of the mass number $A$, the experimental and calculated excitation energies of the $2^+$, $4^+$, and $6^+$ yrast states, respectively. The experimental behavior is well reproduced, expecially for the $J^{\pi}=2^+$ and $4^+$ states, for which the discrepancies do not exceed 350 keV. However, it should be noted that according to our calculations the proton subshell closure, which experimentally occurs at $^{146}$Gd, is not reproduced. As mentioned before, this can be traced to the theoretical position of the $2s_{1/2}$ orbital, which appears to be too low in energy. This is confirmed by the fact [@Andreozzi97] that this level has to be placed at 2.8 MeV in order to reproduce the experimental energy of the $J^{\pi}=\frac{1}{2}^+$ at 2.150 MeV in $^{137}$Cs, which is predominantly of SP nature [@Wildenthal71]. We have verified that, if the $2s_{1/2}$ SP level is placed at 2.8 MeV, we obtain that the $2^+_1$ excitation energy in $^{146}$Gd raises from 1.716 MeV to 1.976 MeV, while in $^{148}$Dy it decreases from 1.827 to 1.758 MeV, thus reproducing the observed subshell closure.
------------ ------- ------------------
Nucleus Calc. Expt.
$^{134}$Te 5.5 $6.3 \pm 2.0$
$^{136}$Xe 9.06 $16.6 \pm 2.4$
$^{138}$Ba 11.1 $10.8 \pm 0.5$
$^{140}$Ce 14.5 $13.8 \pm 0.3$
$^{142}$Nd 16.26 $12.03 \pm 0.22$
$^{144}$Sm 16.6 $11.9 \pm 0.4$
------------ ------- ------------------
: Experimental and calculated $B(E2;2^+_1 \rightarrow 0^+_1)$. The reduced transition probabilities are expressed in W.u..
\[E2\]
To have a more complete test of the theory, we have also calculated the $B(E2;2^+_1 \rightarrow 0^+_1)$ transition rates up to $^{144}$Sm employing an effective operator obtained at third order in perturbation theory, consistently with the derivation of $H_{\rm eff}$. Our results are reported and compared with the experimental data in Table \[E2\]. We see that the agreement is quite good, providing evidence for the reliability of our calculated effective operator which takes into account microscopically core-polarization effects.
Let us now come to the odd-mass isotones. In Figs. \[J5p\_odd\], \[J3p\_odd\], \[J1p\_odd\], and \[J11m\_odd\] we report, as a function of the mass number $A$, the experimental and calculated energies of the $\frac{5}{2}^+$, $\frac{3}{2}^+$, $\frac{1}{2}^+$, and $\frac{11}{2}^-$ yrast states relative to the $\frac{7}{2}^+$ yrast state, respectively. The experimental behavior is well reproduced for the $\frac{5}{2}^+,~\frac{3}{2}^+$, and $\frac{11}{2}^-$ states. For the latter, however, most of the calculated energies are higher than the experimental ones by about 300-400 keV, which reflects the theoretical overestimation of the value of the $0h_{11/2}$ SP energy. Less satisfactory is the comparison between the calculated and experimental behavior of the $\frac{1}{2}^+$ yrast states, which is a further confirmation that the theoretical $2s_{1/2}$ level in $^{133}$Sb lies too low in energy. As a matter of fact, by performing a calculation within the seniority scheme up to $v=3$, we have found that in $^{137}$Cs and $^{139}$La the lowest $\frac{1}{2}^+$ states, which are quite well reproduced by the theory, have significant seniority $v=3$ components. From $A=141$ on, the experimental $(\frac{1}{2}^+)_1$ levels are predominantly of SP nature, as testified by the spectroscopic factors [@nndc], and the corresponding calculated states, which are dominated by $v=1$ components, underestimate their energies.
![(Color online) Experimental and calculated excitation energies of the yrast $J^{\pi}=\frac{5}{2}^+$ states relative to $J^{\pi}=(\frac{7}{2}^+)_1$ for odd-mass $N=82$ isotones.[]{data-label="J5p_odd"}](N82_Fig08.epsi)
![(Color online) Experimental and calculated excitation energies of the yrast $J^{\pi}=\frac{3}{2}^+$ states relative to $J^{\pi}=(\frac{7}{2}^+)_1$ for odd-mass $N=82$ isotones.[]{data-label="J3p_odd"}](N82_Fig09.epsi)
![(Color online) Experimental and calculated excitation energies of the yrast $J^{\pi}=\frac{1}{2}^+$ states relative to $J^{\pi}=(\frac{7}{2}^+)_1$ for odd-mass $N=82$ isotones.[]{data-label="J1p_odd"}](N82_Fig10.epsi)
![(Color online) Experimental and calculated excitation energies of the yrast $J^{\pi}=\frac{11}{2}^-$ states relative to $J^{\pi}=(\frac{7}{2}^+)_1$ for odd-mass $N=82$ isotones.[]{data-label="J11m_odd"}](N82_Fig11.epsi)
As pointed out in the Introduction, the main purpose of this study has been to test the use of a fully realistic shell-model hamiltonian for the description of medium-heavy mass nuclei. In this context, it should be noted that, aside from adopting experimental single-particle energies, earlier calculations on the N=82 isotones differ from the present ones in other respects, like the starting $NN$ potential and the renormalization procedure. For instance, while we employ the CD-Bonn potential renormalized through the $V_{\rm low-k}$ procedure, in the works of Refs. [@Holt97; @Suhonen98] use is made of the Bonn A potential and the Brueckner $G$-matrix approach. A comparison between the results of earlier works and the present one would not therefore be very meaningful. It is worth mentioning, however, that the agreement between experiment and theory achieved in this paper is on the whole comparable with that obtained in the previous calculations employing realistic effective interactions.
Concluding remarks
==================
As discussed in detail in the Introduction, the N=82 isotonic chain has long been considered a benchmark for shell-model calculations. This has resulted in a number of theoretical works which have practically all led to a good description of low-energy properties of these nuclei, evidencing the strong doubly magic character of $^{132}$Sn.
A main step towards a microscopic shell-model description of nuclear structure has been the use of two-body effective interactions derived from the free nucleon-nucleon potential. This approach, in which no adjustable parameter is involved in the calculation of the TBME, has been successfully applied to the N=82 isotones in the last decade, showing the ability of realistic effective interactions to provide an accurate description of nuclear structure properties.
However, some important theoretical questions still remain open. In particular, these concern the calculation of the single-particle energies and the role of three-body correlations. In the present paper, we have tried to investigate the former issue by constructing a realistic effective shell-model hamiltonian for the $N=82$ isotones, where both the SP energies and TBME have been obtained starting from a $V_{\rm low-k}$ derived from the CD-Bonn potential. This is a natural extension of our studies of $p-$shell nuclei [@Coraggio05c] and oxygen isotopes [@Coraggio07b]. Here, we have shown that our effective hamiltonian yields results which are in quite good agreement with experiment along the whole chain of the $N=82$ isotones. As regards the calculated single-particle energies, significant discrepancies with the experimental values occur only for the $0h_{11/2}$ and $2s_{1/2}$ levels. Note that for the latter no direct comparison is possible since the single-particle $2s_{1/2}$ state is still missing in the experimental spectrum of $^{133}$Sb. However, we have verified that increasing the calculated value by 0.8 MeV the proton subshell closure occurs at $^{146}$Gd, in agreement with the experimental findings.
At this point, the question naturally arises as to how contributions from three-body forces would modify the present results, in particular as regards the single-particle energies. While this remains a major task for future investigations, we feel that the present study paves the way toward a better understanding of the microscopic foundations of the nuclear shell model.
This work was supported in part by the U.S. DOE Grant No. DE-FG02-88ER40388.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Linda Chen
title: Poincaré Polynomials of Hyperquot Schemes
---
Introduction
============
In this paper, we find generating functions for the Poincaré polynomials of hyperquot schemes for all partial flag varieties. These generating functions give the Betti numbers of hyperquot schemes, and thus give dimension information for the cohomology ring of every hyperquot scheme.
Let ${{\mathbf F}(n;{\mathbf{s}})}$ denote the partial flag variety corresponding to flags of the form: $$V_1 \subset V_2 \subset ... \subset V_l \subset V={\mathbb{C}}^n$$ with dim $V_i=s_i$. The space ${\mathrm{Mor}}_{\mathbf{d}}({{{\mathbf P}^1}}, {{\mathbf F}(n;{\mathbf{s}})})$ of morphisms from ${{{\mathbf P}^1}}$ to ${{\mathbf F}(n;{\mathbf{s}})}$ of multidegree ${\mathbf{d}}= (d_1,...,d_l)$ can be viewed as the space of successive quotients of $V_{{{\mathbf P}^1}}$ of vector bundles of rank $r_i$ and degree $d_i$, where $r_i:=n-s_i$ and $V_{{{\mathbf P}^1}}:=V\otimes {{\mathcal{O}}}_{{{\mathbf P}^1}}$ is a trivial rank $n$ vector bundle over ${{{\mathbf P}^1}}$. Its compactification, the hyperquot scheme which we denote ${\mathcal{HQ}_{\mathbf{d}}}={{\mathcal{HQ}_{\mathbf{d}}}({{\mathbf F}(n;{\mathbf{s}})})}$, parametrizes flat families of successive quotient sheaves of $V_{{{\mathbf P}^1}}$ of rank $r_i$ and degree $d_i$. It is a generalization of Grothendieck’s Quot scheme [@g].
There has been much interest in compactifications of moduli spaces of maps, for example the stable maps of Kontsevich. Hyperquot schemes are another natural such compactification. Indeed, most of what is known so far about the quantum cohomology of Grassmanians and flag varieties has been obtained by using Quot scheme compactifications. They have been used by Bertram to study Gromov-Witten invariants and a quantum Schubert calculus [@b1] [@b2] for Grassmannians, and by Ciocan-Fontanine and Kim to study Gromov-Witten invariants and the quantum cohomology ring of flag varieties and partial flag varieties [@k] [@cf1] [@cf2] [@fgp], see also [@c1] [@c2].
The paper is organized in the following way:
In section \[quot\], we give some properties of hyperquot schemes, including a description of the Zariski tangent space to ${\mathcal{HQ}_{\mathbf{d}}}$ at a point.
In section \[torus\], we consider a torus action on the hyperquot scheme. By the theorems of Bialynicki-Birula, the fixed points of this action give a cell decomposition of ${\mathcal{HQ}_{\mathbf{d}}}$ [@bb1][@bb2]. The fixed point data is organized to give a generating function for the topological Euler characteristic of ${\mathcal{HQ}_{\mathbf{d}}}$.
In section \[betti\], we use the Zariski tangent space to ${\mathcal{HQ}_{\mathbf{d}}}$ at a point as described in section \[quot\] to compute tangent weights at the fixed points. This gives an implicit formula for the Betti numbers of ${\mathcal{HQ}_{\mathbf{d}}}$. The torus action and techniques are similar to those used by Str[ø]{}mme in the case $l=1$, where ${\mathbf F}(n:s)$ is the Grassmannian ${\mathbf G}_{s}(n)$ and ${\mathcal{HQ}_{\mathbf{d}}}$ is the ordinary Quot scheme [@s].
In section \[poincare\], we reorganize the implicit formula for the Betti numbers in a way that reduces the problem to a purely combinatorial one. In particular, we collect the information into the form of a generating function.
Let ${\mathcal{P}}({\mathbf X}) = \sum_M b_{2M}({\mathbf X})z^M$ denote the Poincaré polynomial of a space ${\mathbf X}$. It is classically known that ${\mathcal{P}}({{\mathbf F}(n;{\mathbf{s}})})$ is equal to the following generating function for the Betti numbers of the partial flag variety: $${\mathcal{P}}({{\mathbf F}(n;{\mathbf{s}})}) =\sum_M b_{2M}({\mathbf F})z^M = \frac{\prod_{i=1}^n (1-z^i)}
{\prod_{j=1}^{l+1} \prod_{i=1}^{s_j-s_{j-1}} (1-z^i) }$$ with $s_{l+1}:=n$ and $s_0:=0$.
Defining $f^{i,j}_k:=1-t_i\cdots t_jz^k$, the main result is:
\[Rpf\] $$\begin{aligned}
\lefteqn{\sum_{{d_1,\ldots,d_l}}{\mathcal{P}}({{\mathcal{HQ}_{\mathbf{d}}}({{\mathbf F}(n;{\mathbf{s}})})})t_1^{d_1}...t_l^{d_l}= }\\
& & {\mathcal{P}}({{\mathbf F}(n;{\mathbf{s}})}) \cdot
\prod_{1\leq i \leq j \leq l}
\prod_{s_{i-1}<k\leq s_i}
\left(\frac{1}{f^{i,j}_{s_j-k}}\right)
\left(\frac{1}{f^{i,j}_{s_{j+1}-k+1}}\right)
$$
In section \[special\], we discuss the special cases of the ordinary Quot scheme and of the hyperquot scheme for complete flags, and provide some specific examples.
Hyperquot Schemes {#quot}
=================
We fix some notation. Let $V={\mathbb{C}}^n$ be a complex n-dimensional vector space. Let ${\mathbf F}:= {{\mathbf F}(n;{\mathbf{s}})}$ denote the partial flag variety corresponding to flags of the form: $$V_1 \subset V_2 \subset ... \subset V_l \subset V$$ where $V_i$ is a complex subspace of dimension $s_i$. We have $s_0:=0 < s_1<...<s_l< s_{l+1}:=n$. Define $r_i:=n-s_i$. As a special case, let ${\mathbf F}(n)={\mathbf F}(n;1,2,...,n-1)$ denote the complete flag variety, with ${{\mathcal{HQ}_{\mathbf{d}}}({\mathbf F}(n))}$ the corresponding hyperquot scheme for complete flags. Also note that the Grassmannian parametrizing $r$-dimensional quotients of $V$, is also a special case, ${\mathbf G}^r(n)={\mathbf F}(n;n-r)$.
For any space $T$, let $V_T$ denote the trivial rank $n$ vector bundle on $T$, i.e. $V_T := V \otimes {{\mathcal{O}}}_T$ .
Consider a functor ${{\mathcal{F}}}_{\mathbf{d}}$ from the category of schemes to the category of sets. For a scheme $T$, ${{\mathcal{F}}}_{\mathbf{d}}(T)$ is defined to be the set of flagged quotient sheaves $$V_{{{{\mathbf P}^1}}\times T}{\twoheadrightarrow}{\mathcal{B}}_1 {\twoheadrightarrow}\cdots {\twoheadrightarrow}{\mathcal{B}}_l$$ with each ${\mathcal{B}}_i$ flat over $T$ with Hilbert polynomial $\chi({{{\mathbf P}^1}},({\mathcal{B}}_i)_t(m)) = (m+1)r_i+d_i$ on the fibers of $\pi_T: {{{\mathbf P}^1}}\times T \rightarrow T$. This last condition requires that ${\mathcal{B}}_i$ be of rank $r_i$ and relative degree $d_i$ over $T$, so that for any $t\in T$, $({\mathcal{B}}_i)_t$ is of degree $d_i$.
It is proven that the functor ${{\mathcal{F}}}_{\mathbf{d}}$ is represented by the projective scheme ${\mathcal{HQ}_{\mathbf{d}}}={{\mathcal{HQ}_{\mathbf{d}}}({{\mathbf F}(n;{\mathbf{s}})})}$ following the ideas of Grothendieck and Mumford [@cf2][@g][@m]. It has also been described in a different way by Kim [@k], as a closed subscheme of a product of Quot schemes. Kim also proves the following result:
\[Rpfproj\] ${{\mathcal{HQ}_{\mathbf{d}}}({{\mathbf F}(n;{\mathbf{s}})})}$ is an irreducible, rational, nonsingular, projective variety of dimension $$\sum_{i=1}^l d_i(s_{i+1}-s_{i-1}) + \dim({{\mathbf F}(n;{\mathbf{s}})})$$
with $s_0=0,s_{l+1}=n$ and $ \dim({{\mathbf F}(n;{\mathbf{s}})}) =\sum_{i=1}^l (s_{i+1}-s_i)s_i.$ In particular, the theorem states that $\dim {{\mathcal{HQ}_{\mathbf{d}}}({\mathbf F}(n))}= 2|d|+ \binom{n}{2}$ where $|d| = \sum_{i=1}^l d_i.$
Associated to ${{\mathcal{HQ}_{\mathbf{d}}}({{\mathbf F}(n;{\mathbf{s}})})}$ is a universal sequence of sheaves on ${{{\mathbf P}^1}}\times {\mathcal{HQ}_{\mathbf{d}}}$ of successive quotients of sheaves, each of which is flat over ${\mathcal{HQ}_{\mathbf{d}}}$. $$V_{{{{\mathbf P}^1}}\times {\mathcal{HQ}_{\mathbf{d}}}}
\twoheadrightarrow B_1 \twoheadrightarrow \cdots \twoheadrightarrow B_l.$$
Define $A_i$ as the kernel of $V_{{{{\mathbf P}^1}}\times {\mathcal{HQ}_{\mathbf{d}}}} {\rightarrow}B_i$. Each $A_i$ is flat over ${\mathcal{HQ}_{\mathbf{d}}}$, and it is an easy consequence of flatness that each $A_i$ is locally free. Thus, we have the following universal sequence on ${{{\mathbf P}^1}}\times {\mathcal{HQ}_{\mathbf{d}}}$: $$\label{univ}
A_1 {\hookrightarrow}A_2 {\hookrightarrow}\cdots {\hookrightarrow}A_l {\hookrightarrow}V_{{{{\mathbf P}^1}}\times {\mathcal{HQ}_{\mathbf{d}}}}
\twoheadrightarrow B_1 \twoheadrightarrow \cdots \twoheadrightarrow B_l.$$ with $A_i$ of rank $s_i$, $B_i$ of rank $n- s_i$. Denote the inclusion maps by $\gamma_i: A_i {\hookrightarrow}A_{i+1}$ and the surjections by $\pi_i:B_{i-1} {\rightarrow}B_i$ for each $1\leq i\leq l$. Here, we define $A_{l+1}=B_0= V _{{{{\mathbf P}^1}}\times {\mathcal{HQ}_{\mathbf{d}}}}$ and $A_0=B_{l+1}=0$. The map $\gamma_i: A_i{\hookrightarrow}A_{i+1}$ is an inclusion of sheaves, not an inclusion of bundles.
The following proposition, proved by Ciocan-Fontanine following the ideas in Kollar’s work on Hilbert schemes, determines the Zariski tangent space of ${\mathcal{HQ}_{\mathbf{d}}}$ at a point [@ko].
\[TR\] Let $x\in {\mathcal{HQ}_{\mathbf{d}}}$ correspond to successive quotients and subsheaves of $V_{{{\mathbf P}^1}}$: $${\mathcal{A}}_1 {\hookrightarrow}\cdots {\hookrightarrow}{\mathcal{A}}_l {\hookrightarrow}V_{{{{\mathbf P}^1}}\times {\mathcal{HQ}_{\mathbf{d}}}}
\twoheadrightarrow {\mathcal{B}}_1 \twoheadrightarrow \cdots \twoheadrightarrow {\mathcal{B}}_l.$$ Then we have the following exact sequence: $$0 {\rightarrow}(T_{{\mathcal{HQ}_{\mathbf{d}}}})_x {\rightarrow}\bigoplus_{i=1}^l {\mathrm{Hom}}({\mathcal{A}}_i,{\mathcal{B}}_i) \stackrel{d}{{\rightarrow}}
\bigoplus_{i=1}^{l-1} {\mathrm{Hom}}({\mathcal{A}}_i,{\mathcal{B}}_{i+1}){\rightarrow}0$$ where $(T_{{\mathcal{HQ}_{\mathbf{d}}}})_x$ is the Zariski tangent space to ${\mathcal{HQ}_{\mathbf{d}}}$ at the point $x$, and $d$ is the restriction of the difference map given by $d(\{\phi_i\})
= \{\pi_{i+1}\circ\phi_i - \phi_{i+1}\circ\gamma_i\}$.
A Torus Action {#torus}
==============
In this section, we use the torus action introduced by Str[ø]{}mme in Theorem 3.6 of [@s]. In the case of the ordinary Quot scheme, we obtain the same description of the fixed points as Str[ø]{}mme, but with slightly different notation. Our description allows us to provide a full description of the fixed point locus of the hyperquot scheme under this torus action.
Consider a maximal (n-dimensional) torus $T$ in $GL(V)$ which acts on $V$ and hence induces an action on subsheaves of $V_{{{\mathbf P}^1}}$. As a ${\mathbb{C}}T$-module, $V$ splits as a direct sum of one-dimensional subspaces, ${\oplus}_{i=1}^n W_i$. Denote ${{\mathcal{O}}}_i := W_i \otimes {{\mathcal{O}}}_{{{\mathbf P}^1}}$.
For $f_k \in H^0({{{\mathbf P}^1}},{{\mathcal{O}}}_{{{\mathbf P}^1}}(d_k))$, a form of degree $d_k$, let $ f_k {{\mathcal{O}}}_i(-d_k) {\hookrightarrow}{{\mathcal{O}}}_i$ denote the sheaf ${{\mathcal{O}}}_i(-d_k)$ defined by the section $f_k$.
A locally free subsheaf $\mathcal{S}\hookrightarrow V_{{{\mathbf P}^1}}$ of rank $s$ and degree $-d$ is fixed by the action of $T$ if and only if it is of the form $$\mathcal{S} = \bigoplus_{k=1}^{s} f_k {{\mathcal{O}}}_{c_k}(-d_k)$$ where $d_k$ are nonnegative integers such that $\sum_k d_k = d$, $1\leq c_1< \cdots <c_s\leq n$, and $f_k$ is a homogeneous form in $X$ and $Y$ of degree $d_k$.
[[*Proof.*]{} ]{}Since $T$ acts with different weights on each ${{\mathcal{O}}}_i$, $\mathcal{S}\hookrightarrow V_{{{\mathbf P}^1}}$ is a fixed point of $T$ if and only if $\mathcal{S} = \bigoplus_{i=1}^n \mathcal{S}_i$ where $\mathcal{S}_i:=\mathcal{S} \cap {{\mathcal{O}}}_i$. Since $\mbox{rank}(\mathcal{S}) = s$, $\mathcal{S}_i \neq \emptyset$ for exactly $s$ such $i$, say $\mathcal{S}_{c_k} \neq \emptyset$ for a sequence $1\leq c_1< \cdots <c_s\leq n$.
Since we have $\mathcal{S}_{c_k}{\hookrightarrow}{{\mathcal{O}}}_{c_k}$, we know that we can write $\mathcal{S}_{c_k}={{\mathcal{O}}}_{c_k}(-d_k)$ for some nonnegative integer $d_k$.
Threfore we have:
$$\mathcal{S}=\bigoplus_{k=1}^s \mathcal{S}_{c_k} =
v\bigoplus_{k=1}^s {{\mathcal{O}}}_{c_k}(-d_k),$$ where deg $\mathcal{S}_{c_k} = -d_k$ with $\sum_{k=1}^s d_k = d$. Since the inclusion of $\mathcal{S}_{c_k}$ is given by some section $f_k$, which is a polynomial of degree $d_k$ in $X$ and $Y$, the lemma is proven.
Let $T'$ be the one-dimensional torus which acts on $H^0({{\mathcal{O}}}_{{{\mathbf P}^1}}(1))$ by $X \mapsto tX$ and $Y \mapsto t^{-1}Y$. Then $T'$ acts on ${{{\mathbf P}^1}}$ and hence on subsheaves of $V_{{{\mathbf P}^1}}$. Thus under the action of the product torus $T\times T'$, the fixed points are subsheaves $\mathcal{S}{\hookrightarrow}V_{{{\mathbf P}^1}}$ fixed by both $T$ and $T'$. A point $(\oplus_{k=1}^s f_k {{\mathcal{O}}}_{c_k}(-d_k){\hookrightarrow}V_{{{\mathbf P}^1}})$ is fixed by $T'$ if and only if each $f_k$ is a monomial in $X$ and $Y$.
Therefore, we have proven:
\[gfix\] A locally free subsheaf $S{\hookrightarrow}V_{{{\mathbf P}^1}}$ of rank $s$ and degree $-d$ is fixed under the action of ${T \times T'}$ if and only if it is of the form $$\bigoplus_{k=1}^s X^{a_k} Y^{b_k} {{\mathcal{O}}}_{c_k}(-a_k-b_k).$$
Here, ${{\mathcal{O}}}_i$ denotes the $i$th component of the trivial rank $n$ vector bundle $V_{{{\mathbf P}^1}}$, and $(a,b,c)$ are sequences of $s$ nonnegative integers satisfying:
1. $\sum_{k=1}^s a_k + b_k = d$
2. $1\leq c_1<\cdots < c_s\leq n$
[[*Remark.*]{} ]{} This combinatorial data is equivalent to the fixed point data of Str[ø]{}mme. For an element $(\alpha,\beta,\delta)$ as in [@s], let $\delta_{c_1}=...\delta_{c_{n-r}} = 1$ be the nonzero elements, with $1\leq c_1 <... < c_{n-r} \leq n$. Then the sequence $(\alpha,\beta,\delta)$ corresponds to the sequence $(a,b,\mathbf{c})$.
A torus action on ${\mathcal{HQ}_{\mathbf{d}}}$ {#torusR}
-----------------------------------------------
Note that a point of ${\mathcal{HQ}_{\mathbf{d}}}$ can be given by successive subsheaves over ${{{\mathbf P}^1}}$ of $V_{{{\mathbf P}^1}}= \oplus_{i=1}^{n} {{\mathcal{O}}}_i$.
Let $T$ and $T'$ be as above. Since the actions of $T$ and $T'$ extend to actions on ${{\mathcal{HQ}_{\mathbf{d}}}({{\mathbf F}(n;{\mathbf{s}})})}$, we have an action of $T\times T'$ on ${\mathcal{HQ}_{\mathbf{d}}}$.
Using the same methods as used in section \[torus\], we find the fixed points of ${\mathcal{HQ}_{\mathbf{d}}}$ under this action.
A point of ${\mathcal{HQ}_{\mathbf{d}}}$ can be given by a sequence of subsheaves $${\mathcal{A}}_1 {\hookrightarrow}\cdots {\hookrightarrow}{\mathcal{A}}_l {\hookrightarrow}V_{{{\mathbf P}^1}}$$ where rank ${\mathcal{A}}_i = s_i$ and $\deg {\mathcal{A}}_i = -d_i$. Let ${\mathcal{A}}$ denote this sequence $\{{\mathcal{A}}_i\}_{i=1}^l$. Then ${\mathcal{A}}$ is fixed by the action of $T\times T'$ if and only if each ${\mathcal{A}}_i{\hookrightarrow}V_{{{\mathbf P}^1}}$ is fixed and the inclusions $ {\mathcal{A}}_i {\hookrightarrow}{\mathcal{A}}_{i+1}$ hold. By Proposition \[gfix\], ${\mathcal{A}}_i$ is fixed when $${\mathcal{A}}_i =\bigoplus_{j=1}^{s_i} {\mathcal{A}}_{i,j}:=
\bigoplus_{j=1}^{s_i} X^{a_{i,j}} Y^{b_{i,j}}
{{\mathcal{O}}}_{c_{i,j}}(-a_{i,j}-b_{i,j})$$ where $\sum_{1\leq j\leq s_i} a_{i,j}+b_{i,j} = d_i$ and $1\leq c_{i,s_{i-1}+1}< \cdots <c_{i,s_i} \leq n$. Here, we denote by ${{\mathcal{O}}}(-a-b)$ the line bundle on ${{{\mathbf P}^1}}$ given by global section $X^aY^b$.
The inclusion $ {\mathcal{A}}_i {\hookrightarrow}{\mathcal{A}}_{i+1}$ holds under exactly the following conditions:
1. $c_{i,j} = c_{i+1,j}$ whenever $1\leq j\leq s_i$.
2. $$\label{subsheaves}
{{\mathcal{O}}}_{c_{i,j}}(-a_{i,j}-b_{i,j}) {\hookrightarrow}{{\mathcal{O}}}_{c_{i,j}}(-a_{i+1,j}-b_{i+1,j})$$
is an inclusion of sheaves.
Let $S:=S(n;s_1,...s_l)$ be the subset of $S_n$ consisting of permutations $\sigma\in S_n$ such that if $\sigma(i)<\sigma(i+1)$ unless $i\in (s_1,...s_l)$. More explicitly, an element $\sigma\in S$ is such that $$\label{sigma}
\sigma(s_{i-1}+1)<\cdots <\sigma (s_i) \mbox{ for } 1\leq i\leq l.$$ Therefore every sequence $\{c_{i,j}\}$ corresponds to an element $\sigma\in S$ by $c_{i,j}=\sigma(j)$, and this correspondence is a bijection by the first condition above.
The second condition gives conditions on the sequences of nonnegative integers $a$ and $b$. The inclusion of sheaves ${{\mathcal{O}}}(-a_{i,j}-b_{i,j}) {\hookrightarrow}{{\mathcal{O}}}$ is given by the global section $X^{a_{i,j}}Y^{b_{i,j}}$ and ${{\mathcal{O}}}(-a_{i+1,j}-b_{i+1,j}) {\hookrightarrow}{{\mathcal{O}}}$ is given by the global section $X^{a_{i+1,j}}Y^{b_{i+1,j}}$. Therefore (\[subsheaves\]) gives an inclusion of subsheaves if and only if that inclusion is given by global sections $$X^{a_{i,j}-a_{i+1,j}}Y^{b_{i,j}-b_{i+1,j}}$$ so that the conditions $a_{i,j}-a_{i+1,j}\geq 0$ and $b_{i,j}-b_{i+1,j}\geq 0$ must hold.
Let $P$ be the set of $(a,b,\sigma)$ such that $a$ and $b$ are sequences of nonnegative integers $a_{i,j}$ and $b_{i,j}$ with $1 \leq i \leq l, 1\leq j\leq s_i$, and $\sigma\in S(n;s_1,...,s_l)$ which satisfy:
1. $a_{i,j}\geq a_{i+1,j}$
2. $b_{i,j}\geq b_{i+1,j}$
3. $\sum_{j=1}^{s_i} a_{i,j} + b_{i,j} = d_i$
For $(a,b,\sigma)\in P$, define $${\mathcal{A}}_{i,j} = \left\{ \begin{array}{cl}
{{\mathcal{O}}}_{\sigma(j)}(-a_{i,j}-b_{i,j}) &
\mbox{ for } 1\leq i\leq l\mbox{ and } 1\leq j \leq s_i\\ 0 &
\mbox{ otherwise} \end{array}\right.$$
This defines sequences of subsheaves $\mathcal{A}_{i,j} {\hookrightarrow}\mathcal{A}_{i+1,j}$. Let ${\mathcal{A}}_i=\oplus_{j=1}^{s_i} \mathcal{A}_{i,j}$. Let $r(a,b,\sigma)\in {\mathcal{HQ}_{\mathbf{d}}}$ be the associated flag of subsheaves of $V_{{{\mathbf P}^1}}$. We have proven:
$r:P{\rightarrow}{\mathcal{HQ}_{\mathbf{d}}}$ is a bijection onto ${\mathcal{HQ}_{\mathbf{d}}}^{T\times T'}$.
Let ${\mathcal{B}}_{i,j}:=V_{{{\mathbf P}^1}}/{\mathcal{A}}_{i,j}$ be the corresponding quotient so that we have the following short exact sequences: $$\label{ses}
0 {\rightarrow}{\mathcal{A}}_{i,j}{\rightarrow}V_{{{\mathbf P}^1}}{\rightarrow}{\mathcal{B}}_{i,j} {\rightarrow}0$$ Similarly, define the sheaves ${\mathcal{B}}_i =\oplus_{j=1}^{s_i} \mathcal{B}_{i,j}$.
Euler characteristic
--------------------
Under a generic choice of one-dimensional subtorus $\Gamma\subset T \times T'$, ${\mathcal{HQ}_{\mathbf{d}}}$ has isolated fixed points under the action of $\Gamma$, with ${\mathcal{HQ}_{\mathbf{d}}}^\Gamma ={\mathcal{HQ}_{\mathbf{d}}}^{T\times T'}$. We know by Theorem \[Rpfproj\] that ${{\mathcal{HQ}_{\mathbf{d}}}({{\mathbf F}(n;{\mathbf{s}})})}$ is a nonsingular complex projective variety. The odd cohomology of ${\mathcal{HQ}_{\mathbf{d}}}$ vanishes since the fixed point locus ${\mathcal{HQ}_{\mathbf{d}}}^\Gamma$ is finite [@bb1] [@bb2]. Therefore, the Euler characteristic is the number of fixed points. The fixed point data has been collected into the combinatorial data of the set $E$, so that the Euler characteristic is the cardinality of $E$. In this section, we find a generating function for the Euler characteristics of the hyperquot schemes. We prove:
\[eulerthm\] $$\sum_{d_1,...d_l} \chi({{\mathcal{HQ}_{\mathbf{d}}}({{\mathbf F}(n;{\mathbf{s}})})}) t_1^{d_1}...t_l^{d_l}= \left(\#(S)\right)
\left(\prod_{1\leq i \leq j \leq l}
\frac{1}{(1-t_i...t_j)^{s_i-s_{i-1}}}
\right)^2$$
The cardinality of the set of permutations $S=S(n;s_1,\cdots,s_l)$ is: $$\#(S) = \prod_{i=0}^l \binom{n-s_{i}}{s_{i+1}-s_{i}}$$ with $s_0=0, s_{l+1}=n$.
[[*Proof.*]{} ]{}Let ${\alpha}_{i,j}=a_{i,j}-a_{i+1,j}$ and ${\beta}_{i,j}=b_{i,j}-b_{i-1,j}$, where we let $a_{l+1,j}=b_{l+1,j}=0$. Then $a_{i,j}= \sum_{k=i}^l {\alpha}_{k,j}$ and $b_{i,j}= \sum_{k=i}^l {\beta}_{k,j}$. Note that while the variables $(a,b)$ satisfy various inequalities, the nonnegative integers of $({\alpha},{\beta})$ are independent. Consider the set $P'$ of elements $({\alpha},{\beta},\sigma)$ with $\sigma\in S$ satisfying: $$\label{rel}
\sum_{k=i}^l \sum_{j=1}^{s_i} {\alpha}_{k,j} + {\beta}_{k,j} = d_i.$$
Let $P''$ denote the set of pairs $({\alpha},{\beta})$ of natural numbers satisfying the linear relations (\[rel\]). We have constructed a bijection between $E$ and $E'$, so that $\chi({{\mathcal{HQ}_{\mathbf{d}}}({{\mathbf F}(n;{\mathbf{s}})})})
= \#(P) = \#(P') = \#(S)\#(P'')$
We see that $$\begin{aligned}
\lefteqn{
\left(\prod_{1\leq i \leq j \leq l}
\frac{1}{(1-t_i...t_j)^{s_i-s_{i-1}}}
\right)^2 } \\
&=&\prod_{1\leq i \leq j \leq l}
\prod_{s_{i-1}<k\leq s_i}
\sum_{\alpha_{k,j}\in\mathbb{N}} (t_i...t_j)^{\alpha_{k,j}}
\sum_{\beta_{k,j}\in\mathbb{N}} (t_i...t_j)^{\beta_{k,j}}\\
&=& \sum_{\alpha_{k,j}} \sum_{\beta_{k,j}}
\prod_{i=1}^l t_i^{\sum_{k=i}^l \sum_{j=1}^{s_i} {\alpha}_{k,j} + {\beta}_{k,j}}\end{aligned}$$ so that each set of natural numbers $({\alpha},{\beta})$ satisfying the relations (\[rel\]) contributes exactly one to the coefficient of $t_1^{d_1}\cdots
t_l^{d_l}$. This proves the theorem.
An implicit formula for the Betti numbers {#betti}
=========================================
Let $N>>0$ be a large integer. Let $\Gamma\subset T \times T'$ be the one-dimensional subtorus which acts on ${{\mathcal{O}}}_i$ by $t\cdot v = t^{Ni}v$ and acts on $H^0({{{\mathbf P}^1}},{{\mathcal{O}}}(1))$ by $X\mapsto tX, Y\mapsto t^{-1}Y$. It is this ${\mathbb{C}}^*$ action of $\Gamma$ on ${{\mathcal{HQ}_{\mathbf{d}}}({{\mathbf F}(n;{\mathbf{s}})})}$ which is used.
In order to find information about the contribution of the various fixed points to the Betti numbers, we must consider our ${\mathbb{C}}^*$ action of $\Gamma$ on ${\mathcal{HQ}_{\mathbf{d}}}$ more carefully.
Fix the following notation. Consider $\Gamma$ a ${\mathbb{C}}^*$-action on a scheme $X$. If $x\in X$ is a fixed point of this action, and $E$ is a bundle over $X$, denote by $E_x^+$ the ${\mathbb{C}}\Gamma$- submodule of $E(x)$ where $\Gamma$ acts with positive weights. In particular, the theorems of Bialynicki-Birula state that for isolated fixed points $\{x_i\} \in X_\Gamma$, there is a cell decomposition of $X$ given by the orbits $X_i$ of $x_i$, with $\dim_{\mathbb{C}}X_i = \dim_{\mathbb{C}}T_X(x_i)^+$ [@bb1] [@bb2].
We first compute the tangent weights at the fixed points of the ${\mathbb{C}}^*$ action of $\Gamma$ on ${\mathcal{HQ}_{\mathbf{d}}}$. For $\sigma\in S$, define $\epsilon_{i,j}^\sigma$ for $1\leq i,j \leq n$ by $\epsilon^\sigma_{i,j} = 1$ if $\sigma(i)<\sigma(j)$, $\epsilon^\sigma_{i,j} = 0$ otherwise. Note that $\epsilon_{i,j}+ \epsilon_{j,i} =1$ for $i\neq j$. Define $\varepsilon_{i(k,k']}^\sigma$ for $1\leq i,k,k' \leq l$ by ${\varepsilon}^\sigma_{i(k,k']} = \sum_{s_k<j\leq s_k'} \epsilon_{i,j}$ and similarly define ${\varepsilon}^\sigma_{(k,k']j}=\sum_{s_k<i\leq s_k'} \epsilon_{i,j}$. If the permutation $\sigma$ is understood, it may be suppressed.
Define a map $h:P{\rightarrow}{\mathbb{Z}}$ by: $$\begin{aligned}
\label{hpf}
h(a,b,\sigma) &=& \sum_{i=1}^l \sum_{k\leq s_i}
(a_{i,k}+b_{i,k}+1){\varepsilon}^{\sigma}_{k(i,i+1]} \\
\nonumber &+&
\sum_{i=1}^l \sum_{k\leq s_i} (a_{i,k}+b_{i,k}){\varepsilon}^{\sigma}_{(i-1,i]k} +
\sum_{i=1}^l \sum_{s_i<k\leq s_{i+1}} b_{i,k}\end{aligned}$$
Then we have the following implicit formula for the Betti numbers of the hyperquot scheme:
\[bpf\] For any (nonnegative) integer $M$, $$b_{2M}({{\mathcal{HQ}_{\mathbf{d}}}({{\mathbf F}(n;{\mathbf{s}})})}) =\mbox{rank } A_M({\mathcal{HQ}_{\mathbf{d}}}) = \#(h^{-1}(M)).$$
[[*Proof.*]{} ]{}If we show that $h(a,b,\mathbf{e}) = \dim_{\mathbb{C}}(T_{{\mathcal{HQ}_{\mathbf{d}}}} (r(a,b,\sigma))^+)$, then by applying the theorems of Bialynicki-Birula, we will be done. Let $\mathcal{A}$ be the sequences of subsheaves associated to $r(a,b,\sigma)\in {\mathcal{HQ}_{\mathbf{d}}}$ as in section \[torusR\]. From the short exact sequence on ${\mathcal{HQ}_{\mathbf{d}}}$ given in Lemma \[TR\] we have:
$$0 {\rightarrow}T_{{\mathcal{HQ}_{\mathbf{d}}}}({\mathcal{A}}) {\rightarrow}\bigoplus_{i=1}^{l} {\mathrm{Hom}}({\mathcal{A}}_i,{\mathcal{B}}_i) {\rightarrow}\bigoplus_{i=1}^{l-1} {\mathrm{Hom}}({\mathcal{A}}_i,{\mathcal{B}}_{i+1}){\rightarrow}0.$$
Therefore the tangent weights of $T_{{\mathcal{HQ}_{\mathbf{d}}}}({\mathcal{A}})$ are those obtained by removing the weights of the quotient term from those of the middle term. In particular, the positive weights are also be obtained this way. More precisely, we can say that $$\begin{aligned}
\lefteqn{ \dim_{\mathbb{C}}(T_{{\mathcal{HQ}_{\mathbf{d}}}}({\mathcal{A}}))^+ = }\\
& & \dim_{\mathbb{C}}\left(\bigoplus_{i=1}^{l} {\mathrm{Hom}}({\mathcal{A}}_i,{\mathcal{B}}_i)\right)^+ -
\, \dim_{\mathbb{C}}\left(\bigoplus_{i=1}^{l-1} {\mathrm{Hom}}({\mathcal{A}}_i,{\mathcal{B}}_{i+1})\right)^+.\end{aligned}$$
Define maps $h_{i,j}^k:P{\rightarrow}{\mathbb{Z}}$ for $0\leq i \leq j\leq l, k=1,2,3$ as follows: $$\begin{aligned}
\label{definehk}
& &h_{i,j}^1(a,b,\sigma) = \sum_{k\leq s_i}
(a_{i,k}+b_{i,k}+1){\varepsilon}^{\sigma}_{k,(j,l+1]}
\\ \nonumber & &h_{i,j}^2(a,b,\sigma) =
\sum_{k\leq s_j}(a_{j,k}+b_{j,k}){\varepsilon}^{\sigma}_{(0,i],k} \\
\nonumber & &h_{i,j}^3(a,b,\sigma) =
\sum_{k\leq s_i} b_{j,k}\end{aligned}$$ where we allow zero maps when appropriate.
The proposition is then an immediate consequence of the following two lemmas.
\[kk\] For any $i\leq j$, $\dim_{\mathbb{C}}({\mathrm{Hom}}({\mathcal{A}}_i,{\mathcal{B}}_j))^+ = (h_{i,j}^1+h_{i,j}^2+h_{i,j}^3)(a,b,\sigma)$.
\[sum\] $$h(a,b,\sigma) = \sum_{i=1}^l (h_{i,i}^1+h_{i,i}^2+h_{i,i}^3)(a,b,\sigma)-
\sum_{i=1}^{l-1} (h_{i,i+1}^1+h_{i,i+1}^2+h_{i,i+1}^3)(a,b,\sigma).$$
We first prove Lemma [\[sum\]]{}.
As an immediate consequence of the definitions (\[definehk\]) we have: $$\begin{aligned}
& &h_{i,i}^1-h_{i,i+1}^1 = \sum_{k\leq s_i}
(a_{i,k}+b_{i,k}+1){\varepsilon}_{k,(i,i+1]}
\\ & &h_{i,i}^2- h_{i-1,i}^2=\sum_{k\leq s_i}
(a_{i,k}+b_{i,k}){\varepsilon}_{(i-1,i],k}
\\ & &h_{i,i}^3 - h_{i-1,i}^3 = \sum_{s_{i-1}<k\leq s_i} b_{i,k}\end{aligned}$$ and since $h_{l,l+1}^1=h_{0,1}^2=h_{0,1}^3=0$, Lemma \[sum\] follows.
It only remains to prove Lemma \[kk\]. We have $$\begin{aligned}
\lefteqn{{\mathrm{Hom}}({\mathcal{A}}_i,{\mathcal{B}}_j) = \bigoplus_{k,m} {\mathrm{Hom}}({\mathcal{A}}_{i,k},{\mathcal{B}}_{j,m}) =}\\
& & \bigoplus_{k\leq s_i}
\bigoplus_{m>s_j}
{\mathrm{Hom}}({{\mathcal{O}}}_{\sigma(k)}(-a_{i,k}-b_{i,k}), {{\mathcal{O}}}_{\sigma(m)}))\\
& &\oplus
\bigoplus_{k\leq s_{i},m\leq s_j}
{\mathrm{Hom}}({{\mathcal{O}}}_{\sigma(k)}(-a_{i,k}-b_{i,k}),
{{\mathcal{O}}}_{\sigma(m)}/ {{\mathcal{O}}}_{\sigma(m)}(-a_{j,m}-b_{j,m}))\end{aligned}$$ We have three situations to consider:
1\. $m>s_j$
${\mathrm{Hom}}({{\mathcal{O}}}_{\sigma(k)}(-a_{i,k}-b_{i,k}), {{\mathcal{O}}}_{\sigma(m)})$ is of rank $a_{i,k}+b_{i,k}+1$, with weights the same sign as $(\sigma(m)-\sigma(k))$ for $N$ large enough. The number of $m>i$ with $\sigma(m)>\sigma(k)$ is $\sum_{i<m\leq n} \epsilon_{k,m}$. In the notation of (\[sigma\]), since $s_{l+1}=n$, this number is $(a_{i,k}+b_{i,k}+1){\varepsilon}_{k,(j,l+1]}$. This gives the term $h^1_{i,j}(a,b,\sigma)$ in Lemma \[kk\].
2\. $k\neq m$
${\mathrm{Hom}}({{\mathcal{O}}}_{\sigma(k)}(-a_{i,k}-b_{i,k}),
{{\mathcal{O}}}_{\sigma(m)}/ {{\mathcal{O}}}_{\sigma(m)}(-a_{j,m}-b_{j,m}))$ is of rank $a_{j,m}+b_{j,m}$, with weights the same sign as $\sigma(m)-\sigma(k)$. Thus the positive weight contribution is $\sum_{k\leq s_i} (a_{j,m}+b_{j,m})\epsilon_{k,m}=
(a_{j,m}+b_{j,m})\epsilon_{(0,i],m}$, which gives the term $h^2_{i,j}$ in Lemma \[kk\]. 3. $k=m$
${\mathrm{Hom}}({\mathcal{A}}_{i,k},{\mathcal{B}}_{j,k})$ sits inside the long exact sequence induced by the short exact sequence (\[ses\]): $$\begin{aligned}
\label{les}
0 &{\rightarrow}& {\mathrm{Hom}}({\mathcal{A}}_{i,k},{\mathcal{A}}_{j,k}){\rightarrow}{\mathrm{Hom}}({\mathcal{A}}_{i,k},{{\mathcal{O}}}_{\sigma(k)})\\
\nonumber & &\mbox{ }{\rightarrow}{\mathrm{Hom}}({\mathcal{A}}_{i,k},{\mathcal{B}}_{j,k})
{\rightarrow}H^1({\mathcal{A}}_{i,k}^*\otimes{\mathcal{A}}_{j,k}) {\rightarrow}0\end{aligned}$$ We have two cases: (a) ${\mathcal{A}}_{i,k}\hookrightarrow {\mathcal{A}}_{j,k}$. Here, equation (\[les\]) becomes $$0 {\rightarrow}{\mathrm{Hom}}({\mathcal{A}}_{i,k},{\mathcal{A}}_{j,k}) {\rightarrow}{\mathrm{Hom}}({\mathcal{A}}_{i,k},{{\mathcal{O}}}_{\sigma(k)})
{\rightarrow}{\mathrm{Hom}}({\mathcal{A}}_{i,k},{\mathcal{B}}_{i,k}) {\rightarrow}0.$$ $\Gamma$ acts on ${\mathrm{Hom}}({\mathcal{A}}_{i,k},{\mathcal{A}}_{j,k})$ by $t\cdot X^rY^s =
X^{r-(a_{i,k}-a_{j,k})}Y^{-s+(b_{i,k}-b_{j,k})}$ and on ${\mathrm{Hom}}({\mathcal{A}}_{i,k},{{\mathcal{O}}}_{\sigma(k)})$ by $t\cdot X^rY^s = X^{r-a_{i,k}}Y^{-s+b_{i,k}}$. Thus, the positive part of this piece of $T_{{\mathcal{HQ}_{\mathbf{d}}}}(r(a,b,\mathbf{\sigma}))$ has dimension $b_{j,k}$. This gives the $h^3_{i,j}$ term. (b) ${\mathcal{A}}_{i,k}\not\hookrightarrow {\mathcal{A}}_{j,k}$. From (\[les\]) we get $$0 {\rightarrow}{\mathrm{Hom}}({\mathcal{A}}_{i,k},{{\mathcal{O}}}_{\sigma(k)})
{\rightarrow}{\mathrm{Hom}}({\mathcal{A}}_{i,k},{\mathcal{B}}_{j,k}) {\rightarrow}H^1({\mathcal{A}}_{i,k}^*\otimes{\mathcal{A}}_{j,k}) {\rightarrow}0.$$ We have $H^1({\mathcal{A}}_{i,k}^*\otimes{\mathcal{A}}_{j,k})=H^1({{\mathcal{O}}}((a_{i,k}-a_{j,k})+
(b_{i,k}-b_{j,k}))).$ By Serre duality and the same arguments as in the previous case, the positive contribution of ${\mathrm{Hom}}({\mathcal{A}}_{i,k},{\mathcal{B}}_{j,k})$ is $(b_{j,k}-b_{i,k})+b_{i,k}=b_{j,k}$. This gives the term $h^3_{i,j}$.
Therefore we have shown that $\dim_{\mathbb{C}}(T_{{\mathcal{HQ}_{\mathbf{d}}}}({\mathcal{A}}))^+ = h(a,b,\sigma)$, so that the proposition is proved.
Poincaré polynomials {#poincare}
====================
We use the implicit formula for the Betti numbers proved in Proposition \[bpf\]. Rewrite (\[hpf\]) as: $$\begin{aligned}
\label{rehpf}
\lefteqn{h(a,b,\sigma) =
\sum_{i=1}^l \sum_{k\leq s_i} {\varepsilon}_{k(i,i+1]}
+ \sum_{i=1}^l \sum_{s_{i-1}<k\leq s_{i}} b_{i,k}
}\\
\nonumber & &\mbox{}+
\sum_{i=1}^l \sum_{k\leq s_i}
(a_{i,k}+b_{i,k})({\varepsilon}_{k(i,i+1]}+ {\varepsilon}_{(i-1,i]k})\end{aligned}$$
Recall the sequences of independent nonnegative integers ${\alpha}$ and ${\beta}$ introduced in the proof of Theorem [\[eulerthm\]]{}, given by ${\alpha}_{i,j}=a_{i,j}-a_{i+1,j},{\beta}_{i,j}=b_{i,j}-b_{i+1,j}$. We see that $a_{i,k} = \sum_{j\geq i} {\alpha}_{j,k}$ and $b_{i,k} = \sum_{j\geq i} {\beta}_{j,k}$. Changing to the variables $({\alpha},{\beta},\sigma)$, the middle sum of (\[rehpf\]) becomes $$\begin{aligned}
\sum_{i\leq j\leq l} \sum_{k\leq s_i}
({\alpha}_{j,k}+{\beta}_{j,k})({\varepsilon}_{k(i,i+1]}+ {\varepsilon}_{(i-1,i]k})\\
= \sum_{i\leq j\leq l} \sum_{s_{i-1}<k\leq s_i}
({\alpha}_{j,k}+{\beta}_{j,k})({\varepsilon}_{k(i,j+1]}+ {\varepsilon}_{(i-1,j]k})\end{aligned}$$
Therefore, we can simplify our expressions to get: $$\begin{aligned}
\label{Hw}
\lefteqn{ H({\alpha},{\beta},\sigma) := h(a,b,\sigma) =
\sum_{i=1}^l \sum_{k\leq s_i} {\varepsilon}_{k(i,i+1]}
+ \sum_{j=1}^l \sum_{k\leq s_j} {\beta}_{j,k} }\\ \nonumber
& &\mbox{} +\sum_{i\leq j\leq l} \sum_{s_{i-1}<k\leq s_i}
({\alpha}_{j,k}+{\beta}_{j,k})(s_j-s_i + {\varepsilon}_{k(j,j+1]}+{\varepsilon}_{(i-1,i]k})\end{aligned}$$
By the definition of $H$, we have shown
\[Hpf\] $$b_{2M}({{\mathcal{HQ}_{\mathbf{d}}}({{\mathbf F}(n;{\mathbf{s}})})}) =\mbox{rank }A_M({\mathcal{HQ}_{\mathbf{d}}}) = \#(H^{-1}(M)).$$
For $w\in S$, define $P(w)$ to be the elements $({\alpha},{\beta},\sigma)\in P$ satisfying $\sigma = w$. Let $H_w$ denote the restriction of $H$ to $F(w)$. Then $\#(H_w^{-1}(M))$ counts the number of sequences of natural numbers $({\alpha},{\beta})$ satisfying the relation $H_w({\alpha},{\beta}) = M$ given by (\[Hw\]) as well as the relations in (\[rel\]). Since all of these relations are linear in the variables ${\alpha}_{k,j}$ and ${\beta}_{k,j}$, by the same reasoning as in the proof of Theorem \[eulerthm\], we have the following generating function:
$$\begin{aligned}
\label{pfw}
\lefteqn{ \sum_{M,d_1,...d_l} \#(H_w^{-1}(M))t_1^{d_1}\cdots t_l^{d_l}z^M = }\\
& &\nonumber z^{\sum_{i\leq l} \sum_{k\leq s_i} {\varepsilon}^w_{k(i,i+1]}}
\prod_{1\leq i \leq j \leq l} \prod_{s_{i-1}<k\leq s_i}
\frac{1}{f^{i,j}_{\rho_{i,j,k}}}
\frac{1}{f^{i,j}_{\rho_{i,j,k}+1}}\end{aligned}$$
where we have defined $f^{i,j}_k= 1-t_i...t_jz^k$ and $\rho_{i,j,k} = s_j-s_i + {\varepsilon}^w_{k(j,j+1]}+k-s_{i-1}-1.$
We are now ready to prove Theorem 1.
By the definitions of $H$ and $H_w$, we have $$\#(H^{-1}(M)) = \sum_{w\in S} \#(H_w^{-1}(M))$$ and hence $b_{2M}({{\mathcal{HQ}_{\mathbf{d}}}({{\mathbf F}(n;{\mathbf{s}})})})= \sum_w \#(H_w^{-1}(M))$. Therefore, by (\[pfw\]), it suffices to prove the following (purely combinatorial) result:
\[comb\] $$\begin{aligned}
\sum_{w\in S(n;s_1,...,s_l)}
z^{\sum_{i\leq l} \sum_{k\leq s_i} {\varepsilon}^w_{k(i,i+1]}}
\prod_{1\leq i \leq j \leq l} \prod_{s_{i-1}<k\leq s_i}
\frac{1}{f^{i,j}_{\rho_{i,j,k}}}
\frac{1}{f^{i,j}_{\rho_{i,j,k}+1}}\\
=
\left(\frac{\prod_{i=1}^n (1-z^i)}{\prod_{j=1}^{l+1}
\prod_{i=1}^{s_j-s_{j-1}} (1-z^i)} \right)
\prod_{1\leq i \leq j \leq l} \prod_{s_{i-1}<k\leq s_i}
\frac{1}{f^{i,j}_{s_j-k }}
\frac{1}{f^{i,j}_{s_{j+1}-k+1}}\end{aligned}$$
We use induction on $n$ to prove the proposition.
For $n=1$, there are two cases:
1. $s_1 = 0$. Here, $S(1;0)=S_1={id}$. Both sides of the equation are equal to $1$, so that the proposition holds.
2. $s_1 = 1$. We have $S(1;1)=S_1={id}$. Both sides of the equation are equal to $\frac{1}{(1-z)(1-tz)}$, so again the proposition holds.
The strategy is to break up $S=S(n;{(s_1,\ldots,s_l)})$ into $l+1$ different permutation groups upon which we can use the inductive hypothesis.
For $1\leq m \leq l+1$, let $S(m)$ denote the subset of $S$ consisting of permutations $w$ such that $w(s_{m-1}+1)=1$. It is clear that $S = \bigcup_m S(m)$ is a disjoint union. Any $w\in S(m)$ satisfies: $\epsilon^w_{s_{m-1}+1,j} = 1$ for $j\neq s_{m-1}+1$ and $\epsilon^w_{i,s_{m-1}}=0$ for all $i$.
For $w\in S(m)$, let $w'\in S':=S(n-1;s'_1,...,s'_l)$ be defined by:
1. $w'(i) = w(i)-1$ for $i\leq s_{m-1}$
2. $w'(j) = w(j+1)-1$ for $s_{m-1}+1<j$.
Let $\epsilon'_{i,j}=\epsilon^{w'}_{i,j}$ and define ${\varepsilon}'_{i(k,k']}$ and $\rho'_{i,j,k}$ accordingly. Note that $s'_i = s_i$ for $i\leq m-1, s'_j = s_j-1$ for $m \leq j$ and $$\label{epsilonsum}
\sum_{i\leq l} \sum_{k\leq s_i} {\varepsilon}_{k(i,i+1]}=
(n-s_m) + \sum_{i\leq l} \sum_{k\leq s_i, k\neq s_{m-1}+1} {\varepsilon}_{k(i,i+1]}.$$
We split the left hand sum over $S$ into sums over $S(m)$ for $1\leq m
\leq l+1$. For each $S(m)$, we can factor out the parts of the product on the left hand side where $i=m$ and $k=s_{m-1}+1$. Using (\[epsilonsum\]), this gives:
$$\begin{aligned}
\lefteqn{ \sum_{w\in S} z^{\sum_{i\leq l} \sum_{k\leq s_i} {\varepsilon}_{k(i,i+1]}}
\prod_{1\leq i \leq j \leq l} \prod_{s_{i-1}<k\leq s_i}
\frac{1}{f^{i,j}_{ \rho_{i,j,k}}}
\frac{1}{f^{i,j}_{ \rho_{i,j,k}+1}} }\\
&=&\sum_{m=1}^{l+1} z^{n-s_m}
\prod_{m\leq j}
\frac{1}{f^{m,j}_{s_{j+1}-s_m}}
\frac{1}{f^{m,j}_{s_{j+1}-s_m+1}} \\
& & \left( \mbox{}\sum_{w\in S(m)}
z^{\sum_{i\leq l} \sum_{k\leq s'_i} {\varepsilon}'_{k(i,i+1]}}
\prod_{i\leq m\leq j} \prod_{s'_{i-1}<k\leq s'_i}
\frac{1}{f^{i,j}_{ \rho'_{i,j,k}+1}}
\frac{1}{f^{i,j}_{ \rho'_{i,j,k}+2}} \right. \\
& & \left.
\prod_{\tiny{ \begin{array}{c} i\leq j \\ m>i \mbox{ or } j<m \end{array} }}
\prod_{s'_{i-1}<k\leq s'_i}
\frac{1}{f^{i,j}_{ \rho'_{i,j,k}}}
\frac{1}{f^{i,j}_{ \rho'_{i,j,k}+1}} \right).\end{aligned}$$
By induction. we may apply the result to $S(m){\stackrel{\sim}{=}}S(n-1;s'_1,\ldots,s'_{l+1})$ to the quantity in parentheses. In particular, by what follows from the proof of Proposition \[comb\] for $S(m)$, the sum becomes:
$$\begin{aligned}
\lefteqn{ \sum_{m=1}^{l+1} z^{n-s_m}
\prod_{m\leq j}
\frac{1}{f^{m,j}_{s_{j+1}-s_m}}
\frac{1}{f^{m,j}_{s_{j+1}-s_m+1}} }\\
& &\left( \frac{\prod_{i=1}^{n-1} (1-z^i)}{\prod_{j=1}^{l+1}
\prod_{i=1}^{s'_j-s'_{j-1}} (1-z^i) }
\prod_{i\leq m \leq j}
\prod_{s'_{i-1}<k\leq s'_i}
\frac{1}{f^{i,j}_{s'_j-k}}
\frac{1}{f^{i,j}_{s'_{j+1}-k+1}} \right. \\
& &\left.
\prod_{\tiny{ \begin{array}{c} i\leq j \\ m>i \mbox{ or } j<m \end{array} }}
\prod_{s'_{i-1}<k\leq s'_i}
\frac{1}{f^{i,j}_{s'_j +1-k}}
\frac{1}{f^{i,j}_{s'_{j+1}+1-k+1}} \right).\end{aligned}$$
Since we have: $$\frac{\prod_{i=1}^n (1-z^i)}{\prod_{j=1}^{l+1}
\prod_{i=1}^{s_j-s_{j-1}} (1-z^i)} =
\frac{\prod_{i=1}^{n-1} (1-z^i)}{\prod_{j=1}^{l+1}
\prod_{i=1}^{s'_j-s'_{j-1}} (1-z^i)} \frac{1-z^n}{1-z^{s_m-s_{m-1}}},$$ we can write this sum explicity as: $$\begin{aligned}
& &\sum_{m=1}^{l+1} z^{n-s_m} \frac{\prod_{i=1}^n (1-z^i)}{\prod_{j=1}^{l+1}
\prod_{i=1}^{s_j-s_{j-1}} (1-z^i)}
\cdot \frac{(1-z^{s_m-s_{m-1}})}{(1-z^n)}
\\ & &\prod_{m\leq j}
\frac{1}{f^{m,j}_{s_{j+1}-s_m}}
\frac{1}{f^{m,j}_{s_{j+1}-s_m+1}}
\cdot \prod_{m\leq j} \prod_{s_{m-1}<k\leq s_m-1}
\frac{1}{f^{m,j}_{s_j-k}}\frac{1}{f^{m,j}_{s_{j+1}+1-k}} \\ & &
\prod_{i\leq m-1}
\prod_{s_{i-1}<k\leq s_i}
\frac{1}{f^{i,m-1}_{s_{m-1}-k}}\frac{1}{f^{i,m-1}_{s_{m}-1-k}}
\prod_{\tiny{ \begin{array}{c} i\leq j \\ i\neq m \\ j \neq m-1
\end{array} }}
\prod_{s_{i-1}<k\leq s_i}
\frac{1}{f^{i,j}_{s_j -k}}\frac{1}{f^{i,j}_{s_{j+1}+1-k}}\end{aligned}$$
By clearing denominators, the proposition follows once we have proven:
\[pfpoly\] $$\begin{aligned}
\lefteqn{ (1-z^n) \prod_{1\leq i \leq j \leq l} f^{i,j}_{s_{j+1}-s_i} =} \\
& &
\sum_{m=1}^{l+1} z^{n-s_m} (1-z^{s_m-s_{m-1}})
\prod_{i \leq m-1} f^{i,m-1}_{s_m-s_{i-1}}
\prod_{m \leq j} f^{m,j}_{s_j-s_m}
\prod_{\tiny{ \begin{array}{c} i\leq j \\ i\neq m \\ j \neq m-1
\end{array} }} f^{i,j}_{s_{j+1}-s_i}\end{aligned}$$
[[*Proof.*]{} ]{}First, we change our notation so that we work with independent variables $x_i$ where we define $x_r = t_1...t_r$, with $x_0=1$. Then for any $i \leq j$, we have $t_i...t_j = x_j/x_{i-1}$. Therefore we have $$f^{i,j}_k = 1- t_i...t_jz^k = \frac{x_{i-1}-x_jz^k}{x_{i-1}}.$$
Substituting into both sides of Lemma \[pfpoly\] and multiplying through by $x_0^{l}x_1^{l-1}...x_l^{0}$ it suffices to prove the following polynomial identity in the ring ${\mathbb{C}}[x_0,...x_n,z]$, where we define $e^{i,j}_{k} = x_i-x_jz^k$: $$\label{xpf}
(1-z^n) \prod_{1\leq i \leq j\leq l} e^{i-1,j}_{s_{j+1}-s_i } =$$ $$\sum_{m=1}^{l+1} z^{n-s_m}(1-z^{s_m-s_{m-1}})
\prod_{i \leq m-1} e^{i-1,m-1}_{s_m-s_{i-1}}
\prod_{m \leq j} e^{m-1,j}_{s_j-s_m}
\prod_{\tiny{ \begin{array}{c} i\leq j \\ i\neq m \\ j \neq m-1
\end{array} }} e^{i-1,j}_{s_{j+1}-s_i}.$$
The polynomial ring ${\mathbb{C}}[x_0,...x_n,z]$ is a unique factorization domain, and that the left side of (\[xpf\]) completely factored except for the term $(1-z^n)$. Since the degree of $z$ matches on both sides, as does the term of $t_1^0...t_{l}^0=1$, namely $(1-z^n)$, it is enough to show the right side vanishes with the relation $e^{i-1,j}_{s_{j+1}-s_i}=0$ for each $i\leq j$, i.e. that $(x_{i-1}-x_jz^{s_{j+1}-s_i})$ is a factor of the right side.
Fix $i\leq j$. Then all but two summands on the right vanish. In particular, after some cancellations in the final terms of the two products, and bringing one of the terms to the other side, we have left to show: $$\begin{aligned}
\label{pfpolyid}
z^{n-s_i} \prod_{p \leq i-1} e^{p-1,i-1}_{s_i-s_{p-1}}
\prod_{i \leq q} e^{i-1,q}_{s_q-s_i}
\prod_{\tiny{ \begin{array}{c} p\leq j \\ p\neq i
\end{array} }} e^{p-1,j}_{s_{j+1}-s_p}
\prod_{ j+1\leq q} e^{j,q}_{s_{q+1}-s_{j+1}} \\
\nonumber =
- z^{n-s_{j+1}} \prod_{p \leq j+1} e^{p-1,j}_{s_{j+1}-s_{p-1}}
\prod_{j+1 \leq q} e^{j,q}_{s_q-s_{j+1}}
\prod_{\tiny{ \begin{array}{c} i\leq q \\ q\neq j
\end{array} }} e^{i-1,q}_{s_{q+1}-s_i}
\prod_{p\leq i-1 } e^{p-1,i-1}_{s_i-s_p}\end{aligned}$$ with the relation $e^{i-1,j}_{s_{j+1}-s_i}=0$.
But now, by doing the substitution $x_{i-1}=x_jz^{s_{j+1}-s_i}$ , we see that we will have proven Lemma \[pfpoly\] once we show that the polynomial identity (\[pfpolyid\]) holds in the (unique factorization domain) ${\mathbb{C}}[x_0,...x_{i-2},x_i,...x_n,z]$. We make the following observations about the substitution:
1. $e^{p-1,i-1}_{s_i-s_{p-1}} \mapsto e^{p-1,j}_{s_{j+1}-s_{p-1}}$
2. $e^{p-1,i-1}_{s_i-s_{p}} \mapsto e^{p-1,j}_{s_{j+1}-s_{p}}$
3. $e^{i-1,q}_{s_q-s_i} \mapsto
\left\{ \begin{array}{ll} z^{s_{j+1}-s_i} e^{j,q}_{s_q-s_{j+1}}
& \mbox{when } j+1\leq q\\
-z^{s_q-s_i} e^{q,j}_{s_{j+1}-s_q} & \mbox{when } q \leq j \end{array} \right. $
4. $e^{i-1,q}_{s_{q+1}-s_i} \mapsto
\left\{ \begin{array}{ll} z^{s_{j+1}-s_i} e^{j,q}_{s_{q+1}-s_{j+1}}
& \mbox{when } j\leq q\\
-z^{s_{q+1}-s_i} e^{q,j}_{s_{j+1}-s_{q+1}} & \mbox{when } q < j \end{array} \right. $
Substituting into both sides of (\[pfpolyid\]), and using the above properties, we have two completely factored polynomials on each side of the identity. It is easy to check that the degree of $z$ in the two terms match, that the sign matches, and that the factors of the form $e^{i,j}_k$ match exactly.
This concludes the proof of Proposition \[comb\] and Theorem \[Rpf\].
For any scheme $X$, setting $z=1$ into the Poincaré polynomial $${\mathcal{P}}(X) = \sum_M (-1)^M b_M(X)z^M$$ gives the Euler characteristic $\chi(X)$. Since odd cohomology of ${\mathcal{HQ}_{\mathbf{d}}}$ vanishes, this substitution into Theorem \[Rpf\] provides another proof of Theorem \[eulerthm\].
Special cases {#special}
=============
Quot scheme
-----------
We can apply Theorem \[Rpf\] to the ordinary Quot scheme, parametrizing rank $r$ degree $d$ quotients of $V_{{{\mathbf P}^1}}$, to get a generating function for the Poincaré polynomials of ${\mathcal{HQ}_d({\mathbf G}^r(n))}$.
$$\begin{aligned}
\lefteqn{\sum_d {\mathcal{P}}({\mathcal{HQ}_d({\mathbf G}^r(n))})t^d = }\\
& & {\mathcal{P}}({\mathbf G}^r(n))\cdot
\prod_{i=1}^{n-r} \left(\frac{1}{1-tz^{i-1}}\right)
\left(\frac{1}{1-tz^{n-i+1}}\right).\end{aligned}$$
where ${\mathcal{P}}({\mathbf G}^r(n)$ is the Poincaré polynomial of ${\mathbf G}^r(n)$, which is the following classical generating function for the Betti numbers for the Grassmannian:
$$\begin{aligned}
{\mathcal{P}}({\mathbf G}^r(n)) =
\sum_M b_{2M}({\mathbf G}^r(n))z^M =
\frac{\prod_{i=1}^n (1-z^i)}
{\prod_{i=1}^{n-r} (1-z^i)\prod_{i=1}^r (1-z^i)} \end{aligned}$$
This was the case studied by Str[ø]{}mme, who found implicit formulas for the Betti numbers, which are the same up to notation as those found in this paper. Generators and relations of the Chow ring of ${\mathcal{HQ}_{\mathbf{d}}}$ are also given in [@s]. However, this set is far from minimal, and is not suited to the study of the Chow rings as the degree $d$ becomes large.
Hyperquot scheme for a complete flag variety
--------------------------------------------
Consider the space ${{\mathcal{HQ}_{\mathbf{d}}}({{\mathbf F}(n;{\mathbf{s}})})}$ where $l=n-1$ and $s_i=i$, i.e. the space ${{\mathcal{HQ}_{\mathbf{d}}}({\mathbf F}(n))}$.
Fix $n$. Then we have the following generating function for the Poincaré polynomials of ${{\mathcal{HQ}_{\mathbf{d}}}({\mathbf F}(n))}$ where ${\mathbf{d}}= (d_1,...,d_{n-1})$.
\[Rcf\] $$\begin{aligned}
\lefteqn{ \sum_{d_1,...d_{n-1}} {\mathcal{P}}({{\mathcal{HQ}_{\mathbf{d}}}({\mathbf F}(n))})t_1^{d_1}...t_{n-1}^{d_{n-1}}= }
\\
& &{\mathcal{P}}({\mathbf F}(n)) \cdot
\prod_{1\leq i \leq j \leq n-1}
\left(\frac{1}{1-t_i...t_jz^{j-i}}\right)
\left(\frac{1}{1-t_i...t_jz^{j-i+2}}\right).\end{aligned}$$
The classical term ${\mathcal{P}}({\mathbf F}(n))$ is given by: $${\mathcal{P}}({\mathbf F}(n))=\sum_M b_{2M}({\mathbf F}(n))z^M = \frac{\prod_{i=1}^n (1-z^i)}
{\prod_{j=1}^{n} (1-z) }.$$
Examples
--------
1. ${\mathbf F}(1;0) = {\mathbf G}^1(1)$ is a point and $\mathcal{HQ}_d({{\mathbf F}(1;0)})$ is a point for $d=0$ and empty for $d>0$. $$\sum_{d}{\mathcal{P}}(\mathcal{HQ}_d({\mathbf F}(1;0)))t^d = 1$$ which is consistent with the theorem.
2. ${\mathbf F}(1;1) = {\mathbf G}^0(1)$ is a point. $\mathcal{HQ}_d({\mathbf F}(1;1))$ parametrizes quotients of ${{\mathcal{O}}}_{{{\mathbf P}^1}}$ of rank $0$ and degree $d$, which are all of the form ${{\mathcal{O}}}{\twoheadrightarrow}{{\mathcal{O}}}_D$, where $D\in \mbox{Sym}^d {{{\mathbf P}^1}}{\stackrel{\sim}{=}}{\mathbf P}^d$. Therefore $\mathcal{HQ}_d({\mathbf F}(1;1)) {\stackrel{\sim}{=}}{\mathbf P}^d$, giving: $$\sum_{d,M} b_{2M}({\mathbf P}^d)t^dz^M = \sum_d t^d
\left( \sum_{M\leq d} z^M \right) = \frac{1}{(1-t)(1-tz)}.$$
3. $\mathcal{HQ}_d({\mathbf G}^{n-1}(n))$ can be viewed as the space of sheaf injections ${{\mathcal{O}}}(-d){\hookrightarrow}{\oplus}_{i=1}^{n} {{\mathcal{O}}}$ up to equivalence. Each inclusion of sheaves is given by $n$ sections in $H^0({\mathbf P}^1,{{\mathcal{O}}}(d))$. Thus, we can view any such inclusion as an element of the vector space ${\oplus}_{i=1}^{n} H^0({{\mathcal{O}}}(d))$ of dimension $n(d+1)$. Two inclusions are equivalent exactly when they differ by a scalar. Hence, $\mathcal{HQ}_d{\stackrel{\sim}{=}}{\mathbf P}^{n(d+1)-1}$ so that ${\mathcal{P}}(\mathcal{HQ}_d) = \sum_{0\leq M\leq n(d+1)} z^M$.
[9]{}
A. Bertram, [*Towards a Schubert calculus for maps from a Riemann Surface to a Grassmannian*]{}, Internat. J. Math., no. 5 (1995), 811-825.
A. Bertram, [*Quantum Schubert calculus*]{}, Adv. Math., no. 128 (1997), 289-305.
A. Bialynicki-Birula, [*Some theorems of algebraic groups*]{}, Ann. Math., 98 (1973), 480-497.
A. Bialynicki-Birula, [*Some properties of the decomposition of algebraic varieties determined by action of a torus*]{}, Bull. Acad. Polon. Sci. Sér. Sci. Math. Astronom. Phys. 24 (1976) no. 9, 667-674.
L. Chen, [*Quantum cohomology of flag manifolds*]{}, preprint.
L. Chen, [*Quantum cohomology of partial flag manifolds*]{}, Ph.D. Dissertation, University of Chicago, 2000, in preparation.
I. Ciocan-Fontanine, [*The quantum cohomology of flag varieties*]{}, Trans. Amer. Math. Soc. [**351**]{} (1999), no. 7, 2695-2729.
I. Ciocan-Fontanine, [*On quantum cohomology rings of partial flag varieties*]{}, Duke Math. J. [**98**]{} (1999), no. 3, 485–524.
S. Fomin, S. Gelfand, and A. Postnikov, [*Quantum Schubert polynomials*]{}, J. Amer. Math. Soc., [**10**]{} (1997), 565-596.
M. Finkelberg and A. Kuznetsov, [*Global intersection cohomology of quasimaps’ spaces*]{}, Internat. Math. Res. Notices, no. 7 (1997), 301–328.
A. Grothendieck, [*Techniques de construction et théorèmes d’existence en géometrie algébrique, IV: Les schémas de Hilbert*]{}, Seminaire Bourbaki, no. 221 (1960/61).
B. Kim, [*Gromov-Witten invariants for flag manifolds*]{}, Ph.D. Dissertation, University of California at Berkeley, 1996.
J. Kollar, [*Rational curves on Algebraic Varieties*]{}, Springer-Verlag, Berlin, 1996.
D. Mumford, [*Lectures on curves on an algebraic surface*]{}, Princeton University Press, Princeton, NJ, 1966.
S.A. Str[ø]{}mme, [*On parametrized rational curves in Grassmann varieties*]{}, Lecture Notes in Math, [**1266**]{} (1987), 251-272.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present our preliminary results for semileptonic form factors of $D$ mesons in unquenched lattice QCD. Simulations are carried out with $n_f=2+1$ dynamical quarks using gauge configurations generated by the MILC collaboration. For the valence quarks, we adopt an improved staggered light quark action and the clover heavy quark action. Our results for $D \to K$ and $D \to \pi$ form factors at $q^2=0$ are in agreement with the experimental values.'
address:
- ' Fermi National Accelerator Laboratory, P.O. Box 500, Batavia, IL 60510'
- ' School of Computer Science, Telecommunications and Information Systems, DePaul University, Chicago, Illinois 60604'
- ' Department of Physics, University of Illinois, Urbana, IL 61801'
- ' Department of Physics, Indiana University, Bloomington, IN 47405'
author:
- 'Masataka Okamoto, Massimo di Pierro, Aida X. El-Khadra, Steven Gottlieb, Andreas S. Kronfeld, Paul B. Mackenzie, Damian P. Menscher, Mehmet B. Oktay, and James N. Simone'
title: 'Semileptonic decays of $D$ mesons in unquenched lattice QCD[^1]'
---
INTRODUCTION
============
In order to extract the CKM matrix element $|V_{ub}|$ accurately with experimental measurement of the semileptonic decay width, a precision lattice QCD calculation of the $B \to \pi$ form factor is required. To check the reliability of lattice calculations of the heavy to light form factors, study of the semileptonic decays of $D$ mesons, such as $D \to K$ and $D \to \pi$, is a good test ground because the corresponding CKM matrices $|V_{cs}|$ and $|V_{cd}|$ are known more accurately than $|V_{ub}|$. Furthermore, forthcoming experiments by the CLEO-c collaboration will provide more stringent checks of lattice calculations in the $D$ meson system.
We have started new lattice calculations of heavy quark physics in unquenched QCD[@spectrum], and here we report our preliminary results for semileptonic form factors of $D$ mesons. We are using unquenched gauge configurations with $n_f=2+1$ improved staggered quarks generated by the MILC collaboration[@milc], with which the systematic errors due to the quenched approximation should be almost absent. For the valence light quarks, we adopt an improved staggered quark action, which allow us to simulate at lighter quark mass than previous studies with the Wilson-type light quarks. Hence, our new calculations should have a better control over the chiral extrapolations.
METHOD {#sec:method}
======
In order to combine the staggered light quark with the Wilson-type heavy quark in heavy-light bilinears, we convert the staggered quark propagator $g(x,y)$ to the “naive” quark propagator $G(x,y)$ according to g(x,y) (x) \^(y) = G(x,y) with $
\Om(x) ~=~ \ga_0^{x_0} \ga_1^{x_1} \ga_2^{x_2} \ga_3^{x_3}
$[@Wingate:2002fh]. The 3-point function for the matrix element is then computed as C\_[3,]{}\^[D]{}(t\_x,t\_y; \_,\_D) = \_[,]{} e\^[i(\_D-\_)- i\_D]{}\
, where subscripts $d,c,u$ denote quark flavors. The matrix element can be extracted from the ratio |V\_|D with the $D$ meson (Wilson-Naive) 2-point function $C_2^{D}$ and the pion (Naive-Naive) 2-point function $C_2^{\pi}$. Care is needed, however, for the overall normalization of amplitude since the naive quark action describes 16 fermions, which can cause the doubling of these correlation functions.
In Ref. [@Wingate:2002fh] it is shown that the Wilson-Naive 2-point function $C_2^{D}$ does not have the doubling because contributions of quarks with momentum $p\sim\O(\pi /a)$ are suppressed by the Wilson term. The same also holds for the 3-point functions which include at least one Wilson propagator such as $C_{3,\mu}^{D\to\pi}$. On the other hand, the Naive-Naive 2-point function $C_2^{\pi}$ should have 16 equivalent contributions. Therefore one has to divide it by 16 to get the physical amplitude; $C_2^{\pi, \rm phys} = C_2^{\pi} / 16$.
$m_l^{\rm sea}/m_s^{\rm sea}$ $m_l^{\rm val}/m_s^{\rm val}$ conf $t_x$
------------------------------- ------------------------------- --------------- -------
0.01/0.05 0.01/0.0415 $552\times 4$ 20
0.02/0.05 0.02/0.0415 460 20
0.03/0.05 0.03/0.0415 358 22
0.01/0.05 0.0415/0.0415 412 26
$\infty /\infty$ 0.0415/0.0415 350 16
: Quark mass, statistics and the sink time.
\[tab:params\]
SIMULATION
==========
Unquenched calculations are performed using $n_f=2+1$ dynamical gauge configurations obtained with an improved staggered quark action on a $20^3\times64$ lattice ($a^{-1}\!\!\approx\!\! 1.58$ GeV)[@milc]. For the valence light quarks we use the same staggered quark action as for the dynamical quarks. The valence light quark ($u,d$) mass $m_l^{\rm val}$ is usually set equal to the dynamical light quark mass $m_l^{\rm sea}$. For the valence charm quark we use the clover action with the Fermilab interpretation[@kkm]. The hopping parameter is fixed to $K_{\rm charm}=0.119$, based on our spectrum study[@spectrum]. The $O(a)$ rotation[@kkm] is performed for the vector current.
The 3-point functions are computed in the $D$ meson rest frame ($\pbf_D=\zerobf$) for the light meson momentum $\pbf_\pi$ up to $(1,1,1)$ in lattice units, using local source and sink. The sink time is fixed to $t_x=20-26$ depending on $m_l^{\rm val}$, whereas the source time is set to $t_0=0$ with an exception at $m_l^{\rm val}=0.01$, where we average over results from four source times $t_0=0,16,32$ and 48. Some simulation parameters are summarized in Table \[tab:params\].
In addition to the unquenched calculations, we also perform a quenched simulation at $m_l^{\rm val}=m_s^{\rm val}=0.0415$ using $\beta=5.9$ ($a^{-1}\!\! \approx \!\! 1.80$ GeV) configurations on a $16^3\times 32$ lattice used in our previous study[@El-Khadra:2001rv]. Comparison between the quenched result with the staggered light quarks and that with the Wilson-type light quarks[@El-Khadra:2001rv] allows us to check the validity of our new calculations.
For the vector current renormalization $Z_{V_\mu}^{cd}$ we follow the method in Ref. [@El-Khadra:2001rv]. We take $
Z_{V_\mu}^{cd}\!\!=\!\!\rho_{V_\mu}(Z_{V}^{cc} Z_{V}^{dd})^{1/2} ,
$ where $Z_{V}^{qq}$ ($q\!\!=\!\!c,d$) is the renormalization constant for the flavor-conserving current, which we compute nonperturbatively from the charge normalization condition $Z_{V}^{qq} \< D(\zerobf)| V_4^{qq} | D(\zerobf)\>\!\! =\!\! 2 m_D$. The $\rho_{V_\mu}$ is set to unity. The one-loop calculation is in progress.
{width="7.5cm"}
\[fig:Ds2eta\]
RESULTS
=======
Form factors are defined through | V\^| D &=& f\_+(q\^2) \^\
&& + f\_0(q\^2) q\^\
&=& with $q = p_D - p_\pi$, $v=p_D/m_D$, $p_\perp=p_\pi-Ev$ and $E=E_\pi$. The second expression using $f_\parallel$ and $f_\perp$ is more convenient when one considers the heavy quark expansion and the chiral limit.
$D_s\to\eta_s$
--------------
In Fig. \[fig:Ds2eta\] we summarize the results of form factors $f_0$ and $f_+$ for the $D_s\to\eta_s(s\bar{s})$ decay obtained with the naive(staggered) light quarks and previous results[@El-Khadra:2001rv] with the Wilson-type light quarks. The lines in the figure are fitting curves with a parametrization by Becirevic and Kaidalov (BK)[@Becirevic:1999kt]. One can see a nice agreement between the quenched result with the naive quarks (squares, dashed line) and that with the Wilson-type quarks (circles, dotted), showing that our new method works well. We also note that the unquenched result (diamonds, solid) is larger than quenched ones for $q^2 > 0$. See also Ref. [@detar03] for a similar comparison for the $B_s\to\eta_s$ form factors.
{width="3.6cm"} {width="3.6cm"}
\[fig:chiral\]
{width="7.cm"} {width="7.cm"}
\[fig:D2piK\]
$D\to\pi$ and $D\to K$
----------------------
To obtain $D\to\pi/K$ form factors at the physical quark mass, we need to perform a chiral extrapolation using data in range of $m_l^{\rm val}=0.01$-0.03. We do this for $f_\parallel$ and $f_\perp$ at fixed pion(kaon) energies $E_{\pi(K)}$ because the chiral perturbation formulas for the heavy to light form factors are given in such a way[@Becirevic:2002sc]. In order to interpolate and extrapolate the results to common values of $E_\pi$, we use a fit with the BK parametrization. We then perform a linear chiral extrapolation in $m_l^{\rm val}$ at nine values of $(aE_\pi)^2$. One example of these procedures is shown in Fig. \[fig:chiral\] for $f_\perp^{D\to\pi}$. Finally $f_\parallel$ and $f_\perp$ are converted to $f_0$ and $f_+$.
The $D\to\pi$ and $D\to K$ form factors are shown in Fig. \[fig:D2piK\] together with experimental values at $q^2=0$[@Hagiwara:fs]. Our results at $q^2=0$ are f\_+\^[D K]{}(0) = 0.75(3) , f\_+\^[D ]{}(0) = 0.64(3) with statistical errors only, whereas experimental values are $
f_+^{D \to K}(0) = 0.73(2)
$ and $
f_+^{D \to \pi}(0) = 0.73(13)
$ with $|V_{cs}|=0.996(13)$ and $|V_{cd}|=0.224(16)$[@Hagiwara:fs]. Our results are in agreement with the experimental values. The analysis including the chiral logarithm and the one-loop renormalization constant is underway.
We thank the MILC collaboration for the use of their configurations, and the Fermilab Computing Division and the SciDAC program for their support. Fermilab is operated by Universities Research Association Inc., under contract with the DOE.
[9]{} See also, P. Mackenzie and J. Simone, these proceedings.
C. Bernard [*et al.*]{} (MILC collaboration), Phys. Rev. D [**64**]{}, 054506 (2001).
M. Wingate [*et al.*]{} Phys. Rev. D [**67**]{}, 054505 (2003) A. X. El-Khadra, A. Kronfeld and P. Mackenzie, Phys. Rev. D [**55**]{}, 3933 (1997).
A. X. El-Khadra [*et al.*]{} Phys. Rev. D [**64**]{}, 014502 (2001) D. Becirevic and A. B. Kaidalov, Phys. Lett. B [**478**]{}, 417 (2000) C. DeTar, these proceedings. D. Becirevic, S. Prelovsek and J. Zupan, Phys. Rev. D [**67**]{}, 054010 (2003) K. Hagiwara [*et al.*]{} \[Particle Data Group Collaboration\], Phys. Rev. D [**66**]{}, 010001 (2002); C. Caso [*et al.*]{} Eur. Phys. J. C [**3**]{}, 1 (1998).
[^1]: Talk presented by M. Okamoto.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We observe that over an algebraically closed field, any finite-dimensional algebra is the endomorphism algebra of an $m$-cluster-tilting object in a triangulated $m$-Calabi-Yau category, where $m$ is any integer greater than $2$.'
address: 'Department of Mathematics, University of Haifa, Mount Carmel, Haifa 31905, Israel'
author:
- Sefi Ladkani
bibliography:
- 'dCYtilt.bib'
title: 'Finite-dimensional algebras are $(m>2)$-Calabi-Yau-tilted'
---
Introduction
============
Cluster categories were introduced by Buan, Marsh, Reineke, Reiten and Todorov [@BMRRT06] as a means to model the combinatorics of cluster algebras with acyclic skew-symmetric exchange matrices within the framework of quiver representations (see also the work of Caldero, Chapoton and Schiffler [@CCS06] for the case of $A_n$ quivers).
The cluster category of an acyclic quiver is the orbit category of the bounded derived category of its path algebra with respect to the autoequivalence $F=\nu \Sigma^{-2}$ where $\nu$ denotes the Serre functor and $\Sigma$ is the suspension functor. By a result of Keller [@Keller05], the cluster category is a triangulated 2-Calabi-Yau category. Particular role is played by the 2-cluster-tilting objects within this category (the precise definitions will be given in Section \[ssec:notat\] below) which model the clusters in the corresponding cluster algebra.
As already shown in [@Keller05], by replacing $\Sigma^{-2}$ by $\Sigma^{-m}$ for $m>2$ and considering the orbit category with respect to the autoequivalence $\nu \Sigma^{-m}$, one gets an $m$-Calabi-Yau triangulated category with $m$-cluster-tilting objects. The categories obtained in this way, called *$m$-cluster categories*, were the subject of many investigations, see [@BaurMarsh08; @Thomas07; @Wraalsen09; @ZhouZhu09].
The endomorphism algebras of 2-cluster-tilting objects in cluster categories are known as *cluster-tilted algebras* and they possess many remarkable representation-theoretic and homological properties [@ABS08; @BMR06; @BMR07; @KellerReiten07]. More generally, consider an *$m$-Calabi-Yau-tilted algebra*, i.e.an algebra $A=\operatorname{End}_{{\mathcal{C}}}(T)$ where ${\mathcal{C}}$ is a $K$-linear, triangulated, $\operatorname{Hom}$-finite, $m$-Calabi-Yau category over a field $K$ and $T$ is an $m$-cluster-tilting object in ${\mathcal{C}}$ for some positive integer $m$. Keller and Reiten have shown the following results in the case $m=2$ (for the first two points, see Sections 2 and 3 of [@KellerReiten07] and for the third one, see [@KellerReiten08 §2]):
- $A$ is Gorenstein of dimension at most $m-1$ (i.e. $\operatorname{id}_A A \leq m-1$ and $\operatorname{pd}_A DA \leq m-1$, where $D(-)=\operatorname{Hom}_K(-,K)$);
- The stable category of Cohen-Macaulay $A$-modules is $(m+1)$-Calabi-Yau;
- If $K$ is algebraically closed, ${\mathcal{C}}$ is algebraic and $A \cong KQ$ for an acyclic quiver $Q$, then ${\mathcal{C}}$ is triangle equivalent to the $m$-cluster category of $Q$.
In addition, they have shown that these results hold also in the case $m>2$ provided that one imposes an additional condition on the $m$-cluster-tilting object $T$ stated in terms of the vanishing of some of its negative extensions, namely $$\label{e:vosnex}
\tag{$\star$}
\operatorname{Hom}_{{\mathcal{C}}}(T, \Sigma^{-i} T)=0 \text{ for any $0 < i < m-1$},$$ see [@KellerReiten08 §4]. Note that by the $m$-Calabi-Yau property of ${\mathcal{C}}$ this condition is equivalent to the vanishing of the positive extensions $\operatorname{Hom}_{{\mathcal{C}}}(T, \Sigma^i T)$ for all $m < i < 2m-1$, whereas these extensions for $0 < i < m$ always vanish since $T$ is $m$-cluster-tilting.
In the terminology of [@Beligiannis15], the condition means that $T$ is $(m-2)$-corigid, and in [@Beligiannis15 Theorem B] Beligiannis presents a more refined result connecting the corigidity property of an $m$-cluster-tilting object with the Gorenstein property of its endomorphism algebra and the Calabi-Yau property of its stable category of Cohen-Macaulay modules. We note also that a condition analogous to , stated for $m$-cluster-tilting subcategories inside bounded derived categories of modules and abbreviated “vosnex”, appears in the works of Amiot-Oppermann [@AmiotOppermann14a Definition 4.9] and Iyama-Oppermann [@IyamaOppermann13 Notation 3.5].
Whereas for $m=2$ the condition is empty and hence automatically holds, this is no longer the case for $m \geq 3$. In particular, there are examples of triangulated $m$-Calabi-Yau categories ${\mathcal{C}}$ and $m$-cluster-tilting objects $T \in {\mathcal{C}}$ for which the vosnex condition does not hold and moreover the endomorphism algebra $\operatorname{End}_{{\mathcal{C}}}(T)$ is a path algebra of a connected acyclic quiver, see Iyama and Yoshino [@IyamaYoshino08 Theorem 9.3] for $m=3$ and [@IyamaYoshino08 Theorem 10.2] for $m$ odd, and also [@KellerReiten08 Example 4.3]. Another example, where the algebra $\operatorname{End}_{{\mathcal{C}}}(T)$ is not Gorenstein, is given in [@KellerReiten07 Example 5.3].
The purpose of this note is to extend this class of examples by showing that given $m \geq 3$, *any* finite-dimensional algebra over an algebraically closed field is the endomorphism algebra of an $m$-cluster-tilting object in an $m$-Calabi-Yau triangulated category. Hence, without any further assumptions on ${\mathcal{C}}$ and $T$, one cannot say too much about the algebras $\operatorname{End}_{{\mathcal{C}}}(T)$.
To this end we invoke the construction of generalized cluster categories due to Amiot [@Amiot09] in the case $m=2$ and generalized by Guo [@Guo11] to the case $m>2$. This construction produces an $m$-Calabi-Yau triangulated category with an $m$-cluster-tilting object from any dg-algebra which is homologically smooth, bimodule $(m+1)$-Calabi-Yau and satisfies additional finiteness conditions. A rich source of such dg-algebras is provided by the deformed Calabi-Yau completions defined and investigated by Keller [@Keller11].
Given a basic finite-dimensional algebra $A$, we choose a dg-algebra $B$ with two properties; firstly, the underlying graded algebra of $B$ is the path algebra of a graded quiver whose arrows are concentrated in degrees $0$ and $-1$, and secondly, $\operatorname{H}^0(B) \cong A$. Such dg-algebra can be constructed from any presentation of $A$ as a quotient of a path algebra of a quiver by an ideal generated by a finite sequence of elements. Conversely, any such dg-algebra arises in this way.
It turns out that for any $m \geq 2$, the $(m+1)$-Calabi-Yau completion of $B$ is a Ginzburg dg-algebra $\Gamma$ of a graded quiver with homogeneous superpotential of degree $2-m$ which can be written explicitly in terms of the quiver and the sequence of elements. In the case $m=2$, the zeroth homology $\operatorname{H}^0(\Gamma)$ is a split extension of $A$, whereas when $m>2$ it is isomorphic to $A$, hence $\Gamma$ satisfies the finiteness conditions required in the construction of [@Amiot09; @Guo11] and thus gives rise to a $\operatorname{Hom}$-finite $m$-Calabi-Yau triangulated category with an $m$-cluster-tilting object whose endomorphism algebra is isomorphic to $A$.
The $m$-cluster-tilting object we get almost never satisfies the vosnex condition . More precisely, that condition holds if and only if $A$ is the path algebra of an acyclic quiver and $B$ is chosen such that $B=A$. Moreover, the flexibility in the choice of the dg-algebra $B$ allows to construct, for certain algebras $A$ and any $m>2$, inequivalent triangulated $m$-Calabi-Yau categories ${\mathcal{C}}\not \simeq {\mathcal{C}}'$ with $m$-cluster-tilting objects $T \in {\mathcal{C}}$ and $T' \in {\mathcal{C}}'$ such that $\operatorname{End}_{{\mathcal{C}}}(T) \cong \operatorname{End}_{{\mathcal{C}}}(T') \cong A$.
Recollections
=============
Notations {#ssec:notat}
---------
We recall the definitions of Calabi-Yau triangulated categories and cluster-tilting objects. Throughout, we fix a field $K$.
Let ${\mathcal{C}}$ be a $K$-linear triangulated category with suspension functor $\Sigma$.
1. We say that ${\mathcal{C}}$ is *$\operatorname{Hom}$-finite* if the spaces $\operatorname{Hom}_{{\mathcal{C}}}(X,Y)$ are finite-dimensional over $K$ for any $X, Y \in {\mathcal{C}}$.
2. Let $m \in {\mathbb{Z}}$. We say that ${\mathcal{C}}$ is *$m$-Calabi-Yau* if ${\mathcal{C}}$ is $\operatorname{Hom}$-finite and there exist functorial isomorphisms $$\operatorname{Hom}_{{\mathcal{C}}}(X,Y) \cong D\operatorname{Hom}_{{\mathcal{C}}}(Y, \Sigma^m X)$$ for any $X, Y \in {\mathcal{C}}$, where $D$ denotes the duality $D(-) = \operatorname{Hom}_K(-, K)$.
For an object $X$ in an additive category ${\mathcal{C}}$, denote by $\operatorname{add}X$ the full subcategory of ${\mathcal{C}}$ whose objects are finite direct sums of direct summands of $X$. For a subcategory ${\mathcal{Y}}$ of ${\mathcal{C}}$, let ${^\perp}{\mathcal{Y}}= \left\{ X \in {\mathcal{C}}\,:\, \operatorname{Hom}_{{\mathcal{C}}}(X,Y)=0
\text{ for any $Y \in {\mathcal{Y}}$} \right\}$.
Let ${\mathcal{C}}$ be a triangulated $m$-Calabi-Yau category for some integer $m \geq 1$. An object $T$ of ${\mathcal{C}}$ is *$m$-cluster-tilting* if:
1. $\operatorname{Hom}_{{\mathcal{C}}}(T, \Sigma^i T) = 0$ for any $0 < i < m$; and
2. if $X \in {\mathcal{C}}$ is such that $\operatorname{Hom}_{{\mathcal{C}}}(X, \Sigma^i T) = 0$ for any $0 < i < m$, then $X \in \operatorname{add}T$.
Equivalently, $\operatorname{add}T = \bigcap_{0 < i < m} {^\perp} \Sigma^i(\operatorname{add}T)$.
Let $m \geq 1$. A $K$-algebra $\Lambda$ is *$m$-Calabi-Yau-tilted* (*$m$-CY-tilted* for short) if there exist a triangulated $m$-Calabi-Yau category ${\mathcal{C}}$ and an $m$-cluster-tilting object $T$ of ${\mathcal{C}}$ such that $\Lambda \cong \operatorname{End}_{{\mathcal{C}}}(T)$.
Generalized cluster categories
------------------------------
In this section we briefly review the construction of triangulated $m$-Calabi-Yau categories with $m$-cluster-tilting object due to Amiot [@Amiot09] (for the case $m=2$) and its generalization by Guo [@Guo11] (for the case $m>2$).
Let $\Gamma$ be a differential graded (dg) algebra over a field $K$. Denote by ${\mathcal{D}}(\Gamma)$ its derived category and by $\operatorname{per}\Gamma$ its smallest full triangulated subcategory containing $\Gamma$ and closed under taking direct summands. Let ${\mathcal{D}}_{\operatorname{fd}}(\Gamma)$ denote the full subcategory of ${\mathcal{D}}(\Gamma)$ whose objects are those of ${\mathcal{D}}(\Gamma)$ with finite-dimensional total homology.
$\Gamma$ is said to be *homologically smooth* if $\Gamma \in \operatorname{per}\Gamma^e$, where $\Gamma^e = \Gamma^{op} \otimes_K \Gamma$.
Let $\Gamma$ be homologically smooth and let $\Omega = \operatorname{\mathbf{R}Hom}_{\Gamma^e}(\Gamma, \Gamma^e)$. By definition, $\Omega \in {\mathcal{D}}((\Gamma^e)^{op})$. We can view $\Omega$ as an object of ${\mathcal{D}}(\Gamma^e)$ via restriction of scalars along the morphism $\tau \colon \Gamma^e \xrightarrow{\sim} (\Gamma^e)^{op}$ given by $\tau(x \otimes y) = y \otimes x$. By [@Keller11 Lemma 3.4] (see also [@Keller08 Lemma 4.1] and [@Guo11 Lemma 2.1]) one has ${\mathcal{D}}_{\operatorname{fd}}(\Gamma) \subseteq \operatorname{per}\Gamma$ and $$\label{e:Serre}
\operatorname{Hom}_{{\mathcal{D}}(\Gamma)}(L \otimes_{\Gamma} \Omega, M) \cong
D \operatorname{Hom}_{{\mathcal{D}}(\Gamma)}(M, L)$$ for any $L \in {\mathcal{D}}(\Gamma)$, $M \in {\mathcal{D}}_{\operatorname{fd}}(\Gamma)$.
Let $m \in {\mathbb{Z}}$. If $\operatorname{\mathbf{R}Hom}_{\Gamma^e}(\Gamma, \Gamma^e) \cong
\Sigma^{-m} \Gamma$ in ${\mathcal{D}}(\Gamma^e)$ we say that $\Gamma$ is *bimodule $m$-Calabi-Yau*.
If $\Gamma$ is homologically smooth and bimodule $m$-Calabi-Yau then the triangulated category ${\mathcal{D}}_{\operatorname{fd}}(\Gamma)$ is $m$-Calabi-Yau. More precisely, yields functorial isomorphisms $$\operatorname{Hom}_{{\mathcal{D}}(\Gamma)}(\Sigma^{-m} L, M) \cong
D \operatorname{Hom}_{{\mathcal{D}}(\Gamma)}(M, L)$$ for any $L \in {\mathcal{D}}(\Gamma)$, $M \in {\mathcal{D}}_{\operatorname{fd}}(\Gamma)$.
\[t:cluster\] Let $m \geq 1$ and let $\Gamma$ be a dg-algebra satisfying the following conditions:
1. $\Gamma$ is homologically smooth;
2. $\operatorname{H}^i(\Gamma)=0$ for any $i>0$;
3. $\dim_K \operatorname{H}^0(\Gamma) < \infty$;
4. $\Gamma$ is bimodule $(m+1)$-Calabi-Yau.
Consider the triangulated category ${\mathcal{C}}_{\Gamma} = \operatorname{per}\Gamma / {\mathcal{D}}_{\operatorname{fd}}(\Gamma)$. Then:
1. ${\mathcal{C}}_{\Gamma}$ is $\operatorname{Hom}$-finite and $m$-Calabi-Yau.
2. For any $i \in {\mathbb{Z}}$, set ${\mathcal{D}}^{\leq i} = \left\{ X \in {\mathcal{D}}(\Gamma) \,:\, \operatorname{H}^p(X)=0
\text{ for all $p>i$} \right\}$ and let ${\mathcal{F}}= {\mathcal{D}}^{\leq 0} \cap {^\perp}{\mathcal{D}}^{\leq -m} \cap \operatorname{per}\Gamma$. The restriction of the canonical projection $\pi \colon \operatorname{per}\Gamma
\to {\mathcal{C}}_{\Gamma}$ to ${\mathcal{F}}$ induces an equivalence of $K$-linear categories ${\mathcal{F}}\xrightarrow{\sim} {\mathcal{C}}_{\Gamma}$.
3. \[it:t:CT\] The image $\pi \Gamma$ of $\Gamma$ in ${\mathcal{C}}_{\Gamma}$ is an $m$-cluster-tilting object and $\operatorname{End}_{{\mathcal{C}}_{\Gamma}}(\pi \Gamma) \cong \operatorname{H}^0(\Gamma)$.
The object $\pi \Gamma$ occurring in part of the theorem is called the *canonical $m$-cluster-tilting* object in ${\mathcal{C}}_{\Gamma}$. The statement of the next lemma concerning negative extension groups of the canonical $m$-cluster-tilting object is implicit in the proof of [@Guo11 Corollary 3.4].
\[l:snex\] Let $\Gamma$ be a dg-algebra satisfying the conditions of Theorem \[t:cluster\] and let $T=\pi \Gamma$ be the canonical $m$-cluster-tilting object in ${\mathcal{C}}_{\Gamma}$. Then $\operatorname{Hom}_{{\mathcal{C}}}(T, \Sigma^{-i} T) \cong \operatorname{H}^{-i}(\Gamma)$ for any $0 \leq i \leq m-1$.
As observed in [@Guo11], the objects $\Gamma, \Sigma \Gamma, \dots,
\Sigma^{m-1} \Gamma$ are in the fundamental domain ${\mathcal{F}}$. Therefore, for any $0 \leq i \leq m-1$, $$\begin{aligned}
\operatorname{Hom}_{{\mathcal{C}}_{\Gamma}}(\pi \Gamma, \Sigma^{-i} \pi \Gamma) &\cong
\operatorname{Hom}_{{\mathcal{C}}_{\Gamma}}(\Sigma^i \pi \Gamma, \pi \Gamma)
= \operatorname{Hom}_{{\mathcal{C}}_{\Gamma}}(\pi \Sigma^i \Gamma, \pi \Gamma) \cong
\operatorname{Hom}_{{\mathcal{D}}(\Gamma)}(\Sigma^i \Gamma, \Gamma) \\
&\cong
\operatorname{Hom}_{{\mathcal{D}}(\Gamma)}(\Gamma, \Sigma^{-i} \Gamma) \cong \operatorname{H}^{-i}(\Gamma) .\end{aligned}$$
Ginzburg dg-algebras
--------------------
Ginzburg dg-algebras were introduced by Ginzburg in [@Ginzburg06]. We recall their definition in the case of a graded quiver with homogeneous superpotential and quote the result of Keller [@Keller11] that they are homologically smooth and bimodule Calabi-Yau (see also the paper [@VandenBergh15] by Van den Bergh). We note that a graded version of quivers with potentials (and their mutations) has also been introduced in [@AmiotOppermann14b] and [@dTdVVdB13], however in these papers one still implicitly considers Ginzburg dg-algebras which are bimodule 3-Calabi-Yau. In contrast, in the setting described below the degree of the superpotential affects the Calabi-Yau dimension of its Ginzburg dg-algebra.
A *quiver* is a finite directed graph. More precisely, it is a quadruple $Q=(Q_0, Q_1, s, t)$, where $Q_0$ and $Q_1$ are finite sets (of *vertices* and *arrows*, respectively) and $s,t \colon Q_1 \to Q_0$ are functions specifying for each arrow its starting and terminating vertex, respectively. A quiver $Q$ is *graded* if we are given a grading $|\cdot| \colon Q_1 \to {\mathbb{Z}}$.
A *path* $p$ in $Q$ is a sequence of arrows $\alpha_1 \alpha_2 \dots \alpha_n$ such that $s(\alpha_{i+1})=t(\alpha_i)$ for all $1 \leq i < n$. For a path $p$ we denote by $s(p)$ its starting vertex $s(\alpha_1)$ and by $t(p)$ its terminating vertex $t(\alpha_n)$. A path $p$ is a *cycle* if it starts and ends at the same vertex, i.e. $s(p)=t(p)$. Any vertex $i \in Q_0$ gives rise to a cycle $e_i$ of length zero with $s(e_i)=t(e_i)=i$.
The *path algebra* $KQ$ has a basis consisting of the paths of $Q$, and the product of two paths $p$ and $q$ is their concatenation $pq$ if $s(q)=t(p)$ and zero otherwise. The path algebra is graded if $Q$ is graded, with the degree of a path being the sum of the degrees of its arrows. The degree of a homogeneous element $x$ of $KQ$ will be denoted by $|x|$.
Consider the $K$-bilinear map $[-,-] \colon KQ \times KQ \to KQ$ whose value on a pair of homogeneous elements $x, y$ is given by their *supercommutator* $[x,y] = xy - (-1)^{|x||y|}yx$. Denote by $[KQ,KQ]$ the linear subspace of $KQ$ spanned by all the supercommutators. The quotient $KQ/[KQ,KQ]$ has a basis consisting of cycles considered up to cyclic permutation “with signs”.
A *superpotential* on $Q$ is a homogeneous element in $KQ/[KQ,KQ]$.
Any arrow $\alpha \in Q_1$ gives rise to a linear map $\partial_\alpha \colon KQ/[KQ,KQ] \to KQ$ (called *cyclic derivative with respect to $\alpha$*) whose value on any cycle $p$ is given by $$\label{e:deriv}
\partial_\alpha p = (-1)^{|\alpha|}
\sum_{p=u \alpha v} (-1)^{|u|(|\alpha|+|v|)} vu$$ where the sum runs over all possible decompositions $p=u \alpha v$ with $u, v$ paths of length $\geq 0$. More explicitly, if $p=\alpha_1 \alpha_2 \dots \alpha_n$ has degree $w$, then $$\partial_\alpha (\alpha_1 \alpha_2 \dots \alpha_n) =
(-1)^{|\alpha|} \sum_{\ell \,:\, \alpha_{\ell} = \alpha}
(-1)^{(w-1)(|\alpha_1|+\dots+|\alpha_{\ell-1}|)}
\alpha_{\ell+1} \dots \alpha_n \alpha_1 \dots \alpha_{\ell-1}$$ is homogeneous of degree $w-|\alpha|$.
\[def:Ginzburg\] Let $Q$ be a graded quiver and let $m \in {\mathbb{Z}}$. Let ${\overline}{Q}$ be the graded quiver whose vertices are those of $Q$ and its set of arrows consists of
- the arrows of $Q$ (with their degree unchanged);
- an arrow $\alpha^* \colon j \to i$ of degree $1-m-|\alpha|$ for each arrow $\alpha \colon i \to j$ of $Q$;
- a loop $t_i \colon i \to i$ of degree $-m$ for each vertex $i \in Q_0$.
Let $W$ be a superpotential on $Q$ of degree $2-m$. The *Ginzburg dg-algebra of $(Q,W)$*, denoted $\Gamma_{m+1}(Q,W)$, is the dg-algebra whose underlying graded algebra is the path algebra $K{\overline}{Q}$ and the differential $d$ is defined by its action on the generators as
- $d(\alpha) = 0$ and $d(\alpha^*) = \partial_{\alpha} W$ for each $\alpha \in Q_1$;
- $d(t_i) = e_i (\sum_{\alpha \in Q_1} [\alpha, \alpha^*]) e_i$ for each $i \in Q_0$.
Note that for each $\alpha \in Q_1$ the element $\partial_\alpha W$ is homogeneous of degree $2-m-|\alpha|$ and the supercommutator $[\alpha,\alpha^*]$ is homogeneous of degree $1-m$, hence the definition of the differential $d$ makes sense. Moreover, as $[\alpha,\alpha^*]=\alpha \alpha^* -
(-1)^{|\alpha||\alpha^*|} \alpha^* \alpha$ for any $\alpha \in Q_1$, by using the sign conventions in one verifies that $d^2(t_i)=0$ for any $i \in Q_0$, so that $d$ is indeed a differential. Note also that in [@Keller11] the differential of $t_i$ is $(-1)^{m-1}$ times the one given here, but of course by replacing each $t_i$ by $(-1)^{m-1}t_i$ one sees that the dg-algebras are isomorphic.
\[t:GinzburgCY\] Let $m \in {\mathbb{Z}}$ and let $W$ be a superpotential of degree $2-m$ on a graded quiver $Q$. Then $\Gamma_{m+1}(Q,W)$ is homologically smooth and bimodule $(m+1)$-Calabi-Yau.
We record two useful observations. Let $Q$ be a graded quiver, $m \in {\mathbb{Z}}$ an integer and $W$ a homogeneous superpotential on $Q$ of degree $2-m$.
\[l:reparrow\] Suppose that $\alpha \colon i \to j$ is an arrow in $Q$ such that no term of $W$ contains $\alpha$. Define a graded quiver $Q'$ by $Q'_0 = Q_0$ and $Q'_1 = Q_1 \setminus \{\alpha\} \cup \{\alpha^*\}$ where $\alpha^* \colon
j \to i$ has degree $1-m-|\alpha|$. Then $W$ can be naturally viewed as a superpotential $W'$ on $Q'$ and $\Gamma_{m+1}(Q,W) \cong \Gamma_{m+1}(Q',W')$.
Since no term of $W$ contains $\alpha$, we can view $W$ as an element $W'$ in the path algebra $KQ'$. The graded quivers ${\overline}{Q'}$ and ${\overline}{Q}$ are isomorphic by the map $\varphi \colon {\overline}{Q'} \to {\overline}{Q}$ sending $(\alpha^*)^*$ to $\alpha$ and fixing all other arrows. Moreover, since $\partial_\alpha W = 0 = \partial_{\alpha^*} W'$ and $\partial_\beta W = \partial_\beta W'$ for any $\beta \in Q_1 \cap Q'_1$, the map $\varphi$ induces an isomorphism of the Ginzburg dg-algebras $\Gamma_{m+1}(Q',W') \cong \Gamma_{m+1}(Q,W)$.
For a subset of arrows $\Omega \subseteq Q_1$, let $Q_{\Omega}$ be the subquiver of $Q$ with $(Q_{\Omega})_0=Q_0$ and $(Q_{\Omega})_1 = \Omega$. For the definition of Calabi-Yau completion, see [@Keller11 §4].
\[l:CYcompl\] Suppose that $\Omega \subseteq Q_1$ is a set of arrows such that $W = \sum_{\beta \in Q_1 \setminus \Omega} \beta \omega_\beta$ with $\omega_\beta \in KQ_{\Omega}$ for each $\beta \not \in \Omega$. Consider the subquiver $Q'$ of ${\overline}{Q}$ defined by $Q'_0 = Q_0$ and $Q'_1 = \Omega \cup \{\beta^* : \beta \in Q_1 \setminus \Omega\}$. Let $d$ be the differential on $\Gamma_{m+1}(Q,W)$. Then:
1. $d(KQ') \subseteq KQ'$, hence $B=(KQ',d)$ is a sub-dg-algebra of $\Gamma_{m+1}(Q,W)$.
2. $\Gamma_{m+1}(Q,W)$ is isomorphic to the $(m+1)$-Calabi-Yau completion of $B$.
The first claim holds since $d(\alpha)=0$ for any $\alpha \in \Omega \subseteq Q_1$ and $d(\beta^*) = \partial_\beta W = (-1)^{|\beta|} \omega_\beta \in K Q_{\Omega}
\subseteq KQ'$ for any $\beta \not \in \Omega$. This argument also shows that the differential $d$ on $KQ'$ satisfies the condition in [@Keller11 §3.6] via the filtration $\varnothing \subseteq \Omega \subseteq Q'$ of the set of arrows. Hence we can use [@Keller11 Proposition 6.6] to compute the $(m+1)$-Calabi-Yau completion of $B$ and get that it is isomorphic to $(K{\overline}{Q'},d')$ with the differential $d'$ given on the generators by $$\begin{aligned}
d'(\alpha) = \partial_{\alpha^*} W' = 0
&,& d'((\beta^*)^*) = \partial_{\beta^*} W' = 0
&,& d'(\alpha^*) = \partial_\alpha W'
&,& d'(\beta^*) = \partial_{(\beta^*)^*} W'\end{aligned}$$ and $d(t_i) = (-1)^{m+1} e_i (\sum_{\gamma \in Q'_1} [\gamma, \gamma^*]) e_i$, where $W' \in K{\overline}{Q'}$ is the element $$W' = \sum_{\alpha \in \Omega} (-1)^{|\alpha|} \alpha^* d(\alpha)
+ \sum_{\beta \in Q_1 \setminus \Omega} (-1)^{|\beta^*|} (\beta^*)^* d(\beta^*)
= \sum_{\beta \in Q_1 \setminus \Omega} (-1)^{m-1} (\beta^*)^* \omega_\beta .$$
Finally, the isomorphism $\varphi \colon K{\overline}{Q'} \to K{\overline}{Q}$ defined on the generators by $$\varphi(\gamma) = \begin{cases}
(-1)^{m-1} \gamma & \text{if $\gamma=\alpha^*$ for some $\alpha \in Q_1$}, \\
\beta & \text{if $\gamma=(\beta^*)^*$ for some
$\beta \in Q_1 \setminus \Omega$}, \\
\gamma & \text{otherwise}
\end{cases}$$ induces an isomorphism $(K{\overline}{Q'},d') \cong \Gamma_{m+1}(Q,W)$ of dg-algebras.
Non-positively graded Ginzburg dg-algebras
------------------------------------------
In this section we restrict attention to Ginzburg dg-algebras which are concentrated in non-positive degrees and quote the construction of Guo [@Guo11] (generalizing that of Amiot in [@Amiot09] for the case $m=2$) of the generalized cluster category associated to a quiver with superpotential.
Let $Q$ be a graded quiver and let $w \in {\mathbb{Z}}$. Denote by $Q^{(w)}$ the subquiver of $Q$ consisting of the arrows of degree $w$, in other words, $Q^{(w)}_0 = Q_0$ and $Q^{(w)}_1 = \{ \alpha \in Q_1 \,:\, |\alpha|=w \}$.
Now fix a graded quiver $Q$ and let $m \in {\mathbb{Z}}$. Let ${\overline}{Q}$ denote the quiver constructed in Definition \[def:Ginzburg\]. Fix a homogeneous superpotential $W$ on $Q$ of degree $2-m$ and let $\Gamma=\Gamma_{m+1}(Q,W)$ be the Ginzburg dg-algebra. We immediately observe:
$\Gamma^i=0$ for any $i>0$ if and only if all the arrows of ${\overline}{Q}$ have non-positive degrees. This condition implies that $\operatorname{H}^i(\Gamma)=0$ for any $i>0$.
\[l:h0\] Assume that all the arrows of ${\overline}{Q}$ have non-positive degrees. Then $\operatorname{H}^0(\Gamma) \cong K{{\overline}{Q}}^{(0)}/(d\alpha : \alpha \in {\overline}{Q}^{(-1)}_1)$.
Since there are no arrows of ${\overline}{Q}$ of positive degree, the graded piece $\Gamma^0$ equals the path algebra of ${\overline}{Q}^{(0)}$ and the graded piece $\Gamma^{-1}$ is spanned by the elements of the form $u \alpha v$ where $u,v \in K{{\overline}{Q}}^{(0)}$ and $\alpha \in {\overline}{Q}^{(-1)}_1$. As $d(u \alpha v) = u(d\alpha)v$, the claim follows.
\[l:wbQnonpos\] The following conditions are equivalent:
1. Each arrow of ${\overline}{Q}$ has non-positive degree;
2. $m \geq 0$ and for any $\alpha \in Q_1$ one has $1-m \leq |\alpha|$ and $|\alpha| \leq 0$.
If $m<0$ then the each of the loops $t_i$ in ${\overline}{Q}$ has positive degree $-m$. In addition, each other arrow of ${\overline}{Q}$ is either $\alpha$ or $\alpha^*$ for some arrow $\alpha \in Q_1$, and its degree is $|\alpha|$ or $1-m-|\alpha|$, respectively. These observations imply the statement of the lemma.
In the next two examples we discuss the cases $m=0$ and $m=1$.
Assume that $m=0$. Lemma \[l:wbQnonpos\] implies that the arrows of ${\overline}{Q}$ have non-positive degrees if and only if $Q$ has no arrows. In this case ${\overline}{Q}$ is a disjoint union of graded quivers of the form $$\xymatrix{
{\bullet} \ar@{-}@(ur,dr)[]^{t}
}$$ with $|t|=0$ and $dt=0$. Hence $\Gamma$ is concentrated in degree $0$ and it is a finite direct product of polynomial rings $k[t]$. The algebra $k[t]$ is $1$-Calabi-Yau, see [@Keller08 §4.2]
Assume that $m=1$. Lemma \[l:wbQnonpos\] implies that the arrows of ${\overline}{Q}$ have non-positive degrees if and only if all the arrows of $Q$ are in degree $0$, so we can regard $Q$ as an ungraded quiver. Since the superpotential on $Q$ is homogeneous of degree $1$, it must vanish. In this case all the arrows $\alpha$ and $\alpha^*$ of ${\overline}{Q}$ are in degree $0$ and the differential of each loop $t_i$, whose degree is $-1$, is given by $d(t_i) = e_i (\sum_{\alpha \in Q_1} [\alpha,\alpha^*]) e_i$. By Lemma \[l:h0\], $\operatorname{H}^0(\Gamma)$ is isomorphic to the *preprojective algebra* of the quiver $Q$.
When $m \geq 2$, the next lemma shows that we may assume that $Q^{(1-m)}$ has no arrows.
Assume that $m \geq 2$ and that $1-m \leq |\alpha| \leq 0$ for any $\alpha \in Q_1$. Then there exist a graded quiver $Q'$ with $2-m \leq |\alpha'| \leq 0$ for each $\alpha' \in Q'_1$ and a homogeneous superpotential $W'$ on $Q'$ of degree $2-m$ such that $\Gamma_{m+1}(Q,W) \cong \Gamma_{m+1}(Q',W')$.
We define the graded quiver $Q'$ as a subquiver of ${\overline}{Q}$; we set $Q'_0=Q_0$ and $$Q'^{(w)}_1 = \begin{cases}
Q_1^{(0)} \cup \{\alpha^* : \alpha \in Q_1^{(1-m)}\} &
\text{if $w=0$,} \\
Q_1^{(w)} & \text{if $1-m < w < 0$,}
\end{cases}$$ with $Q'^{(w)}_1$ empty for any other $w \in {\mathbb{Z}}$. Then $2-m \leq |\alpha'| \leq 0$ for any arrow $\alpha' \in Q'_1$ by construction. Moreover, since the superpotential $W$ is of degree $2-m$ and all the arrows of $Q$ have non-positive degrees, no term of $W$ can contain any arrows of $Q^{(1-m)}$. The result now follows by iterated application of Lemma \[l:reparrow\] for each of the arrows in $Q^{(1-m)}$.
In particular, when $m=2$ one can always reduce to the classical setting of an ungraded quiver with potential.
Let $Q$ be a graded quiver such that $-1 \leq |\alpha| \leq 0$ for any $\alpha \in Q_1$ and let $W$ be a homogeneous superpotential on $Q$ of degree $0$. Then there exist a quiver $Q'$ concentrated in degree $0$ and a superpotential $W'$ on $Q'$ such that $\Gamma_3(Q,W) \cong \Gamma_3(Q',W')$.
The next lemma generalizes [@KellerYang11 Lemma 2.11].
\[l:h0Q\] Assume that $m \geq 2$ and $2-m \leq |\alpha| \leq 0$ for any $\alpha \in Q_1$. Then $$\operatorname{H}^0(\Gamma_{m+1}(Q,W)) \cong
KQ^{(0)}/(\partial_\alpha W : \alpha \in Q^{(2-m)}_1) .$$
Observe that since the arrows of $Q$ have non-positive degrees and $W$ is of degree $2-m$, the cyclic derivative $\partial_\alpha W$ with respect to any arrow $\alpha$ of degree $2-m$ lies in the path algebra of $Q^{(0)}$ and the quotient in the right hand side makes sense.
The arrows of ${\overline}{Q}$ have non-positive degrees by Lemma \[l:wbQnonpos\], hence we may apply Lemma \[l:h0\] to deduce that $\operatorname{H}^0(\Gamma_{m+1}(Q,W)) \cong K {\overline}{Q}^{(0)}/(d \alpha :
\alpha \in {\overline}{Q}_1^{(-1)})$. Observe that ${\overline}{Q}^{(0)} = Q^{(0)}$ since $Q_1^{(1-m)}$ is empty by assumption and ${\overline}{Q}^{(-1)}_1 = Q^{(-1)}_1 \cup \{\alpha^* : \alpha \in Q^{(2-m)}_1 \}$ with $d(\alpha)=0$ and $d(\alpha^*) = \partial_\alpha W$ for any $\alpha \in Q_1$, hence the claim follows.
\[t:mCY\] Let $m \geq 1$ be an integer, let $Q$ be a graded quiver, let $W$ be a superpotential on $Q$ of degree $2-m$ and denote by $\Gamma = \Gamma_{m+1}(Q,W)$ the Ginzburg dg-algebra. Assume that:
1. \[it:t:deg\] $1 - m \leq |\alpha| \leq 0$ for all $\alpha \in Q_1$;
2. \[it:t:fd\] $\operatorname{H}^0(\Gamma)$ is finite-dimensional.
Then the triangulated category ${\mathcal{C}}_{(Q,W)} = \operatorname{per}\Gamma / {\mathcal{D}}_{\operatorname{fd}}(\Gamma)$ is $\operatorname{Hom}$-finite and $m$-Calabi-Yau. Moreover, the image of $\Gamma$ in ${\mathcal{C}}_{(Q,W)}$ is an $m$-cluster-tilting object whose endomorphism algebra is isomorphic to $\operatorname{H}^0(\Gamma)$.
The condition implies that $\operatorname{H}^i(\Gamma)=0$ for all $i>0$. Now Theorem \[t:GinzburgCY\] and condition imply that $\Gamma$ satisfies the conditions of Theorem \[t:cluster\] and the result is now a consequence of that theorem.
The construction
================
Superpotentials from quivers with relations
-------------------------------------------
In this section we construct, given a quiver $Q$, a finite sequence $R$ of relations on $Q$ and an integer $m \geq 2$, a graded quiver with homogeneous superpotential of degree $2-m$. The construction generalizes that of Keller in [@Keller11 §6.9] for the case where $m=2$ and $KQ/(R)$ has global dimension $2$. The idea of adding, for each relation, an arrow in the opposite direction appears already in the description of relation-extension algebras by Assem, Brüstle and Schiffler [@ABS08].
Let $Q$ be a quiver and let ${\mathfrak{r}}$ be the ideal of $KQ$ generated by all the arrows. A *relation* on $Q$ is an element of $e_i {\mathfrak{r}}e_j$ for some $i, j \in Q_0$. In other words, a relation is a linear combination of paths of positive lengths starting at $i$ and ending at $j$.
We start with some preparations concerning split extensions of algebras which will be needed for the case $m=2$.
An algebra ${\widetilde}{A}$ is a *split extension* of an algebra $A$ if there exist algebra homomorphisms $\iota \colon A \to {\widetilde}{A}$ and $\pi \colon {\widetilde}{A} \to A$ such that $\pi \iota = id_A$.
Let $Q$ be a quiver and let ${\widetilde}{Q}$ be a quiver such that ${\widetilde}{Q}_0=Q_0$ and $Q_1 \subseteq {\widetilde}{Q}_1$ (in other words, ${\widetilde}{Q}$ is obtained from $Q$ by adding arrows). Then the path algebra $K{\widetilde}{Q}$ is a split extension of the path algebra $KQ$. Indeed, there are algebra homomorphisms $$KQ \xrightarrow{\iota_{Q,{\widetilde}{Q}}} K{\widetilde}{Q} \xrightarrow{\pi_{{\widetilde}{Q},Q}} KQ$$ whose values on the generators are given by $$\begin{aligned}
\iota_{Q,{\widetilde}{Q}}(\alpha) = \alpha \quad (\alpha \in Q_1)
\qquad \text{and} \qquad
\pi_{{\widetilde}{Q},Q}(\alpha) = \begin{cases}
\alpha & \text{if $\alpha \in Q_1$,} \\
0 & \text{if $\alpha \in {\widetilde}{Q}_1 \setminus Q_1$.}
\end{cases}\end{aligned}$$
Denote by ${\mathfrak{r}}'$ the ideal of $K{\widetilde}{Q}$ generated by the arrows in the set ${\widetilde}{Q}_1 \setminus Q_1$. Let $R$ be a set of relations in $KQ$ and let ${\widetilde}{R}$ be a set of relations in $K{\widetilde}{Q}$ such that $R \subseteq {\widetilde}{R}$ (via the natural embedding $\iota_{Q,{\widetilde}{Q}} \colon KQ \hookrightarrow K{\widetilde}{Q}$).
\[l:splitext\] Assume that ${\widetilde}{R} \setminus R \subseteq {\mathfrak{r}}'$. Then the algebra $K{\widetilde}{Q}/({\widetilde}{R})$ is a split extension of the algebra $KQ/(R)$.
Consider the composition $KQ \xrightarrow{\iota_{Q,{\widetilde}{Q}}} K{\widetilde}{Q} \twoheadrightarrow K{\widetilde}{Q}/({\widetilde}{R})$. Since $R \subseteq {\widetilde}{R}$, the image of any relation $\rho \in R$ vanishes and we get an algebra homomorphism $\iota \colon KQ/(R) \to K{\widetilde}{Q}/({\widetilde}{R})$. Consider now the composition $K{\widetilde}{Q} \xrightarrow{\pi_{{\widetilde}{Q},Q}} KQ \twoheadrightarrow KQ/(R)$ and let $\rho \in {\widetilde}{R}$. If $\rho \in R$, then its image obviously vanishes. Otherwise, $\rho \not \in R$ and our assumption that ${\widetilde}{R} \setminus R
\subseteq {\mathfrak{r}}'$ implies that $\pi_{{\widetilde}{Q},Q}(\rho)=0$, so its image vanishes as well. Hence we get a well defined algebra homomorphism $\pi \colon K{\widetilde}{Q}/({\widetilde}{R}) \to KQ/R$. The composition $\pi \iota$ maps the image of any arrow $\alpha \in Q$ in $KQ/(R)$ to itself, therefore it is the identity on $KQ/(R)$.
Any quiver with a finite sequence of relations gives rise to a dg-algebra whose underlying graded algebra is the path algebra of a graded quiver with arrows concentrated in degrees $0$ and $-1$. The details are given in the construction below.
\[con:dgQR\] Let $(Q,R)$ be a pair where $Q$ is a quiver and $R = \bigcup_{i,j \in Q_0} R_{i,j}$, where each $R_{i,j}$ is a finite sequence of relations inside $e_i {\mathfrak{r}}e_j$ and $R$ is the concatenation of these sequences (there may be repetitions inside each sequence $R_{i,j}$ and moreover the zero element can appear inside several such sequences). For a relation $\rho \in R_{i,j}$, set $s(\rho)=i$ and $t(\rho)=j$.
We define a graded quiver $Q'$ as follows:
- The set of vertices of $Q'$ equals that of $Q$;
- The set of arrows consists of
- the arrows of $Q$, with their degree set to $0$;
- an arrow $\eta_\rho \colon s(\rho) \to t(\rho)$ of degree $-1$ for each relation $\rho \in R$;
and denote by $B(Q,R)$ the dg-algebra whose underlying graded algebra is the path algebra $KQ'$ with the differential acting on the generators by
- $d(\alpha)=0$ for any $\alpha \in Q_1$;
- $d(\eta_\rho)=\rho$ for any $\rho \in R$.
Obviously, $B(Q,R)$ is concentrated in non-positive degrees and $\operatorname{H}^0(B(Q,R))$ is isomorphic to $KQ/(R)$, but in general $B(Q,R)$ is not quasi-isomorphic to its zeroth homology.
If $B=(KQ',d)$ is any dg-algebra whose underlying graded algebra is the path algebra of a graded quiver with arrows concentrated in degrees $0$ and $-1$ and the image of the differential $d$ lies in the ideal generated by the arrows, then $B \cong B(Q,R)$ for some $(Q,R)$. Indeed, we can take $Q=Q'^{(0)}$ and for each $i,j \in Q_0$ let $R_{i,j}$ be the list of $d(\alpha)$ where $\alpha$ runs over the arrows in $Q'^{(-1)}_1$ starting at $i$ and ending at $j$.
It turns out (see Lemma \[l:preprojB\] below) that for any $m \geq 2$, the $(m+1)$-Calabi-Yau completion of $B(Q,R)$ is a Ginzburg dg-algebra of a graded quiver with homogeneous superpotential of degree $2-m$ whose construction is described below.
\[con:QW\] Let $(Q,R,m)$ be a triple where $(Q,R)$ is as in Construction \[con:dgQR\] and $m \geq 2$ is an integer. We construct a graded quiver ${\widetilde}{Q}$ with homogeneous superpotential $W$ of degree $2-m$ as follows.
- The set of vertices of ${\widetilde}{Q}$ equals that of $Q$;
- The set of arrows of ${\widetilde}{Q}$ consists of
- the arrows of $Q$, with their degree set to $0$;
- an arrow ${\varepsilon}_\rho \colon t(\rho) \to s(\rho)$ of degree $2-m$ for each relation $\rho \in R$;
- The superpotential $W$ is the image of the element $\sum_{\rho \in R} {\varepsilon}_\rho \rho$ in $K{\widetilde}{Q}/[K{\widetilde}{Q},K{\widetilde}{Q}]$.
We denote by $\Gamma(Q,R,m)$ the Ginzburg dg-algebra $\Gamma_{m+1}({\widetilde}{Q},W)$.
The quiver with superpotential $({\widetilde}{Q},W)$ depends on the particular choice of the sequence $R$ and not only on the two-sided ideal $(R)$ it generates in $KQ$, as we shall see in Example \[ex:nonuniq\].
For the rest of this section, we fix a triple $(Q,R,m)$ where $Q$ is a quiver, $R$ is a finite sequence of relations on $Q$ and $m \geq 2$. We denote by $({\widetilde}{Q},W)$ the graded quiver with superpotential of degree $2-m$ associated to $(Q,R,m)$ as in Construction \[con:QW\], and by $\Gamma = \Gamma(Q,R,m) = \Gamma_{m+1}({\widetilde}{Q},W)$ its Ginzburg dg-algebra. We denote the elements of $R$ by $\rho_1, \rho_2, \dots, \rho_n$ and write $|R|=n$. For simplicity, we denote by ${\varepsilon}_k$ the arrow ${\varepsilon}_{\rho_k}$ of degree $2-m$ in ${\widetilde}{Q}$ corresponding to the relation $\rho_k$, so that $W$ is the image of $\sum_{k=1}^n {\varepsilon}_k \rho_k$ modulo $[K{\widetilde}{Q},K{\widetilde}{Q}]$. Similarly, denote by $\eta_k$ the arrow $\eta_{\rho_k}$ and let $B=B(Q,R)$ be the dg-algebra of Construction \[con:dgQR\].
We start by describing the graded quiver ${\overline}{{\widetilde}{Q}}$ underlying $\Gamma$ occurring in Definition \[def:Ginzburg\].
\[l:QQarrows\] The arrows of the graded quiver ${\overline}{{\widetilde}{Q}}$, their degrees and their differentials are as given in Table \[tab:arrows\].
The description of the arrows and their degrees is evident from Definition \[def:Ginzburg\]. For the differentials, note that since none of the arrows ${\varepsilon}_k$ occur in any $\rho \in R$, we have $$\partial_{{\varepsilon}_k}W =
\partial_{{\varepsilon}_k} \Bigl(\sum_{\ell=1}^{n} {\varepsilon}_{\ell} \rho_{\ell} \Bigr) =
\partial_{{\varepsilon}_k} ({\varepsilon}_k \rho_k) = (-1)^{|{\varepsilon}_k|} \rho_k =
(-1)^m \rho_k$$ for any $1 \leq k \leq n$.
$$\begin{array}{cccl}
\text{arrow} & \text{degree} & \text{differential} \\ \hline
\alpha & 0 & 0 & (\alpha \in Q_1) \\
{\varepsilon}_k^* & -1 & (-1)^m \rho_k & (1 \leq k \leq n) \\
{\varepsilon}_k & 2-m & 0 & (1 \leq k \leq n) \\
\alpha^* & 1-m & \partial_{\alpha} W & (\alpha \in Q_1) \\
t_i & -m & d(t_i) & (i \in Q_0)
\end{array}$$
\[l:H0\] Consider the algebras $A = KQ/(R)$ and ${\widetilde}{A} = \operatorname{H}^0(\Gamma)$.
1. \[it:l:mgt2\] If $m>2$ then ${\widetilde}{A} \cong A$.
2. \[it:l:meq2\] If $m=2$ then ${\widetilde}{A}$ is a split extension of $A$.
By Lemma \[l:h0Q\], ${\widetilde}{A} \cong K{\widetilde}{Q}^{(0)}/(\partial_\alpha W :
\alpha \in {\widetilde}{Q}_1^{(2-m)})$.
If $m>2$ then ${\widetilde}{Q}^{(0)} = Q$, the arrows of ${\widetilde}{Q}^{(2-m)}$ are ${\varepsilon}_1, {\varepsilon}_2, \dots, {\varepsilon}_n$ and $\partial_{{\varepsilon}_k} W = (-1)^m \rho_k$ according to Table \[tab:arrows\]. Hence ${\widetilde}{A} \cong KQ/(\rho_1, \rho_2, \dots, \rho_n)$. This shows part .
If $m=2$ then all the arrows of the quiver ${\widetilde}{Q}$ have degree $0$, and we can think of ${\widetilde}{Q}$ as an ungraded quiver consisting of the arrows of $Q$ and the arrows ${\varepsilon}_k$ for $1 \leq k \leq n$. In other words, ${\widetilde}{Q} = {\widetilde}{Q}^{(0)}$ and $Q \subseteq {\widetilde}{Q}$ with ${\widetilde}{Q} \setminus Q = \{{\varepsilon}_1, {\varepsilon}_2, \dots, {\varepsilon}_n\}$. Moreover, ${\widetilde}{A} = K{\widetilde}{Q}/({\widetilde}{R})$ for ${\widetilde}{R} = \{\partial_\alpha W :
\alpha \in {\widetilde}{Q}_1\}$. Now $R \subseteq {\widetilde}{R}$ since $\partial_{{\varepsilon}_k} W = \rho_k$ for each $1 \leq k \leq n$. In addition, for any $\alpha \in Q_1$ the element $\partial_\alpha W = \sum_{k=1}^n \partial_\alpha ({\varepsilon}_k \rho_k)$ lies in the ideal generated by ${\varepsilon}_1, {\varepsilon}_2, \dots, {\varepsilon}_n$ in ${\widetilde}{Q}$. Part now follows from Lemma \[l:splitext\].
The next two lemmas relate the dg-algebra $B$ and the Ginzburg dg-algebra $\Gamma$. For a similar result in the case $m=2$, see [@Keller11 §6.7 and Proposition 6.8].
\[l:preprojB\] $\Gamma$ is the $(m+1)$-Calabi-Yau completion of $B$.
This is a consequence of Lemma \[l:CYcompl\]. Indeed, we may take $\Omega = Q_1$ so that ${\widetilde}{Q}_1 \setminus \Omega =
\{{\varepsilon}_1, \dots, {\varepsilon}_n\}$ and the superpotential $W = \sum_{k=1}^n
{\varepsilon}_k \rho_k$ has the required form. We only note that the differential of $\Gamma$ restricts to the path algebra of the quiver with arrows $Q_1 \cup \{{\varepsilon}^*_1, \dots, {\varepsilon}^*_n\}$ and the resulting dg-algebra is isomorphic to $B$ by mapping each arrow ${\varepsilon}^*_k$ to $(-1)^m \eta_{\rho_k}$ and sending each $\alpha \in Q_1$ to itself.
$\operatorname{H}^{-i}(\Gamma) \cong \operatorname{H}^{-i}(B)$ for any $i < m-2$.
From Table \[tab:arrows\] we see that $\Gamma^{-i} \cong B^{-i}$ for any $i < m-2$, hence $\operatorname{H}^{-i}(\Gamma) \cong \operatorname{H}^{-i}(B)$ for any $i < m-3$. The graded piece $\Gamma^{2-m}$ can be decomposed as $\Gamma^{2-m} \cong B^{2-m} \oplus F$, where the space $F$ is spanned by the elements of the form $u {\varepsilon}_k v$ where $1 \leq k \leq n$ and $u, v$ are paths in $Q$. Since the differential of each such element vanishes, one has $d(F)=0$, hence $d(\Gamma^{2-m}) \cong d(B^{2-m})$ and therefore $\operatorname{H}^{3-m}(\Gamma)
\cong \operatorname{H}^{3-m}(B)$ as well.
\[l:Qempty\] If $R$ is empty then $\operatorname{H}^{-i}(\Gamma)=0$ for any $1 \leq i \leq m-2$.
If $R$ is empty, then by Lemma \[l:QQarrows\] the arrows in the graded quiver ${\overline}{{\widetilde}{Q}}$ have degrees $0$, $1-m$ or $-m$. Hence the graded piece $\Gamma^{-i}$ vanishes for each $1 \leq i \leq m-2$ and the claim follows.
The next lemma provides a partial converse to Lemma \[l:Qempty\].
\[l:H2mR\] If $m>2$ and $R \subseteq {\mathfrak{r}}^2$, then $\dim_K \operatorname{H}^{2-m}(\Gamma) \geq
|R|$.
The graded piece $\Gamma^{2-m}$ is spanned by two types of elements:
1. \[it:typ01\] $u_1 {\varepsilon}_{k_1}^* u_2 {\varepsilon}_{k_2}^* u_3 \dots {\varepsilon}_{k_{m-2}}^* u_{m-1}$, where $u_1, u_2, \dots, u_{m-1}$ are paths in $Q$ and $1 \leq k_j \leq n$ for each $1 \leq j \leq m-2$;
2. \[it:typ02\] $u {\varepsilon}_k v$, where $u, v$ are paths in $Q$ and $1 \leq k \leq n$;
As a $K$-vector space, we may thus decompose $\Gamma^{2-m}$ into a direct sum $E \oplus E'$, where $E$ is the $n$-dimensional subspace $E = \{\lambda_1 {\varepsilon}_1 + \lambda_2 {\varepsilon}_2 + \dots + \lambda_n {\varepsilon}_n :
(\lambda_1, \lambda_2, \dots, \lambda_n) \in K^n\}$ and $E'$ is spanned by all the elements of type and those of type such that at least one of $u, v$ has positive length.
We will show that $d(\Gamma^{1-m}) \subseteq E'$ and hence no non-zero element in $E$ lies in the image of the differential $d$ acting on the graded piece $\Gamma^{1-m}$. Since $d$ vanishes on $E$, this will yield an $n$-dimensional subspace inside $\operatorname{H}^{2-m}(\Gamma)$.
If $m>3$, the graded piece $\Gamma^{1-m}$ is spanned by three types of elements:
1. \[it:typ1\] $u_1 {\varepsilon}_{k_1}^* u_2 {\varepsilon}_{k_2}^* u_3 \dots {\varepsilon}_{k_{m-1}}^* u_m$, where $u_1, u_2, \dots, u_m$ are paths in $Q$ and $1 \leq k_j \leq n$ for each $1 \leq j \leq m-1$;
2. \[it:typ2\] $u {\varepsilon}_k v {\varepsilon}_l^* w$ and $u {\varepsilon}_k^* v {\varepsilon}_l w$, where $u, v, w$ are paths in $Q$ and $1 \leq k, l \leq n$;
3. \[it:typ3\] $u \alpha^* v$, where $u, v$ are paths in $Q$ and $\alpha \in Q_1$.
If $m=3$, in addition to the elements above there is a fourth type
4. \[it:typ4\] $u {\varepsilon}_k v {\varepsilon}_l w$, where $u, v, w$ are paths in $Q$ and $1 \leq k, l \leq n$.
It suffices to prove that the differential of any of these elements belongs to $E'$. This is clear for the elements of the type since $$d(u_1 {\varepsilon}_{k_1}^* u_2 {\varepsilon}_{k_2}^* u_3 \dots {\varepsilon}_{k_{m-1}}^* u_m) =
\sum_{j=1}^{m-1} \pm u_1 {\varepsilon}_{k_1}^* u_2 \dots {\varepsilon}_{k_{j-1}}^* u_j \rho_j
u_{j+1} {\varepsilon}_{k_{j+1}}^* \dots {\varepsilon}_{k_{m-1}}^* u_m$$ and none of the arrows ${\varepsilon}_1, {\varepsilon}_2, \dots, {\varepsilon}_n$ can appear in the right hand side. This is also clear for the elements of type since their differential vanishes.
Consider an element of type . Then $d(u {\varepsilon}_k v {\varepsilon}_l^* w) = u {\varepsilon}_k v \rho_l w$, hence the differential is spanned by elements of the form $u' {\varepsilon}_k v'$ where $u', v'$ are paths in $Q$ and $v'$ has positive length. The case of $d(u {\varepsilon}_k^* v {\varepsilon}_l w)$ is similar.
Consider an element of type . Then $d(u \alpha^* v) = u (\partial_{\alpha} W) v = \sum_{k=1}^n u
\partial_\alpha ({\varepsilon}_k \rho_k) v$. Our assumption that $R \subseteq {\mathfrak{r}}^2$ implies that each $\partial_\alpha ({\varepsilon}_k \rho_k)$ is a linear combination of terms $u'' {\varepsilon}_k v''$ where at least one of the paths $u'', v''$ has positive length.
If $m>2$ and $R \subseteq {\mathfrak{r}}^2$, then $\operatorname{H}^{2-m}(\Gamma)=0$ if and only if $R$ is empty.
Consider the triangulated category ${\mathcal{C}}_{(Q,R,m)} = \operatorname{per}\Gamma(Q,R,m) / {\mathcal{D}}_{\operatorname{fd}}(\Gamma(Q,R,m))$.
\[t:mCYtilted\] Let $Q$ be a quiver, let $R$ be a finite sequence of relations on $Q$ such that the algebra $A=KQ/(R)$ is finite-dimensional and let $m>2$ be an integer. Then:
1. The category ${\mathcal{C}}= {\mathcal{C}}_{(Q,R,m)}$ is $\operatorname{Hom}$-finite and $m$-Calabi-Yau.
2. The image $T$ of $\Gamma(Q,R,m)$ in ${\mathcal{C}}$ is an $m$-cluster-tilting object with $\operatorname{End}_{{\mathcal{C}}}(T) \cong A$.
3. If $R \subseteq {\mathfrak{r}}^2$ then $\dim_K \operatorname{Hom}_{{\mathcal{C}}}(T, \Sigma^{-(m-2)} T) \geq |R|$.
Recall that $\Gamma(Q,R,m)= \Gamma_{m+1}({\widetilde}{Q},W)$ where $({\widetilde}{Q},W)$ is the graded quiver with superpotential of degree $2-m$ associated to the triple $(Q,R,m)$ as in Construction \[con:QW\]. We claim that $({\widetilde}{Q},W)$ satisfies the conditions of Theorem \[t:mCY\]. Indeed, condition holds since the degree of any arrow in ${\widetilde}{Q}$ is either $0$ or $2-m$, and condition holds since by Lemma \[l:H0\], $\operatorname{H}^0(\Gamma) \cong A$ and the algebra $A$ is assumed to be finite-dimensional.
Hence, by Theorem \[t:mCY\], the triangulated category ${\mathcal{C}}= {\mathcal{C}}_{(Q,R,m)}
= {\mathcal{C}}_{({\widetilde}{Q},W)}$ is $\operatorname{Hom}$-finite, $m$-Calabi-Yau and the image $T$ of $\Gamma = \Gamma(Q,R,m)$ under the canonical projection $\operatorname{per}\Gamma \to
{\mathcal{C}}$ is an $m$-cluster-tilting object whose endomorphism algebra is $\operatorname{End}_{{\mathcal{C}}}(T) \cong \operatorname{H}^0(\Gamma) \cong A$. The last assertion follows from Lemma \[l:snex\] and Lemma \[l:H2mR\].
The $m$-Calabi-Yau category ${\mathcal{C}}$ of Theorem \[t:mCYtilted\] depends on $Q$ and $R$ and not only on the algebra $A$. There exist quivers $Q$ and sequences of relations $R$ and $R'$ on $Q$ such that the algebras $KQ/(R)$ and $KQ/(R')$ are finite-dimensional and isomorphic but for any $m>2$ the categories ${\mathcal{C}}_{(Q,R,m)}$ and ${\mathcal{C}}_{(Q,R',m)}$ are not equivalent, see Example \[ex:nonuniq\] below.
\[rem:endproj\] Keep the notations and assumptions of Theorem \[t:mCYtilted\] and write $A$ as $A = \oplus_{i \in Q_0} P_i$ where $P_i = e_i A$ are the indecomposable projective right $A$-modules. The decomposition $\Gamma = \oplus_{i \in Q_0}
e_i \Gamma$ for $\Gamma=\Gamma(Q,R,m)$ induces a decomposition $T = \oplus_{i \in Q_0} T_i$ with $T_i$ being the image of $e_i \Gamma$ under the canonical projection to ${\mathcal{C}}$. Hence for any finitely generated projective $A$-module $P = \oplus_{i \in Q_0} P_i^{e_i}$ such that $e_i>0$ for all $i \in Q_0$ there exists an $m$-cluster-tilting object $T_P = \oplus_{i \in Q_0} T_i^{e_i}$ in ${\mathcal{C}}$ with $\operatorname{End}_{{\mathcal{C}}}(T_P) \cong \operatorname{End}_A(P)$.
The next result shows that the $m$-cluster category of any acyclic quiver can be realized as a category of the form ${\mathcal{C}}(Q,R,m)$. Recall that a quiver is *acyclic* if it has no cycles of positive length. We refer to [@Guo11 Corollary 3.4] for a related result.
\[p:vosnex\] Let $Q$ be a quiver, let $R$ be a finite sequence of relations on $Q$ such that $R \subseteq {\mathfrak{r}}^2$ and the algebra $KQ/(R)$ is finite-dimensional, and let $m>2$ be an integer. Then the following conditions are equivalent, where ${\mathcal{C}}$ denotes the $m$-Calabi-Yau category ${\mathcal{C}}_{(Q,R,m)}$ and $T$ is the canonical $m$-cluster-tilting object in ${\mathcal{C}}$.
1. \[it:p:KQ\] The quiver $Q$ is acyclic and the sequence $R$ is empty;
2. \[it:p:B0\] The dg-algebra $B(Q,R)$ is concentrated in degree $0$ and has finite total dimension;
3. \[it:p:vosnex\] $\operatorname{Hom}_{{\mathcal{C}}}(T, \Sigma^{-i} T) = 0$ for any $0 < i < m-1$;
4. \[it:p:m2\] $\operatorname{Hom}_{{\mathcal{C}}}(T, \Sigma^{-(m-2)} T) = 0$.
Moreover, if any of these equivalent conditions holds and the field $K$ is algebraically closed, then ${\mathcal{C}}$ is triangle equivalent to the $m$-cluster category of $Q$.
The equivalence of and is clear. Let $\Gamma = \Gamma(Q,R,m)$. For the implication $\Rightarrow$ , note that if $Q$ is acyclic and $R$ is empty then $\operatorname{Hom}_{{\mathcal{C}}}(T, \Sigma^{-i} T) \cong \operatorname{H}^{-i}(\Gamma) = 0$ for any $0 < i < m-1$ by Lemma \[l:snex\] and Lemma \[l:Qempty\]. The implication $\Rightarrow$ is clear. For the implication $\Rightarrow$ , note that by Lemma \[l:snex\] and Lemma \[l:H2mR\] $$\dim_K \operatorname{Hom}_{{\mathcal{C}}}(T, \Sigma^{-(m-2)} T) = \dim_K \operatorname{H}^{2-m}(\Gamma) \geq |R|,$$ hence if $\operatorname{Hom}_{{\mathcal{C}}}(T, \Sigma^{-(m-2)} T)$ vanishes $R$ must be empty and then $Q$ is acyclic by our assumption that $KQ/(R)$ is finite-dimensional.
If any of these conditions holds, then ${\mathcal{C}}$ is an algebraic $\operatorname{Hom}$-finite, $m$-Calabi-Yau triangulated category with an $m$-cluster-tilting object $T$ such that $\operatorname{End}_{{\mathcal{C}}}(T) \cong KQ$ for an acyclic quiver $Q$ and $\operatorname{Hom}_{{\mathcal{C}}}(T, \Sigma^{-i} T) = 0$ for any $0 < i < m-1$. By the characterization of higher cluster categories of Keller and Reiten [@KellerReiten08 Theorem 4.2]) if $K$ is algebraically closed, then ${\mathcal{C}}$ is triangle equivalent to the $m$-cluster-category of the quiver $Q$.
Systems of relations
--------------------
Let $Q$ be a quiver and let ${\mathfrak{r}}$ be the two-sided ideal of $KQ$ generated by the arrows of $Q$. An ideal $I$ of $KQ$ is *admissible* if there exists some $N \geq 2$ such that ${\mathfrak{r}}^N \subseteq I \subseteq {\mathfrak{r}}^2$.
Let $I$ be an ideal of $KQ$. A *system of relations for $I$* is a set $R$ of relations such that $R$, but no proper subset of it, generates $I$ as a two-sided ideal.
The following statement is well-known.
\[l:sysrel\] If $I$ is admissible then there exists a finite system of relations for $I$.
By assumption, there is some $N \geq 2$ such that ${\mathfrak{r}}^N \subseteq I$. The algebra $KQ/{\mathfrak{r}}^N$ is finite-dimensional, as it is spanned by all the paths of $Q$ of length smaller than $N$. Therefore the space $I/{\mathfrak{r}}^N$ is also finite-dimensional. For each $i,j \in Q_0$, choose a basis of $e_i (I/{\mathfrak{r}}^N) e_j$ and choose a set $R_{i,j}$ inside $e_i I e_j$ whose image modulo ${\mathfrak{r}}^N$ equals that basis. Then $I=(R)$ for the finite set $R$ given by $$R = \{\text{the paths of length $N$ in $Q$}\} \cup
\bigcup_{i,j \in Q_0} R_{i,j} .$$
If $R$ is not a system of relations for $I$ then there is a proper subset $R'$ of $R$ such that $I=(R')$. In this way we can repeatedly remove elements and still have a set generating $I$. Since $R$ is finite, this process must terminate and we eventually end with a system of relations for $I$.
Under some conditions, another approach to the construction of systems of relations involves lifting of basis elements of the space $I/(I {\mathfrak{r}}+ {\mathfrak{r}}I)$, see the discussion in [@BIKR08 §7].
\[l:fdIrrI\] If $I$ is admissible, then $I/(I {\mathfrak{r}}+ {\mathfrak{r}}I)$ is finite-dimensional.
Let $N \geq 2$ be such that ${\mathfrak{r}}^N \subseteq I$. Then ${\mathfrak{r}}^{N+1} \subseteq (I {\mathfrak{r}}+ {\mathfrak{r}}I)$ and we have an inclusion and a surjection $$KQ/{\mathfrak{r}}^{N+1} \supset I/{\mathfrak{r}}^{N+1} \twoheadrightarrow I/(I {\mathfrak{r}}+ {\mathfrak{r}}I).$$ The claim now follows since the quotient $KQ/{\mathfrak{r}}^{N+1}$ is finite-dimensional.
\[l:minrel\] Let $I$ be an ideal of $KQ$ and let $R$ be a set of relations inside $I$.
1. \[it:l:span\] If $I=(R)$ then the image of $R$ modulo $I {\mathfrak{r}}+ {\mathfrak{r}}I$ spans the vector space $I/(I {\mathfrak{r}}+ {\mathfrak{r}}I)$.
2. \[it:l:lift\] Assume that the ideal (R) is admissible and the image of $R$ modulo $I {\mathfrak{r}}+ {\mathfrak{r}}I$ spans the vector space $I/(I {\mathfrak{r}}+ {\mathfrak{r}}I)$. Then $I=(R)$.
3. \[it:l:liftbasis\] Assume that the ideal (R) is admissible and the image of $R$ modulo $I {\mathfrak{r}}+ {\mathfrak{r}}I$ is a basis of the vector space $I/(I {\mathfrak{r}}+ {\mathfrak{r}}I)$. Then $R$ is a system of relations for $I$.
For part , observe that since each $\rho \in R$ is a relation, multiplying it from the left or from the right by an element of the form $\sum_{i \in Q_0} \lambda_i e_i$ gives a scalar multiple of $\rho$.
For part , we slightly modify the argument in [@BMR06 Lemma 3.6]. Let $N \geq 2$ such that ${\mathfrak{r}}^N \subseteq (R)$ and let $x \in I$. By assumption, we can write $$\label{e:IbyR}
x = \lambda_1 \rho_1 + \dots + \lambda_n \rho_n + x'_1 r'_1 + \dots +
x'_k r'_k + r''_1 x''_1 + \dots + r''_{\ell} x''_{\ell}$$ with scalars $\lambda_1, \dots, \lambda_n \in K$, relations $\rho_1, \dots, \rho_n \in R$ and $x'_1, \dots, x'_k, x''_1, \dots, x''_{\ell} \in I$, $r'_1, \dots, r'_k, r''_1, \dots, r''_{\ell} \in {\mathfrak{r}}$. Writing each of the elements $x'_1, \dots, x'_k, x''_1, \dots, x''_{\ell}$ using and repeating this process $N$ times, we conclude that $x = \rho + r$ where $\rho \in (R)$ and $r \in {\mathfrak{r}}^N$, hence $x \in (R)$.
Finally, part follows from and . Moreover, $R$ is finite by Lemma \[l:fdIrrI\].
\[ex:minrel\] The assumption in parts and that the ideal $(R)$ is admissible cannot be dropped. For example, consider the quiver $Q$ given by $$\xymatrix{
{\bullet} \ar@(ul,dl)[]_{\alpha} \ar@(ur,dr)[]^{\beta}
}$$ and let $I=(\alpha^2-\beta \alpha \beta, \beta^2 - \alpha \beta \alpha, \alpha^2 \beta)$. One can check that ${\mathfrak{r}}^5 \subseteq I \subseteq {\mathfrak{r}}^2$ hence the ideal $I$ is admissible. The $8$-dimensional algebra $KQ/I$ is an algebra of quaternion type in the sense of Erdmann [@Erdmann90]. When $K$ is algebraically closed of characteristic $2$, this algebra is isomorphic to the group algebra of the quaternion group.
The image of $\alpha^2 \beta$ in $I {\mathfrak{r}}+ {\mathfrak{r}}I$ vanishes, as the following calculation shows: $$\alpha^2 \beta = (\alpha^2 - \beta \alpha \beta) \beta + \beta \alpha
(\beta^2 - \alpha \beta \alpha) + \beta \alpha^2 \beta \alpha \in I {\mathfrak{r}}+ {\mathfrak{r}}I,$$ hence $I/(I {\mathfrak{r}}+ {\mathfrak{r}}I)$ is spanned by the images of the elements $\alpha^2 - \beta \alpha \beta$ and $\beta^2 - \alpha \beta \alpha$. Nevertheless, the ideal $I' = (\alpha^2-\beta \alpha \beta, \beta^2 - \alpha \beta \alpha)$ is not equal to $I$. Indeed, by letting $\alpha$ and $\beta$ act on $K$ as the identity we get a one-dimensional module over $KQ/I'$ with a non-zero action of $\alpha^2 \beta$, hence $\alpha^2 \beta \not \in I'$.
Finite-dimensional algebras are $(m>2)$-CY-tilted
-------------------------------------------------
In this section we assume that the field $K$ is algebraically closed. A finite-dimensional algebra $A$ over $K$ is called *basic* if $A_A \cong P_1 \oplus \dots \oplus P_r$ where $P_1, \dots, P_r$ are representatives of the isomorphism classes of the indecomposable projective right $A$-modules.
\[t:Gabriel\] Let $A$ be a basic, finite-dimensional algebra over $K$. Then there exist a quiver $Q$ and an admissible ideal $I$ of $KQ$ such that $A \cong KQ/I$.
See [@Gabriel80 §4.3].
Let $A$ be a basic, finite-dimensional algebra. By Theorem \[t:Gabriel\], we can write $A=KQ/I$ for a quiver $Q$ and an admissible ideal $I$ of $Q$. We denote by $S_i$ the simple $A$-module corresponding to a vertex $i \in Q_0$ and consider the $A$-module $S = \bigoplus_{i \in Q_0} S_i$.
\[l:mincardR\] If $R$ is a system of relations for $I$ then $|R| \geq \dim_K \operatorname{Ext}^2_A(S,S)$.
We have $|R| \geq \dim_K I/(I {\mathfrak{r}}+ {\mathfrak{r}}I) = \dim_K \operatorname{Ext}^2_A(S,S)$ where the left inequality is a consequence of Lemma \[l:minrel\] and the right equality is [@Bongartz83 Corollary 1.1].
Note that Example \[ex:minrel\] shows that the inequality in Lemma \[l:mincardR\] can be strict.
\[t:mCYbasic\] Let $A$ be a basic finite-dimensional algebra. Then the set of pairs $(Q,R)$ consisting of a quiver $Q$ and a sequence of relations $R \subseteq {\mathfrak{r}}^2$ such that $A \cong KQ/(R)$ is not empty. For any such pair $(Q,R)$ and any integer $m>2$, the triangulated category ${\mathcal{C}}={\mathcal{C}}_{(Q,R,m)}$ is $\operatorname{Hom}$-finite, $m$-Calabi-Yau and its canonical $m$-cluster-tilting object $T$ satisfies $\operatorname{End}_{{\mathcal{C}}}(T) \cong A$ and $\dim_K \operatorname{Hom}_{{\mathcal{C}}}(T, \Sigma^{-(m-2)}T) \geq \dim_K \operatorname{Ext}^2_A(S,S)$.
The first claim is a consequence of Theorem \[t:Gabriel\] and Lemma \[l:sysrel\]. The second claim is a consequence of Theorem \[t:mCYtilted\] and Lemma \[l:mincardR\].
A finite-dimensional algebra over an algebraically closed field is $m$-CY-tilted for any $m>2$.
We can write $A \cong \operatorname{End}_{\bar{A}}(P)$ for a finite-dimensional, basic algebra $\bar{A}$ and a finitely generated projective $\bar{A}$-module $P$ containing as direct summands all the indecomposable projective $\bar{A}$-modules. Hence the claim is a consequence of Theorem \[t:mCYbasic\] and Remark \[rem:endproj\].
If $\operatorname{gldim}A \geq 2$, then for any of the $m$-Calabi-Yau categories ${\mathcal{C}}$ of Theorem \[t:mCYbasic\], the small negative extension $\operatorname{Hom}_{{\mathcal{C}}}(T, \Sigma^{-(m-2)}T)$ of the canonical $m$-cluster-tilting object $T$ cannot vanish. Had it vanished, Proposition \[p:vosnex\] would then imply that $A$ is the path algebra of an acyclic quiver, a contradiction.
Examples
--------
Our first example is similar in spirit to [@Guo11 Example 3.5].
Consider the algebra $A=KQ/I$ where $Q$ is the left quiver $$\begin{aligned}
\xymatrix@=1.5pc{
& {\bullet} \ar[dr]^{\beta} \\
{\bullet} \ar[ur]^{\alpha} \ar[dr]_{\gamma} && {\bullet} \\
& {\bullet} \ar[ur]_{\delta}
}
& &
\xymatrix@=1.5pc{
& {\bullet} \ar[dr]^{\beta} \\
{\bullet} \ar[ur]^{\alpha} \ar[dr]_{\gamma} \ar[rr]^{\eta} && {\bullet} \\
& {\bullet} \ar[ur]_{\delta}
}
& &
\xymatrix@=1.5pc{
& {\bullet} \ar[dr]^{\beta} \\
{\bullet} \ar[ur]^{\alpha} \ar[dr]_{\gamma} && {\bullet} \ar[ll]^{{\varepsilon}} \\
& {\bullet} \ar[ur]_{\delta}
}\end{aligned}$$ and $I=(R)$ for the system of relations $R=\{\alpha \beta - \gamma \delta\}$. The algebra $A$ has global dimension $2$, hence it cannot be 2-CY-tilted by [@KellerReiten07 Corollary 2.1].
Let $B=B(Q,R)$ be the dg-algebra of Construction \[con:dgQR\]. Its graded quiver is shown in the middle; the arrows $\alpha, \beta, \gamma$ and $\delta$ have degree $0$ while $\eta$ has degree $-1$ and $d(\eta)=\alpha \beta
- \gamma \delta$. Observe that the dg-algebra $B$ is quasi-isomorphic to $A \cong \operatorname{H}^0(B)$.
Let $m \geq 2$ and let $({\widetilde}{Q},W)$ be the graded quiver with homogeneous superpotential of degree $2-m$ of Construction \[con:QW\]. The graded quiver ${\widetilde}{Q}$ is shown on the right; the degrees of the arrows $\alpha, \beta, \gamma, \delta$ are $0$, that of ${\varepsilon}$ is $2-m$ and the superpotential is $W={\varepsilon}(\alpha \beta - \gamma \delta)$.
The Ginzburg dg-algebra $\Gamma = \Gamma_{m+1}({\widetilde}{Q},W)$ is the $(m+1)$-Calabi-Yau completion of $B$ (Lemma \[l:preprojB\]). Since $B$ is quasi-isomorphic to $A$ and $A$ is derived equivalent to the path algebra of the Dynkin quiver $D_4$, the Morita invariance of Calabi-Yau completions [@Keller11 Proposition 4.2] implies that $\Gamma$ is Morita equivalent to $\Gamma_{m+1}(D_4,0)$, hence by [@Guo11 Corollary 3.4] (or Proposition \[p:vosnex\]) the generalized $m$-cluster category ${\mathcal{C}}_{\Gamma}$ is triangle equivalent to the $m$-cluster category of type $D_4$.
If $m=2$ then $\operatorname{H}^0(\Gamma) \cong K{\widetilde}{Q}/({\widetilde}{R})$, where ${\widetilde}{R}=\{\alpha \beta - \gamma \delta, {\varepsilon}\alpha, \beta {\varepsilon}, {\varepsilon}\gamma,
\delta {\varepsilon}\}$. This is a cluster-tilted algebra [@BMR06] of type $D_4$, which is the relation-extension [@ABS08] of the tilted algebra $A$. If $m>2$ then $\operatorname{H}^0(\Gamma) \cong A$ and $A$ is $m$-CY-tilted by Theorem \[t:mCYtilted\].
\[ex:nonuniq\] Let $K$ be algebraically closed and consider the algebra $A=K$ whose quiver $Q$ is $\bullet$. Consider the two sequences of relations $R=\{\}$ and $R'=\{0\}$ on $Q$.
Let $m > 2$. The Ginzburg dg-algebras $\Gamma=\Gamma(Q,R,m)$ and $\Gamma'=\Gamma(Q,R',m)$ are given by the following graded quivers with differentials $$\begin{aligned}
\begin{array}{lc}
\begin{array}{l}
_{|t|=-m} \\
_{d(t)=0}
\end{array}
&
\begin{array}{c}
\xymatrix{
{\bullet} \ar@(ur,dr)[]^t
}
\end{array}
\end{array}
&&
\begin{array}{cl}
\begin{array}{c}
\xymatrix{
{\bullet} \ar@(ur,dr)[]^t \ar@(l,u)[]^{{\varepsilon}^*} \ar@(l,d)[]_{{\varepsilon}}
}
\end{array}
&
\begin{array}{l}
_{|{\varepsilon}^*|=-1,\, |{\varepsilon}|=-(m-2),\, |t|=-m} \\
_{d({\varepsilon}) = d({\varepsilon}^*) = 0} \\
_{d(t) = {\varepsilon}{\varepsilon}^* - (-1)^m {\varepsilon}^* {\varepsilon}}
\end{array}
\end{array}\end{aligned}$$
A basis for the graded piece $\Gamma'^i$ is given by $({\varepsilon}^*)^i$ if $0 < i < m-2$; $({\varepsilon}^*)^{m-2}, {\varepsilon}$ if $i = m-2$; $({\varepsilon}^*)^{m-1}, {\varepsilon}{\varepsilon}^*, {\varepsilon}^* {\varepsilon}$ if $i=m-1$ and $m>3$; and $({\varepsilon}^*)^{m-1}, {\varepsilon}{\varepsilon}^*, {\varepsilon}^* {\varepsilon}, {\varepsilon}^2$ if $i=m-1$ and $m=3$, hence one computes $$\begin{aligned}
\label{e:hiGamma}
\dim_K \operatorname{H}^{-i}(\Gamma) =
\begin{cases}
1 & \text{if $i=0$,} \\
0 & \text{if $0 < i < m$,}
\end{cases}
&&
\dim_K \operatorname{H}^{-i}(\Gamma') =
\begin{cases}
1 & \text{if $0 \leq i < m-2$,} \\
2 & \text{if $i=m-2$,} \\
2 + \delta_{3,m} & \text{if $i=m-1$.}
\end{cases}\end{aligned}$$
Let ${\mathcal{C}}={\mathcal{C}}(Q,R,m)$ and ${\mathcal{C}}'={\mathcal{C}}(Q,R',m')$ be the corresponding $m$-Calabi-Yau categories and let $T \in {\mathcal{C}}$ and $T' \in {\mathcal{C}}'$ be the canonical $m$-cluster-tilting objects. We have $\operatorname{End}_{{\mathcal{C}}}(T) \cong \operatorname{End}_{{\mathcal{C}}'}(T') \cong K$.
By Proposition \[p:vosnex\], the category ${\mathcal{C}}$ is triangle equivalent to the $m$-cluster category of $Q$. Hence the $m$-cluster-tilting objects in ${\mathcal{C}}$ are the indecomposable objects $\Sigma^j T$ for $0 \leq j < m$. Since $\operatorname{Hom}_{{\mathcal{C}}}(T, \Sigma^{-i} T) = 0$ for any $0 < i < m$, this remains true if we replace $T$ by any of the other $m$-cluster-tilting objects $\Sigma^j T$ in ${\mathcal{C}}$. On the other hand, $T'$ is an $m$-cluster-tilting object in ${\mathcal{C}}'$ with $\operatorname{Hom}_{{\mathcal{C}}'}(T', \Sigma^{-i} T') \neq 0$ for any $0 < i < m$ by , hence ${\mathcal{C}}'$ cannot be triangle equivalent to ${\mathcal{C}}$.
The previous example can be generalized as follows. A *Dynkin quiver* is a quiver obtained by orienting the edges of a Dynkin diagram of type $A_n$ ($n \geq 1$), $D_n$ ($n \geq 4$) or $E_n$ ($n=6,7,8$).
Let $Q$ be a Dynkin quiver and let $m > 2$. There exists an $m$-Calabi-Yau triangulated category ${\mathcal{C}}'$ with an $m$-cluster-tilting object $T'$ such that $\operatorname{End}_{{\mathcal{C}}'}(T') \cong KQ$ but ${\mathcal{C}}'$ is not triangle equivalent to the $m$-cluster category of $Q$.
Let ${\mathcal{C}}$ be the $m$-cluster category of $Q$. Since $Q$ is Dynkin, the category ${\mathcal{C}}$ has only finitely many indecomposable objects, hence the number of $m$-cluster-tilting objects $T$ in ${\mathcal{C}}$ such that $\operatorname{End}_{{\mathcal{C}}}(T)
\cong KQ$ is finite. Therefore there exists an integer $n$ such that $\dim_K \operatorname{Hom}_{{\mathcal{C}}}(T, \Sigma^{-(m-2)} T) < n$ for any such $T$.
Let $R = \{0, 0, \dots, 0\}$ be a sequence consisting of $n$ zero elements (it does not matter which starting and ending vertex we assign to each zero element) and let ${\mathcal{C}}' = {\mathcal{C}}_{(Q,R,m)}$. By Theorem \[t:mCYtilted\], the category ${\mathcal{C}}'$ is $\operatorname{Hom}$-finite and $m$-Calabi-Yau with an $m$-cluster-tilting object $T'$ satisfying $\operatorname{End}_{{\mathcal{C}}'}(T') \cong KQ$ and $\dim_K \operatorname{Hom}_{{\mathcal{C}}'}(T', \Sigma^{-(m-2)} T') \geq n$.
If $F \colon {\mathcal{C}}' \simeq {\mathcal{C}}$ were a triangulated equivalence, then $T=FT'$ would be an $m$-cluster-tilting object in ${\mathcal{C}}$ with $\operatorname{End}_{{\mathcal{C}}}(T) \cong KQ$ and $\dim_K \operatorname{Hom}_{{\mathcal{C}}}(T, \Sigma^{-(m-2)}T) \geq n$, a contradiction.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present some congruences modulo $p^{6-r}$ for sums of the type $\sum_{k=0}^{(p-3)/2}x^k{2k\choose k}/(2k+1)^r$, for $r=1,2,3$ where $p>5$ is a prime. This is a preliminary draft and more refined versions will follow.'
author:
- |
[Roberto Tauraso]{}\
Dipartimento di Matematica\
Università di Roma “Tor Vergata”, Italy\
[[email protected]]{}\
[http://www.mat.uniroma2.it/$\sim$tauraso]{}
title: '**More congruences from Apéry-like formulae**'
---
Introduction
============
In proving the irrationality $\zeta(3)$, Apèry [@Ap:79] mentioned the identities $$\sum_{k=1}^{\infty}{1\over k^2}{2k \choose k}^{-1}={1\over 3}\, \zeta(2),\qquad
\sum_{k=1}^{\infty}{(-1)^k\over k^3}{2k \choose k}^{-1}=-{2\over 5}\, \zeta(3)
.$$ Later Koecher [@Ko:80], and Leshchiner [@Le:81], found several analogous results for other $\zeta(r)$. In particular, in [@Le:81], the author gives several proofs of four identities. Two of them, namely (3b) and (4b), are connected with the series expansion of odd powers of the complex function $\arcsin(z)$ (see [@BC:07]): for $|z|<2$ and for any odd integer $r>0$ then $${\arcsin(z/2)^r}={r!\over 2}
\sum_{k=0}^{\infty}{{2k\choose k}\over 16^k}\cdot {z^{2k+1}\over (2k+1)}\cdot \HH_k(\{2\}^{r-1\over 2}).$$ where $$\HH_n(a_1,a_2,\dots,a_d)=
\sum_{0\leq k_1<k_2<\dots<k_r<n}\; {1\over (2k_1+1)^{a_1}(2k_2+1)^{a_2}\cdots (2k_r+1)^{a_r}},$$ with $(a_1,a_2,\dots,a_r)\in (\NN^*)^r$.
By revisiting the combinatorial proof (due to D. Zagier) presented in Section 5 in [@Le:81], it is easy to obtain the finite versions of these identities: if $r$ is a positive odd integer then $$\begin{aligned}
\label{idodd}\nonumber
\sum_{k=0}^{n-1}{{2k\choose k}\over 16^k}&
\left(\sum_{j=0}^{{r-1\over 2}}{(-1)^j\HH_k(\{2\}^j)\over (2k+1)^{r-2j}}
-{(-1)^{{r-1\over 2}}\over 4}\cdot{\HH_k(\{2\}^{{r-1\over 2}})\over (2k+1)}\right)\\
&=\sum_{k=0}^{n-1}{(-1)^k\over (2k+1)^r}+{(-1)^{{r-1\over 2}}\over 4}
\sum_{k=0}^{n-1}{{2k\choose k}\over 16^k{n+k\choose 2k+1}}\cdot{(-1)^{n-k}\HH_k(\{2\}^{{r-1\over 2}})\over (2k+1)},\end{aligned}$$ and if $r$ is a positive even integer then $$\begin{aligned}
\label{ideven}\nonumber
\sum_{k=0}^{n-1}{{2k\choose k}\over (-16)^k}&
\left(\sum_{j=0}^{{r\over 2}-1}{(-1)^j\HH_k(\{2\}^j)\over (2k+1)^{r-2j}}
+{(-1)^{{r\over 2}-1}\over 4}\cdot{\HH_k(\{2\}^{{r\over 2}-1})\over (2k+1)^2}\right)\\
&=\sum_{k=0}^{n-1}{1\over (2k+1)^r}+{(-1)^{{r\over 2}-1}\over 4}
\sum_{k=0}^{n-1}{{2k\choose k}\over (-16)^k{n+k\choose 2k+1}}\cdot{\HH_k(\{2\}^{{r\over 2}-1})\over (2k+1)^2}.\end{aligned}$$
After some preliminary results, starting from for $r=1$, and for $r=2$, that is $$\begin{aligned}
\label{iduno}
{3\over 4}\sum_{k=0}^{n-1}{{2k\choose k}\over 16^k(2k+1) }&=
\sum_{k=0}^{n-1}{(-1)^k\over (2k+1)}+{(-1)^n\over 4}
\sum_{k=0}^{n-1}{{2k\choose k}\over (-16)^k{n+k\choose 2k+1}}\cdot{1\over 2k+1},\\
\label{iddue}
{5\over 4}\sum_{k=0}^{n-1}{{2k\choose k}\over (-16)^k(2k+1)^2}&=
\sum_{k=0}^{n-1}{1\over (2k+1)^2}+{1\over 4}
\sum_{k=0}^{n-1}{{2k\choose k}\over (-16)^k{n+k\choose 2k+1}}\cdot{1\over (2k+1)^2},\end{aligned}$$ we prove that for any prime $p>5$ $$\begin{aligned}
\sum_{k=0}^{{p-3\over 2}}{{2k \choose k}\over 16^k(2k+1)}&\equiv
(-1)^{{p-1\over 2}}\left({H_{p-1}(1)\over 12}+{3\over 160}\,p^4 B_{p-5}\right)
\pmod{p^5},\\
\sum_{k=0}^{{p-3\over 2}}{{2k \choose k}\over (-16)^k(2k+1)^2}&\equiv
{H_{p-1}(1)\over 5p}+{7\over 20}\,p^3 B_{p-5}
\pmod{p^4},\end{aligned}$$ where $H_n(r)=\sum_{k=1}^{n}{1\over k^r}$ and $B_n$ denotes the $n$-th Bernoulli number. In the last section we consider the case $r=3$. These congruences confirm as conjectures in [@ZWS:11 Conjecture 5.1]. Notice that by taking the limit as $n$ goes to infinity in and we obtain $$\begin{aligned}
\sum_{k=0}^{\infty}{{2k \choose k}\over 16^k(2k+1)}&=
{4\over 3}\sum_{k=0}^{\infty}{(-1)^k\over (2k+1)}={4\over 3}\,\arctan(1)={\pi\over 3},\\
\sum_{k=0}^{\infty}{{2k \choose k}\over (-16)^k(2k+1)^2}&=
{4\over 5}\sum_{k=0}^{\infty}{1\over (2k+1)^2}={3\over 5}\,\zeta(2)={\pi^2\over 10},\end{aligned}$$ which allow us to evaluate the analogy between the finite and the infinite sum. For more results of this flavour, involving Apéry-like formulae, see for example [@PP:11; @MT:11; @Ta:10; @ZWS:11]. In the recent preprint [@PP:11b], the authors prove congruences modulo $p^{4-r}$ for sums of the general form $\sum_{k=0}^{(p-3)/2}{2k\choose k}x^k/(2k+1)^r$ for $r=1,2$ in terms of the finite polylogarithms.
Results concerning multiple harmonic sums
=========================================
We define the [*multiple harmonic sum*]{} as $$H_n(a_1,a_2,\dots,a_r)=
\sum_{0< k_1<k_2<\dots<k_r\leq n}\; {1\over k_1^{a_1}k_2^{a_2}\cdots k_r^{a_r}},$$ where $n\geq r>0$ and $(a_1,a_2,\dots,a_r)\in (\NN^*)^r$. The values of many harmonic sums modulo a power of prime $p$ are well known. Here there is a list of results that we will need later.
1. ([@Sunzh:00 Theorem 5.1]) for any prime $p>r+2$ we have $$\begin{aligned}
H_{p-1}(r)\equiv
\begin{cases}
-\frac{r(r+1)}{2(r+2)}\,p^2\,B_{p-r-2} \pmod{p^3} &\mbox{if $r$ is odd,}\\
\frac{r}{r+1}\,p\,B_{p-r-1} \pmod{p^2} &\mbox{if $r$ is even;}
\end{cases}\end{aligned}$$
2. ([@Zh:06 Theorem 3.1]) for $r,s>0$, and for any prime $p>r+s$, we have $$H_{p-1}(r,s)\equiv {(-1)^s\over r+s}{r+s\choose s} \,B_{p-r-s}\pmod{p};$$
3. ([@Zh:06 Theorem 3.5]) for $r,s,t>0$, and for any prime $p>r+s+t$ such that $r+s+t$ is odd, we have $$H_{p-1}(r,s,t)\equiv {1\over 2(r+s+t)}\left((-1)^r{r+s+t\choose r}-(-1)^t{r+s+t\choose t}\right) \,B_{p-r-s-t}\pmod{p};$$
4. ([@Ta:10 Theorem 2.1]) for any prime $p>5$, $$H_{p-1}(1)\equiv -{1\over 2}\,H_{p-1}(2)-{1\over 6}\,H_{p-1}(3)\pmod{p^5};$$
5. ([@PP:11 Lemma 3]) for any prime $p>5$, $$H_{p-1}(1,2)\equiv -3\,{H_{p-1}(1)\over p^2}+{1\over 2}\,p^2B_{p-5}\pmod{p^3};$$
6. ([@Sunzh:00 Theorem 5.2]) for any prime $p>r+4$ we have $$\begin{aligned}
H_{{p-1\over 2}}(r)\equiv
\begin{cases}
-\frac{2^r-2}{r}\,B_{p-r} \pmod{p} &\mbox{if $r>1$ is odd,}\\
{r(2^{r+1}-1)\over 2(r+1)}\,p\,B_{p-r-1} \pmod{p^2} &\mbox{if $r$ is even.}
\end{cases}\end{aligned}$$
The rest of this section is dedicated to proving some congruences involving $H_{{p-1\over 2}}(r)$ and $H_{{p-1\over 2}}(r,s)$.
\[Ldue\] Let $r,a>0$ then for any prime $p>r+2$ $$\begin{aligned}
\label{C1}
H_{p-1}(r)&\equiv H_{{p-1\over 2}}(r)+(-1)^r\sum_{k=0}^a{r-1+k\choose k} H_{{p-1\over 2}}(r+k)p^k\pmod{p^{a+1}},
\\\label{C3}\nonumber
H_{{p-1\over 2}}(r)&\equiv H_{\lfloor{p\over 4}\rfloor}(r)+
(-2)^r\sum_{k=0}^a{r-1+k\choose k} H_{{p-1\over 2}}(r+k)p^k\\
&\qquad -(-1)^r\sum_{k=0}^a{r-1+k\choose k}{H_{\lfloor{p\over 4}\rfloor}(r+k)\over 2^k}\, p^k \pmod{p^{a+1}}.\end{aligned}$$ Moreover, if $r,s>0$ such that $r+s$ is odd then for any prime $p>r+s$ $$\label{C2}
H_{{p-1\over 2}}(r,s)\equiv {B_{p-r-s}\over 2(r+s)}\left((-1)^s{r+s \choose s}+2^{r+s}-2\right)\pmod{p}.$$
The congruence follows from the identity $$H_{{p-1}}(r)=H_{{p-1\over 2}}(r)+\sum_{k=1}^{{p-1\over 2}}{1\over k^r(1-p/k)^r}.$$ Moreover the identity $$H_{{p-1\over 2}}(r)=
H_{\lfloor{p\over 4}\rfloor}(r)+2^r\left(
\sum_{k=1}^{(p-1)/2}{1\over (p-k)^r}-\sum_{k=1}^{\lfloor p/4\rfloor}({1\over (p-2k)^r}\right)$$ yields . Now we show . $$\begin{aligned}
H_{{p-1}}(r,s)&=\!\!\!\!
\sum_{1\leq i<j\leq {p-1\over 2}}{1\over i^r j^s}
+\sum_{1\leq i,j\leq {p-1\over 2}}{1\over i^r (p-j)^s}
+\sum_{1\leq j<i\leq {p-1\over 2}}{1\over (p-i)^r (p-j)^s}\\
&\equiv
H_{{p-1\over 2}}(r,s)+(-1)^s\,H_{{p-1\over 2}}(r)H_{{p-1\over 2}}(s)+(-1)^{r+s}H_{{p-1\over 2}}(s,r)\\
&\equiv
H_{{p-1\over 2}}(r,s)-H_{{p-1\over 2}}(s,r) \pmod{p}.\end{aligned}$$ By the stuffle product property $$H_{{p-1\over 2}}(r,s)+H_{{p-1\over 2}}(s,r)+H_{{p-1\over 2}}(r+s)=H_{{p-1\over 2}}(r)H_{{p-1\over 2}}(s)\equiv 0\pmod{p}.$$ Therefore $$2H_{{p-1\over 2}}(r,s)\equiv H_{p-1}(r,s)-H_{{p-1\over 2}}(r+s)\pmod{p}$$ and by applying (ii) and (vi) we obtain .
\[T21\] For any prime $p>2$ $$\label{Wmezzo}
H_{{p-1\over 2}}(2)+{7\over 6}\, p\,H_{{p-1\over 2}}(3) +{5\over 8}\, p^2\,H_{{p-1\over 2}}(4)\equiv 0
\pmod{p^4}.$$
Let $m=\varphi(p^4)=p^3(p-1)$ and let $B_n(x)$ the $n$-th Bernoulli polynomial. For $r=2,3,4$, Faulhaber’s formula implies $$\begin{aligned}
\sum_{k=1}^{(p-1)/2}k^{m-r}&={B_{m-r+1}\left({p-1\over 2}\right)-B_{m-r+1}\over m-r+1}\\
&= {B_{m-r+1}\left({p-1\over 2}\right)-B_{m-r+1}\left({1\over 2}\right)\over m-r+1}+{B_{m-r+1}\left({1\over 2}\right)-B_{m-r+1}\over m-r+1}\\
&= \sum_{k=r}^m{m-r\choose k-r}{B_{m-k}\left({1\over 2}\right)\over k-r+1}\cdot \left({p\over 2}\right)^{k-r+1}
+{B_{m-r+1}\left({1\over 2}\right)-B_{m-r+1}\over m-r+1}.\end{aligned}$$ By Euler’s theorem, $H_{{p-1\over 2}}(r)\equiv \sum_{k=1}^{(p-1)/2}k^{m-r}\pmod{p^4}$. Since $m$ is even, $B_{m-k}=0$ when $m-k>1$ and $k$ is odd. Moreover, $pB_{m-k}$ is $p$-integral and $$B_{m-k}\left({1\over 2}\right)=(2^{1-m+k}-1)B_{m-k}\equiv (2^{1+k}-1)B_{m-k}\pmod{p^4}.$$ Hence $$\begin{aligned}
\sum_{r=2}^4 \alpha_r\, p^{r-2} H_{{p-1\over 2}}(r)&\equiv
{p\over 2}\left(\alpha_2 -{6\over 7}\,\alpha_3\right)B_{m-2}\left({1\over 2}\right)\\
&\qquad +{p^3\over 8}\left(\alpha_2 -3\,\alpha_3+4\alpha_4\right)B_{m-4}\left({1\over 2}\right)
\pmod{p^4}.\end{aligned}$$ The right-hand side vanishes if we let $\alpha_2=1$, $\alpha_3=7/6$, and $\alpha_4=5/8$.
For any prime $p>5$ $$\begin{aligned}
\label{Hp2}
H_{{p-1}}(2)&\equiv -2\,{H_{{p-1}}(1)\over p}+{2\over 5}\,p^3 B_{p-5}\pmod{p^4},\\
\label{H2}
H_{{p-1\over 2}}(2)&\equiv -7\,{H_{{p-1}}(1)\over p}+{17\over 10}\,p^3 B_{p-5}\pmod{p^4},\\
\label{H3}
H_{{p-1\over 2}}(3)&\equiv 6\,{H_{{p-1}}(1)\over p^2}-{81\over 10}\,p^2 B_{p-5}\pmod{p^3},\\
\label{unoduepiuunotre}
H_{{p-1\over 2}}(1,2)+p\,H_{{p-1\over 2}}(1,3)
&\equiv -{9\over 2}\,{H_{p-1}(1)\over p^2} -{49\over 20}\,\, p^2 B_{p-5}
\pmod{p^{3}}.\end{aligned}$$
The congruence follows from (iv) and (i). By and (vi) we have $$H_{{p-1\over 2}}(2)+{7\over 6}pH_{{p-1\over 2}}(3)+{31\over 4}\,B_{p-5}\,p^3\equiv 0\pmod{p^4}.$$ From and (vi), we deduce $$H_{p-1}(2)=2H_{{p-1\over 2}}(2)+2pH_{{p-1\over 2}}(3)+{66\over 5}\,B_{p-5}\,p^3\equiv 0\pmod{p^4}.$$ By solving these two congruences with respect to $H_{{p-1\over 2}}(2)$ and $H_{{p-1\over 2}}(3)$, and by using , we get the result. Now we show . We consider the identity $$H_{p-1}(1,2)=H_n(1,2)+H_n(1)\sum_{j=1}^n{1\over (p-j)^2} +\sum_{1\leq j<i\leq n}{1\over (p-i)(p-j)^2}$$ and by expanding the sums like in Lemma \[Ldue\], we get $$\begin{aligned}
H_{p-1}(1,2)&\equiv H_n(1,2)+H_n(1)(H_n(2)+2p\,H_n(3)+3p^2\,H_n(4))-H_n(2,1)\\
&\qquad -p\,(2H_n(3,1)+H_n(2,2))-p^2\,(3H_n(4,1)+2H_n(3,2)+H_n(2,3))
\pmod{p^{3}}.\end{aligned}$$ By applying the stuffle product for $H_n(1)H_n(2)$ and $H_n(1)H_n(3)$, and (i), (ii) and (vi), we obtain $$H_n(1,2)+p\,H_n(3,1)\equiv {1\over 2}\,H_{p-1}(1,2)-{1\over 2}\,H_n(3)-{27\over 4}\,p^2 B_{p-5}
\pmod{p^{3}}.$$ Thus the proof of is complete as soon as we apply (v) and .
For any prime $p>5$ $$\begin{aligned}
\nonumber
2(-1)^{{p-1\over 2}}\sum_{k=0}^{{p-3\over 2}}{(-1)^k\over 2k+1}
&\equiv \HH_{{p-1\over 2}}(1) -p\HH_{{p-1\over 2}}(2)-p^2\HH_{{p-1\over 2}}(2,1)\\\label{alts}
&\quad+p^3\HH_{{p-1\over 2}}(2,2) +p^4\HH_{{p-1\over 2}}(2,2,1) \pmod{p^{5}},
\\\label{altsb}
&\equiv -{1\over 2}\,H_{{p-1\over 2}}(1)+{11\over 16}\, H_{p-1}(1)-{57\over 1280}\, p^4 B_{p-5}
\pmod{p^{5}}.\end{aligned}$$
The following identities hold $$\sum_{k=0}^n{(-16)^k{n+k\choose 2k}\over (2k+1){2k\choose k}}
=2(-1)^{n}\sum_{k=0}^{n-1}{(-1)^k\over 2k+1}+{1\over 2n+1}\,
\quad\mbox{and}\quad
\sum_{k=0}^n{(-16)^k{n+k\choose 2k}\over (2k+1)^2{2k\choose k}}={1\over (2n+1)^2}.$$ Let $n=(p-1)/2$. Notice that for $k=0,\dots,n$ $$\label{CCC}
{(-16)^k{n+k\choose 2k}\over {2k\choose k}}=\prod_{j=0}^{k-1}\left(1-{p^2\over (2j+1)^2}\right)=
\sum_{j=0}^{k-1} (-1)^jp^{2j}\,\HH_k(\{2\}^j).$$ Thus the two identities yield $$\begin{aligned}
\HH_n(1)-p^2\HH_n(2,1)+p^4\HH_n(2,2,1)&\equiv 2(-1)^{n}\sum_{k=0}^{n-1}{(-1)^k\over 2k+1}+{1-(-1)^n 4^{p-1}{2n\choose n}^{-1}\over p}\pmod{p^{6}},\\
\HH_n(2)-p^2\HH_n(2,2)+p^4\HH_n(2,2,2)&\equiv {1-(-1)^n 4^{p-1}{2n\choose n}^{-1}\over p^2}\pmod{p^{6}},\end{aligned}$$ and, by subtracting $p$ times the second congruence from the first one, we get .
As regards , let $m=\lfloor p/4\rfloor$, then $$\begin{aligned}
(-1)^{n}\sum_{k=0}^{n-1}{(-1)^k\over 2k+1}&=
\sum_{k=1}^{n}{(-1)^{k}\over p-2k}
=\sum_{k=1}^{m}{1\over p-4k}-\sum_{k=1}^{m}{1\over p-(4k-2)}\\
&=2\sum_{k=1}^{m}{1\over p-4k}-\sum_{k=1}^{n}{1\over p-2k}
\equiv -{1\over 2}\sum_{k=0}^4 {H_m(k+1)\over 4^k}\,p^k+{1\over 2}\sum_{k=0}^4 {H_n(k+1)\over 2^k}\,p^k\pmod{p^{5}}.\end{aligned}$$ By taking a suitable linear combination of the equations for $r=1,2,4$ we find that $$\sum_{k=0}^4 {H_m(k+1)\over 4^k}\,p^k
\equiv
{3\over 2}\,H_n(1)+{13\over 16}\,pH_n(2)+{1\over 2}\,p^2H_n(3)+{143\over 512}\,p^3 H_n(4)+{1\over 8}\,p^4H_n(5)
\pmod{p^{5}}.$$ Hence $$(-1)^{n}\sum_{k=0}^{n-1}{(-1)^k\over 2k+1}\equiv
-{1\over 4}\,H_n(1)-{5\over 32}\,pH_n(2)-{1\over 8}\,p^2H_n(3)-{79\over 1024}\,p^3 H_n(4)-{1\over 32}\,p^4H_n(5)
\pmod{p^{5}},$$ and follows by applying , , and (vi).
Note that by using $\eqref{alts}$, $\eqref{altsb}$ and (iii), it can be shown that for any prime $p>5$, $$H_{{p-1\over 2}}(2,1,2)\equiv 2\,H_{{p-1\over 2}}(1,2,2)\equiv 10\,H_{{p-1\over 2}}(2,2,1)\equiv -{15\over 4}\, B_{p-5} \pmod{p}.$$ Moreover, another consequence of the previous theorem is a generalization of Morley’s congruence [@Ca:53]: for any prime $p>5$, $${(-1)^{{p-1\over 2}}\over 4^{p-1}}{p-1\choose {p-1\over 2}}\equiv
1-{1\over 4}\,p\,H_{p-1}(1)-{1\over 80}\,p^5 B_{p-5}\pmod{p^6}.$$
Proof of the main result
========================
We are finally ready to prove the congruences announced in the introduction.
\[TM\] For any prime $p>5$ we have $$\begin{aligned}
\label{mc1}
\sum_{k=0}^{{p-3\over 2}}{{2k \choose k}\over 16^k(2k+1)}&\equiv
(-1)^{{p-1\over 2}}\left({H_{p-1}(1)\over 12}+{3\over 160}\,p^4 B_{p-5}\right)
\pmod{p^5},\\
\label{mc2}
\sum_{k=0}^{{p-3\over 2}}{{2k \choose k}\over (-16)^k(2k+1)^2}&\equiv
{H_{p-1}(1)\over 5p}+{7\over 20}\,p^3 B_{p-5}
\pmod{p^4}.\end{aligned}$$
From , it follows that $$\begin{aligned}
\label{conbin}\nonumber
{{2k\choose k}\over (-16)^k{n+k\choose 2k+1}}&={(2k+1){2k\choose k}\over (-16)^k(n-k){n+k\choose 2k}}
\equiv -{2\over \left(1-{p\over 2k+1}\right)\left(1-p^2\HH_k(2)+p^4\HH_k(4)-p^6\HH_k(6)\right)}\\\nonumber
&\equiv -2\left(1+{p\over 2k+1}+
\left({1\over (2k+1)^2}+\HH_k(2)\right)\,p^2+\left({1\over (2k+1)^3}+{\HH_k(2)\over 2k+1}\right)\,p^3
\right.\\
&\qquad\left.
\left({1\over (2k+1)^4}+{\HH_k(2)\over (2k+1)^2}+\HH_k(4)+\HH_k(2,2)\right)\,p^4
\right)\pmod{p^5}.\end{aligned}$$ Therefore, by , and , we get $$\begin{aligned}
\label{a2}\nonumber
{3(-1)^n\over 2}\sum_{k=0}^{n-1}{{2k\choose k}\over 16^k(2k+1)}&\equiv
-2p\HH_{n}(2) -p^2(\HH_{n}(3)+2\HH_{n}(2,1))-p^3\HH_{n}(4)\\
&\qquad-p^4(\HH_{n}(5)+\HH_{n}(2,3)+\HH_{n}(4,1)))\pmod{p^5}.\end{aligned}$$ Now we replace the terms $\HH_n$ with the corresponding expressions involving $H_{p-1}$ and $H_n$: $$\begin{aligned}
\HH_n(r)&=H_{p-1}(r)-{H_n(r)\over 2^r},\\
\HH_n(r,s)&=\sum_{0<j<i\leq n}{1\over (2(n-i)+1)^r(2(n-j)+1)^s}=
{1\over(-2)^{r+s}}\sum_{0<j<i\leq n}{1\over j^s i^r\left(1-{p\over 2i}\right)^r\left(1-{p\over 2j}\right)^s}\\
&\equiv {1\over(-2)^{r+s}}\left(H_n(s,r)+{p\over 2}\left(r H_n(s,r+1)+s H_n(s+1,r)\right)\right.\\
&\qquad\left.+{p^2\over 4}\left({r+1\choose 2} H_n(s,r+2)+rs H_n(s+1,r+1)+{s+1\choose 2} H_n(s+2,r)\right)\right)
\pmod{p^3}.\end{aligned}$$ So the right-hand side of becomes $$\begin{aligned}
&-2pH_{p-1}(2)-p^2H_{p-1}(3)-p^3H_{p-1}(4)-p^4H_{p-1}(5)\\
&\qquad +{p\over 2}H_n(2)+{p^2\over 8}(H_n(3)+2H_n(1,2))
+{p^3\over 16}\left(H_n(4)+4H_n(1,3)+2H_n(2,2)\right)\\
&\qquad-{p^4\over 32}\left(H_n(5)+4H_n(2,3)+3H_n(3,2)+7H_n(1,4)\right) \pmod{p^5}\end{aligned}$$ Since $$H_n(2,2)={1\over 2}\left( H_n(2)^2-H_n(4)\right)\equiv -{31\over 5}\, pB_{p-5} \pmod{p^2},$$ by (i), (ii), and (vi), the above expression simplifies to $$-2p H_{p-1}(2)+{1\over 2}\,pH_n(2)+{1\over 8}\,p^2 H_n(3)
+{1\over 4}\,p^2\left(H_n(1,2)+pH_n(1,3)\right)+{513\over 320}\,p^4 B_{p-5}
\pmod{p^5}.$$ Finally we apply , , and $${H_{p-1}(1)\over 8}+{9\over 320}\,p^4 B_{p-5}\pmod{p^5}$$ which concludes our proof of .
As regards , by and , we have $$\begin{aligned}
\label{a1}\nonumber
{5}\sum_{k=0}^{n-1}{{2k\choose k}\over (2k+1)^2(-16)^k}&\equiv
4\HH_{n}(2)-2(\HH_{n}(2)+p\HH_{n}(3)+p^2(\HH_{n}(4)+\HH_{n}(2,2))\\
&\qquad +p^3(\HH_{n}(5)+\HH_{n}(2,3)))\pmod{p^4}.\end{aligned}$$ As before, after replacing the terms $\HH_n$, the right-hand side of becomes $$\begin{aligned}
&2H_{p-1}(2)-2pH_{p-1}(3)-2p^2H_{p-1}(4)-2p^3H_{p-1}(5)\\
&\qquad -{1\over 2}H_n(2)+{p\over 4}H_n(3)+{p^2\over 8}\left(H_n(4)- H_n(2,2)\right)
+{p^3\over 16}\left(H_n(5)-2H_n(2,3)-H_n(3,2)\right) \pmod{p^4}.\end{aligned}$$ By (i), (ii), and (vi), it simplifies to $$2H_{p-1}(2)-{1\over 2}H_n(2)+{1\over 4}p H_n(3) +{9\over 4}p^3B_{p-5} \pmod{p^4}$$ Finally we apply , and and we get $${H_{p-1}(1)\over p}+{7\over 4}\,p^3B_{p-5} \pmod{p^4}.$$
The case $r=3$
==============
From for $r=3$, we have $$\begin{aligned}
\label{idtre}
\sum_{k=0}^{n-1}{{2k\choose k}\over 16^k}\left({1\over (2k+1)^3 }
-{3\over 4}{\HH_k(2)\over (2k+1)}\right)&=
\sum_{k=0}^{n-1}{(-1)^k\over (2k+1)^3}-{(-1)^n\over 4}
\sum_{k=0}^{n-1}{{2k\choose k}\HH_k(2)\over (-16)^k{n+k\choose 2k+1}}\cdot{1\over 2k+1}.\end{aligned}$$ Let $n=(p-1)/2$ and let $m=\lfloor p/4\rfloor$, then, by taking a suitable linear combination of the equations for $r=3,5$ we find that $$\begin{aligned}
(-1)^{n}\sum_{k=0}^{n-1}{(-1)^k\over (2k+1)^3}
&=2\sum_{k=1}^{m}{1\over (p-4k)^3}-\sum_{k=1}^{n}{1\over (p-2k)^3}\\
&\equiv -{1\over 32}\left(
H_m(3)+ {3H_m(3)\over 4}\,p+{3H_m(4)\over 8}\,p^2\right)\\
&\qquad+{1\over 8}\left(
H_n(3)+ {3H_n(3)\over 2}\,p+{3H_n(4)\over 2}\,p^2\right)\\
&=-{1\over 64}\,H_n(3)-{3\over 16}\,pH_n(4)-{189\over 512}\,p^2H_n(5)\\
&\equiv -{3\over 32}\,{H_{n}(1)\over p^2}
+{21\over 1280}\, p^2 B_{p-5}\pmod{p^{3}}.\end{aligned}$$ Moreover, acting as in Theorem \[TM\], we obtain that $$\begin{aligned}
\sum_{k=0}^{n-1}{{2k\choose k}\HH_k(2)\over (-16)^k{n+k\choose 2k+1}}\cdot{1\over 2k+1}
&\equiv 2\HH_n(2,1)+2p\HH_n(2,2)\\
&\qquad +2p^2(\HH_n(2,3)+\HH_n(4,1)+2\HH_n(2,2,1))\\
&\equiv -{9\over 8}\,{H_{p-1}(1)\over p^2}
+{9\over 320}\, p^2 B_{p-5}\pmod{p^{3}}.\end{aligned}$$ Hence $$\label{mc3e21}
(-1)^n\sum_{k=0}^{n-1}{{2k\choose k}\over 16^k}\left({1\over (2k+1)^3}
-{3\over 4}{\HH_k(2)\over (2k+1)}\right)\equiv {3\over 16}\,{H_{p-1}(1)\over p^2}
+{3\over 320}\, p^2 B_{p-5}\pmod{p^{3}}.$$ On the other hand, since $3$ does not divide $2n+1$, by , we have $$0=\sum_{k=0}^{n-1}{n+k\choose 2k+1}{(-1)^k\over 2k+1}\equiv
\sum_{k=0}^{n-1}{{2k\choose k}\over 16^k(2k+1)}
\left(1-p^2\,\HH_k(2)+p^4\HH_k(2,2)\right)\pmod{p^{5}}.$$ Since (the proof will be given in the next version) $$\sum_{k=0}^{n-1}{{2k\choose k}\HH_k(2,2)\over 16^k(2k+1)}\equiv
(-1)^n {5\over 864}\, B_{p-5} \pmod{p},$$ by , it follows that $$\label{mc21}
\sum_{k=0}^{n-1}{{2k\choose k}\HH_k(2)\over 16^k(2k+1)}
\equiv (-1)^n\left({H_{p-1}(1)\over 12p^2}+{53\over 2160}\,p^2 B_{p-5}\right) \pmod{p^3}.$$ Finally, by , we get $$\label{mc21}
\sum_{k=0}^{n-1}{{2k\choose k}\over 16^k(2k+1)^3}
\equiv (-1)^n\left({H_{p-1}(1)\over 4p^2}+{1\over 36}\,p^2 B_{p-5}\right) \pmod{p^3}.$$ Note that the values of the corresponding infinite series are known (see [@PP:10]): $$\sum_{k=0}^{n-1}{{2k\choose k}\over 16^k(2k+1)^3}={7\pi^3\over 216}
\quad\mbox{and}\quad
\sum_{k=0}^{\infty}{{2k\choose k}\HH_k(2)\over 16^k(2k+1)}={\pi^3\over 648}.$$
[99]{} R. Apèry, [*Irrationalité de $\zeta(2)$ et $\zeta(3)$. Journees arithmétiques de Luminy*]{}, Astérisque [**61**]{} (1979), 11–13.
J. M. Borwein and M. Chamberland, [*Integer powers of $\arcsin$*]{}, Int. J. Math. Math. Sciences (2007), Art. ID 19381.
L. Carlitz, [*A theorem of Glaisher*]{}, Canadian J. Math. [**5**]{} (1953), 306–316.
M. Koecher, [*Letter (German)*]{}, Math. Intelligencer, [**2**]{} (1979/1980), no. 2, 62–64.
D. H. Leshchiner, [*Some new identities for $\zeta(k)$*]{}, J. Number Theory [**13**]{} (1981), 355–362.
S. Mattarei, R. Tauraso, [*Congruences for central binomial sums and finite polylogarithms*]{}, preprint, arXiv:1012.1308v6 (2011).
Kh. Hessami Pilehrood, T. Hessami Pilehrood, [*Series acceleration formulas for beta values*]{}, Discrete Math. Theor. Comput. Sci. [**12**]{} (2010), 223–236.
Kh. Hessami Pilehrood, T. Hessami Pilehrood, [*Congruences arising from Apéry-type series for zeta values*]{}, preprint, arXiv:1108.1893v2 (2011).
Kh. Hessami Pilehrood, T. Hessami Pilehrood, [*Congruences concerning Jacobi polynomials*]{}, preprint, arXiv:1110.5308v1 (2011).
Z. H. Sun, [*Congruences concerning Bernoulli numbers and Bernoulli polynomials*]{}, Discrete Appl. Math. [**105**]{} (2000), 193–223.
Z. W. Sun, [*On congruences related to central binomial coefficients*]{}, J. Number Theory [**131**]{} (2011), 2219–2238.
R. Tauraso, [*More congruences for central binomial coefficients*]{}, J. Number Theory [**130**]{} (2010), 2639–2649.
J. Zhao, [*Wolstenholme type theorem for multiple harmonic sums*]{}, Int. J. Number Theory [**4**]{} (2008), 73–106.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We demonstrate that the asymmetric distribution of M31 satellites cannot be produced by tides from the Milky Way as such effects are too weak. However, loosely bound associations and groups of satellites can fall into larger haloes and give rise to asymmetries. We compute the survival times for such associations. We prove that the survival time is always shortest in Keplerian potentials, and can be $\sim3$ times longer in logarithmic potentials. We provide an analytical formula for the dispersal time in terms of the size and velocity dispersion of the infalling structure. We show that, if an association of $\sim 10$ dwarfs fell into the M31 halo, its present aspect would be that of an asymmetric disk of satellites. We also discuss the case of cold substructure in the Andromeda II and Ursa Minor dwarfs.'
author:
- 'A. Bowden, N.W. Evans, V. Belokurov'
title: On Asymmetric Distributions of Satellite Galaxies
---
Introduction
============
Asymmetric distributions of satellite galaxies are seemingly common. @Mc06 found that all bar one of the known 16 satellites of M31 lie on the hemisphere of the M31 sky facing the Milky Way Galaxy. The more recent investigation of @Co13 using a larger sample has soothed worries that this was an artefact of small sample size or observational bias. @Ib13 analyzed the radial velocities of the satellites and argued that 13 of the satellites possess a coherent rotational motion, with 12 of the 13 lying on the side of M31 nearest to the Milky Way Galaxy.
For our Galaxy, there are also tantalizing hints of a similar asymmetric distributions, though our ‘fishbowl’ viewpoint makes such claims hard to evaluate. Plots of the locations of the known Milky Way satellites in the sky seem to suggest that the northern Galactic hemisphere is over-endowed compared to the southern. Most of the northern Galactic hemisphere lies on the far side of the Milky Way as judged from an M31 perspective. In other words, the Milky Way satellites are also asymmetrically distributed with a preponderance on the far side from M31.
What causes such asymmetries in satellite distributions to be set up and how do they maintain themselves? At first sight, tidal forces suggest themselves as a possibility. Suppose the Milky Way and M31 have mass $M$ and separation $D$ in the $x$-direction, then the tidal potential in the frame of the host is $$\Phi_{\rm tidal} = -\frac{GM}{\sqrt{(D-x)^2 + y^2 + z^2}}.$$ Using a Taylor expansion in the limit $x,y,z \ll D$ for a distant perturber, the tidal force is $$F_x = \frac{GM}{D} \left(\frac{1}{D} + \frac{2x}{D^2} + \frac{3x^2}{D^3} - \frac{3y^2}{2D^3} - \frac{3z^2}{2D^3}\right).$$ The first term here is $GM/D^2$. As this calculation is done in the frame of the host galaxy, this cancels with the acceleration due to the non-inertial frame. We then have several terms which are linear in $x$ and these cannot lead to an asymmetry in the satellite population. The lowest order term that can generate any asymmetry is the $3GMx^2/D^4$ term. The magnitude of this effect for a Milky Way-M31 pair is roughly $\sim 1$ kpc, as can be verified by simulations in the center of mass frame of the Milky Way and M31. This is clearly insufficient to explain the asymmetry in the population of M31 satellites.
This leads us to consider the possibility that such asymmetries arise from the initial conditions of infall of satellites. Significant fractions of satellites may be accreted from a similar direction in groups, or in loosely bound associations or clumps and this can lead to asymmetries in the satellite distributions [e.g., @Li08; @Do08]. Motivated by the example of M31, @Sh13, @Go13 and @Sa14 provide scenarios in which groups of satellite galaxies may be accreted in preferred directions. @Li11 demonstrate that in simulations of the Local Group, satellite infall is not spherical in nature and that preferred directions are observed. @Ba14 show that planar features are not uncommon in cosmological simulations, however the features they describe are transient in nature. The extent to which such associations and phase space structures can persist over a Hubble time without dispersal through phase-mixing is not immediately an obvious.
  
The Dispersal Theorem
=====================
An observable asymmetry may indicate that an initially clumpy population has not had sufficient time to fully phase mix. This suggests that we examine the dispersal of such clumps. The most natural coordinate system in which to do so is action-angles. Here, actions are conserved while angles evolve linearly with time [see e.g., @Go80] $$\theta = \theta_0 + \omega t.$$ Let us consider the dispersal of a clump in action-angle space. For our application, the clump is composed of a loose association of satellite galaxies, though the theorem we prove holds more generally. As the clump disperses, the spread in actions (which are adiabatic invariants) is conserved, whilst the spread in angles is $$\Delta\theta = \delta\theta_0 + \delta\omega t.$$ So, the dispersal of the clump in time therefore only depends on $\delta\omega$, the difference in frequencies as actions are changed by a small amount, that is, the Hessian matrix multiplied by the difference in actions. This is similar to the related problem in stream dynamics [c.f., @Sa13].
We can compute the Hessian for scale-free power-law potentials [@Ev94], which take the form: $$\Phi(r) = Ar^{\alpha},\qquad A = {v_0^2\over \alpha r_0^\alpha}.
\label{eq:powlaw}$$ Here, $v_0$ is the circular velocity at radius $r_0$. When $\alpha
=0$, this becomes the isothermal sphere, whilst when $\alpha = -1$ this is of course the Kelperian case. The outer parts of galaxy haloes lie between these two limiting cases.
In a spherical potential, there are two actions, namely the radial action $J_r$ and the azimuthal action $J_\phi$ which may be taken as the angular momentum $L$ [see e.g., @Go80]. The Hamiltonian as a function of the actions takes the form [@Wi14] $$H(L,J_r) = C (J_\phi + D J_r)^\beta,$$ Here, $\beta = 2\alpha/(\alpha+2)$ and the constants $C$ and $D$ are given in Williams et al (2014). Differentiation gives the angular frequency as $$\omega_\phi = (A\alpha)^{2/(\alpha+2)}(J_\phi + D J_r)^{(\alpha-2)/(\alpha+2)}.$$ One more differentiation gives the components of the Hessian $$\frac{\partial^2H}{\partial J_\phi\partial J_i} = (A\alpha)^{2/(\alpha+2)}
\frac{\alpha-2}{\alpha+2} a_i (J_\phi + D J_r)^{-4/(\alpha+2)},$$ where $a_i$ is $D$ for $J_r$ and $1$ for $J_\phi$. This is a decreasing function of $\alpha$ if the power-law models are normalised to the same enclosed mass. [*So, for a fixed enclosed mass, the closer the potential is to Keplerian, the larger the magnitude of the Hessian and the faster the dispersal.*]{} This is the dispersal theorem.
We may be interested in the number of orbits $n_{\rm orb}$ it takes for a clump to spread out. This means we divide out the frequency $\omega_\phi$ by the Hessian matrix, $$n_{\rm orb} \propto \omega_\phi (\frac{\partial^2H}{\partial
J_\phi\partial J_i})^{-1} = \frac{\alpha+2}{\alpha-2}
\frac{(J_\phi+DJ_r)}{a_i}.
\label{eq:norb}$$ As $D$ increases as a function of $\alpha$, it is straightforward to establish that the complete expression is an increasing function of $\alpha$. In other words, the Keplerian case is again the most efficient at mixing.
An assumption in these calculations is that the spread in angles dominates the evolution of the clump, whilst the spread in actions is small. This assumption is likely to hold good for small clumps. Therefore, it is important to test the results against simulations, which we proceed to do in the next section.
The Dispersal of Groups of Satellites
=====================================
Keplerian Case
--------------
To describe the degree of phase mixing as a clump disperses, we use the mean angular separation between galaxies in the association. This can be computed both analytically and numerically. We consider phase mixing to be complete when this value reaches $2\pi$.
To start with, we assume the simplest case of a Keplerian galactic host potential. In this potential, orbits have periods given by $$P_{\rm orb} = \frac{\pi GM}{\sqrt{2}} \epsilon^{-3/2},$$ where $\epsilon$ is the energy and $M$ is the mass of the host galaxy. Provided the orbit is not highly eccentric, then the angular phase is $$\phi \approx \frac{2\pi t}{P_{\rm orb}} = \frac{2\sqrt{2}}{GM} \epsilon^{3/2} t.$$ The magnitude of the difference in $\phi$ between two such satellites is $$(\Delta\phi)^2 = A (\epsilon_1^{3/2} - \epsilon_2^{3/2})^2 = A (\epsilon_1^{3} + \epsilon_2^{3} - 2\epsilon_1^{3/2}\epsilon_2^{3/2}),$$ where $A = 8t^2/(G^2M^2)$. To find the mean, we integrate over phase space co-ordinates after multiplying by a probability distribution. We model the group of satellites as a (truncated) Gaussian with dispersions in position and velocity $\sigma_r$ and $\sigma_v$ around means $\mu_r$ and $\mu_v$ $$\begin{aligned}
P &=& \frac{1}{(2\pi\sigma_r\sigma_v)^2} \exp-\bigg(
\frac{(r_1 - \mu_r)^2}{2\sigma_r^2} \\ &+& \frac{(r_2 -
\mu_r)^2}{2\sigma_r^2} + \frac{(v_1 - \mu_v)^2}{2\sigma_v^2} +
\frac{(v_2 - \mu_v)^2}{2\sigma_v^2} \bigg).\nonumber
\label{eq:gaussclump}\end{aligned}$$ We thus evaluate numerically $$\overline{(\Delta\phi)^2} = A \int P (\epsilon_1^{3} + \epsilon_2^{3}
- 2\epsilon_1^{3/2}\epsilon_2^{3/2}) dr_1 dr_2 dv_1 dv_2.
\label{eq:analytic}$$ truncating our probability distribution at $4\sigma$.
![Plot showing the number of orbits required for a Gaussian clump to phase mix as a function of power-law index for scale-free spherical potentials. The mean position and rotational velocity of the clump is the same as in Fig. 1. Different lines correspond to different orbital eccentricities for the clump. We observe that the efficiency of phase mixing is an increasing function of power-law index $\alpha$, with the Keplerian case as the fastest dispersing. This vindicates the dispersal theorem.[]{data-label="fig:fixapoorbs"}](fixapoorbs){width=".45\textwidth"}
1.5cm
Logarithmic Case
----------------
For arbitrary spherical potentials, such as the flat rotation curve or logarithmic case, we must evaluate $$\overline{(\Delta\phi)^2} = C \int P \left(\frac{1}{T^2_{\phi_1}} +
\frac{1}{T^2_{\phi_2}} - \frac{1}{T_{\phi_1}T_{\phi_2}}\right) dr_1
dr_2 dv_1 dv_2.$$ Here, $T_\phi$ is the azimuthal period, given by $T_\phi = 2\pi
T_r/|\Delta\phi|$, with $$T_r = 2 \int^{r_2}_{r_1} \frac{dr}{\sqrt{2[E - \Phi(r)] - L^2/r^2}},$$ and $$\Delta\phi = 2L \int^{r_2}_{r_1} \frac{dr}{r^2\sqrt{2[E - \Phi(r)] - L^2/r^2}}.$$ The validity of these approximations can be tested via direct comparison to numerical simulations. The simulations involve drawing test particles from a 6-D Gaussian centered on the mean position and velocity. These particles are integrated forwards in time in the host galactic potential, assuming no self-gravity, and their mean angular separation evaluated.
Fig. \[fig:numvsana\] shows the time taken in the Keplerian and logarithmic cases for the mean $\Delta\phi$ to reach $2\pi$, for a range of values of $\sigma_r$ and $\sigma_v$. In each case, the clump is centered on an orbit with $\mu_r = 200$ kpc and $\mu_v = 150$ kms$^{-1}$, as might be appropriate for the outer parts of a large galaxy like M31. The enclosed mass of the host galaxy is $M = 2.2
\times 10^{12} M_\odot$, comparable to estimates of the mass of M31 [see e.g., @Ev00; @Di14]. The analytic approximation reproduces well the shape of the contours, however it slightly underestimates the time taken for a clump to disperse. The likely sources of the discrepancy are the approximations we have made, particularly the assumption that the angular velocity of the particles is constant in time, allowing us to perform our integrals in terms of the angular period. Further, we modeled our probability distribution as Gaussian in the magnitude $r$ and $v$. This requires the implicit assumption that our perturbations about the mean clump velocity are in the direction of motion, and perturbations about the position are radial. In actuality, we should perform an integral over the 6-D phase space. However, perturbations to the relevant quantities or $r$ and $v$ from a six-dimensional Gaussian are dominated by the contributions from the radial position terms and the direction of motion velocity terms.
Fig. \[fig:fixapoorbs\] shows, for the same clump mean velocities, the number of orbits required for the mean $\Delta \phi$ to reach $2\pi$ for a variety of power-law potentials as given in eq (\[eq:powlaw\]). This demonstrates that the number of orbital periods needed for dispersal is an increasing function of $\alpha$, as suggested by our calculation in Section 2. Note that the simulations have really pushed beyond the frozen action approximation used there. Comparing the Keplerian and Logarithmic potentials – which are the likely bounding cases for the outer parts of galaxy haloes – an identical clump takes more orbital periods to disperse in the logarithmic halo by a factor of $\sim 3$.
One assumption underlying all the work so far is that the disk of satellites lies in a nearly spherical potential. This is a natural choice, as orbits in spherical potentials lie in a plane. However, the disk of satellites could lie in the equatorial plane of an axisymmetric potential, or in one of the two stable principal planes of a triaxial potential [@Bo13]. We have verified by numerical calculations that the dispersal time and the number of orbits for $\Delta \phi$ to reach $2\pi$ is close to the spherical case in these instances.
Whilst our numerical simulations are simple in nature, they display qualitatively similar results to more complex simulations such as those in @Li11, where Local Group substructure is shown to persist over long timescales.
Applications to Galactic Substructure
=====================================
In order to apply our results to the expected survival timescales of observed substructure, we desire a simple analytic formula for estimating a lower limit (Keplerian case) for the time taken for a clump of satellites to disperse. From eq (\[eq:norb\]), we see that $$n_{\rm orb} \propto \frac{J}{|\delta J|},$$ where $\Delta J$ is the spread in the sum of the actions. Using our expression for the Hamiltonian, we can show $$\delta J = \frac{GM \delta\epsilon}{2\sqrt{2}\epsilon^{\frac{3}{2}}},$$ where $\epsilon$ is reduced energy. Using $$\delta\epsilon = -\frac{GM\delta r}{r^2} - v\delta v,$$ we fit the functional form for the dispersal time $$T \propto n_{\rm orb}P_{\rm orb} = A\frac{(GM)^{\frac{2}{3}} P_{orb}^{\frac{1}{3}}}
{\sqrt{ (\frac{GM \sigma_r}{ \mu^2_r})^2 + (\mu_v \sigma_v)^2}}.$$ To determine the parameter $A$, we simultaneously fit five different orbits each with a 10x10 grid of $\sigma_r$ between $2-20$ kpc and $\sigma_v$ between $2-20$ kms$^{-1}$. The values were calculated analytically using the method described in Section 3.1. The orbital parameters for $\mu_r$ /kpc, $\mu_v$ /kms$^{-1}$ were $[200,150]$,$[100,250]$,$[100,180]$,$[150,180]$,$[150,150]$. Using Monte Carlo methods, we recover a best fit value of $A = 0.738$. The mean absolute fractional error for the fit is $0.065$, and the maximum is $0.326$. The high maximum errors occurred in cases two and three, when $\mu_r = 100$ kpc and $\sigma_r = 20$ kpc. This is when we expect our approximations to break down – the Taylor expansion for calculating $\delta\omega$ is no longer valid in this regime. We only expect our formula to be applicable in the case when $\Delta J/J$ is much less than unity, ruling out for example the dispersion of structure on highly radial orbits.
The M31 satellites
------------------
One application of this formula is for the dispersal of a clump of satellites falling into M31. We take an M31 mass of $2 \times 10^{12}
M_{\odot}$. Let us assume that an association of dwarfs of total mass of $10^{10} M_\odot$ fell into the M31 halo and was the progenitor of $\sim 10$ dwarf galaxies in the present-day disc of satellites. Note that the total mass in the association is rather modest, as it is smaller by a factor of $\sim 10$ as compared to nearby associations of dwarf galaxies studied observationally by [@Tu06]. We take $\sigma_r$ as $20$ kpc, which, for a loosely bound clump with potential energy shape factor of $~0.3$, fixes $\sigma_v$ at $20$ kms$^{-1}$. In order to mimic the properties of the M31 galaxies, we place the association on an orbit with apocenter at $200$ kpc and pericentre at $100$ kpc. This gives an apocenter velocity of $~170$ kms$^{-1}$. In these units, this orbit has a period of $3.93$ kpckm$^{-1}$s. Our analytic formula then gives a lower limit of $~10$ Gyr for the clump to disperse. This is of order the age of the universe, and we would expect this timescale to be roughly 3 times larger in a logarithmic halo.
In other words, if an association of $\sim 10$ dwarfs galaxies was accreted by M31 sometime within the last $~10$ Gyr, then we expect the association to persist rather easily and the present-day distribution would retain memory of the initial conditions. The satellites would not be fully phase-mixed and would be spread out to lie in an asymmetric disk, much as observed.
Substructure in And II and UMi
------------------------------
Another application is to the cold substructure in dwarf spheroidal galaxies, such as Andromeda II and Ursa Minor. These are probably the the result of the engulfment of smaller dwarfs and clusters within the larger dwarf spheroidal itself.
The cold stellar stream in @Am14 is at a distance of roughly $1.3$ kpc with a dispersion of order $0.1$ kpc. The velocity dispersion is given as less than $3$ kms$^{-1}$. The enclosed mass interior to the stream is $2.5$x$10^8 M_{\odot}$. This gives a circular velocity at $1.3$ kpc of $28.8$ kms$^{-1}$. This orbit has a period of $0.3$ kpc km$^{-1}$s. A lower limit to the dispersal timescale for these values is $450$ Myr. Even for a logarithmic halo, this would suggest that, if there is a detectable angular asymmetry in the substructure, then it is relatively young. The data at this time certainly indicate that the angular extent of the stream is not a full $2\pi$ around Andromeda II, although this may be partly a consequence of the spectroscopic selection function.
The cold clump in Ursa Minor was discovered by @Kl03, who suggested that it may be a cluster that had fallen into the dwarf galaxy and was in the process of dissolving. As the velocity offset between the cold clump and Ursa Minor is small, this suggests that the orbit is either radial or is circular and viewed almost face-on. @Kl03 investigated the former possibility and showed that the dissolution timescale depended on the properties of the central parts of the potential. Here, we follow the latter possibility, namely that that cold substructure is on a near circular orbit at 150 pc from the center. The mass enclosed within the orbit is $1.25 \times 10^7
M_\odot$. The size and the velocity dispersion of the clump are given by @Kl03 as $0.012$ pc and $0.5$ kms$^{-1}$ respectively, which gives a phase-mixing timescale of $\sim 130$ Myr. This is comparable to the disruption timescale in a cusped potential found by @Kl03 if the clump is on a radial orbit. Note however we have probably pushed our dispersal formula to beyond its formal domain of applicability in studying disruption of such a tightly bound clump.
Conclusions
===========
Our main conclusion is a general formula for the timescale for phase-mixing of orbits in nearly spherical potentials. Survival times are always shortest in Keplerian potentials and can be $\sim$ 3 times longer in logarithmic potentials. We use the formula to demonstrate that the asymmetric distribution of the disk of satellites of M31 can be maintained without undue difficulty in nearly spherical potentials. This is because the outer parts of such galaxies have enormously long memories. If a loose association of dwarf galaxies were accreted together, then the substructure can survive for timescales longer than the age of the Universe. The accreted dwarfs would still show appreciable spatial asymmetries, much as the M31 satellites do.
AB thanks STFC for financial support. We thank the referee for a careful and thoughtful reading of the manuscript.
Amorisco, N. C., Evans, N. W., & van de Ven, G. 2014, , 507, 335
Bahl, H., & Baumgardt, H. 2014, , 438, 2916
Bowden, A., Evans, N. W., & Belokurov, V. 2013, , 435, 928
Conn, A.R., et al., 2013, ApJ, 766 120
Diaz, J., Koposov, S., Irwin, M.J., Belokurov, V., Evans, N.W., 2014, MNRAS, 443, 1688
D’Onghia, E., & Lake, G. 2008, , 686, L61
Evans, N. W., 1994, , 267, 333
Evans, N. W., & Wilkinson, M. I. 2000, , 316, 929
Goerdt, T., & Burkert, A. 2014, MNRAS, in press, arXiv:1307.2102
Goldstein, H., 1980, Classical Mechanics, Addison-Wesley
Kleyna, J. T., Wilkinson, M. I., Gilmore, G., & Evans, N. W. 2003, , 588, L21
Ibata, R.A., Lewis, G.F., Conn, A.R., et al. 2013, , 493, 62
Li, Y.-S., & Helmi, A. 2008, , 385, 1365
Libeskind, N. I., Knebe, A., Hoffman, Y., et al. 2011, MNRAS, 411, 1525
McConnachie, A. W., Irwin, M. J. 2006, MNRAS, 365, 902
Sadoun, R., Mohayaee, R., & Colin J. 2014, MNRAS, 442, 160
Sanders, J. L., & Binney, J. 2013, , 433, 1813
Shaya, E. J., & Tully, R. B. 2013, , 436, 2096
Tully, R. B., Rizzi, L., Dolphin, A. E., et al. 2006, , 132, 729
Williams, A., Evans, N.W., Bowden, A. 2014, MNRAS, 442, 1405
| {
"pile_set_name": "ArXiv"
} |
[**Nel’s category theory based\
\
differential and integral Calculus,\
\
or\
\
Did Newton know category theory ?**]{}\
\
Elemér E Rosinger\
Department of Mathematics\
and Applied Mathematics\
University of Pretoria\
Pretoria\
0002 South Africa\
[email protected]\
\
[**Abstract**]{}\
In a series of publications in the early 1990s, L D Nel set up a study of non-normable topological vector spaces based on methods in category theory. One of the important results showed that the classical operations of derivative and integral in Calculus can in fact be obtained by a rather simple construction in categories. Here we present this result in a concise form. It is important to note that the respective differentiation does [*not*]{} lead to any so called generalized derivatives, for instance, in the sense of distributions, hyperfunctions, etc., but it simply corresponds to the classical one in Calculus.\
Based on that categorial construction, Nel set up an infinite dimensional calculus which can be applied to functions defined on non-convex domains with empty interior, a situation of great importance in the solution of partial differential equations.\
\
[**1. The Setup**]{}\
The presentation follows mainly Nel \[1\], where further details as well as the proofs can be found. For convenience, a few basic concepts needed from Category Theory are recalled in short in the Appendix.\
Let $I \subset \mathbb{R}$ be any compact interval and let $E$ be a real Banach space. We consider the Banach space structure on the set ${\cal C} ( I, E )$ of all continuous functions $f : I \longrightarrow E$, induced by the ${\cal L}^\infty$ norm $||~||_\infty$.\
Our first aim is to define a [*derivative*]{} for functions $f$ in a suitable subset of ${\cal C} ( I, E )$, and do so by means of a simple construction in categories. Among such functions $f$ for which a derivative can be defined by categorial means are those called [*paths*]{} in $E$, and they are characterized as follows.\
Let us consider in ${\cal C} ( I, E )$ the subset\
(1.1) $ Path~{\cal C} ( I, E ) $\
of [*paths*]{} in $E$, given by all $f \in {\cal C} ( I, E )$, for which there exists $h_f \in
{\cal C} ( I \times I, E )$, such that\
(1.2) $ f(y) - f(x) ~=~ (y-x)h_f(x,y),~~~ x,~ y \in I $\
We give several useful properties of such paths. For that purpose, similar with ${\cal C}
( I, E )$, we first define a Banach space structure on the set ${\cal C} ( I \times I, E )$ of all continuous functions $h : I \times I \longrightarrow E$.\
Let us now consider in ${\cal C} ( I \times I, E )$ the following [*closed*]{} subspace\
(1.3) $ ad~{\cal C} ( I \times I, E ) $\
given by all $h \in {\cal C} ( I \times I, E )$, such that\
(1.4) $ (y-x)h(x,y) + (z-y)h(y,z) + (x-z)h(z,x) ~=~ 0,~~~ x,~ y,~ z \in I $\
It is easy to see that, with the notation in (1.2), we have\
(1.5) $ \begin{array}{l}
\forall~~~ f \in path~{\cal C} ( I, E ) ~: \\ \\
~~~~~~~ *)~~~ h_f \in ad~{\cal C} ( I \times I, E ) \\ \\
~~~~ **)~~~ h_f ~~\mbox{is {\it unique}}
\end{array} $\
Furthermore, if $f \in path~{\cal C} ( I, E )$, then the classical derivative of Calculus, namely, $f^\prime ( x )$ exists for every $x \in I$, and\
(1.6) $ \begin{array}{l}
f^\prime ( x ) ~=~ h_f ( x, x ) ~=~ \\
~~~~=~ \lim_{~y \to x,~z \to x,~ y \neq z}~ ( f ( z ) - f ( y ) ) /
( z - y ),~~~ x \in I
\end{array} $\
This property, however, will [*not*]{} be used in the categorial definition of the derivative given in the sequel, and instead, it will follow from that definition.\
\
[**Remark**]{}\
1) Regardless of (1.6), in the [*purely algebraic*]{} condition (1.4) which defines the elements in $ad~{\cal C} ( I \times I, E )$, there is [*no*]{} involvement of any kind of operation of limit, derivation, or even merely of a division with a quantity which may eventually vanish. And as we shall see in the sequel, that purely algebraic condition (1.4) is [*sufficient*]{} in order to define and effectively compute both the derivative and the integral by using a simple [*diagonal*]{} construction in a suitable category.\
2) However, the idea in the rather unusual definition (1.3), (1.4) is, in view of (1.2), (1.6), quite straightforward. Namely, for every function $f : I \longrightarrow E$, we can define the function\
(1.7) $ h : I \times I \longrightarrow E $\
by the expression\
(1.8) $ h ( x, y ) ~=~ \begin{array}{|r} ~~
( f ( y ) - f ( x ) ) / ( y - x )~~~ \mbox{if}~~ x,~ y \in I,~~ x \neq y \\ \\
\mbox{an arbitrary element}~ a_x \in E~~~ \mbox{if}~~ x = y \in I
\end{array} $\
and then, clearly, $h$ satisfies (1.2) and (1.4).\
The issue, however, is whether $h$ is continuous on $I \times I$, more precisely, whether in (1.8) one can choose $a_x \in E$, with $x \in I$, so that $h$ becomes continuous. And obviously, this can be done, if and only if $h \in ad~{\cal C} ( I \times I, E )$, in which case $f$ is a path and $h = h_f$, thus (1.6) holds.
$\Box$\
Let us note two further useful properties of the functions in\
$ad~{\cal C}
( I \times I, E )$.\
First, it is easy to see that every $h \in ad~{\cal C} ( I \times I, E )$ is [*symmetric*]{}.\
Second, given two compact intervals $I=[a,b],~ J=[b,c] \subset \mathbb{R}$ together with functions $h \in ad~{\cal C} ( I \times I, E ),~ k \in ad~{\cal C} ( J \times J, E )$. If $h(b,b)=k(b,b)$, then there exists a [*unique*]{} function $l \in ad~{\cal C} ( ( I \cup J )
\times ( I \cup J ), E )$, such that $l_{I \times I} = h$ and $l_{J \times J} = k$.\
\
[**2. A Few Related Categories**]{}\
Let us denote by ${\cal B}an$ the category of Banach spaces and their continuous linear mappings.\
For every compact interval $I \subset \mathbb{R}$, we obtain two [*covariant functors*]{}\
(2.1) $ \begin{array}{l}
{\cal C} ( I, - ) : {\cal B}an \longrightarrow {\cal B}an \\ \\
{\cal C} ( I \times I, - ) : {\cal B}an \longrightarrow {\cal B}an
\end{array} $\
acting as follows. Given two Banach spaces $E$ and $F$ and a continuous linear mapping $u :
E \longrightarrow F$, then\
(2.2) $ \begin{array}{l}
{\cal C} ( I, u ) : {\cal C} ( I, E ) \ni f \longmapsto u \circ f \in
{\cal C} ( I, F ) \\ \\
{\cal C} ( I \times I, u ) : {\cal C} ( I \times I, E ) \ni h \longmapsto
u \circ h \in {\cal C} ( I \times I, F )
\end{array} $\
Consequently, in view of (1.1), we can define the [*covariant functor*]{}\
(2.3) $ ad~{\cal C} ( I \times I, - ) : {\cal B}an \longrightarrow {\cal B}an $\
as the restriction of the functor ${\cal C} ( I \times I, - ) : {\cal B}an \longrightarrow
{\cal B}an$, namely, for $u : E \longrightarrow F$ as above, we have\
(2.4) $ ad~{\cal C} ( I \times I, u ) : ad~{\cal C} ( I \times I, E ) \ni h \longmapsto
u \circ h \in ad~{\cal C} ( I \times I, F ) $\
Finally, let us define the [*diagonal evaluation*]{} mapping\
(2.5) $ ed_{~I,~E} : ad~{\cal C} ( I \times I, E ) \longrightarrow {\cal C} ( I, E ) $\
by\
(2.6) $ ( ed_{~I,~E}~ ( h ) ) ( x ) ~=~ h ( x, x ) $\
for $h \in ad~{\cal C} ( I \times I, E )$ and $x \in I$.\
The basic result is given in\
[**Theorem on Diagonal Evaluation (L D Nel)**]{}\
For every compact interval $I \subset \mathbb{R}$, the mapping $ed_{I,-}$ is a [*natural isomorphism*]{} and [*isometry*]{} between the functors $ad~{\cal C} ( I \times I, - )$ and ${\cal C} ( I, - )$.\
\
[**3. The Derivative**]{}\
Given a compact interval $I \subset \mathbb{R}$ and a Banach space $E$, the [*derivative*]{} $D$ is defined as the mapping\
(3.1) $ D : Path~{\cal C} ( I, E ) \longrightarrow {\cal C} ( I, E ) $\
where for $f \in Path~{\cal C} ( I, E )$ we have, see (1.5)\
(3.2) $ D f ~=~ ed_{~I,~E}~ ( h_f ) $\
Here the [*important*]{} point to note is that, as mentioned at pct. 1 in the Remark above, this definition of derivative does [*not*]{} refer in any way to limits, derivation, or even merely to division with a quantity which may eventually vanish. Yet, as seen next, this concept of derivative [*recovers*]{} the usual derivative in Calculus for functions $f : I
\longrightarrow E$, see Nel \[1\] for further details.\
Indeed, in view of (2.6) and (1.6), we have\
(3.3) $ ( D f ) ( x ) ~=~ h_f ( x, x ) ~=~ f^\prime ( x ),~~~ x \in I $\
Therefore it follows that for every path $f \in Path~{\cal C} ( I, E )$ and $x \in I$, we have\
(3.4) $ ( D f ) ( x ) ~=~ \lim_{~y \to x,~z \to x,~ y \neq z}~ ( f ( z ) - f ( y ) ) /
( z - y ) $\
which is the classical definition of derivative in Calculus.\
Here we give a simple example which shows that the derivative in (3.1), (3.2) does [*not*]{} lead to any concept of derivation is a generalized sense, like for instance, distributional. Instead, it corresponds to the classical derivative in Calculus.\
Let $I = [ - 1, 1]$, $E = \mathbb{R}$, and let us take the function $f \in {\cal C} ( I, E )$ defined by $f = x_{+}$, that is, $f(x)=0$, for $x \in [-1,0]$, while $f(x)=x$, for $x \in
[0,1]$.\
Then $f \in {\cal C} ( I, E )$, however $f$ is [*not*]{} a path, therefore, the derivation in (3.1) - (3.3) [*cannot*]{} be applied to it.\
Indeed, assume that $f$ is a path. Then (1.4) gives $h_f \in ad~{\cal C} ( I \times I, E )$ such that $f(y)-f(x)=(y-x)h_f(x,y)$, for $x,~y \in I$. Consequently (3.4), (3.5) imply that
$$\lim_{~x \to 0,~x \neq 0}~ f(x) / x ~=~ h_f(0,0)$$
which is absurd, since the limit in the left hand term does not exist.\
It follows that the that categorial definition of derivative in (3.1), (3.2) rather corresponds to the usual derivative of ${\cal C}^1$-smooth functions.\
Let us further illustrate the above categorial concept of derivative in (3.1), (3.2) in the case of some simple functions.\
A function $f \in {\cal C} ( I, E )$ is called [*affine*]{}, if and only if, for suitable $c, d \in E$, we have\
(3.5) $ f ( x ) = c + x d,~~~ x \in I $\
We denote by $Aff~( I, E )$ the subset of all affine functions $f \in {\cal C} ( I, E )$.\
A function $f \in {\cal C} ( I, E )$ is called [*polygonal*]{}, if and only if it is piece-wise affine on $I$. We denote by $Poly~( I, E )$ the subset of all polygonal functions $f \in {\cal C} ( I, E )$.\
It is well known, Dieudonne, that the set of polygonal functions $Poly~( I, E )$ is [*dense*]{} in ${\cal C}( I, E )$.\
Now for every affine function $f \in Aff~( I, E )$ let us define the corresponding [*averaging*]{} function\
(3.6) $ av_f \in {\cal C} ( I \times I, E ) $\
by\
(3.7) $ av_f ( x, y ) ~=~ c + ( ( x + y ) / 2 ) d,~~~ x, y \in I $\
Then obviously\
(3.8) $ av_f \in ad~{\cal C} ( I \times I, E ),~~~ ed_{~I,~E}~ ( av_f ) ~=~ f $\
It follows therefore, Nel \[1, p.53\], that one can in a piece-wise manner extend the averaging function in (3.6), (3.7) to all polygonal functions $f \in Poly~( I, E )$, and obtain relations similar to those in (3.8).\
Moreover, since the polygonal functions are dense in ${\cal C} ( I, E )$, it follows that one can further extend the averaging function in (3.6), (3.7) to the whole of ${\cal C} ( I, E )$, with the preservation of the relations in (3.8).\
\
[**4. The Integral**]{}\
Based on the categorial concept of derivative in (3.1), (3.2), we can define as well a notion of [*integral*]{}. Indeed, such an integral is given by the mapping\
(4.1) $ \int : I \times I \times {\cal C} ( I, E ) \ni ( a, b, f ) ~\longmapsto~
\int_a^b f ~\stackrel{def}=~ ( b - a )\, av_f ( a, b ) \in E $\
Now, the following two properties result :\
For every path $f \in Path~( I, E )$ and every $a, b \in I$, we have the classical Newton-Leibniz formula of Calculus\
(4.2) $ \int^b_a f^\prime ~=~ f ( b ) - f ( a ) $\
where the derivative $f^\prime$ is given by the categorial concept in (3.1), (3.2).\
Conversely, for every given $f \in {\cal C} ( I, E )$ and $a \in I$, let us define $F \in
{\cal C} ( I, E )$ by\
(4.3) $ F ( x ) ~=~ \int^x_a f,~~~ x \in I $\
Then $F$ is a path in ${\cal C} ( I, E )$, that is, $F \in Path~( I, E )$, and with the categorial concept of derivative in (3.1), (3.2), we have\
(4.4) $ F^\prime ~=~ f $\
\
[**Appendix**]{}\
One way to define a [*category*]{} ${\cal A}$ is as being a directed graph which has the following properties :\
C1) The vertices are called [*objects*]{}, while the directed edges are called [*morphisms*]{}, and given two objects $A$ and $B$, the totality of morphisms $A \stackrel{f}{\longrightarrow}
B$ is a set denoted by $Hom_{\cal A}(A,B)$.\
C2) For every object $A$ there exists a morphism $A \stackrel{id_A}{\longrightarrow} A$ with property C3.2) below.\
C3) For every two morphisms $A \stackrel{f}{\longrightarrow} B \stackrel{g}{\longrightarrow}
C$, there exists a well defined morphism $A \stackrel{g \circ f }{\longrightarrow} C$, such that :\
C3.1) the operation $\circ$ is associative\
C3.2) for every morphism $A \stackrel{f}{\longrightarrow} B$, we have $f \circ
id_A = f = id_B \circ f$\
\
A [*functor*]{} $F$ acts between two categories ${\cal A}$ and ${\cal X}$, namely\
$~~~~~~~~~~~~~~~~~~~~~~~~ {\cal A} \stackrel{F}{\longrightarrow} {\cal X} $\
as follows :
$\setlength{\unitlength}{0.2cm}
\thicklines
\begin{picture}(90,32)
\put(5,29){$A$}
\put(25,26){$f$}
\put(8,29.5){\vector(1,0){36.5}}
\put(46,29){$B$}
\put(5.5,27){\vector(0,-1){20}}
\put(3,16){$F$}
\put(46.5,27){\vector(0,-1){20}}
\put(48,16){$F$}
\put(2,4){$X = F ( A )$}
\put(13,4.5){\vector(1,0){29}}
\put(26.5,16){$F$}
\put(23,7){$\xi = F ( f )$}
\put(44,4){$Y = F ( B )$}
\put(25.5,24){\vector(0,-1){14.5}}
\put(60,29){${\cal A}$}
\put(60.5,27){\vector(0,-1){20}}
\put(60,4){${\cal X}$}
\put(62,16){$F$}
\end{picture}$
That is, $F$ takes objects $A$ in category ${\cal A}$ into objects $X = F ( A )$ in category ${\cal X}$.\
Further, $F$ takes morphisms $A \stackrel{f}{\longrightarrow} B$ in category ${\cal A}$ into morphisms $X = F ( A ) \stackrel{\xi = F ( f )}{\longrightarrow} Y = F ( B )$ in category ${\cal X}$.\
The above actions of the functor $F$ have the properties :\
For every object $A$ in category ${\cal A}$, we have in category ${\cal X}$ the relation\
F1) $ F ( id_A ) ~=~ id_{F ( A )} $\
Also, for every two morphisms $A \stackrel{f}{\longrightarrow} B \stackrel{g}
{\longrightarrow} C$ in category ${\cal A}$, we have in category ${\cal X}$ the relation\
F2a) $ F ( g \circ f ) ~=~ F ( g ) \circ F ( f ) $\
in which case the functor $F$ is called [*covariant*]{}.\
If on the other hand, we have in category ${\cal X}$ the relation\
F2b) $ F ( g \circ f ) ~=~ F ( f ) \circ F ( g ) $\
then the functor $F$ is called [*contravariant*]{}.\
A [*natural transformation*]{} $~\nu~$ acts between two functors $F$ and $G$ which, on their turn, act between the same two categories ${\cal A}$ and ${\cal X}$, namely
$\setlength{\unitlength}{0.2cm}
\thicklines
\begin{picture}(90,22)
\put(10,20){${\cal A}$}
\put(10.5,18){\vector(0,-1){14}}
\put(12,10.5){$F$}
\put(10,1){${\cal X}$}
\put(15,11.5){\vector(1,0){20}}
\put(24,13){$\nu$}
\put(37,10.5){$G$}
\put(40,20){${\cal A}$}
\put(40.5,18){\vector(0,-1){14}}
\put(40,1){${\cal X}$}
\end{picture}$\
And the action of $\nu$ is as follows :\
N1) To every object $A$ in category ${\cal A}$ corresponds in category ${\cal X}$ a morphism\
$\setlength{\unitlength}{0.2cm}
\thicklines
\begin{picture}(90,21)
\put(8,20){$F(A)$}
\put(9.6,18){\vector(0,-1){15}}
\put(6,10.5){$\nu_A$}
\put(8,0){$G(A)$}
\end{picture}$\
N2) Every morphisms $A \stackrel{f}{\longrightarrow} B$ in category ${\cal A}$ generates the [*commutative diagram*]{} in category ${\cal X}$, given by\
$\setlength{\unitlength}{0.2cm}
\thicklines
\begin{picture}(90,62)
\put(10,60){$A$}
\put(13,60.5){\vector(1,0){20}}
\put(35,60){$B$}
\put(50,60){${\cal A}$}
\put(23,57.5){$f$}
\put(8,40){$F(A)$}
\put(14,40.5){\vector(1,0){18}}
\put(22,42){$F(f)$}
\put(34,40){$F(B)$}
\put(50,40){${\cal X}$}
\put(23.5,55){\vector(0,-1){9.5}}
\put(50.5,58){\vector(0,-1){15}}
\put(52,50){$F$}
\put(10,38){\vector(0,-1){15}}
\put(6,30){$\nu_A$}
\put(36,38){\vector(0,-1){15}}
\put(38,30){$\nu_B$}
\put(18,30){$\mbox{commutes}$}
\put(10,0){$A$}
\put(13,0.5){\vector(1,0){20}}
\put(35,0){$B$}
\put(50,0){${\cal A}$}
\put(23,2.5){$f$}
\put(8,20){$G(A)$}
\put(14,20.5){\vector(1,0){18}}
\put(22,17.5){$G(f)$}
\put(34,20){$G(B)$}
\put(50,20){${\cal X}$}
\put(23.5,6){\vector(0,1){9.5}}
\put(50.5,3){\vector(0,1){15}}
\put(52,10){$G$}
\put(54.5,50.5){\line(1,0){7}}
\put(61.5,50.55){\line(0,-1){40}}
\put(61.5,10.5){\vector(-1,0){7}}
\put(63,30){$\nu$}
\end{picture}$\
We recall that $\nu_A$ and $\nu_B$ above are morphisms in the category ${\cal X}$.\
A natural transformation $~\nu~$ between categories ${\cal A}$ and ${\cal X}$ is called a [*natural isomorphism*]{}, if and only if for every object $A$ in category ${\cal A}$, the morphism\
$\setlength{\unitlength}{0.2cm}
\thicklines
\begin{picture}(90,22)
\put(8,20){$F(A)$}
\put(9.6,18){\vector(0,-1){15}}
\put(6,10.5){$\nu_A$}
\put(8,0){$G(A)$}
\end{picture}$\
is an isomorphism in category ${\cal X}$, that is, there exists a morphism $G(A)
~\stackrel{\mu_A}\longrightarrow~ F(A)$ in category ${\cal X}$, such that\
$~~~~~~~~~~ \mu_A \circ \nu_A ~=~ id_{F(A)},~~~\nu_A \circ \mu_A ~=~ id_{G(A)} $\
\
[99]{}
Dieudonne J : Foundations of Modern Analysis. Acad. Press, New York, 1969
Herrlich H, Strecker G E : Category Theory. Heldermann, Berlin, 1973
Nel L D \[1\] : Differential calculus founded on an isomorphism. Applied Categorical Structures, Vol. 1, 1993, 51-57
Nel L D \[2\] : Foundational isomorphism in continuous differentiation theory. Applied Categorical Structures, 1977, Vol. 5, 1-9
Nel L D \[3\] : Infinite dimensional calculus allowing for nonconvex domains with empty interior. Mh. Math., Vol. 110, 1990, 145-166
Nel L D \[4\] : Newton’s method and Frobenius-Dieudonné theorem in nonnormable spaces. Applied Categorical Structures, Vol. 5, 1997, 205-216
Nel L D \[5\] : Nonlinear existence theorems in nonnormable analysis. In (Eds. Herrlich H, Porst H-E) Category Theory at Work, Heldermann, Berlin, 1991, 343-365
Nel L D \[6\] : Effective categories of complete separated spaces. In Recent Developments General TOpology and its Applications. International Conference in Memory of Felix Hausdorff. Mathematical Research 67, Berlin 1992, 242-251
Nel L D \[7\] : Categorical Differential Calculus for Infinite Dimensional Spaces. Cahiers de Topologie et Geometrie Differentielle Categoriques, Vol. 29-4, 1988, 257-286
Nel L D \[8\] : Introduction to Categorical Methods. Parts 1 and 2. Carleton-Ottawa Mathematical Lecture Note Series, No. 11, August 1991
Nel L D \[9\] : Introduction to Categorical Methods. Part 3. Carleton-Ottawa Mathematical Lecture Note Series, No. 12, March 1992
Monadi A, Nel L D \[10\] : Holomorphy in Convergence Spaces. Applied Categorical Structures, Vol. 1, 1993, 233-245
Monadi A, Nel L D \[11\] : Holomorphic Maps on Domains with Empty Interior. Quaestiones Mathematicae, Vol. 19, 1996, 409-415
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study polynomial systems whose equations have as common support a set ${\mathcal{C}}$ of $n+2$ points in ${{\mathbb{Z}}}^n$ called a circuit. We find a bound on the number of real solutions to such systems which depends on $n$, the dimension of the affine span of the minimal affinely dependent subset of ${\mathcal{C}}$, and the rank modulo $2$ of ${\mathcal{C}}$. We prove that this bound is sharp by drawing so-called dessins d’enfant on the Riemann sphere. We also obtain that the maximal number of solutions with positive coordinates to systems supported on circuits in ${{\mathbb{Z}}}^n$ is $n+1$, which is very small comparatively to the bound given by the Khovanskii fewnomial theorem.'
address: |
Laboratoire de Mathématiques\
Université de Savoie\
73376 Le Bourget-du-Lac Cedex\
France
author:
- Frederic Bihan
title: 'Polynomial systems supported on circuits and dessins d’enfants'
---
Introduction and statement of the main results {#introduction-and-statement-of-the-main-results .unnumbered}
==============================================
The support of a multivariate polynomial is the set of exponent vectors of the monomials appearing in the polynomial. A polynomial system of $n$ polynomial equations in $n$ variables is supported on ${\mathcal{A}}\subset {{\mathbb{Z}}}^n$ if ${\mathcal{A}}$ is the common support of each polynomial in the system. A theorem due to Kouchnirenro gives that the number of (simple) solutions in the complex torus $({{\mathbb{C}}}^*)^n$ to a generic polynomial system supported on ${\mathcal{A}}$ is the volume $v({\mathcal{A}})$ of the convex hull of ${\mathcal{A}}$, normalized so that the unit cube $[0,1]^n$ has volume $n!$.
Here, we consider generic polynomial systems with real coefficients, and are interested in their numbers of real solutions in the real torus $({{\mathbb{R}}}^*)^n$. Contrary to the complex case, the number of real solutions depends on the coefficients of the system, and one of the most important question is to find a sharp upper bound (for related results, see [@BeBiS; @LRW03; @So; @So1] for example). A trivial bound is given by $v({\mathcal{A}})$, the number of complex solutions. Another bound due to Khovanskii depends only on the number $n$ of variables and the cardinality of the support ${\mathcal{A}}$. The above Kuchnirenko result shows immediately that such a bound belongs to the “real world”, that is, cannot exist as a bound on the number of complex solutions. The Khovanskii bound is $$2^n 2 ^{\binom{|{\mathcal{A}}|}{2}}
\cdot (n+1)^{|{\mathcal{A}}|},$$ and gives an easy way to construct supports ${\mathcal{A}}$ for which the Kuchnirenko bound $v({\mathcal{A}})$ is not sharp. The Khovanskii bound can be stated alternatively without the term $2^n$ as a bound on the number of [*positive solutions*]{}, which are solutions with positive coordinates.
Since we are interested only in the solutions in the real torus, we have the freedom to translate the support ${\mathcal{A}}$ by an integral vector and also to choose a basis of the lattice ${{\mathbb{Z}}}^n$. A translation by an integral vector corresponds to multiplying each polynomial of the system by a monomial, while a change of basis for the lattice ${{\mathbb{Z}}}^n$ corresponds to a monomial change of coordinates (with coefficients $1$) for the torus. These operations do not change the number of positive, real and complex solutions. In particular, we will always assume that $0 \in {\mathcal{A}}$. The [*rank modulo 2*]{} of ${\mathcal{A}}$ is the rank of the reduction modulo $2$ of the matrix obtained by putting in columns the non zero elements of ${\mathcal{A}}$. We will denote it by $rk \, (\bar{{\mathcal{A}}})$.
If ${\mathcal{A}}\subset {{\mathbb{Z}}}^n$ and $|{\mathcal{A}}| <n+1$, then $v({\mathcal{A}})=0$, so that supports with such numbers of elements are not interesting. The first case where $v({\mathcal{A}})$ is different from $0$ arises when $|{\mathcal{A}}|=n+1$, and the convex hull of ${\mathcal{A}}$ is an $n$-dimensional simplex. This case is easy. The number of positive solutions to such systems is at most $1$, so that the number of real solutions is at most $2^n$. In fact, the maximal number of real solutions is $2^{n-rk \, (\bar{{\mathcal{A}}})}$ ([@S], see also Lemma \[L:simplex\]).
The goal of this paper is to obtain similar results in the first non trivial case, when the support is [*a circuit*]{}. A circuit is a set ${\mathcal{C}}\subset {{\mathbb{Z}}}^n$ of $n+2$ points which affinely span ${{\mathbb{R}}}^n$. It contains an unique minimal affinely dependent subset, and we denote by $m({\mathcal{C}})$ the dimension of the affine span of this subset. If $m({\mathcal{C}})=n$, that is, if any proper subset of ${\mathcal{C}}$ is affinely independent, then ${\mathcal{C}}$ is called a [*non degenerate circuit*]{}. We have $m({\mathcal{C}})=1$ for example when three points of ${\mathcal{C}}$ lie on a same line. In [@BeBiS], the authors consider circuits ${\mathcal{C}}$ (and also more general supports called near circuits) such that the index of ${{\mathbb{Z}}}{\mathcal{C}}$ in ${{\mathbb{Z}}}^n$ is odd. This corresponds to circuits ${\mathcal{C}}\subset {{\mathbb{Z}}}^n$ with $rk \, (\bar{{\mathcal{C}}})=n$. They obtain the upper bound $2m({\mathcal{C}})+1$ on the number of real solutions and also prove that this bound is sharp among systems supported on circuits ${\mathcal{C}}\subset {{\mathbb{Z}}}^n$ with $rk \, (\bar{{\mathcal{C}}})=n$. In particular, this gives the sharp bound $2n+1$ which can be attained only with non degenerate circuits. We generalize the results in [@BeBiS] to arbitrary values of the rank modulo $2$. The first result gives sharp bounds on the number of positive solutions. Obviously, multiplying by $2^n$ such a bound gives a bound on the number of real solutions, which is sharp among systems supported on circuits ${\mathcal{C}}$ with $rk \, (\bar{{\mathcal{C}}})=0$.
[**Theorem A.**]{}
*The number of positive solutions to a generic real polynomial system supported on a circuit ${\mathcal{C}}\subset {{\mathbb{Z}}}^n$ is at most $$m({\mathcal{C}})+1.$$ Therefore, the number of real solutions to a generic real polynomial system supported on a circuit ${\mathcal{C}}\subset {{\mathbb{Z}}}^n$ is at most $$2^n(m({\mathcal{C}})+1).$$ Moreover, the first bound, and thus the second bound, is sharp. Namely, for any integer $m$ such that $1 \leq m \leq n$, there exist a circuit ${\mathcal{C}}\subset {{\mathbb{Z}}}^n$ with $m({\mathcal{C}})=m$, and a system supported on ${\mathcal{C}}$ which has $m+1$ positive solutions.*
As a consequence, the number of positive solutions to a generic system supported on a circuit in ${{\mathbb{Z}}}^n$ is at most $$n+1$$ while its number of real solutions is at most $$2^n \cdot (n+1)$$ and these bounds are sharp and can be attained only with non degenerate circuits.
In particular, as in the case of supports forming a simplex, the Khovansky bound is far from being sharp among systems supported on circuits. Theorem A follows from Proposition \[P:boundpositive\] and Theorem \[T:positivesharp\].
[**Theorem B.**]{}
*The number, $N$, of real solutions to a generic real polynomial system supported on a circuit ${\mathcal{C}}\subset {{\mathbb{Z}}}^n$ satisfies*
1. If $ {{rk \, ({\bar {\mathcal{C}}})}} \leq m({\mathcal{C}})$, then $$N \leq 2^{n-{{rk \, ({\bar {\mathcal{C}}})}}} \cdot \left(m({\mathcal{C}})+ {{rk \, ({\bar {\mathcal{C}}})}}+1\right).$$
2. if $ {{rk \, ({\bar {\mathcal{C}}})}} \geq m({\mathcal{C}})$, then $$N \leq 2^{n-{{rk \, ({\bar {\mathcal{C}}})}}} \cdot \left(2m({\mathcal{C}})+1\right).$$
Moreover, both these bounds are sharp. Namely, let $n,m,R$ be integers such that $1 \leq m \leq n$ and $0 \leq R \leq n$. If $R \leq m$ (resp. $R \geq m$), there exist a circuit ${\mathcal{C}}\subset {{\mathbb{Z}}}^n$ with $m({\mathcal{C}})=m$, ${{rk \, ({\bar {\mathcal{C}}})}}=R$, and a system supported on ${\mathcal{C}}$ whose number of real solutions is the bound in (1) (resp. the bound in (2)).
The bounds in Theorem B look like the bound $2^{n-rk \, (\bar{{\mathcal{A}}})}$ when ${\mathcal{A}}$ forms an $n$-dimensional simplex. However, there is an essential difference in that the sharp bounds in Theorem B do not provide the maximal number of real solutions to systems supported on a [*given*]{} circuit. Theorem B follows from Theorem \[T:upper bound\] and Theorem \[T:sharp\].
We consider the eliminant defined in [@BeBiS] of a system supported on a circuit ${\mathcal{C}}\subset {{\mathbb{Z}}}^n$. Assuming that ${\mathcal{C}}=\{0,\ell e_1,w_1,\dotsc,w_n\} \subset {{\mathbb{Z}}}^n$, this eliminant is a univariate polynomial of the form $$f(x)=x^{\lambda_0}\prod_{i=1}^t (g_i(x))^{\lambda_i} - \prod_{i=t+1}^{\nu} (g_i(x))^{\lambda_i},$$ where each polynomial $g_i$ has the form $g_i(x)=a_i+b_ix^{\ell}$ with $a_i$ and $b_i$ non zero real numbers. Here, the coefficients $a_i$ and $b_i$ come from the coefficients of the system while the other numbers are determined by ${\mathcal{C}}$. The number $\nu$ is equal to $m({\mathcal{C}})$ if $0$ and $\ell e_1$ belong to the minimal affinely dependent subset of ${\mathcal{C}}$. If $rk \, (\bar{{\mathcal{C}}}) \neq 0$, then the integer $\ell$ can be assumed to be odd. We choose to distinguish the case $rk \, (\bar{{\mathcal{C}}}) \neq 0$ from the case $rk \, (\bar{{\mathcal{C}}})=0$ for which the number of real solutions is given by the number of positive ones. When $rk \, (\bar{{\mathcal{C}}})=n$, it is proved in [@BeBiS] that the real solutions to the system are in bijection with the real roots of $f$ via the projection onto the first coordinate axis. In general, the real solutions project onto the real roots of $f$ which satisfy sign conditions involving products of polynomials $g_i$. Moreover, such a real root is the image of $2^{n-{{rk \, ({\bar {\mathcal{C}}})}}}$ solutions to the system, at least when $\ell$ is odd. The positive solutions project bijectively onto the positive roots of $f$ at which all $g_i$ are positive. We determine the above sign conditions, and prove the upper bounds in Theorem A and B using essentially Rolle’s theorem.
Writing the eliminant as $f=P-Q$, we see that the number of real (and positive) solutions is closely related to the arrangement of the roots of $P$, $Q$ and $f$, their multiplicities being determined by ${\mathcal{C}}$. We consider the rational function $$\phi=f/Q=P/Q-1: {{\mathbb{C}}}P^1 \rightarrow {{\mathbb{C}}}P^1.$$ The roots of $P$, $Q$ and $f$ are the inverse images of $-1$, $\infty$ and $0$, respectively. These roots lie on the graph $$\Gamma=\phi^{-1}({{\mathbb{R}}}P^1) \subset {{\mathbb{C}}}P^1$$ and we see $\Gamma$ as an example of so-called [*dessin d’enfant*]{}. We use then the observation made in [@Br; @O] (see also [@NSV]) that polynomials $P$, $Q$ and $f=P-Q$ with prescribed arrangement and multiplicities of their real roots can be constructed drawing so-called [*real rational graphs*]{} on ${{\mathbb{C}}}P^1$. The sharpness of the bounds in Theorem A and B is then proved by means of such graphs.
The paper is organized as follows. In Section 1, we define the eliminant $f$ of a system and explain the relation between its real roots and the real (resp. positive) solutions to the system. In Section 2, we prove the upper bounds in Theorem A and B, while in Section 3 we introduce real rational graphs and achieve constructions proving the sharpness of these upper bounds.
[**Acknowledgements.**]{} I’m greatly indebted to Ilia Itenberg and Frank Sottile for their support during these last years. Thanks to Frank Sottile and Erwan Brugallé for useful comments on the first version of this paper. Thanks also to Benoit Bertrand and J. Maurice Rojas for their interest in this work.
Elimination {#S:elimination}
===========
A real polynomial system supported on a circuit $${\mathcal{C}}=\{w_{-1},w_0,\cdots,w_n\} \subset {{\mathbb{Z}}}^n$$ is called [*generic*]{} if it has $v({\mathcal{C}})$ solutions in ${({{\mathbb{C}}}^*)}^n$. This forces each solution to be a simple solution. We are interested in the real solutions to such a system, by which we mean solutions in the real torus ${({{\mathbb{R}}}^*)}^n$.
Translating ${\mathcal{C}}$ by an integral vector and choosing a basis for the lattice ${{\mathbb{Z}}}^n$ if necessary, we may assume without loss of generality that $w_{-1}=0$, and $w_0=\ell e_1$ for some positive integer $\ell$. Perturbing slightly the system, and using Gaussian elimination, it becomes equivalent to a system of the form $$\label{E:reduced}
S \; : \qquad x^{w_i}=g_i(x_1) \;, \quad i=1,\dotsc,n,$$ where $g_i(x_1)=a_i+b_ix_1^{\ell}$ with $a_i,b_i$ non zero real numbers for $i=1,\dotsc,n$ and $g_1,\dotsc,g_n$ have distinct roots. Reordering the vectors $w_1,\dotsc,w_n$ if necessary, the affine primitive relation on $\{0,e_1,w_1,\dotsc, w_n\}$ can be written as $$\label{E:primitiveaffinerelation}
\lambda_0 \, e_1+\sum_{i=1}^t \lambda_iw_i = \sum_{i=t+1}^{\nu} \lambda_i w_i,$$ where $1 \leq \nu \leq n$, $\lambda_0,\dotsc,\lambda_{\nu}$ are coprime integers, $\lambda_0 \geq 0$ and $\lambda_1,\dotsc,\lambda_{\nu} >0$. Note that here we could have $t=0$ or $t=\nu$ so that one of the two sums collapses to $0$. The integer $$\delta:= \lambda_0+\sum_{i=1}^t \lambda_i-\sum_{i=t+1}^{\nu} \lambda_i$$ is (up to sign) the coefficient of $w_{-1}=0$ in (the term $\delta w_{-1}$ does not appear in since it is $0$). Multiplying by $\ell$ gives an affine relation on ${\mathcal{C}}$. Set $$\label{E:minimal dimension}
m({\mathcal{C}}):=\nu -\chi(\lambda_0=0)-\chi(\delta=0)$$ where $\chi(Y)$ is the boolean truth value of $Y$: $\chi(Y)=1$ if $Y$ is true and $\chi(Y)=0$ otherwise. Then $m({\mathcal{C}})+2$ is the number of non zero coefficients in the primitive affine relation on ${\mathcal{C}}$, so that $m({\mathcal{C}})$ is the dimension of the affine span of the minimal affinely dependent subset of ${\mathcal{C}}$. The case $m({\mathcal{C}})=n$ arises when ${\mathcal{C}}$ is a non degenerate circuit, that is, when any proper subset of ${\mathcal{C}}$ is affinely independent.
Define the [*eliminant*]{} of $S$ to be the following univariate polynomial $$\label{E:eliminant}
f(x_1)=x_1^{\lambda_0}\prod_{i=1}^t (g_i(x_1))^{\lambda_i} - \prod_{i=t+1}^{\nu} (g_i(x_1))^{\lambda_i}.$$
From and , it follows immediately that if $x=(x_1,\dotsc,x_n)$ is a solution to $S$ then $x_1$ is a root of the eliminant $f$.
Write each vector $w_i=l_ie_1+v_i$ with $v_i \in {{\mathbb{Z}}}^{n-1}$. We get the relations $$\label{E:primitiveaffinerelationv}
\sum_{i=1}^t \lambda_iv_i = \sum_{i=t+1}^{\nu} \lambda_i v_i$$ and $$\label{E:primitiveaffinerelationl}
\lambda_0+\sum_{i=1}^t \lambda_il_i = \sum_{i=t+1}^{\nu}\lambda_i l_i.$$
The fact that $\lambda_0,\dotsc,\lambda_{\nu}$ are coprime and gives immediately that $\lambda_1,\dotsc,\lambda_{\nu}$ are also coprime, so that the relation is in fact the primitive affine relation on $\{0,v_1,\dotsc,v_n\} \subset {{\mathbb{Z}}}^{n-1}$. In particular, at least one integer among $\lambda_1,\dotsc,\lambda_{\nu}$ should be odd.
\[P:Systemreduced\] Assume that $\lambda_j$ is odd with $j \in \{1,\dotsc,{\nu}\}$. Then $x=(x_1,\dotsc,x_n)$ is a real solution to the system $S$ if and only if $x_1$ is a real root of $f$ and $(x_2,\dotsc,x_{n})$ is a real solution to the system $$y^{v_i}=g_i(x_1)/x_1^{l_i} \;, \quad i \in \{1,\dotsc,n\} \setminus \{j\}.$$
For simplicity, we will often assume that $\lambda_1$ is odd and thus apply Proposition \[P:Systemreduced\] with $j=1$. Define $$B:=(v_1,
v_2, \dotsc , v_{n}),$$ the $n-1$ by $n$ matrix whose columns are the vectors $v_1,v_2,\dotsc,v_{n}$. Let $\bar{B} \in M_{n-1,n}({\mathbb{Z}}/2)$ be the the reduction modulo $2$ of $B$ and $rk \,(\bar{B})$ be the rank of $\bar B$.
\[P:real\] The number of real solutions to $S$ is equal to $2^{n-1-rk \,(\bar{B})}$ times the number of real roots $r$ of $f$ subjected to the sign conditions $${(g_1(r)/r^{l_1})}^{\epsilon_1} \cdots {(g_{n}(r)/r^{l_{n}})}^{\epsilon_{n}}>0$$ for any $\epsilon=(\epsilon_1,\dotsc,\epsilon_n)$ such that $\bar{\epsilon} \in \mbox{Ker} \, \bar{B}$.
It is worth noting that the number of sign conditions can be reduced to $n-rk \,(\bar{B})$ using a basis of $\mbox{Ker} \, \bar{B}$.
Before giving the proof, we present the following well-known lemma (see [@S]).
\[L:simplex\] Suppose that $0,u_1,\dotsc,u_n \in {{\mathbb{Z}}}^n$ form an $n$-dimensional simplex , and let $\bar{U} \in M_n({\mathbb{Z}}/2)$ denote the matrix whose columns contains in this order the reductions modulo $2$ of $u_1,\dotsc,u_n$. Then the number of real solutions in the torus to the system $$\label{E:intilasyst1}
x^{u_i}=c_i,\;, \quad i=1,\dotsc,n,$$ where $c_i$ is a non zero real number, is
1. $0$ or $2^{n-rk(\bar{U})}$ if the volume of the simplex is even,
2. $1$ if this volume is odd.
Moreover, we have $2^{n-rk(\bar{U})}$ real solutions in the first case if and only if $$c_1^{\epsilon_1}\cdots c_n^{\epsilon_n} >0$$ for any $(\epsilon_1,\dotsc,\epsilon_n) \in {{\mathbb{Z}}}^n$ whose reduction modulo $2$ belong to the kernel of $\bar{U}$.
Note that the case $2)$ is in fact redundant since if the volume of the simplex is odd then the kernel of $\bar{U}$ is reduced to $0$, hence $2^{n-rk(\bar{U})}=1$, and the sign conditions are empty.
[**Proof.**]{} There exist two matrices $L,R \in GL_{n}({\mathbb{Z}})$ and integers $a_1,\dotsc,a_{n}$ such that $LUR$ is the diagonal matrix $D=\mbox{diag} \, (a_1,\dotsc,a_{n})$. Let $\tilde{u}_1,\dotsc,\tilde{u}_{n}$ be the columns vectors, in this order, of the matrix $LU$. The matrix $L$ provides the multiplicative change of coordinates $\tilde{y_i}= \prod_{j=1}^{n} y_j^{(L^{-1})_{ji}}$, $i=1,\dotsc,n$, of the complex torus. This change of coordinates transforms our system to the system $\tilde{y}^{\tilde{u}_i}=c_i \;, \quad i=1,\dotsc,n$, which has the same number of real (and complex) solutions. Multiplication on the right by $R$ transforms this system to the system
$$\label{E:diagsyst}
\tilde{y_i}^{a_i}=d_i \quad \mbox{with} \quad d_i:=\prod_{j=1}^{n}c_j^{R_{ji}}\;, \quad i=1,\dotsc,n.$$
This system is equivalent to the previous one since $\tilde{y_i}^{a_i}=\prod_{j=1}^{n}(\tilde{y}^{\tilde{u}_j})^{R_{ji}}$ and $R \in GL_{n}({\mathbb{Z}})$. In particular, the initial system and the system have the same number of real solutions. This number of real solutions is $0$ if $d_i<0$ for some even $a_i$, or $2^{n-rk(\bar{D})}$ if $d_i>0$ for all even $a_i$. The conclusion is now obvious noting that $rk(\bar{D})=rk(\bar{U})$ and that the reductions modulo $2$ of the vector columns $(R_{1i},\dotsc,R_{n,i})$ for which $a_i$ is even generate the kernel of $\bar{U}$. $\Box$
[**Proof of Proposition \[P:real\]**]{} Assume that $\lambda_1$ odd. According to Proposition \[P:Systemreduced\], we have to count, for any root $r$ of $f$, the number of real solutions to $$\label{E:intilasyst}
y^{v_i}=g_i(r)/r^{l_i} \;, \quad i=2,\dotsc,n.$$ Set $c_i(r):=g_i(r)/r^{l_i}$, $i=1,\dotsc,n$. Note that $0,v_2,\dotsc,v_n$ form an $(n-1)$-dimensional simplex for otherwise there would be an additional affine relation on ${\mathcal{C}}$ (so that the convex hull of ${\mathcal{C}}$ would have dimension $<n$). Let $\bar{U} \in M_{n-1}({\mathbb{Z}}/2)$ denote the matrix whose columns contains in this order the reductions modulo $2$ of $v_2,\dotsc,v_n$. According to Lemma \[L:simplex\], the number of solutions to is either $0$, or $2^{n-1-rk(\bar{U})}$. Moreover, it is $2^{n-1-rk(\bar{U})}$ if and only if $$c_2(r)^{\epsilon_2} \cdots c_n(r)^{\epsilon_n}>0$$ for any $(\epsilon_2,\dotsc,\epsilon_n)$ whose reduction modulo $2$ belongs to the kernel of $\bar{U}=(\bar{v_2}, \; \dotsc, \bar{v_n})$.
It follows from the relation that the reduction modulo $2$ of $\lambda:=(\lambda_1,\dotsc,\lambda_{\nu},
0,\dotsc,0)$ belongs to the kernel of $\bar B$. Moreover, since $\lambda_1$ is odd, the kernel of $\bar B$ is the direct sum of $({\mathbb{Z}}/2) \cdot \bar{\lambda}$ and the kernel (in $\{0\} \times ({\mathbb{Z}}/2)^{n-1}$) of $\bar{U}$. This also implies that $rk(\bar{U})=rk(\bar{B})$. To finish, it remains to see that $f(r)=0 \Rightarrow c_1(r)^{\lambda_1} \cdots c_{\nu}(r)^{\lambda_{\nu}} >0$. $\Box$
\[R:redundant\] As explained at the end of the proof, the sign condition $c_1(r)^{\lambda_1} \cdots c_{\nu}(r)^{\lambda_{\nu}} >0$ given by the affine relation on $\{0,v_1,\dotsc,v_n\}$ is in fact redundant in Proposition \[P:real\].
A [*positive*]{} solution is a solution with positive coordinates.
\[P:positive\] The number of positive solutions to $S$ is equal to the number of roots $r$ of $f$ subjected to the sign conditions $$r>0 \; , \; g_1(r)>0 \; , \,\dotsc \, , \; g_n(r) >0.$$
[**Proof.**]{} Assume that $\lambda_1$ is odd. According to Proposition \[P:Systemreduced\] $(x_1,\dotsc,x_n)$ is a positive solution to $S$ if and only if $x_1$ is a positive root $r$ of $f$ and $(x_2,\dotsc,x_n)$ is a positive solution to $$y^{v_i}=g_i(r)/r^{l_i} \;, \quad i=2,\dotsc,n.$$ As it is well-known, the previous system has a positive solution if and only if $g_i(r)/r^{l_i}>0$ for $i=2,\dotsc,n$, and in this case it has exactly one positive solution. To finish, it remains to note that $f(r)=0$, $r>0$ and $g_i(r)>0$ for $i=2,\dotsc,n$ implies that $g_1(r)>0$. $\Box$
Upper bounds
============
Let $S$ be a system as in Section \[S:elimination\] and write the eliminant as $$f=P-Q,$$ where $P(x_1)=x_1^{\lambda_0}\prod_{i=1}^t (g_i(x_1))^{\lambda_i}$ and $Q(x_1)=\prod_{i=t+1}^{\nu} (g_i(x_1))^{\lambda_i}$. Consider the rational function $$\label{E:rationalfunction}
\phi:=f/Q=P/Q-1.$$ Recall that $m({\mathcal{C}})$ is the dimension of the affine span of the minimal affinely dependent subset of ${\mathcal{C}}$.
\[P:boundpositive\] The number of positive solutions to $S$ is no more than $m({\mathcal{C}})+1$.
[**Proof.**]{} Let ${{\mathcal{R}}}_+$ denote the set of positive roots $r$ of $f$ satisfying $g_1(r)>0 \, , \,\dotsc \, , \,g_n(r) >0$ and let $N$ denote the number of elements of ${{\mathcal{R}}}_+$. The number of positive solutions to $S$ is equal to $N$ by Proposition \[P:positive\]. Consider an open interval $I$ formed by two consecutive elements of ${{\mathcal{R}}}_+$. This interval contains no root of $P$ and $Q$ (since each $g_i$ has simple real roots). As a consequence, the function $\phi$ is bounded on $I$. Moreover, $\phi$ takes the same value $0$ on the endpoints of $I$. Therefore, by Rolle’s theorem, the derivative $\phi'$ has at least one real root in $I$. This gives $N-1$ distinct positive roots of $\phi'$, which are different from the roots of $P$ and $Q$. The roots of $\phi'$ are the roots of $P'Q-PQ'$. As noticed in ([@BeBiS], Proof of Proposition 4.3.), the quotient
$$\label{E:quotient}
\frac{P'(x)Q(x)-P(x)Q'(x)}{x^{\lambda_0-1}\prod_{i=1}^{\mu} (g_i(x))^{\lambda_i-1}},$$
or, if $\lambda_0=0$, the same quotient but with $\ell-1$ in place of $\lambda_0 -1$, is a polynomial $H$ which can be written as $H(x)=h(x^{\ell})$ where $h$ is real polynomial of degree $m({\mathcal{C}})$ and with non zero constant term. Thus the $N-1$ positive roots of $\phi'$ different from the roots of $P$ and $Q$ should be roots of $H$. But $H$ has at most $m({\mathcal{C}})$ distinct positive roots, hence $N-1 \leq m({\mathcal{C}})$. $\Box$
J. Maurice Rojas informed us that Proposition \[P:boundpositive\] can be obtained from [@LRW03] (Lemma 2, Section 3).
Define the matrix $A$ as the matrix obtained by putting (in this order) the vectors $w_{0},\dotsc,w_n$ in columns
$$\label{E:matrixcircuit}
A:=(w_0, \dotsc , w_{n})
\in M_{n,n+1}({\mathbb{Z}})$$
We have by definition ${{rk \, ({\bar {\mathcal{C}}})}}={{rk \, ({\bar A})}}$, where $\bar{A}$ is the reduction modulo $2$ of $A$. Recall that $w_{0}={\ell}e_1$ and $w_i=l_ie_1+v_i$ for $i=1,\dotsc,n$, so that
$$\label{E:matrixcircuit2}
A=
\left(
\begin{array}{cccc}
{\ell} & l_1 & \dotsc & l_n \\
0 & & & \\
\vdots & v_1 & \dotsc & v_n \\
0 & & &
\end{array}
\right)$$
The lower-right matrix $(v_1, \dotsc, v_n)$ is the matrix $B$ that we already defined.
\[T:upper bound\] Let $N$ be the number of real solutions to $S$.
1. If $ {{rk \, ({\bar {\mathcal{C}}})}} \leq m({\mathcal{C}})$, then $$N \leq 2^{n-{{rk \, ({\bar {\mathcal{C}}})}}} \cdot \left(m({\mathcal{C}})+ {{rk \, ({\bar {\mathcal{C}}})}}+1\right).$$
2. if $ {{rk \, ({\bar {\mathcal{C}}})}} \geq m({\mathcal{C}})$, then $$N \leq 2^{n-{{rk \, ({\bar {\mathcal{C}}})}}} \cdot \left(2m({\mathcal{C}})+1\right).$$
[**Proof.**]{} Consider first the case ${{rk \, ({\bar {\mathcal{C}}})}}=0$. This corresponds to the case ${\mathcal{C}}\subset 2{{\mathbb{Z}}}^n$. Setting $\tilde{w}_i=w_i/2$, a system supported on ${\mathcal{C}}$ can be rewritten as a system supported on the circuit $\tilde{{\mathcal{C}}}=\{0,\tilde{w}_{-1},\tilde{w}_0,\dotsc,\tilde{w}_n\}$ so that the number of real solutions (in the real torus) to the system supported on ${\mathcal{C}}$ is equal to $2^n$ times the number of positive solutions to the system supported on $\tilde{{\mathcal{C}}}$. Obviously, we have $m(\tilde{{\mathcal{C}}})=m({\mathcal{C}})$. The bound $2^n(m({\mathcal{C}})+1)$ for the number of real solutions to $S$ follows then from Proposition \[P:boundpositive\].
Suppose now that ${{rk \, ({\bar {\mathcal{C}}})}} \neq 0$, that is, ${\mathcal{C}}$ is not contained in $2{{\mathbb{Z}}}^n$. Then, it is easy to show the existence of points $w_i,w_j \in {\mathcal{C}}$ such that $w_i-w_j \notin 2{{\mathbb{Z}}}^n$, and $w_i$ belongs to the minimal affinely dependent subset of ${\mathcal{C}}$. Thus, reordering the points of ${\mathcal{C}}$ so that $w_i$ becomes $w_{-1}$ and $w_j$ becomes $w_0$ if necessary, we can assume that the number ${\ell}$ is an odd number, and the coefficient $\delta$ is non zero. The fact that $\ell$ is odd gives that $\bar{A}$ and $\bar{B}$ have isomorphic kernels, so that ${{rk \, ({\bar {\mathcal{C}}})}}={{rk \, ({\bar A})}}=rk \,(\bar{B})+1$.
Define $$c(r):=(c_1(r),\dotsc,c_n(r)) \quad \mbox{with} \quad
c_i(r):={g_{i}(r)/r^{l_{i}}} \;, \quad i=1,\dotsc,n.$$ Let $\pi_1,\dotsc,\pi_{n-{{rk \, ({\bar {\mathcal{C}}})}}+1}$ be vectors in ${{\mathbb{Z}}}^{n-1}$ whose reductions modulo $2$ form a basis of $\mbox{Ker} \, \bar{B}$. Then Proposition \[P:real\] gives that the number of real solutions to $S$ is $2^{n-{{rk \, ({\bar {\mathcal{C}}})}}}$ times the number of real roots $r$ of $f$ subjected to the sign conditions
$$\label{E:cond1}
(c(r))^{\pi_i}>0 \quad \mbox{for} \quad i=1,\dotsc,n-rk \, ({\bar {\mathcal{C}}})+1.$$
Consider the ${n-{{rk \, ({\bar {\mathcal{C}}})}}}+1$ by $n$ matrix $\Pi$ whose rows are given by the exponent vectors of the sign conditions . Taking the Hermite normal form of the reduction modulo $2$ of $\Pi$, we obtain another basis of $\mbox{Ker} \, \bar{B}$ producing equivalent sign conditions
$$\label{E:normalcond}
\begin{array}{cccc}
c_{s_1}(r) \cdot \prod_{i>s_1, i \notin {{\mathcal{S}}}} \, (c_i(r))^* & > & 0 \\
& & & \\
c_{s_2}(r) \cdot \prod_{i>s_2, i \notin {{\mathcal{S}}}} \, (c_i(r))^* & > & 0 \\
& & & \\
\vdots & \vdots & \vdots \\
& & & \\
c_{s_{n-{{rk \, ({\bar {\mathcal{C}}})}}+1}}(r) \cdot \prod_{i>s_{n-{{rk \, ({\bar {\mathcal{C}}})}}+1}, i \notin {{\mathcal{S}}}} \, (c_i(r))^* & > & 0
\end{array}$$
where $1 \leq s_1 < s_2 < \dotsc < s_{n-{{rk \, ({\bar {\mathcal{C}}})}}+1} \leq n$, ${{\mathcal{S}}}=\{s_1,\dotsc,s_{n-{{rk \, ({\bar {\mathcal{C}}})}}+1}\}$ and the stars are exponents taking values $0$ or $1$.
It remains to show that the set ${{\mathcal{R}}}$ of real roots $r$ of $f$ satisfying the sign conditions has at most $m({\mathcal{C}})+1+{{rk \, ({\bar {\mathcal{C}}})}}$ elements if ${{rk \, ({\bar {\mathcal{C}}})}} \leq m({\mathcal{C}})$, and at most $2m({\mathcal{C}})+1$ elements if ${{rk \, ({\bar {\mathcal{C}}})}} \geq m({\mathcal{C}})$.
First note that the integer ${\ell}$ being odd, each $g_i$ has only one real root, that we denote by $\rho_i$. Moreover, the polynomial $H$ defined in has at most $m({\mathcal{C}})$ real roots since $H(x_1)=h(x_1^{\ell})$ with $h$ polynomial of degree $m({\mathcal{C}})$. Let $I$ denote an open (bounded) interval formed by two consecutive elements of ${{\mathcal{R}}}$. We claim that $I$ should contain at least one element among $0$, $\rho_1,\dotsc,\rho_{\nu}$ and the real roots of $H$. Indeed, if the function $\phi=f/Q$ is bounded on $I$, then $I$ should contain a real critical point (with finite critical value) of $\phi$ by Rolle’s theorem. The critical points of $\phi$ with finite critical values are contained in the set formed of $0$, the roots $\rho_1,\dotsc,\rho_t$ of $P$ and the roots of $H$ (see the proof of Proposition \[P:boundpositive\]). If $\phi$ is not bounded on $I$, then obviously $I$ should contain a root of $Q$, that is, a root $\rho_i$ for some $i=t+1,\dotsc,\nu$.
There is at most one interval $I$ containing $0$ and at most $m({\mathcal{C}})$ intervals $I$ containing a real root of $H$. Let us concentrate now on the intervals $I$ which do no contain $0$ but contain a root $\rho_i$ with $i=1,\dotsc,\nu$. The number of such intervals is obviously no more than $\nu$. We claim that this number of intervals is also no more than ${{rk \, ({\bar {\mathcal{C}}})}}-1$. Indeed, such an interval $I$ should contain a $\rho_i$ with $i \in \{1,\dotsc,n\} \setminus {{\mathcal{S}}}$, for otherwise some sign condition in would be violated at the endpoints of $I$. The claim follows then as the cardinality of $\{1,\dotsc,n\} \setminus {{\mathcal{S}}}$ is ${{rk \, ({\bar {\mathcal{C}}})}}-1$.
Therefore, the number of intervals $I$ formed by two consecutive elements of ${{\mathcal{R}}}$ is no more than $1+m({\mathcal{C}})+{{rk \, ({\bar {\mathcal{C}}})}}-1=m({\mathcal{C}})+{{rk \, ({\bar {\mathcal{C}}})}}$, and no more than $1+m({\mathcal{C}})+\nu$. Obviously, the number of elements of ${{\mathcal{R}}}$ is equal to the number of these intervals $I$ added by one. The bound $m({\mathcal{C}})+{{rk \, ({\bar {\mathcal{C}}})}}$ on the number of intervals $I$ gives thus $$\label{E:firstbound}
N \leq 2^{n-{{rk \, ({\bar {\mathcal{C}}})}}} \cdot \left(m({\mathcal{C}})+ {{rk \, ({\bar {\mathcal{C}}})}}+1\right).$$
Consider now the bound $1+m({\mathcal{C}})+\nu$ on the number of intervals $I$. We see that if this bound is attained, then the polynomial $H$ has $m({\mathcal{C}})$ real roots, and each interval $I$ contains exactly one element among the real roots of $H$, the roots $\rho_1,\dotsc,\rho_{\nu}$, and $0$. This forces each of these points to be a critical point of $\phi$ with even multiplicity (and unbounded critical values for the roots of $Q$). The $m({\mathcal{C}})$ real roots of $H$ are simple roots of $H$, hence critical points with multiplicity $2$ of $\phi$. The multiplicity of $0$ with respect to $\phi $ is $\lambda_0$, while the multiplicity of $\rho_i$, $i=1,\dotsc,\nu$, with respect to $\phi$ is $\pm \lambda_i$. Consequently, if the number of intervals $I$ is $1+m({\mathcal{C}})+\nu$, then $\lambda_0,\lambda_1,\dotsc,\lambda_{\nu}$ are even: a contradiction since these numbers are coprime. Hence, the number of intervals $I$ is no more than $m({\mathcal{C}})+\nu$. If $\lambda_0$ and $\delta$ are both non zero, then $\nu=m({\mathcal{C}})$ so that the number of intervals $I$ is no more than $2m({\mathcal{C}})$. This gives the bound $$\label{E:secondbound}
N \leq 2^{n-{{rk \, ({\bar {\mathcal{C}}})}}} \cdot \left(2m({\mathcal{C}})+1\right).$$ Recall that we have assumed at the beginning that $\delta \neq 0$. Hence, it remains to see what happens when $\lambda_0=0$. In this case, we have $m({\mathcal{C}})=\nu-1$ and $0$ cannot be a critical point of $\phi$. Arguing as before, we obtain then that the number of intervals $I$ is no more than $m({\mathcal{C}})+\nu$. Moreover, if this number is equal to $m({\mathcal{C}})+\nu$, then $\lambda_1,\dotsc,\lambda_{\nu}$ are even: a contradiction since $\lambda_1,\dotsc,\lambda_{\nu}$ are coprime. Hence, the number of intervals $I$ is no more than $m({\mathcal{C}})+\nu-1=2m({\mathcal{C}})$, and we retrieve the bound . $\Box$
Real rational graphs and sharpness of bounds
============================================
Real rational graphs
--------------------
This subsection comes from [@Br] (see also [@Brt; @O; @NSV]). Consider the function $$\phi =f/Q=P/Q-1$$ as a rational function ${{\mathbb{C}}}P^1 \rightarrow {{\mathbb{C}}}P^1$. The degree of $\phi$ coincides with that of $f$. Define $$\Gamma:=\phi^{-1}({{\mathbb{R}}}P^1).$$ This is a real graph on ${{\mathbb{C}}}P^1$ (invariant with respect to the complex conjugation) and which contains ${{\mathbb{R}}}P^1$. Each vertex of $\Gamma$ has even valency, and the multiplicity of a critical point with real critical value of $\phi$ is half its valency. The graph $\Gamma$ contains the inverse images of $-1$, $\infty$ and $0$ which are the sets of roots of $P$, $Q$ and $f$, respectively. Denote by the same letter $p$ (resp. $q$ and $r$) the points of $\Gamma$ which are mapped to $-1$ (resp. $\infty$ and $0$). Orient the real axis on the target space via the arrows $-1 \rightarrow \infty \rightarrow 0 \rightarrow -1$ (orientation given by the decreasing order in ${\mathbb{R}}$) and pull back this orientation by $\phi$. The graph $\Gamma$ becomes an oriented graph, with the orientation given by arrows $p \rightarrow q \rightarrow r \rightarrow p$.
Clearly, the arrangement of the real roots of $f$, $P$ and $Q$ together with their multiplicities can be extracted from the graph $\Gamma$. We encode this arrangement together with the multiplicities by what is called a root scheme.
A root scheme is a $k$-uple $((l_1,m_1),\dotsc,(l_k,m_k)) \in (\{p,q,r\} \times {\mathbb{N}})^k$. A root scheme is realizable by polynomials of degree $d$ if there exist real polynomials $P$ and $Q$ such that $f=P-Q$ has degree $d$ and if $\rho_1< \dotsc < \rho_k$ are the real roots of $f$, $P$ and $Q$, then $l_i=p$ (resp. $q$, $r$) if $\rho_i$ is root of $P$ (resp. $Q$, $f$) and $m_i$ is the multiplicity of $\rho_i$.
Conversely, suppose we are given a real graph $\Gamma \subset {{\mathbb{C}}}P^1$ together with a real continuous map $\varphi: \Gamma \rightarrow {{\mathbb{R}}}P^1$. Denote the inverse images of $-1$, $\infty$ and $0$ by letters $p$, $q$ and $r$, respectively, and orient $\Gamma$ with the pull back by $\varphi$ of the above orientation of ${{\mathbb{R}}}P^1$. This graph is called [*a real rational graph*]{} [@Br] if
- any vertex of $\Gamma$ has even valence,
- any connected component of ${{\mathbb{C}}}P^1 \setminus \Gamma$ is homeomorphic to an open disk,
Then, for any connected component $D$ of ${{\mathbb{C}}}P^1 \setminus \Gamma$, the map $\varphi_{| \partial D}$ is a covering of ${{\mathbb{R}}}P^1$ whose degree $d_D$ is the number of letters $p$ (resp. $q$, $r$) in $\partial D$. We define the [*degree*]{} of $\Gamma$ to be half the sum of the degrees $d_D$ over all connected components of ${{\mathbb{C}}}P^1 \setminus \Gamma$. Since $\varphi$ is a real map, the degree of $\Gamma$ is also the sum of the degrees $d_D$ over all connected components $D$ of ${{\mathbb{C}}}P^1 \setminus \Gamma$ contained in one connected component of ${{\mathbb{C}}}P^1 \setminus {{\mathbb{R}}}P^1$.
\[P:rootgraph\] A root scheme is realizable by polynomials of degree $d$ if and only if it can be extracted from a real rational graph of degree $d$ on ${{\mathbb{C}}}P^1$.
Let us explain how to prove the if part in Proposition \[P:rootgraph\] (see [@Br; @Brt; @O; @NSV]). For each connected component $D$ of ${{\mathbb{C}}}P^1 \setminus \Gamma$, extend $\varphi_{|\partial D}$ to a, branched if $d_D>1$, covering of one connected component of ${{\mathbb{C}}}P^1 \setminus {{\mathbb{R}}}P^1$, so that two adjacent connected components of ${{\mathbb{C}}}P^1 \setminus \Gamma$ project to differents connected components of ${{\mathbb{C}}}P^1 \setminus {{\mathbb{R}}}P^1$. Then, it is possible to glue continuously these maps in order to obtain a real branched covering $\varphi:{{\mathbb{C}}}P^1 \rightarrow {{\mathbb{C}}}P^1$ of degree $d$. The map $\varphi$ becomes a real rational map of degree $d$ for the standard complex structure on the target space and its pull-back by $\varphi$ on the source space. There exist then real polynomials $P$ and $Q$ such that $f=P-Q$ has degree $d$ and $\varphi =f/Q=P/Q-1$, so that the points $p$ (resp. $q$, $r$) correspond to the roots of $P$ (resp. $Q$, $f$) and $\Gamma=\varphi^{-1}({{\mathbb{R}}}P^1)$.
As in [@Br], we will abbreviate a sequence $S$ repeated $u$ times in a root scheme by $S^u$. If $u=0$, then $S^u$ is the empty sequence.
Constructions
-------------
We are going to prove the existence of root schemes by constructing real rational graphs $\Gamma$ on ${{\mathbb{C}}}P^1$. Since these graphs are real, only half of them will be drawn and the horizontal line will represent the real axis ${{\mathbb{R}}}P^1$.
\[P:rootschemeR=0\]
For any even integer $n=2k>0$, the root scheme $$\left([(q,2 ),(p,2)]^{k}, (q,1), (r,1)^{2k+1}, (p,1) \right)$$ is realizable by polynomials of degree $n+1$.
For any odd integer $n=2k+1>0$, the root scheme $$\left([(q,2 ),(p,2)]^{k}, (q,2),(p,1)(r,1)^{2k+2},(p,1)\right)$$ is realizable by polynomials of degree $n+1$.
[**Proof.**]{} According to Proposition \[P:rootgraph\], it suffices to construct a real rational graph $\Gamma_{n+1}$ of degree $n+1$ on ${{\mathbb{C}}}P^1$ from which the desired root scheme can be extracted. This is done in Figure 1 for $n=2,4$ and in Figure 2 for $n=3,5$. These figures provide the induction step $n \rightarrow n+2$ for constructing suitable graphs $\Gamma_{n+1}$ for any positive integer $n$. $\Box$
$$\PandocStartInclude{BihRojfig2.pstex_t}\PandocEndInclude{input}{737}{27}$$
[Figure 1.]{}
$$\PandocStartInclude{BihRojfig1.pstex_t}\PandocEndInclude{input}{743}{27}$$
[Figure 2.]{}
Let $n$ and $R$ be integers such that $0<R \leq n$. If $n-R$ is even, define $I$ and $J$ by $$\begin{array}{lll}
I & := & [(p,2),(q,2)]^{\frac{n-R}{2}},(q,3),(r,1),(p,2),(r,1)^{n-R+1}\\
& & \\
J & := &\left\{\begin{array}{ll}
(r,1), (p,1) & \mbox{if $R=1$} \\
& \\
(r,1)^{R},(q,2),[(r,1),(p,4),(r,1),(q,4)]^{\frac{R}{2}-1},(r,1),(p,3)& \mbox{if $R$ is even}\\
& \\
(r,1)^{R},(p,2),[(r,1),(q,4),(r,1),(p,4)]^{\frac{R-3}{2}},(q,4),(r,1),(p,3) & \mbox{if $R$ is odd}
\end{array}
\right.
\end{array}$$ If $n-R$ is odd, define $$\begin{array}{lll}
I & := & [(p,2),(q,2)]^{\frac{n-R-1}{2}},(p,2),(q,3),(r,1),(p,2),(r,1)^{n-R+1}\\
& & \\
J & := & \left\{\begin{array}{ll}
(r,1), (q,1) & \mbox{if $R=1$} \\
& \\
(r,1)^{R},(p,2),[(r,1),(q,4),(r,1),(p,4)]^{\frac{R}{2}-1},(r,1),(q,3)& \mbox{if $R$ is even}\\
& \\
(r,1)^{R},(q,2),[(r,1),(p,4),(r,1),(q,4)]^{\frac{R-3}{2}},(p,4),(r,1),(q,3) & \mbox{if $R$ is odd}
\end{array}
\right.
\end{array}$$ Note that the definitions of $J$ with $n-R$ odd and with $n-R$ even are obtained from each other by permuting $p$ and $q$.
$$\PandocStartInclude{Bihfig3.pstex_t}\PandocEndInclude{input}{782}{24}$$
[Figure 3: graph $\Gamma_I$.]{}
\[P:rootschemeRnot0\] For any integers $n$ and $R$ such that $0<R \leq n$, the root scheme $(I \, , \, J)$ is realizable by polynomials of degree $n+R+1$.
[**Proof.**]{} First, we contruct graphs $\Gamma_I$ and $\Gamma_j$, which are not real rational graphs, but from which the sequences $I$ and $J$ can be extracted, respectively. Figure 3 shows $\Gamma_I$ for $n-R=0,1,2,3$, and indicates how to construct $\Gamma_I$ for any integer value of $n-R$ with $0 \leq n-R <n$. Similarly, Figure 4 shows $\Gamma_J$ for $R=1,2,3,4$ and $n-R$ even, and indicates how to construct $\Gamma_J$ for any integers $n$ and $R$ such that $0<R \leq n$ and $n-R$ is even. The graph $\Gamma_J$ for $n-R$ odd is obtained from the graph $\Gamma_J$ with $n-R$ even and the same value of $R$ by permuting $p$ and $q$ and reversing all the arrows. For any integers $n$ and $R$ such that $0<R \leq n$ we can glue the corresponding $\Gamma_I$ and $\Gamma_J$ in order to obtain a real rational graph of degree $n+R+1$ whose root scheme is $(I \, , \, J)$ (see Figure 5 for $n=4$ and $R=3$). The result follows then from Proposition \[P:rootgraph\]. $\Box$
$$\PandocStartInclude{Bihfig4.pstex_t}\PandocEndInclude{input}{812}{24}$$
[Figure 4: graph $\Gamma_J$.]{}
$$\PandocStartInclude{Bihfig6.pstex_t}\PandocEndInclude{input}{820}{24}$$
[Figure 5: gluing of $\Gamma_I$ and $\Gamma_J$ for $n=4$, $R=3$.]{}
Sharpness of bounds
-------------------
\[P:positiveexistence\] For any positive integer $n$, there exist a non degenerate circuit ${\mathcal{C}}\subset {{\mathbb{Z}}}^n$ and a generic real polynomial system supported on ${\mathcal{C}}$ whose number of positive solutions is equal to $n+1$.
[**Proof.**]{} We prove only the case $n$ even since the case $n$ odd is similar (the case $n=1$ is trivial). So assume that $n=2k>0$ is even. According to Proposition \[P:rootschemeR=0\], there exist polynomials $P$ and $Q$ such that $f=P-Q$ has degree $n+1$ and the corresponding root scheme is $$\left([(q,2), (p,2)]^k, (q,1), (r,1)^{2k+1}, (p,1)\right).$$
Composing on the right $\phi=f/Q=P/Q-1$ with a real automorphism of ${{\mathbb{C}}}P^1$ (a real rational map of degree $1$), we can choose three values for the points is in this root scheme. Reading the above scheme from the left to the right, we choose $p=0$ for the last but one $p$, a positive value for the last $q$, and $p=\infty$ for the last $p$. This gives the following arrangement of the roots of $P$, $Q$ and $f$ (where we naturally omit the root at infinity) $$q_1 < p_1 < \dotsc < q_k < p_k=0 < q_{k+1} < r_1 <r_2 < \dotsc < r_{2k+1},$$ so that $f=P-Q$ with $$P(x)=a x^2(x-p_1)^2 \cdots (x-p_{k-1})^2 \; , \quad
Q(x)=b (x-q_{k+1}) (x-q_1)^2 \cdots (x-q_k)^2,$$ and $f=P-Q$ has $n+1=2k+1$ positive roots which are bigger than the roots of $P$ and $Q$. Note that $ab >0$ for otherwise the polynomial $f=P-Q$ could not have roots bigger than the roots of $P$ and $Q$. Dividing $f$ by $a$ and setting $g_1(x)=x-p_1, \dotsc, g_{k-1}=x-p_{k-1}$, $g_k(x)=x-q_1,\dotsc, g_{n-1}(x)=x-q_k$ and $g_n(x)=b(x-q_{k+1})/a$, we obtain a polynomial $f$ of the form with ${\ell}=1$, the exponents are coprime, and such that ${{\mathcal{R}}}_+$ has $n+1$ elements. Let us now give explicitely a non degenerate circuit ${\mathcal{C}}\subset {{\mathbb{Z}}}^n$ and then a system whose eliminant is $f$. Take the circuit ${\mathcal{C}}$ with $w_{-1}=0$, $w_0=e_1$, $w_1=e_{2},\dotsc,w_{n-1}=e_{n}$ and $w_n=2(e_1+e_2+\dotsc+e_{k}-e_{k+1}-\dotsc-e_{n})$. The primitive affine relation on $\{e_1,w_1,\dotsc,w_n\}$ is $$2e_1+2(w_1+\dotsc +w_{k-1})=2(w_k+\dotsc+w_{n-1})+w_n$$ and has the desired coefficients. Hence $f(x_1)$ is the eliminant of the system $x^{w_i}=g_i(x_1)$, $i=1,\dotsc,n$, given explicitely by $$\left\{
\begin{array}{l}
x_2 = x_1-p_1 \quad , \; \dotsc \; , \; x_{k} = x_1-p_{k-1} \\
x_{k+1} = x_1-q_1 \; , \; \dotsc \; , \; x_{n} = x_1-q_{k} \\
{\left(
\frac{x_1 \cdots x_{k}}{x_{k+1} \cdots x_{n}}
\right)}^2 = b(x_1-q_{k+1})/a.
\end{array}
\right.$$ According to Proposition \[P:positive\], this system has $n+1$ positive solutions. $\Box$
It may be interesting to note that the previous system has in fact only these $n+1$ positive solutions as solutions in the complex torus. To see this, one can compute that the volume of the circuit is $n+1$, and use Kouchnirenko theorem.
\[P:existence\] For any integers $n$ and $R$ such that $0 \leq R \leq n$, there exist a non degenerate circuit ${\mathcal{C}}\subset {{\mathbb{Z}}}^n$ with $rk \, (\bar{{\mathcal{C}}})=R$, and a generic real polynomial system supported on ${\mathcal{C}}$ whose number of real solutions is equal to $2^{n-R}(n+R+1)$.
[**Proof.**]{} Note that for $R=0$ this follows from Proposition \[P:positive\] (see the proof of Theorem \[T:upper bound\]). We will give the proof only when $n$ and $R$ are even since the other cases are similar (the case $n=1$ is trivial).
So let $n$ and $R$ be even integers such that $0 < R \leq n$. According to Proposition \[P:rootschemeRnot0\], there exist polynomials $P$ and $Q$ such that $f=P-Q$ has degree $n+R+1$ and the corresponding root scheme is [$$\left([(p,2),(q,2)]^{\frac{n-R}{2}},(q,3),(r,1),(p,2),(r,1)^{n+1},
(q,2),[(r,1),(p,4),(r,1),(q,4)]^{\frac{R}{2}-1},(r,1),(p,3) \right).$$ ]{} We may choose three values for the roots in this root scheme. Choose $p=0$ for the $p$ in the sequence $(q,3),(r,1),(p,2)$, a negative value for the $r$ in the same sequence, and the infinity value for the last $p$ on the right in the root scheme. Then, the polynomials $P$ and $Q$ can be written as $$P(x)=x^2 \prod_{i=1}^{\frac{n-R}{2}} (g_i(x))^2 \cdot \prod_{i=\frac{n-R}{2}+1}^{\frac{n}{2}-1}((g_i(x))^4,$$ $$Q(x)=(g_{\frac{n}{2}}(x))^3 \cdot \prod_{i=\frac{n}{2}+1}^{n-\frac{R}{2}+1} (g_i(x))^2 \cdot
\prod_{i=n-\frac{R}{2}+2}^{n}((g_i(x))^4$$ so that
--------------- ---------------------------------------------------------------------------------------------------------------------------
($\star$)$\;$ [*$f=P-Q$ has $n+R+1$ (non zero) real roots, all but one are positive, and*]{}
[*$g_i$ is positive at the real roots of $f$ for $i=1,\dotsc,\frac{n-R}{2}$ and $i=\frac{n}{2},\dotsc,n-\frac{R}{2}$.*]{}
--------------- ---------------------------------------------------------------------------------------------------------------------------
Here the $g_i$ have all, but one, the form, $g_i(x)=x-\rho_i$ with $\rho_i$ a finite non zero root $p$ if $i < \frac{n}{2}$ and a root $q$ otherwise. The last one is of the form $c(x-\rho_i)$ where $c$ is a positive real number due to the fact $f$ has a real root bigger than the roots of $P$ and $Q$ (see the proof of Proposition \[P:positiveexistence\]). The polynomials $g_i$ mentioned in ($\star$) correspond to the sequence $[(p,2),(q,2)]^{\frac{n-R}{2}},(q,3)$ on the left of the above root scheme.
Now, define $v_{n-\frac{R}{2}+1}=-(e_2+\dotsc+e_{n-1})-3e_n$, $v_{\frac{n}{2}}=2(e_2+\dotsc+e_n)$, and $$\begin{array}{lll}
v_i & = & 2e_{i+1} \;, \quad i=1,\dotsc,\frac{n-R}{2}\\
& & \\
v_i & = & e_{i+1} \;, \quad i=\frac{n-R}{2}+1,\dotsc,\frac{n}{2}-1\\
& & \\
v_i & = & -2e_{i} \;, \quad i=\frac{n}{2}+1,\dotsc,n-\frac{R}{2}\\
& & \\
v_i & = & -e_{i-1} \;, \quad i=n-\frac{R}{2}+2,\dotsc,n
\end{array}$$ The primitive affine relation on $\{0,v_1,\dotsc,v_n\}$ is the desired one $$2\sum_{i=1}^{\frac{n-R}{2}}v_i+4\sum_{i=\frac{n-R}{2}+1}^{\frac{n}{2}-1}v_i=
3v_{\frac{n}{2}}+2\sum_{i=\frac{n}{2}+1}^{n-\frac{R}{2}+1}v_i+4\sum_{i=n-\frac{R}{2}+2}^{n}v_i,$$ and it’s easy to see that the rank of the matrix $\bar{B}=(\bar{v}_1, \dotsc, \bar{v}_n)$ is $R-1$. We want to find integers $l_1,\dotsc,l_n$ verifying the relation and such that the sign conditions in Proposition \[P:real\] are verified at each root $r$ of $f$. Here, these sign conditions are given by the $v_i$ which belong to $2{{\mathbb{Z}}}^n$ $$\label{E:signproof}
\begin{array}{lll}
g_i(r)/r^{l_i} & > 0& , \quad i=1,\dotsc,\frac{n-R}{2} \quad \mbox{and} \quad
i=\frac{n}{2},\dotsc,n-\frac{R}{2}
\end{array}$$ In view of property $(\star)$, we want to have $l_i$ even for $i=1,\dotsc,\frac{n-R}{2}$ and $i=\frac{n}{2},\dotsc,n-\frac{R}{2}$ since one root of $f$ is negative. In fact we can find integers $l_i$ which are all even. Indeed, in our situation, the relation is
$$\label{E:relationproof}
2+2\sum_{i=1}^{\frac{n-R}{2}}l_i+4\sum_{i=\frac{n-R}{2}+1}^{\frac{n}{2}-1}l_i=
3l_{\frac{n}{2}}+2\sum_{i=\frac{n}{2}+1}^{n-\frac{R}{2}+1}l_i+4\sum_{i=n-\frac{R}{2}+2}^{n}l_i.$$
Replacing the first $2$ on the left by $1$, the existence of integers $l_i$ verifying the resulting relation is due to the fact that $2,4,3$ are coprime. Multiplying then by $2$ the members of this modified relation gives then the existence of even integers $l_i$ verifying .
So, let us choose even integers $l_i$ verifying so that the sign conditions are satisfied. The proof is almost finished.
Consider the system $$\label{E:systemproof}
x^{w_i}=g_i(x_1) \; , \quad i=1,\dotsc,n,$$ where $w_i:=l_ie_1+v_i$. This system is equivalent to a (generic) system supported on the circuit ${\mathcal{C}}=\{w_{-1}=0,w_0=e_1,w_1,\dotsc,w_n\}$. The primitive affine relation on ${\mathcal{C}}$ is $$2w_0+2\sum_{i=1}^{\frac{n-R}{2}}w_i+4\sum_{i=\frac{n-R}{2}+1}^{\frac{n}{2}-1}w_i=
3w_{\frac{n}{2}}+2\sum_{i=\frac{n}{2}+1}^{n-\frac{R}{2}+1}w_i+4\sum_{i=n-\frac{R}{2}+2}^{n}w_i,$$ so that $f$ is the eliminant of . The rank modulo $2$ of ${\mathcal{C}}$ is ${{rk \, ({\bar {\mathcal{C}}})}}=rk \,(\bar{B})+1=R$. The eliminant $f$ has $n+R+1$ real roots and the sign conditions are satisfied at each real root of $f$. Therefore, according to Proposition \[P:real\] , the system has $2^{n-R}(n+R+1)$ real solutions. $\Box$
The case $R=n$ in Proposition \[P:existence\] has already been obtained in [@BeBiS] by different methods.
\[T:positivesharp\] For any integers $n, m$ such that $1 \leq m \leq n$, there exist a circuit ${\mathcal{C}}\subset {{\mathbb{Z}}}^n$ with $m({\mathcal{C}})=m$ and a generic real polynomial system supported on ${\mathcal{C}}$ whose number of positive solutions is $m+1$
[**Proof.**]{} By Proposition \[P:positiveexistence\], there exists a circuit ${{\mathcal{C}}}'=\{0,w_0=\ell e_1,w_1,\dotsc,w_{m}\} \subset {{\mathbb{Z}}}^{m}$ with $m({{\mathcal{C}}}')=m$ and polynomials $g_1,\dotsc,g_{m}$ such that the system $x^{w_i}=g_i(x_1)$, $i=1,\dotsc, m$, has ${m}+1$ positive solutions. By Proposition \[P:positive\], this means that the eliminant $f$ of this system has $m+1$ positive roots at which $g_1,\dotsc,g_m$ are positive. Define $$w_i:=e_i \; , \quad i={m} +1,\dotsc,n,$$ and choose polynomials $g_{m+1},\dotsc,g_n$ (as in ) which are positive at the real roots of $f$. Then, the system $x^{w_i}=g_i(x_1)$, $i=1,\dotsc, n$, is equivalent to a system supported on the circuit ${{\mathcal{C}}}=\{0,w_0=\ell e_1,w_1,\dotsc,w_n\} \subset {{\mathbb{Z}}}^{n}$. The eliminant of this system is also $f$ and we have $m({{\mathcal{C}}})=m({{{\mathcal{C}}}'})=m$. Proposition \[P:positive\] and our choice of $g_{{m}+1},\dotsc,g_n$ implies that this system has also ${m}+1$ positive solutions. $\Box$
\[T:sharp\] For any integers $n, m, R$ such that $0 \leq R \leq n$ and $1 \leq m \leq n$, there exist a circuit ${\mathcal{C}}\subset {{\mathbb{Z}}}^n$ with $m({\mathcal{C}})=m$, ${rk \, ({\bar {\mathcal{C}}})}=R$, and a generic real polynomial system supported on ${\mathcal{C}}$ whose number of real solutions, $N$, verifies:
1. if $R \leq m$, then $$N= 2^{n-{R}} \cdot \left({m}+ {R}+1\right).$$
2. if ${R} \geq {m}$, then $$N = 2^{n-{R}} \cdot \left(2{m}+1\right).$$
[**Proof.**]{} The case $R=0$ is an immediate consequence of Theorem \[T:positivesharp\]. Suppose that $0 < R \leq {m}$. By Proposition \[P:existence\], there exist a circuit ${{\mathcal{C}}}'=\{0,w_0=\ell e_1,w_1,\dotsc,w_{m}\} \subset {{\mathbb{Z}}}^{m}$ with $m({{\mathcal{C}}}')=m$, $rk \, (\bar{{{\mathcal{C}}}'})=R$, and polynomials $g_1,\dotsc,g_{m}$ such that the system $$\label{E:system1}
x^{w_i}=g_i(x_1) \; , \quad i=1,\dotsc, m,$$ has $2^{m-{R}} \cdot \left({m}+ {R}+1\right)$ real solutions. Here and in the rest of the proof $g_i$ is a polynomial as in . Since $ rk \, (\bar{{{\mathcal{C}}}'}) \neq 0$, we may assume that $\ell$ is odd. Define $$w_i:=2e_i \; , \quad i={m} +1,\dotsc,n,$$ and choose polynomials $g_{m+1},\dotsc,g_n$ which are positive at the real roots of the eliminant $f$ of . Then, the system $$\label{E:system2}
x^{w_i}=g_i(x_1) \; , \quad i=1,\dotsc,n,$$ is equivalent to a system supported on the circuit ${{\mathcal{C}}}=\{0,w_0=\ell e_1,w_1,\dotsc,w_n\} \subset {{\mathbb{Z}}}^{n}$. We have $m({\mathcal{C}})=m({{\mathcal{C}}'})=m$, ${rk \, ({\bar {{\mathcal{C}}}})}=rk \, (\bar{{{\mathcal{C}}}'})=R$ and $f$ is also the eliminant of . Moreover the sign conditions in Proposition \[P:real\] which correspond to are obtained from those corresponding to by adding $g_i(r)>0$ for $i=m+1,\dotsc,n$ (note that $l_{m+1}=\dotsc=l_n=0$). But since has $2^{m-{R}} \cdot \left({m}+ {R}+1\right)$ real solutions, and $\ell$ is odd, it follows from Proposition \[P:real\] that the sign conditions corresponding to are satisfied at ${m}+ {R}+1$ real roots of $f$. By our choice of $g_{m+1},\dotsc,g_n$, the sign conditions corresponding to are satisfied at the same ${m}+ {R}+1$ real roots of $f$. Hence, again by Proposition \[P:real\], the system has $2^{n-{R}} \cdot \left({m}+ {R}+1\right)$ real solutions.
Finally, suppose that ${R}> {m}$. The idea is exactly as before. By Proposition \[P:existence\], there exist a circuit ${{\mathcal{C}}}'=\{0,w_0=\ell e_1,w_1,\dotsc,w_{m}\} \subset {{\mathbb{Z}}}^{m}$ with $m({{\mathcal{C}}}')=m$, $rk \, (\bar{{{\mathcal{C}}}'})=m$ and polynomials $g_1,\dotsc,g_{m}$ such that the system has $2{m}+1$ real solutions. According to Proposition \[P:real\] (see also Remark \[R:redundant\]), this means that the eliminant $f$ of this system has $2{m}+1$ real roots. Define $$w_i:=e_i \; , \quad i={m} +1,\dotsc,{R}, \quad w_i:=2e_i \; , \quad i={R} +1,\dotsc,n$$ and choose polynomials $g_{R+1},\dotsc,g_n$ which are positive at the roots of the eliminant $f$. Then, the system is equivalent to a system supported on the circuit ${{\mathcal{C}}}=\{0,w_0=\ell e_1,w_1,\dotsc,w_n\} \subset {{\mathbb{Z}}}^{n}$. We have $m({\mathcal{C}})={m}$, $ rk \, (\bar{{\mathcal{C}}})=R$ and has $2^{n-R} \cdot \left(2{m}+1\right)$ real solutions. $\Box$
[1]{}
B. Bertrand, F. Bihan, and F. Sottile, *Polynomial Systems with Few Real Zeroes*, ArXiv:math.AG/0502051, to appear in Math. Zeitschrift.
E. Brugallé, *Real plane algebraic curves with asymptotically maximal number of even ovals*, ArXiv:math.AG/0411097, to appear in Duke Mathematical Journal.
E. Brugallé, *Courbes algébriques réelles et courbes pseudoholomorphes réelles dans les surfaces réglées*, Phd thesis, University of Rennes 1, France, (2004), (in english), http://tel.ccsd.cnrs.fr/documents/archives0/00/00/86/52/.
A.G. Kouchnirenko, *A [N]{}ewton polyhedron and the number of solutions of a system of $k$ equations in $k$ unknowns*, Usp. Math. Nauk. **30** (1975), 266–267.
Tien-Yien Li, J. Maurice Rojas, and Xiaoshen Wang,*Counting real connected components of trinomial curve intersections and [$m$]{}-nomial hypersurfaces*, Discrete Comput. Geom. **30** (2003), no. 3, 379–414. [MR ]{}[2 002 964]{}
S. Yu Orevkov, *Riemann existence theorem and construction of real algebraic curves*, Ann. Fac. Sci. Toulouse Math. (6) [**12**]{} (2003), no. 4, 517–531.
S. Natanzon, B. Shapiro, and A. Vainshtein, *Topological classification of generic real rational functions*, J. Knot Theory Ramifications [**11**]{} (2002), no. 7, 1063–1075.
F. Sottile, *Enumerative real algebraic geometry*, Algorithmic and quantitative real algebraic geometry (Piscataway, NJ, 2001), 139–179, DIMACS Ser. Discrete Math. Theoret. Comput. Sci., 60, Amer. Math. Soc., Providence, RI, 2003.
E. Soprunova, F. Sottile, *Lower bounds for real solutions to sparse polynomial systems*, ArXiv:math.AG/0409504, to appear in Advances in Mathematics.
B. Sturmfels, *On the number of real roots of a sparse polynomial system*, Hamiltonian and gradient flows, algorithms and control, 137–143, Fields Inst. Commun., 3, Amer. Math. Soc., Providence, RI, 1994.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report on the idea to use colours to distinguish syntax and semantics as an educational tool in logic classes. This distinction gives also reason to reflect on some philosophical issues concerning semantics.'
author:
- Reinhard Kahle
- Wilfried Keller
bibliography:
- 'svs.bib'
title: Syntax versus Semantics
---
=1
Introduction
============
When the first author attended as a first year student a logic course at a philosophy department, he was wondering why the professor was repeating, after introducing the relation $\vDash$, “the same” again but using the symbol ‘$\vdash$’. It was, most likely, the negligence of the student, and not the presentation of the professor, which caused this fundamental misunderstanding. In retrospect, there is one question which might make one wonder: how was it possible to miss one of the most fundamental distinctions in modern logic? On a more mature stage, as third year student, the problem repeated itself on a different level: speaking about Gödel’s first incompleteness theorem with a teaching assistant, he refused to continue the discussion when the question was raised what its meaning is from the perspective of “proper Mathematics”. Again, the distinction of syntax and semantics wasn’t clearly seen and the discussion stalled when the participants did not realize that reference of a term like “proper Mathematics” need to be better specified while entering into a discussion of Gödel’s theorem.
Based on the experiences above, when starting to teach logic courses by himself, the first author took it as a particular challenge to present the problems concerning the distinction of syntax and semantics in a persuasive way. In this paper, we present the chosen solution: the use of colours to distinguish the syntactical and semantical role of logical text. We also report on some of the educational insights one might win with this approach together with certain philosophical questions which surface (again) in this context.
We assume that the reader is familiar with first-order logic and has, at least, an idea of Gödel’s incompleteness theorems. Without formal introduction, we use the standard notations of first-order logic and Peano Arithmetic, using typically Greek letters from the end of the alphabet to denote formulae.
First-order logic
=================
Let us go *in media res* and recall the definition of the satisfiability relation $\models$ for first-order logic as given by Barwise in [@Bar77 p. 21].
\[def1\]
Let [[$\mathfrak{M}$]{}]{} be an L-structure. We define a relation $${\ensuremath{\mathfrak{M}}}\models \phi[s],$$ (read: the assignment $s$ satisfies the formula $\phi$ in ${\ensuremath{\mathfrak{M}}}$) for all assignments $s$ and all formulae $\phi$ as follows.
(i) ${{\ensuremath{\mathfrak{M}}}\models {(t_1=t_2)}[s]}$ iff ${{t_1}^{{\ensuremath{\mathfrak{M}}}}(s) = {t_2}^{{\ensuremath{\mathfrak{M}}}}(s)}$,
(ii) ${{\ensuremath{\mathfrak{M}}}\models {R(t_1,\dots,t_n)}[s]}$, iff ${({t_1}^{{\ensuremath{\mathfrak{M}}}}(s),\dots,{t_n}^{{\ensuremath{\mathfrak{M}}}}(s)) \in {R}^{{\ensuremath{\mathfrak{M}}}}}$,
(iii) ${{\ensuremath{\mathfrak{M}}}\models {(\neg \phi)}[s]}$ iff not ${{\ensuremath{\mathfrak{M}}}\models {\phi}[s]}$,
(iv) ${{\ensuremath{\mathfrak{M}}}\models {(\phi \wedge \psi)}[s]}$ iff ${{\ensuremath{\mathfrak{M}}}\models {\phi}[s]}$ and ${{\ensuremath{\mathfrak{M}}}\models {\psi}[s]}$,
(v) ${{\ensuremath{\mathfrak{M}}}\models {(\phi \vee \psi)}[s]}$ iff ${{\ensuremath{\mathfrak{M}}}\models {\phi}[s]}$ or ${{\ensuremath{\mathfrak{M}}}\models {\psi}[s]}$,
(vi) ${{\ensuremath{\mathfrak{M}}}\models {(\phi \rightarrow \psi)}[s]}$ iff either not ${{\ensuremath{\mathfrak{M}}}\models {\phi}[s]}$ or else ${{\ensuremath{\mathfrak{M}}}\models {\psi}[s]}$,
(vii) ${{\ensuremath{\mathfrak{M}}}\models {(\exists v.\phi)}[s]}$ iff there is an ${a\in M}$ such that ${{\ensuremath{\mathfrak{M}}}\models {\phi}[s(^{{a}}_{{v}})]}$,
(viii) ${{\ensuremath{\mathfrak{M}}}\models {(\forall v.\phi)}[s]}$ iff for all ${a \in M}$, ${{\ensuremath{\mathfrak{M}}}\models {\phi}[s(^{{a}}_{{v}})]}$.
Barwise then states: “There is nothing surprising here. It is just making sure that each of our symbols means what we want it to mean.”
This is certainly correct, but it might be worth to reflect shortly what this definition actually does.[^1] Looking, for instance, to the clauses for conjunction and disjunction, it simply “lifts” the symbols ‘$\wedge$’ and ‘$\vee$’ to the *meta-level* using the natural language expressions “and” and “or”. Thus, who understands “and” and “or” is supposed to understand how to obtain the truth value of a formula using ‘$\wedge$’ or ‘$\vee$’. The *meta-level* comes into play, because the natural language terms are outside of the $\models$ relation, while the logical symbols are inside. Nothing surprising here, as said. We will see later in § \[sec-int\] that the situation for the negation is actually more subtle than it looks (and this is a first educational pitfall!). But the issue we like to start with is unambiguously expressed by Barwise (in continuation of the citation above):
> There is one possibly confusing point, in (i), caused by our using $=$ for both the real equality (on the right-hand side) and the symbol for equality (on the left). Many authors abhor this confusion of use and mention and use something like $\equiv$ or $\approx$ for the symbol.
It is this confusion[^2], which will repeat itself when one comes to the language of arithmetic, we like to address. As Barwise remarks, many authors try to resolve the problem by the use of different symbols for syntactical and semantical equality. This would be a way out, but it comes at the price that the formal language appears artificial, in particular, if this is extended to arithmetical terms. Let us mention the clumsy notation $\textbf{k}_n$ for the numeral representing the number $n$ in the classical textbook of Shoenfield [@Sho67]. With respect to equality, Sernadas and Sernadas [@SS08], for instance, use $\cong$ as special sign on the syntactical side. Cori and Lascar [@CL00] use, quite generally, a bar over the syntactical symbol to refer to its *semantic* interpretation. However, in general (and with good reasons), the syntactic symbols are chosen to match exactly with their intended meaning; the resulting “overloading” is just a problem when the difference of syntax and semantics is the subject under discussion.
Our solution is to use the same symbols, but different colours to distinguish the use of symbols on the syntactical and semantical side. So, let us choose red for syntax and blue for semantics.[^3] After giving a definition of first-order language with all syntactic expressions in red, and the definition of a structure with blue for the elements of the structure, the definition above reads as follows:
Let [[$\mathfrak{M}$]{}]{} be an L-structure. We define a relation $${{\ensuremath{\mathfrak{M}}}\models {\phi}[s]},$$ (read: the assignment ${s}$ satisfies the formula ${\phi}$ in ${{\ensuremath{\mathfrak{M}}}}$) for all assignments ${s}$ and all formulae ${\phi}$ as follows.[^4]
(i) ${{\ensuremath{\mathfrak{M}}}\models {(t_1=t_2)}[s]}$ iff ${{t_1}^{{\ensuremath{\mathfrak{M}}}}(s) = {t_2}^{{\ensuremath{\mathfrak{M}}}}(s)}$,
(ii) ${{\ensuremath{\mathfrak{M}}}\models {R(t_1,\dots,t_n)}[s]}$, iff ${({t_1}^{{\ensuremath{\mathfrak{M}}}}(s),\dots,{t_n}^{{\ensuremath{\mathfrak{M}}}}(s)) \in {R}^{{\ensuremath{\mathfrak{M}}}}}$,
(iii) ${{\ensuremath{\mathfrak{M}}}\models {(\neg \phi)}[s]}$ iff not ${{\ensuremath{\mathfrak{M}}}\models {\phi}[s]}$,
(iv) ${{\ensuremath{\mathfrak{M}}}\models {(\phi \wedge \psi)}[s]}$ iff ${{\ensuremath{\mathfrak{M}}}\models {\phi}[s]}$ and ${{\ensuremath{\mathfrak{M}}}\models {\psi}[s]}$,
(v) ${{\ensuremath{\mathfrak{M}}}\models {(\phi \vee \psi)}[s]}$ iff ${{\ensuremath{\mathfrak{M}}}\models {\phi}[s]}$ or ${{\ensuremath{\mathfrak{M}}}\models {\psi}[s]}$,
(vi) ${{\ensuremath{\mathfrak{M}}}\models {(\phi \rightarrow \psi)}[s]}$ iff either not ${{\ensuremath{\mathfrak{M}}}\models {\phi}[s]}$ or else ${{\ensuremath{\mathfrak{M}}}\models {\psi}[s]}$,
(vii) ${{\ensuremath{\mathfrak{M}}}\models {(\exists v.\phi)}[s]}$ iff there is an ${a\in M}$ such that ${{\ensuremath{\mathfrak{M}}}\models {\phi}[s(^{{a}}_{{v}})]}$,
(viii) ${{\ensuremath{\mathfrak{M}}}\models {(\forall v.\phi)}[s]}$ iff for all ${a \in M}$, ${{\ensuremath{\mathfrak{M}}}\models {\phi}[s(^{{a}}_{{v}})]}$.
Now, the difference of ${=}$ and ${=}$ is obvious; the “possible confusion” will even catch the student’s eye, and this is of particular educational value. In addition, the quantifier clauses allow to distinguish better between the syntactical object variable ${v}$ and the semantical *element* ${a}$—even if ${a}$ is here, of course, a meta-variable for “real elements” of ${M}$.
Introducing now, as counterpart to the semantic consequence relation, a derivation relation ${\vdash}$ which exclusively, deals with red objects, syntactic and semantic reasoning is already distinguished by colours. It is our teaching experience, that these colours are of great help for the students, providing some kind of orientation in rather technical proofs like these for the compactness and the completeness theorem. As prime examples for the use of colours we may also mention the Skolem paradox, which, like the completeness theorem, can be regarded as a corollary of the compactness theorem.
And so the question of soundness and completeness is the next one to be addressed. Even in a compact course on Gödel’s incompleteness theorems these results of Gödel’s dissertation should be discussed.[^5] Today, they are usually not proven along the lines of Gödel’s original work, but rather with the well-known technique introduced by Henkin. This technique is standard for many completeness proofs for quite a couple of logics and, therefore, constitute an integral tool in the toolbox of the mathematical logician: it proceeds by constructing a blue world (validating the negation of a formula, that cannot be proven in the syntactic calculus) out of the red material itself. (The situation is somewhat dual to the more demanding case of (set-theoretic) forcing, where one, so to speak, bestows properties of a language on certain blue sets, thereby emulating the red world with originally blue objects.)
Arithmetic
==========
The use of the colours exhibits its real potential only, when we study Arithmetic and want to motivate, for instance, Gödel’s incompleteness theorems for Peano Arithmetic. In its original form, Gödel’s proof actually entirely refrains from the semantic realm and in some sense requires only “red” reasoning. It, indeed, is a theorem about the red world, but one that operates on a meta-level. It speaks about the axiomatic system, but it does not reason about its meaning, quite like Hilbert’s attempted consistency proofs do. And likewise it uses only finitistically acceptable means.\[Hilbert\] In [@Bea02 fn. 1, p. 337] the editors explicitly attribute Gödel’s avoiding of the concept of truth in his 1931 paper to the “respect for the ‘prejudice’ of the Hilbert school.” Further reflections can be found in [@Fef84].
But, of course, the first incompleteness theorem can—and probably should—be motivated by the question whether Peano Arithmetic is complete with respect to the *standard model*. Peano Arithmetic may be defined over the set $\{{0},{S},{+},{\cdot}\}$ of non-logical symbols. The standard model, as semantic object, can be given in blue as follows: ${{\ensuremath{\mathfrak{N}}}= \langle \mathbb{N},0,S,+,\cdot\rangle}.$ It goes without saying that the red symbols should be interpreted by their respective blue counterparts. Gödel’s first incompleteness theorem tells us, that the red world, as long as a consistent recursive axiomatic system is chosen, will miss some formula *and* its negation[^6], it cannot decide all given formulae by proof. However, the blue world so to speak decides formulae: Either a formula or its negation is satisfied (in a model by a valuation); this is a direct consequence of clause (iii) in Def. \[def1\] (see also §\[sec-int\] below). There is no gap in the blue world—but there will always be one in the (sensibly chosen, i.e., recursive and consistent) red one. This then not only entails that no red world (as above) fixes the blue world of Arithmetic as its only model—this could already be concluded from Löwenheim–Skolem—but that not all red statements that are true about the blue world can be proven in the red realm.
So, *how* are the blue objects in ${{\ensuremath{\mathfrak{N}}}= \langle
\mathbb{N},0,S,+,\cdot\rangle}$ actually given? ${\mathbb{N}}$ as the set of natural numbers ${\{0,1,2,\dots\}}$ can be taken for granted, and with it, of course, also the object ${0}$. But what *is* ${+}$? It should be, of course, the set of all triples $({a},{b},{c})$ such that ${a} +
{b} = {c}$. But this doesn’t help us, as we just moved the question to another level, writing $+$ in black (this plus sign can, of course, not be blue if we don’t want to run into an immediate circle). To avoid it, one could get back to use of dots and say that ${+}$ is the set: $$\begin{aligned}
\{&{(0,0,0)}, {(0,1,1)}, {(0,2,2)}, {(0,3,3)}, \dots
\\
& {(1,0,1)}, {(1,1,2)}, {(1,2,3)}, {(1,3,4)}, \dots \\
& {(2,0,2)}, {(2,1,3)}, {(2,2,4)}, {(2,3,5)}, \dots \\
& \qquad \vdots \hspace*{150pt} \}.\end{aligned}$$
In fact, we see only two ways out of this situation: First, we appeal to the common platonism in mathematics and stipulate that ${+}$ exists in the structure ${{\ensuremath{\mathfrak{N}}}}$ in the way we expect it and with the properties we want it to have. Or, secondly, we treat ${+}$ as a proper set-theoretic object and construct it within our favorite axiomatic set theory, let’s say $\textsf{ZFC}$.
While the second one is a solution concerning the level of Arithmetic, it simply shifts the problem to set theory. Within $\textsf{ZFC}$, the defined ${+}$ would be a red object (relative to $\textsf{ZFC}$), and we would have to ask now, how the models of $\textsf{ZFC}$ look like—which is, not surprising, a much harder question than to look for structures for Arithmetic, not only in a technical sense, but also in a philosophical one; in particular, it leads us simply back to the situation we are discussing for ${+}$ just for other, now set-theoretical, objects.
For the first option, the sheer existence of ${+}$ presupposes a certain form of platonism. This platonism allows to dispose of any further technical questions, but one enters a dangerous philosophical terrain.
We may leave aside here the preferred way out (this is a philosophical question—but one which is naturally asked in this context and can nicely be brought out and talked about with the help of our colours); what we consider as an important lesson for the teaching of logic is, that any discussion of the syntax/semantics distinction in the context of Arithmetic—and, *a fortiori*, in relation to Gödel’s Incompleteness Theorems—has to presuppose some “blue objects”. And this presupposition is far from being trivial. As a minimal conclusion, we like to state that an unreflected “identification” of, for instance, the “syntactical addition” (symbol) ${+}$ with the “semantical addition” (function) ${+}$ prevents a student from any understanding of the problem which is at issue here.
Gödel’s Incompleteness Theorems
===============================
Gödel’s Incompleteness Theorems are, without doubt, the most important results in mathematical logic. They are, however, clouded by continuing philosophical misunderstandings,[^7] and a pure formal presentation of the results can easily support such prejudices.[^8] We like to address only one problem coming from the fact that many presentations make use of the *Chinese remainder theorem*. Apparently, here proper mathematics is used to prove something about formal systems for mathematics and one may wonder whether there is some kind of vicious circle in the back.[^9] This is, of course, not the case; in fact, the use of this theorem is only relevant for a certain technical step, in particular in the case of Peano Arithmetic, and it is irrelevant for other axiomatization or codings (see [@Smo77 §3.2.6]). Hao Wang reported this from interviews with Gödel [@Wan81 p. 653]: “He enjoyed much the lectures by Furtw\[ä\]ngler on number theory and developed an interest in this subject which was, for example, relevant to his application of the Chinese remainder theorem in expressing primitive recursive functions in terms of addition and multiplication.” Thus, what is at issue here is the representation of primitive recursive functions in the formal arithmetical theory. Albeit, it is, of course, possible to develop primitive recursion *in* such theories, a proper educational approach *can* “outsource” primitive recursion.
We did that by setting up a new realm for primitive recursion (formally a functional algebra, consisting of function symbols and equalities)—and using a new colour for it, as, for the moment, it can be located outside of the red and blue world. After providing a sufficient stock of primitive recursive functions, one simply proves the representation theorem showing that every such function can be represented in the formal theory under consideration (as Peano Arithmetic, for instance). In the particular case of Peano Arithmetic, this proof may use the Chinese remainder theorem—but it is obvious that it is not relevant if one would provide another proof, in particular for other theories. In addition, it should become clear that Gödel’s theorems can be carried out in essentially the same way for all theories allowing for the appropriate representation theorem.
With respect to the Chinese Remainder Theorem and the like the diagnosis is as following: The Chinese Remainder Theorem is—as far as Gödel’s first incompleteness theorem is concerned—at the level of meta theory, but not necessarily in the blue world. It deals with the recursion-theoretic realm, concerning codings, and can be outsourced and imported if needed.
The situation changes however, if one wants to proceed to the second Incompleteness Theorem: There one repeats the proof of the first completely inside the red theory. For that case it is of vital importance that one uses codings and theorems that can be reproduced in the object theory itself—this is the sense in which it is “occasionally not just important \[in logic\] what you prove, but how you prove it” [@rautenberg2006concise p. 190].
“Semantics comes first”?
========================
The dictum “Semantics comes first” is attributed to Tarski.[^10] And it has some rationale: in first-order logic as well as in Arithmetic, we presuppose the *meanings* of the logical and arithmetical expressions before we start to manipulate them. It is, in general, seen as a task of the axiomatization to provide calculi which capture such meaning—and the completeness theorems should show us that the axiomatization was successful.
As much as completeness theorems are concerned, they give us—*a posteriori*—the possibility to identify the red with the blue objects. But, to make sense out of a completeness theorem—and to prove it—the difference of the syntactical and semantical expression has to be seen clearly. In this perspective, we could even propose that one could forget of the red and blue colour for, let say, equality, *after* the proof of the completeness theorem for first-order logic.
But there is still an issues to discuss which might call Tarski’s dictum in question.
First, the dictum makes sense only, if the semantics is *understood before* the syntax. It is doubtful that semantics could serve its purpose if it is, by itself, not fully understood. For first-order logic and Arithmetic it should be the case that their semantics is clear (although we mentioned the problem to precisely articulate it in the case of ${+}$, for instance). And this holds probably also for most of the mathematical theories we are familiar with.
The situation changed drastically, when Computer Science entered the stage. Semantics of programming languages is a rather challenging topic in CS. And here, the syntax is essentially always prior to the semantics. Instead of finding an axiomatization for a given semantics, now one looks for a semantics for a given programming language. But it is not only that such a semantics is no longer “first”, it might be the case that proposed semantics are, in a technical sense, more complicated than the programming language in itself (as example, we may refer to [@Aea11]). In some cases, one might wonder in which sense such semantics provide any *meaning*.[^11]
From the educational perspective, this only means that logic cannot any longer be taught with Tarski’s dictum as a rock solid starting point. The differentiation of syntax and semantics becomes even more significant. What it means to put “syntax first” we will discuss in the next section.
Intuitionism {#sec-int}
============
When Brouwer conceived his mathematical philosophy of *intuitionism* he had definitely nothing formal in mind. However, based on the “successful” axiomatization of intuitionistic logic by Heyting, today, intuitionism is often presented in an axiomatic context; for propositional logic, for instance, just like a calculus for classical logic, but without *tertium-non-datur*. In this way, it can be recast totally in our red world. An “additional” semantics, may it be informal or set-theoretically, would be alien to intuitionism.[^12] One could go a step further and say—instead of that there is no blue world—that the red world should be its own blue world. This is often paraphrased by saying that, from an intuitionistic perspective, the meaning of a formula is given by its proof (better: the set of all its proofs). Today, this idea is reflected in *proof-theoretic semantics*, an approach which “attempts to locate the meaning of propositions and logical connectives not in terms of interpretations …but in the role that the proposition or logical connective plays within the system of inference.”[^13]
In view of our discussion above, the formal framework of intuitionism undermines substantially Tarski’s dictum. In particular, when we go back to Definition \[def1\] of a L-structure, one can easily observe that clause (iii) for negation *builds in* classical logic into any structure. It is worth to reflect a little bit more on this clause; it looks as innocent as any other clause, that for conjunction and disjunction, for instance, “just making sure that each of our symbols means what we want it to mean.”, to repeat the citation from Barwise [@Bar77 p. 21]. But it is not only the case, that Brouwer would not agree that this is what we want negation to mean; it is the problem that the right-hand side—not ${{\ensuremath{\mathfrak{M}}}\models {\phi}[s]}$—can, in general, not practically be verified. It is the full purpose of an axiomatization to give a *positive* approach to negation, i.e., the symbol for negation ${\neg}$ has to stay always on the right-hand side of the derivation sign ${\vdash}$, and we must not make use of a “meta-negation” ${\nvdash}$. That this is possible is far from being trivial—the complexity is visible in the proof of Gödel’s completeness theorem—but it makes clear the fundamental difference between syntax (in form of an axiomatization, providing an—executable, albeit not decidable—derivation relation) and semantics, which is, in general, inherently non-constructive.
As much as it comes to Brouwer’s criticism of the usual semantic understanding of negation, it is worth to cite Bernays [@Ber79 p. 4 (our translation)], who replied to it as follows:
> As one knows, the use of the “tertium-non-datur” in relation to infinite sets, in particular in Arithmetic, was disputed by <span style="font-variant:small-caps;">L. E. J. Brouwer</span>, namely in the form of an opposition to the traditional logical principle of the excluded middle. Against this opposition it is to say that it is just based on a reinterpretation of the negation. <span style="font-variant:small-caps;">Brouwer</span> avoids the usual negation non-A, and takes instead “A is absurd”. It is then obvious that the general alternative “Every sentence A is true or absurd” is not justified.
Following Bernays, the controversy about classical and intuitionistic negation could even be boiled down to a question of words. What is of importance for us here, is that the semantic determination of (classical) negation closes the road to understand Brouwer’s (intuitionistic) negation. To say it differently: the “standard approach” of teaching logic based on Tarski’s dictum (used for Tarskian semantics as given in Def. \[def1\]) blocks a proper transition to intuitionistic logic, at least, if the students are “semantically biased”. As a matter of fact, intuitionistic logic can—and maybe: has to—be developed purely in the red world.[^14] It is not our aim to motivate an approach that exclusively focusses on intuitionism or the like.[^15] But it is important that an introduction to classical logic should not come with a semantical bias which blocks the road to non-classical logic from the start.[^16] Here the colours may help.
Conclusion
==========
It is our experience from courses on mathematical logic—regular university classes, block courses, and summer school courses—that the use of colours gives the students, indeed, a “tool” at hand to see better the fundamental distinction of syntax and semantics.[^17] Although the use of colours requires some additional effort in the preparation of the student material, it is a rather inexpensive investment to obtain significant educational added value. At a certain stage—for instance, as mentioned, after the proof of the completeness theorem—, when the student has already internalized the difference of syntax and semantics one can even easily go back to the common identification of the syntactical and semantical expression using them just in black; it provides even a good test template, as one can ask the students in all kind of instances whether a certain expression would have to be coloured red or blue.
In addition, while the colouring can, of course, not solve any philosophical problem concerning syntax and semantics, it helps to sharpen the sensibility for the underlying problems. In any case, the success of a logic course will not only depend on the use of tools, but on the interaction between teacher, tutor(s) and students. The use of a new dimension—like colour—underlines the importance of the distinction between syntax and semantics, it brings various (philosophical) topics into focus and gives a convenient way of talking about these issues—a way that often is readily accepted by the students.
The message of this paper is not: By all means use colours to distinguish syntax from semantics! But rather: Distinguish syntax and semantics clearly and discuss some of the issues related to that distinction—the use of colour will help you doing so!
[^1]: Here a disclaimer is in order. In the following we don’t intend to criticise Barwise’s presentation; first, the citation is taken from a handbook article and not from a textbook (and we may add: from an excellent article in an excellent handbook); secondly, mathematically there is no issue here, the definition is surely the correct one; we are just looking for aspects which are of interest in an educational perspective.
[^2]: And it might be doubted, whether the difficulty at hand is best couched in terms of use and mention. It seems to us rather clear, however, that it is educationally not best dealt with by introducing subtleties of quotation theory.
[^3]: The particular choice of the colours is arbitrary; it is, however, known that one should avoid yellow and green for beamer presentations.
[^4]: For sure, as one referee pointed put, colouring these expressions does not remove entirely the possibility of misunderstanding: The functional expression ${t_1}^{{{\ensuremath{\mathfrak{M}}}}}$, for instance, takes the syntactic argument ${t_1}$ and evaluates it relative to the semantic model ${{\ensuremath{\mathfrak{M}}}}$—so the value of the whole expression of course denotes a blue item. The interpretation function itself does not appear in this notation, and therefore, luckily, we do not have to decide on its colour. However, with interested students we had some interesting discussions about its colour.
[^5]: In introductory logic courses in philosophy seminars these results will usually not be proven, but, of course, deserve at least to be mentioned and discussed.
[^6]: Of course on pain of inconsistency it should *not* prove any formula together with its negation.
[^7]: For a detailed discussion, in particular in view of Gentzen’s later results, we refer to [@Kah15].
[^8]: An extreme case was reported from a philosophical seminar: the participants, after studying carefully—and probably painfully—a formal presentation of Gödel’s proof, came to the conclusion that, yes, after all the results seem to be correct *if one has a language with 7 symbols; but, of course, one doesn’t know whether it still holds when one would have 8 symbols*.
[^9]: The first author remembers an elaboration of a Math student of a course presentation of Gödel’s theorems which focussed nearly entirely on the use of the Chines remainder theorem; apparently it was the only part of the proof the student was properly understanding based on his Math classes.
[^10]: The well-known computer software package *Tarski’s World* [@BPBE08] illustrates this slogan quite appropriately when the student can literally manipulate the semantical objects.
[^11]: A colleague expressed this in a sarcastic statement: “They put these brackets $\llbracket \cdot \rrbracket$ around something and believe they have achieved something.”
[^12]: The first author remembers a remark of an author to the effect that intuitionism was not succeeding due to the lack of a proper semantics, until this deficit was cured by the introduction of Kripke semantics. In our view, this sounds like claiming that a protestant church is not a church as long as it doesn’t have a pope.
[^13]: `http://en.wikipedia.org/wiki/Proof-theoretic_semantics`, accessed Febr. 20, 2015. For more on proof-theoretic semantics, see [@KSH06; @SH14].
[^14]: As said at the beginning, this concerns intuitionistic logic based on Heyting’s axiomatization; Brouwer’s original intuitionism, in particular if considered in the mathematical area of Analysis, should not come in a formal livery at all.
[^15]: It is worth mentioning that *dialogical logic* has a rather interesting status concerning the syntax and semantics distinction, which doesn’t seem to be properly explored yet, see Kahle [@Kah08; @Kah15l].
[^16]: For instance, substructural logics are best approached via proof theory, because—quite like intuitionism—there lies their motivation; on the other hand their semantics usually turn out to be rather complicated.
[^17]: We have, of course, not developed an empirical study to “prove” that the use of colours improves the teaching; it is rather an informed impression backed by feedback from our students. And although we do not have statistically valid data comparing the effects of this method with other approaches, the test results of the courses were very positive (but we had few, but rather good and understanding students).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate effects of pairing and of quadrupole deformation on two sorts of nuclear excitations, $\gamma$-vibrational $K^{\pi}=2^+$ states and dipole resonances (isovector dipole, pygmy, compression, toroidal). The analysis is performed within the quasiparticle random-phase approximation (QRPA) based on the Skyrme energy functional using the Skyrme parametrization SLy6. Particular attention is paid to i) the role of the particle-particle (pp) channel in the residual interaction of QRPA, ii) comparison of volume pairing (VP) and surface pairing (SP), iii) peculiarities of deformation splitting in the various resonances. We find that the impact of the pp-channel on the considered excitations is negligible. This conclusion applies also to any other excitation except for the $K^{\pi}=0^+$ states. Furthermore, the difference between VP and SP is found small (with exception of peak height in the toroidal mode). In the low-energy isovector dipole (pygmy) and isoscalar toroidal modes, the branch $K^{\pi}=1^-$ is shown to dominate over $K^{\pi}=0^-$ one in the range of excitation energy $E <$ 8–10 MeV. The effect becomes impressive for the toroidal resonance whose low-energy part is concentrated in a high peak of almost pure $K^{\pi}=1^-$ nature. This peculiarity may be used as a fingerprint of the toroidal mode in future experiments. The interplay between pygmy, toroidal and compression resonances is discussed, the interpretation of the observed isoscalar giant dipole resonance is partly revised.'
author:
- 'A. Repko$^{1}$, J. Kvasil$^{2}$, V.O. Nesterenko$^{3,4}$, and P.-G. Reinhard$^{5}$'
title: Pairing and deformation effects in nuclear excitation spectra
---
Introduction {#intro}
============
Pairing is known to play an important role in low-energy nuclear excitations [@SSS; @Ri80]. For the ground state, it is usually treated within the Bardeen-Cooper-Shrieffer (BCS) scheme or Hartree-Fock-Bogoliubov (HFB) method, see e.g. [@SSS; @Ri80; @Do84; @Do96; @Re97; @Be00; @Ta04]. Nuclear excitations are described with the Quasiparticle Random Phase Approximation (QRPA) where the pairing-induced particle-particle (pp) channel is added to the particle-hole (ph) residual interaction [@SSS; @Ri80; @Te05; @Li08]. During last decades, both static and dynamical pairing aspects were thoroughly explored in schematic and self-consistent models, see e.g. [@SSS; @Do84; @Do96; @Re97; @Be00; @Ta04; @Te05; @Li08; @Iud05; @Se08; @Te10; @Te11g]. The pp-channel was found important in electric modes with positive parity (monopole, quadrupole) where it generates pairing vibrations, see e.g. [@SSS; @Ri80; @Li08; @Iud05]. In deformed axial nuclei, the pp-channel is important in states with $K^{\pi}=0^+$ where $K$ is the projection of the angular momentum [@SSS; @Iud05; @Te10; @Te11g].
In spite of extensive studies, some pairing effects in nuclear excitations still deserve further exploration, especially in deformed nuclei. For example, it is yet unclear if we indeed need the pp-channel in the self-consistent description of excitations with $K^{\pi} \ne
0^+$. These excitations include all negative-parity states (dipole, octupole, ...) as well as positive-parity states with $K>0$. Clarification of this point is important because the pp-channel considerably complicates QRPA calculations. A further point of interest is to check the influence of surface pairing (SP) as compared to volume pairing (VP).
These two problems are addressed in the present study for the case of the standard monopole $\delta$-force pairing. The calculations are performed within the self-consistent QRPA method [@Re16] with the Skyrme force SLy6 [@Cha97]. As typical examples we consider $^{152,154,156}$Sm which are well deformed having still axially symmetric shape. In a first step, we inspect the lowest $\gamma$-vibrational states with $K^{\pi}=2^+$. The previous systematic self-consistent explorations of $\gamma$-vibrational states were performed with [@Te11g] and without [@Ne16g] the pp-channel. We will show that the effect of pp-channel is very small in these modes. At least it is much weaker than the difference between the results obtained with VP and SP.
In a second step, we inspect exotic parts of the dipole excitation spectrum which attract high attention presently (see, e.g., [@Pa07; @Kv11; @Re13]), namely low-lying isovector strength also coined pygmy dipole resonance (PDR), toroidal dipole resonance (TDR), and compression dipole resonance (CDR). These modes are located, fully or partly, at rather low energies and so may be affected by pairing. Here again we have found the pp-channel negligible. Altogether, our analysis suggests that the pp-channel can be ignored in QRPA calculations for [*all* ]{}excitations with $K^{\pi} \ne 0^+$.
The next point of our study concerns deformation effects in dipole excitations. We analyze the ordinary isovector giant dipole resonance (GDR), PDR, and isoscalar TDR and CDR. The TDR and CDR are often considered as low- and high-energy parts of the isoscalar giant dipole resonance (ISGDR) [@Pa07; @Har]. In axial nuclei, all these resonances exhibit a deformation splitting into $K=0$ and $K=1$ branches. However, the features of these branches depend very much on the actual type of resonance. For example, our previous studies have shown that the main low-energy peak of the TDR is, somewhat surprisingly, dominated by $K=1$ strength instead of $K=0$ one as would have expected [@Kv13Sm; @Kv14Yb; @Ne17]. In fact, the TDR produces the opposite sequence of $K$-branches as compared to the GDR. Since the TDR shares the same energy region as the PDR and is, in fact, the source of the PDR [@Re13], this anomaly becomes especially interesting because it can be related to the deformation splitting of the PDR. In this connection, we propose here a further comparative analysis of the deformation splitting of various dipole modes with the main accent to the low-energy region embracing PDR and TDR.
The paper is organized as follows. In Sec. 2, the calculation scheme is outlined paying particular attention to the treatment of pairing. The main factors determining the impact of the pp-channel are discussed comparing analytically the pp- and ph-channels. In Sec. 3, results of the QRPA calculations for $K^{\pi}=2^+$ states and dipole modes are presented and discussed. Some experimental perspectives are commented. In Sec. 4, conclusions are given. In Appendix A, the Skyrme functional used in this paper is sketched. In Appendix B, the actual HF+BCS scheme is outlined. In Appendix C, the basic QRPA equations with pp-channel are presented.
Calculation scheme and theoretical background {#sec-1-theor}
=============================================
Method and numerical details {#subsec21}
----------------------------
The calculations are performed within QRPA [@Ri80] based on the Skyrme energy functional [@Skyrme; @Vau; @Be03] $$\mathcal{E}(\rho,\tau, {\bf J}, \vec{j}, \vec{s}, \vec{T}, \tilde{\rho})
= \mathcal{E}_{\rm kin} + \mathcal{E}_{\rm Sk}
+ \mathcal{E}_{\rm Coul} + \mathcal{E}_{\rm pair}
\label{Efunc}$$ including kinetic, Skyrme, Coulomb and pairing parts. The Skyrme part $\mathcal{E}_{\rm{Sk}}$ depends on the following local densities and currents: density $\rho(\vec{r})$, kinetic-energy density $\tau(\vec{r})$, spin-orbit density $\mathbf{J}(\vec{r})$, current $\vec{j}(\vec{r})$, spin density $\vec{s}(\vec{r})$, and spin kinetic-energy density $\vec{T}(\vec{r})$. The Coulomb exchange term is treated in Slater approximation. The pairing functional is derived from monopole contact $\delta$-force interaction [@Be00]. This functional depends on the pairing density $\tilde{\rho}(\mathbf{r})$ and, for the surface pairing, also on the normal density $\rho(\mathbf{r})$. More details on the functional (\[Efunc\]) can be found in the next subsection and in Appendix A.
The mean-field Hamiltonian and the QRPA residual interaction are determined through the first and second functional derivatives of the total energy (\[Efunc\]). The approach is fully self-consistent since: i) both the mean field and residual interaction are obtained from the same Skyrme functional, ii) time-even and time-odd densities are involved, iii) the residual interaction includes all the terms of the initial Skyrme functional as well as the Coulomb direct and exchange terms, iv) both ph- and pp-channels in the residual interaction are taken into account. Details of our mean field and QRPA schemes are given in Appendices B and C. The present version of the Skyrme QRPA scheme is implemented in the two-dimensional (2D) QRPA code [@Re16].
The calculations are performed with the Skyrme force SLy6 [@Cha97] which was successfully used earlier for a systematic exploration of the GDR in rare-earth, actinide and superheavy nuclei [@Kl08]. The QRPA code employs a 2D mesh in cylindrical coordinates. The calculation box extends over three times the nuclear radius. The mesh size is 0.4 fm. The single-particle spectrum taken into account embraces all levels from the bottom of the potential well up to +30 MeV. The (axial) equilibrium quadrupole deformation is determined by minimization of the total nuclear energy. This yields the deformation parameters $\beta$ = 0.311, 0.339, and 0.361 for $^{152,154,156}$Sm, respectively, which are sufficiently close to the experimental values $\beta_{\rm exp}$ = 0.308 for $^{152}$Sm and $\beta_{\rm exp}$ = 0.339 for $^{154}$Sm [@bnl_exp].
Pairing in the ground state is treated with the HF+BCS scheme, see next subsection. To cope with the principle divergency of short-range forces, we implement in the HF+BCS equations and pairing transition densities a soft, energy-dependent cutoff factor [@Be00] $$f^q_i
=
\frac{1}
{1+\exp\Big[\frac{e^q_i-\lambda_{q}-\Delta E_q}{\eta_{q}}\Big]}
\label{f-cut}$$ where $q=\{n,p\}$, $i$ denotes single-particle states, $e^q_i$ is the single-particle energy, and $\lambda_q$ is the chemical potential. The cutoff parameter $\Delta E_q$ and width $\eta_{q}=\Delta E_{q}/10$ are chosen self-adjusting to the actual level density in the vicinity of the Fermi energy [@Be00]. Here $\Delta E_q$ lies around 6 MeV. In the pp-channel, two-quasiparticle (2qp) states $(i,j)$ are limited to those with $\sqrt{f^q_if^q_j} > 10^{-4}$ to remove high-energy particle states. More information on the pairing formalism is given in the next subsection.
To minimize the calculational expense in ph-channel, we employ only states with $|u_iv_j+v_iu_j| > 10^{-4}$, where $u_i,v_i$ are coefficients of the special Bogoliubov transformation. This cut-off still leaves the 2qp configuration space large enough to exhaust the energy-weighted sum rules (EWSR). Note also that a large configuration space in ph-channel is crucial for a successful description of low-energy vibrational excitations [@Ne16g] and extraction of the spurious admixtures [@Re16; @Rei92d]. For example, in calculations of $\gamma$-vibrational modes in $^{154}$Sm, we deal with $\approx
6300$ proton and $\approx 12500$ neutron pairs and exhaust the isoscalar quadrupole $\mathrm{EWSR}=(\hbar e)^2/(8\pi m)\cdot 50A\langle
r^2\rangle_A$ by $\approx$ 97 $\%$. The Thomas-Reiche-Kuhn sum rule for dipole excitations is exhausted in the interval 2–45 MeV by $\approx$ 94 $\%$.
Treatment of pairing {#subsec22}
--------------------
The pairing part in the functional (\[1\]) reads $$\mathcal{E}_{\rm pair} = \frac{1}{4} \: \sum_{q=n,p} V_q \: \int \mathrm{d}\vec{r} \:
|\tilde{\rho}_q (\vec{r})|^2 \: G_q(\vec{r})
\;.
%\Big( 1 - \eta \: \Big(\frac{\rho(\vec{r})}{\rho_0}\Big)^{\gamma} \Big)
\label{Epair}$$ It is motivated by the zero-range pairing interaction [@Be00] $$V^q_{\rm pair}(\vec{r},\vec{r}') = G_q(\vec{r})\delta(\vec{r}-\vec{r}')
\label{Vpair}$$ with $$G_q(\vec{r})= V_q \Big[ 1 - \eta \:
\Big(\frac{\rho(\vec{r})}{\rho_0}\Big)^{\!\gamma}\Big] .
\label{Gpair}$$ Here $\rho(\vec{r})=\rho_p(\vec{r})+\rho_n(\vec{r})$ is the sum of normal proton and neutron densities, $\tilde\rho_q(\vec{r})$ are the pairing densities, $V_q$ are pairing strength constants, and $\rho_0$=0.16 ${\rm fm}^{-1}$ is the saturation density of symmetric nuclear matter. The coefficient $\gamma$ regulates the density dependence. Following common practice, we fix here $\gamma$=1. We obtain so-called volume pairing (VP) for $\eta$=0 and the (density-dependent) surface pairing (SP) for $\eta$=1. Of course, $\eta$ may be varied freely and take values in between 0 and 1, when optimized properly [@Klu09a]. We consider here VP and SP as limiting cases to explore the sensitivity to the pairing model.
Pairing is treated within the HF+BCS scheme described in Appendix B. In this scheme, the nucleon and pairing densities in axial nuclei read [@Be03] $$\begin{aligned}
\label{rho}
\rho_q(\vec{r})&=& 2 \sum_{i \epsilon q} v_i^2 |\psi_i (\vec{r})|^2 ,
%\: \rho(\vec{r})=\rho_p(\vec{r})+\rho_n(\vec{r})
\\
\tilde\rho_q(\vec{r}) &=& -2 \sum_{i \epsilon q} f_i^q v_i u_i |\psi_i (\vec{r})|^2
\label{rho_pair}\end{aligned}$$ where $\psi_i (\vec r)$ are single-particle wave functions and $f^q_i$ is the energy-dependent cut-off weight (\[f-cut\]).
Pairing induces the pp-channel in the QRPA residual interaction, see Appendix C. Its contributions to the QRPA matrices A and B are given by expressions (\[C.1\])-(\[C.4\]). For $\gamma$=1, these contributions read: $$\begin{aligned}
A^{\rm pair}_{qq'}
(i{\bar j}, i'{\bar j}')
&=&
\frac{1}{2} \delta_{q q'} V_q \int \mathrm{d}\vec{r}
\Big[ 1 - \eta \frac{\rho({\vec r})}{\rho_0} \Big]
\delta\tilde{\rho}^q_{i{\bar j}}(\vec{r})
\delta\tilde{\rho}^{q'}_{i'{\bar j}'}(\vec{r})
\nonumber
\\
&-& \frac{\eta}{2 \rho_0} V_q
\int \mathrm{d}\vec{r} \tilde{\rho}_q(\vec{r})
\delta\tilde{\rho}^q_{i\bar{j}}(\vec{r})
\delta{\rho}^{q'}_{i'\bar{j}'}(\vec{r})
\nonumber\\
&-& \frac{\eta}{2 \rho_0} V_{q'}
\int \mathrm{d}\vec{r} \tilde{\rho}_{q'}(\vec{r})
\delta\rho^q_{i\bar{j}}(\vec{r})
\delta\tilde{\rho}^{q'}_{i'\bar{j}'}(\vec{r})
\label{RPA_App}\end{aligned}$$ with $$B^{\rm pair}_{i {\bar j} i' {\bar j}'} = -A^{\rm pair}_{i {\bar j}\: i' {\bar j}'} .
\label{B}$$ Here we assume that $i{\bar j} \in q, \: i'{\bar j}' \in q'$. Furthermore $$\delta\rho^q_{i{\bar j}}(\vec{r})=
(u_i v_j + u_j v_i) \psi^*_i(\vec{r})\psi_j(\vec{r})
\label{TDn}$$ $$\delta\tilde{\rho}^q_{i{\bar j}}(\vec{r})= \sqrt{f^q_i f^q_j}\:
(u_i u_j - v_j v_i) \psi^*_i(\vec{r})\psi_j(\vec{r})
\label{TDp}$$ are 2qp nucleon and pairing transition densities.
The contributions (\[RPA\_App\])-(\[B\]) constitute the pairing-induced pp-channel. The first term in (\[RPA\_App\]) exists for both VP and SP. It is diagonal in $q$ since we do not consider a proton-neutron pairing. The next two terms which mix nucleon and pairing density are caused by the density-dependence of SP and so exist only in SP case. In the present study we consider only quadrupole states $K^{\pi}
= 2^+$ and dipole states. Then the transition densities (\[TDn\]) and (\[TDp\]) include only non-diagonal 2qp configurations with $i \ne j$.
It is instructive to compare the contributions to QRPA matrices from ph- and pp-channels. The ph-contribution to the A-matrix from the Skyrme functional is dominated by [@Ne06] $$\frac{\delta\mathcal{E}_{\rm Sk}}{\delta\rho^q \delta\rho^{q'}}= b_0 -b_0'\delta_{qq'}$$ where $b_0$ and $b_0'$ are coefficients of the Skyrme interaction given in the Appendix A. This results in $$A^{\rm ph}_{qq'}
(i{\bar j}, i'{\bar j}')
=
(b_0 -b_0'\delta_{qq'}) \int d\vec{r}
\delta\rho^q_{i{\bar j}}(\vec{r})
\delta\rho^{q'}_{i'{\bar j}'}(\vec{r})$$ to be compared for $q=q'$ with the pp-contribution (\[RPA\_App\]). For states $i$ and $j$ near the Fermi level, the transition densities (\[TDn\]) and (\[TDp\]) are of the same order of magnitude. Then the ratio between ph and pp contributions is mainly determined by the ratio of the channel strengths $$R_q = \frac{2(b_0 -b_0')}{V_q} .$$ For SLy6, $b_0=-3502.3$ MeV, $b_0'=-3285.3$ MeV. Further, $V_p=-298.8$ MeV fm$^3$, $V_n=-288.5$ MeV fm$^3$ for VP and $V_p=-1053.12$ MeV fm$^3$, $V_n=-864.2$ MeV fm$^3$ for SP [@Fle04]. Then the ratios are $R_n=1.50$, $R_p=1.45$ for VP and $R_n=0.50$, $R_p=0.41$ for SP. In average, $R_q \sim 1$, i.e. Skyrme amplitudes in ph-channel and pairing amplitudes in pp-channel are similar (though, for SP, one may expect somewhat larger effect of pp-channel than for VP).
The observations that $\rho^q_{ij} \sim \tilde\rho^q_{ij}$ and $R_q
\sim $1 do not yet suffice to conclude that ph- and pp-channels are of the same order of magnitude. We should also take into account that in ph-channel the main interaction is the proton-neutron one (with the amplitude $b_0$). This interaction is much stronger than its proton-proton and neutron-neutron counterparts characterized by the amplitudes $b_0-b_0'$. Indeed, for SLy6, $b_0/(b_0-b_0')=16.1$. Pairing, on the other hand, has no neutron-proton interaction at all. Besides, in the pairing pp-channel, the contribution of 2qp configurations with high-energy states is suppressed by the cut-off (\[f-cut\]). So in general one may expect that the ph-channel is much stronger than the pp-channel.
Note that the above qualitative analysis mainly concerns states with $K^{\pi} \ne 0^+$. For $K^{\pi} = 0^+$ states, the situation becomes much more complicated. Here the diagonal 2qp configurations $i
\bar{i}$, resulting in the pairing vibrations, come into play. Thus the pp-channel, even being still weaker than its ph-counterpart, becomes important for low-energy excitations [@SSS; @Ri80].
QRPA strength functions {#subsec23}
-----------------------
The energy-weighted strength function for electric transitions of multipolarity $\lambda\mu$ between QRPA ground state $|{\rm gr}\rangle$ and excited state $|\nu\rangle$ has the form $$S_X(E\lambda\mu, T; E) = \sum_{\nu} E_{\nu} \: \big| \langle \nu|\:
\hat{M}^X_{\lambda \mu}(T) \: |{\rm gr} \rangle \big|^2 \: \xi_{\Delta}(E-E_{\nu})
\label{13}$$ where $E$ is the excitation energy, $\hat{M}^X_{\lambda \mu}(T)$ is the transition operator, $X=\{\rm el, com, tor\}$ marks the type of the excitation (isovector E1(T=1), compression isoscalar E1(T=0), toroidal isoscalar E1(T=0)), $E_{\nu}$ is the energy of the QRPA state. For excited states $K^{\pi}_{\nu}$, we assume $K=\mu$, $\pi=(-1)^{\lambda}$. The total strength functions read $$S_X(E\lambda, T; E) = \sum_{\mu \ge 0 } (2-\delta_{\mu,0}) S_X(E\lambda\mu, T; E) .
\label{13a}$$ The strengths are weighted by the Lorentz function $$\xi_{\Delta}(E-E_{\nu}) = \frac{1}{2 \pi}
\frac{\Delta (E)}{(E-E_{\nu})^2 + [\Delta(E)/2]^2}
\label{17}$$ with the energy-dependent folding [@Kv13Sm] $$\Delta(E) =
\left\{
\begin{array}{ll}
\Delta_0 & \:{\rm for} \:\:\: E\leq E_0, \\
\Delta_0 + a\: (E-E_0) &\: {\rm for} \:\:\: E> E_0.
\end{array} \right.
\label{18}$$ The weight $\xi_{\Delta}(E-E_{\nu})$ is used to simulate escape width and coupling to the complex configurations. Since these two effects grow with increasing the excitation energy, we employ an energy dependent folding width $\Delta(E)$. The parameters $\Delta_0$=0.1 MeV, $E_0$=8.3 MeV and $a=$0.45 are chosen so as to reproduce the smoothness of the experimental photoabsorprion in $^{154}$Sm. The value $E_0$ roughly corresponds to the neutron and proton separation energies in $^{154}$Sm: $S_n$=8.0 MeV and $S_p$=9.1 MeV [@bnl_exp].
The operator for the electric $E1\mu (T=1)$ transition reads $$\hat{M}^{el}_{1\mu}(T=1) = e \sum_{i=1}^{A}
e^{\rm eff}_i \: r_i Y_{1\mu}(\Omega_i)
\label{14}$$ where $Y_{1\mu}(\Omega_i)$ is the spherical harmonic and $e^{\rm eff}_i$ are effective charges equal to $N/A$ for protons and $-Z/A$ for neutrons. The operators for the compression and toroidal $E1\mu (T=0)$ transitions are [@Kv11] $$\label{Tcom}
\hat{M}^{\rm com}_{1\mu} (T=0)
= \frac{e}{10} \:
\sum_{i=1}^{A} [ r_i^3
%-\kappa_\mu\langle r^2(3\cos^2\theta-1) \rangle_0 \nonumber\\
%&&
%\qquad\qquad\qquad
-\frac{5}{3}\langle r^2 \rangle_0 r_i ]
Y_{1 \mu}(\Omega_i) ,$$ $$\begin{aligned}
&&
\hat{M}^{\rm tor}_{1\mu}(T=0) =
- \frac{1}{2c\sqrt{3}}
\int d\vec{r} \:
\hat{{\bf j}}_{\rm nuc}(\vec{r}-\vec{r}_i)
\nonumber \\
&&
\cdot
\big[
\frac{\sqrt{2}}{5} r^2_i {\bf Y}_{12\mu}(\Omega_i)
+ (r^2_i - \langle r^2 \rangle_0)
{\bf Y}_{10\mu}(\Omega_i)
\big]
\label{15}\end{aligned}$$ where ${\bf Y}_{10\mu}(\Omega_i)$, ${\bf Y}_{12\mu}(\Omega_i)$ are vector spherical harmonics; $
\hat{\vec{j}}_{\rm nuc}(\vec{r}-\vec{r}_i)
= -ie\hbar/(2m) \sum_{i=1}^A [\delta(\vec{r}-\vec{r}_i)\vec{\nabla}_i
+\vec{\nabla}_i\delta(\vec{r}-\vec{r}_i)]
$ is operator of the isoscalar convection nuclear current. The terms with the squared ground-state radius $\langle r^2 \rangle_0$ represent the center-of-mass-corrections. Note that the compression operator (\[Tcom\]) is also the transition (probe) operator for ISGDR [@Har].
The photoabsorption [@Ri80] is related to the strength function (\[13a\]) as $$\sigma (E) \: [{\rm fm}^2] \: \approx 0.402 \: S_{\rm el}(E1, T=1;\:E) .
\label{16}$$
Results and discussion {#sec-3}
======================
$\gamma$-vibrational states {#subsec31}
---------------------------
Figs. \[fig:1\]-\[fig:4\] collect results of our calculations for $\gamma$-vibrational states in $^{152,154,156}$Sm. Fig. \[fig:1\] shows energies and $B(E2;\: gr \rightarrow I^{\pi}K=2^+2)$ values computed with and without the pp-channel for VP and SP cases. The reduced transition probabilities are obtained with the effective charges $e^{\rm eff} =1$ for protons and $e^{\rm eff} =0$ for neutrons. Fig. \[fig:1\] indicates that the present QRPA description is not fully satisfying: the calculations essentially overestimate the experimental energies in $^{152,154}$Sm and $B(E2)$ in $^{154}$Sm. Note that such significant overestimation of energies of $\gamma$-vibrational states in $^{152,154}$Sm has been already observed in previous self-consistent Skyrme QRPA calculations [@Te11g; @Ne16g]. It was shown [@Ne16g] that this problem can be partly solved by taking into account the pairing blocking effect and using a more appropriate Skyrme parametrization. Besides that, the coupling with complex (two-phonon) configurations could be here important [@SSS].
As seen from Fig. \[fig:1\], VP and SP results are rather similar though VP performance is somewhat better. What is remarkable, for both VP and SP, the impact of the pp-channel is very small, which actually means that the pp-channel is much weaker than the ph-one (at least its impact is much smaller than the difference between VP and SP results).
This conclusion can also be drawn from Fig. \[fig:2\] which shows the pairing and nucleon transition densities (TD) for the $\gamma$-vibrational state ($\lambda\mu\nu$ = 221) in $^{154}$Sm, using VP. We see that, at the nuclear surface, nucleon TD $\delta\rho^q_{221}$ are much larger than pairing TD $\delta\tilde{\rho}^q_{221}$, especially for neutrons. The same takes place in Fig. \[fig:3\] for the case of SP.
The next question is: what is the origin of the difference in the nucleon and pairing one-phonon TD? To clarify this point, we should numerically analyze the main ingredients of the contributions (\[RPA\_App\])–(\[B\]) of the pp-channel to the RPA matrices. First of all, let’s consider nucleon (\[TDn\]) and pairing (\[TDp\]) TD for the main 2qp component of the $\gamma$-vibrational state. This is the proton $F,F+3$ configuration $pp[413\downarrow-411\downarrow]$ (here we use asymptotic Nilsson quantum numbers, $F$ denotes the Fermi level). The TD for this configuration are shown in Fig. \[fig:4\] for the case of VP. It is seen that the nucleon and pairing TD are very similar. This is not surprising since both TD have the same coordinate dependence and deviate only by the pairing factors which are of the same order of magnitude for the states near the Fermi level. As shown in Sec. \[subsec22\], the strength coefficients of the pp- and ph-channels are also similar. So indeed, following the discussion in Sec. \[subsec22\], the weakness of the pp-channel is perhaps mainly caused by the absence of the neutron-proton interaction (which plays a dominant role in ph-channel).
The conclusion that the pp-channel plays a negligible role can, in principle, be generalized to arbitrary states with $K^{\pi}
\ne 0^+$ in even-even nuclei, e.g. to all states of negative parity. This allows to ignore the pp-channel from the onset and thus considerably decrease the numerical expense of QRPA.
Dipole excitations {#subsec32}
------------------
The previous subsection concluded that pp-channel has negligible impact for low-energy $K^{\pi} \ne 0^+$ states. This should be even more the case for excitations with higher energies, e.g. for the various dipole resonances. For these modes, the pairing gap $\Delta_q
\sim$ 2 MeV is much smaller than the resonance energies and so the impact of pairing should be even weaker than for low-energy excitations. Note that low-energy excitations are composed from 2qp pairs whose single-particle states $\{i,j\}$ are placed near the Fermi level and so are strongly affected by pairing. Instead, in high-energy excitations, at most one single-particle state from the 2qp pair can originate from the Fermi level region. Thus in general high-energy 2qp excitations are much less affected by pairing than low-energy ones and so, in principle, should be only weakly affected by the pairing-induced pp-channel. For the same reason, the difference between VP and SP in the resonance region is also expected to be negligible. This will be confirmed below by our QRPA calculations for GDR, PDR, TDR and CDR in $^{154}$Sm. In this subsection, we will consider peculiarities of the deformation splitting of the dipole resonances in this nucleus. The main attention will be paid to a comparison of their $K=0$ and $K=1$ branches in the low-lying energy interval 3–9 MeV, embracing PDR and TDR. In what follows, it is assumed that $K=1$ branch includes both $K=\pm $1 contributions.
Use of a large configuration space allows to keep the spurious center-of-mass mode at low energies $<2$ MeV. The elimination of this mode from the dipole strength functions is additionally ensured by using appropriate effective charges and correction terms, see Sec. 2.3. In the following, we discuss dipole strength functions only for the excitation energy $E > $ 3 MeV, which is at the safe side with respect to the center-of-mass mode.
Fig. \[fig:5\] compares the QRPA photoabsorption strength with the experimental data [@GDRexp]. We see an excellent agreement between the theory and experiment. As said above, an appropriate modeling of the GDR shape is achieved by energy-dependent averaging described in Sec. 2.3. However, a correct description of peak position and, in a large extent, of fragmentation widths is achieved by QRPA with the parametrization SLy6. Altogether, the obtained agreement validates our approach for the further studies. Fig. \[fig:5\] shows that VP and SP give almost identical results. The pp-channel does not have any effect (not shown since curves with and without the pp-channel almost coincide). The GDR exhibits deformation splitting into $K=0$ and $K=1$ branches. The sequence of branches is typical for GDR in prolate nuclei: $K=0$ branch lies below of $K=1$ one. Though $K=0$ branch involves about twice less strength than $K=1$ one, its peak height is a bit larger. This is explained by a stronger fragmentation of the $K=1$ branch, which broadens the distribution.
Fig. \[fig:6\] shows the calculated photoabsorption in the PDR region. Following plots c)-d), the influence of the pp-channel is very weak for VP and SP (though for SP there are some small differences for the structures around 8 MeV). In the plot b), results from VP and SP deviate only in detail. Plot a) shows that $K=0$ strength exceeds $K=1$ strength at $E \ge$ 8 MeV but becomes smaller at $E \le$ 8 MeV. A similar effect was found for Mo isotopes [@Mo]. This is the result of the competition between GDR tails from $K=0$ and $K=1$ branches.
Fig. \[fig:7\] shows the QRPA isoscalar dipole strength (ISGDR) as a fraction of the computed energy-weighted sum rule EWSR(E1,T=0)=$\int dE S_{\rm com}(E1, T=0; E)$ (a similar presentation but in terms of experimental EWSR is used for experimental data [@CDRexp]). We see that, in agreement with experiment, the computed ISGDR includes two broad bumps at 10-17 and 17-40 MeV. Similar bumps were also observed in other (spherical) nuclei [@CDRexp; @Young90Zr; @Uch04]. They are often treated as TDR and CDR respectively, see e.g. [@CDRexp]. However, our previous studies [@Re13; @Kv13Sm; @Kv14Yb; @Ne17; @NeDre14] for different nuclei and present analysis for $^{154}$Sm suggest that both broad bumps in ISGDR belong to CDR. Comparison of Figs. \[fig:7\] and \[fig:9\] shows that even spikes at 4–6 MeV can hardly be considered as being toroidal because they are almost absent in the toroidal distributions shown in Figs. \[fig:9\] and \[fig:10\].
Fig. \[fig:7\] demonstrates that strengths obtained with VP and SP are almost identical. The effect of the pp-channel for the strength at $E >$ 10 MeV is negligible (not shown). Concerning the deformation effects, $K=0$ and $K=1$ branches show the familiar sequence (like in GDR) at E=10-40 MeV, namely that the $K=0$ branch lies below the $K=1$ branch.
The ISGDR strength at lower energies is shown in Fig. \[fig:8\]. Plot a) demonstrates that $K=1$ strength dominates at E=8–10 MeV but becomes weaker at 3–8 MeV. So, at 3–8 MeV, the ratio of $K=0$ and $K=1$ strengths is opposite to that in PRD. Plots b)-d) of Fig. \[fig:8\]) also show very close results for VP as compared to SP and an almost negligible effect of the pp-channel.
Fig. \[fig:9\] shows the toroidal isoscalar strength function obtained with the transition operator (\[15\]) as a fraction of the integral toroidal strength (EWSR). We see that the strength has a broad plateau at 10–40 MeV. Most probably such a broad distribution is caused by the coupling between TDR and CDR [@Kv11]. In this plateau, the $K=1$ strength dominates at $E >$ 25 MeV. Below, at 12–25 MeV, the $K=0$ and $K=1$ strengths are about the same. Even lower, at $E <$ 12 MeV, the $K=1$ strength becomes strongly dominant. In particular a strong toroidal peak at 7–8 MeV is almost fully built from $K=1$ strength. Note that SP makes this peak essentially higher than VP. Perhaps this happens because the toroidal $K$=1 motion is strongest just at the nuclear surface, see Fig. \[fig:11\] below. So the description of the TDR peak depends on the sort of the pairing. This is the only place where we could observes a significant difference between VP and SP.
The TDR strength is presented in more detail in Fig. \[fig:10\]. Plot b) demonstrates a significant difference between VP and SP results for the highest TDR peak. Plots c) and d) show a negligible contribution of the pp-channel. Further, plot a) confirms that the TDR strength is almost fully provided by $K$=1 contribution. This peculiarity of the TDR corresponds to the PDR picture (where $K$=1 strength also dominates at $E <$ 8 MeV) but is opposite to low-energy part of the CDR with dominant $K$=0 strength. It is interesting that the PDR peaks roughly correspond to the PDR structures at 7–8 MeV but have no counterpart in the low-energy CDR located at 4-6 MeV. This is one more indication for the intimate relation between TDR and PDR, discussed in [@Re13]. Note that the peaked TDR strength appears even in the unperturbed (without residual interaction) strength function [@NeDre14]. The isoscalar residual interaction only downshifts the structure by $\sim$ 2 MeV.
To illustrate the mainly toroidal nature of dipole states at 6–9 MeV, we show in Fig. \[fig:11\] their isoscalar average current transition densities (CTD) for the branches $K=0$ and $K=1$. The method of calculation of the average CTD in QRPA is described in detail in [@Re13]. Remind that CTD depend only on the structure of QRPA states and are not affected by the probing external field. In Fig. \[fig:11\], we see for both $K=0$ and $K=1$ basically toroidal flow. The toroidal axes are directed along z-axis for $K=0$ and x-axis for $K=1$. It is interesting that the toroidal flow resembles a vortex-antivortex configuration. Fig. \[fig:11\] shows that for $K=1$ the flow is much stronger, especially at the nuclear surface. This supports the previous finding that the TDR is mainly realized in the $K=1$ branch. Perhaps, the nuclear geometry in the $K=1$ case is more suitable for hosting the toroidal flow and thus makes the flow more energetically favorable. Our analysis shows that the strong $K=1$ flow at the nuclear surface is mainly provided by neutrons. The strong motion of the neutron surface against the remaining core corresponds to a typical PDR picture. So Fig. \[fig:11\] confirms our previous suggestion [@Re13] that PDR is mainly a local boundary manifestation of the toroidal nuclear flow.
Finally, Fig. \[fig:12\] shows the currents for the dipole excitations at 3–6 MeV (exhibited at Fig. 8a). The flow considerably deviates from the toroidal picture.
Altogether, Figs. \[fig:5\]–\[fig:12\] show that in the dipole resonances (GDR, PDR, TDR, CDR): i) the pp-channel is negligible; ii) the difference between VP and SP results is small (except for the strongest TDR peak at 7–8 MeV); iii) at $E>$ 10 MeV, the familiar sequence of $K=$0 and $K=$1 branches is observed; iv) at $E<$ 10 MeV, the ratio between the branches depends on the resonance ($K=1$ strength dominates in TDR and PDR but becomes minor in CDR); v) The $K=1$ strength forms an impressively narrow low-energy peak in the TDR (the latter feature may serve as a fingerprint of the TDR in future experiments); vi) current transition densities (CTD) confirm the mainly toroidal nature of dipole excitations at 6–9 MeV.
Figs. \[fig:5\]–\[fig:12\] allow to do some additional comments concerning origin of PDR and ISGDR, deformation splitting of these resonances, and perspectives of experimental observation of TDR.
1\) The TDR operator (\[15\]) has the radial dependence $\sim r^2$ and so surface reactions should be sensitive probes of this resonance. In particular, the ($\alpha,\alpha'\gamma$) reaction in deformed nuclei looks suitable to excite the isoscalar TDR and determine, using Alaga rules [@Ala55], a dominant K=1 low-energy TDR peak. The ratio of intensities of $E1K$-transitions from a given dipole QRPA $K$-state to the members $I^{\pi}K=0^+0$ and $2^+0$ of the ground-state rotational band is determined through Clebsh-Gordan coefficients as $$A_K=|\frac{\langle 1K 1 -K |00\rangle}{\langle 1K 1-K|20\rangle}|^2.$$ This yields $A_0$=0.5 and $A_1$=2. Thus $K=0$ and $K=1$ branches can be distinguished. Although these dipole transitions are isoscalar, they should be strong enough to be detectable since the peaked TDR is a collective mode [@Kv14Yb; @Ne17; @NeDre14]. Note that in $^{154}$Sm the TDR peak lies below the neutron and proton threshold energies, which should favor its $\gamma$-decay.
2\) The relevance of the ($\alpha,\alpha'\gamma$) reaction to observe the TDR is not so evident. Note that, despite the TDR/CDR relation, the irrotational ISGDR strength in Figs. \[fig:7\]-\[fig:8\] does not match the vortical toroidal strength exhibited in Figs. \[fig:9\]-\[fig:10\]. There is a general question if the ($\alpha,\alpha'$) reaction is, in principle, able to probe vortical modes. We need here further analysis. From the theoretical side, DWBA calculations for ($\alpha,\alpha'$) angular distributions should be performed, using QRPA input formfactors for the TDR low-energy $K=1$ peak.
3\) Following our results, the ISGDR observed at 10-33 MeV in $(\alpha,\alpha')$ scattering to small angles [@CDRexp] is perhaps a fully irrotational mode dominated by compression-like dipole excitations. In other words, the observed ISGDR does not include the TDR at all. Indeed both observed broad bumps at 10-17 MeV and 17-33 MeV are located higher than the TDR peak predicted at 7-8 MeV. Moreover, in this experiment, only excitations at $E>$ 10MeV are measured, i.e. the actual TDR energy region is fully omitted. In the same experiment for other nuclei ($^{116}$Sn, $^{144}$Sm and $^{208}$Pb) the TDR low-energy region was also missed [@CDRexp]. At the same time, some low-energy peaks (which can be attributed to the TDR) were measured in $^{90}$Zr, $^{116}$Sn, $^{144}$Sm and $^{208}$Pb in an alternative $(\alpha,\alpha')$ experiment [@Uch04] where the interval $E>$ 8 MeV was inspected. These peaks can match a part of the TDR. However we still need measurements for the lower energy interval which fully embraces the predicted TDR.
4\) As discussed in [@Re13], the PDR is probably a complicated mixture of various dipole excitations (toroidal, compression, GDR tail) with a dominant contribution of the toroidal fraction. Following Figs. \[fig:6\], \[fig:8\], \[fig:10\], the dipole response crucially depends on the probe. The PDR structure in deformed nuclei (K-branches) also should depend on the reaction applied. For example, following Figs. \[fig:8\] and \[fig:10\], $K$-branches of the PDR can arise separately as responses to different probes: $K$=0 in the CDR/ISGDR response and $K=1$ in the toroidal response. If so, then investigation of CDR/ISGDR and TDR could be useful to understand the deformation splitting of the PDR.
Conclusions {#concl}
===========
Effects of nuclear deformation and monopole pairing in nuclear excitations with $K^{\pi} \ne 0^+$ were investigated within the self-consistent Quasiparticle Random Phase Approximation (QRPA) method [@Re16] using the Skyrme parametrization SLy6 [@Cha97]. Mean field and pairing were treated within the HF+BCS scheme. As representative examples the axially deformed nuclei $^{152,154,156}$Sm were considered. One of the main aims was to investigate the role of the particle-particle (pp) channel in the residual QRPA interaction. This channel is known to be important for $K^{\pi} = 0^+$ states [@SSS; @Ri80] but its role in $K^{\pi} \ne 0^+$ excitations was yet unclear. We scrutinized this problem for the lowest $\gamma$-vibrational $K^{\pi}=2^+$ states and for dipole low- and high-energy excitations: isovector giant dipole resonance (GDR), isovector pygmy dipole resonance (PDR), toroidal dipole resonance (TDR), and compression dipole resonance (CDR). We also compared QRPA results for these excitations with volume and surface pairing. Last but not least, we analyzed deformation effects in the dipole resonances, especially in the low-energy interval where PDR is located. The following results have been obtained:\
1) The effect of the pp-channel in $K^{\pi} \ne 0^+$ excitations is negligible. This can be explained by the absence of the proton-neutron part in the pairing forces (though the proton-neutron interaction dominates in the ph-channel). As a result, the pp-channel turns out to be much weaker than its ph-counterpart and finally can be safely neglected in description of $K^{\pi} \ne 0^+$ modes. This result is important from the practical side because omitting the pp-channel considerably simplifies QRPA calculations, which is especially beneficial for deformed nuclei where QRPA calculations are time-consuming. Since the effect of the pp-channel is fairly small, its omission does not spoil the consistency of QRPA. In our study, this result was obtained for low-energy $K^{\pi}=2^+$ states and dipole resonances. However it is likely to be valid for any nuclear excitation with $K^{\pi} \ne 0^+$ (dipole, quadrupole, octupole).\
2) Our calculations have shown that, as a rule, the difference between the QRPA results obtained with from volume pairing (VP) and from surface pairing (SP) is small. A noticeable difference was found only for the collective low-energy peak of the TDR.\
3) We analyzed deformation effects arising due to competition between $K=0$ and $K=1$ branches in the dipole excitations, comparing the behavior of the branches in various dipole resonances (GDR, PDR, TDR, CDR). It was shown that, in all the resonances, the familiar sequence of $K=0$ and $K=1$ branches takes place for excitation energy $E >$ 10 MeV, namely the $K=0$ structure lies lower than the $K=1$ one as it should be in prolate nuclei. However, at lower excitation energies $E<$ 10 MeV, the behavior of the branches depends on the resonance type. In particular, in the PDR and TDR, the $K=1$ strength becomes dominant. For TDR, this effect manifests itself in the appearance of very strong peak at $\sim$ 7–8 MeV, which is almost completely built from $K=1$ strength. This remarkable feature can be used as a fingerprint of the TDR in future experiments. The ($\alpha,\alpha'\gamma$) reaction in deformed nuclei looks most promising. The peaked $K=1$ strength can be identified inspecting intensities of the $\gamma$-decay in terms of Alaga rules.\
4) Following our results, the ISGDR observed in $(\alpha,\alpha')$ scattering to forward angles [@CDRexp] is basically irrotational and perhaps does not include at all the low-energy vortical TDR. To search the TDR, the lower excitation energies should be experimentally inspected.\
5)The nuclear flow in the PDR energy region is basically toroidal. The deformation structure of PDR can be related to $K$ branches in TDR and CDR.
The work was partly supported by Votruba-Blokhincev (Czech Republic-BLTP JINR) and Heisenberg-Landau (Germany - BLTP JINR) grants. The support of the Czech Scinece Foundation Grant Agency (project P203-13-07117S) is appreciated. A.R. is grateful for support from Slovak Research and Development Agency under Contract No. APVV-15-0225.
Skyrme functional {#app-A}
=================
The Skyrme energy functional (\[Efunc\]) has the kinetic-energy, Skyrme, Coulomb and pairing parts [@Be03; @Rei92d] with ($q=n,p$) $$%&&
\mathcal{E}_{\rm kin} = \int\mathrm{d}\vec{r}
\sum_q \frac{\hbar^2}{2m_q}\tau_q(\vec{r}) ,
\label{A.2}$$ $$\begin{aligned}
&&
\mathcal{E}_{\rm Sk} = \int\mathrm{d}\vec{r}
\bigg\{ \frac{b_0}{2}\rho^2 - \frac{b_0'}{2}\sum_q\rho_q^2
+\frac{b_2}{2}(\vec{\nabla}\rho)^2
\label{A.3}
\\
&&
- \frac{b_2'}{2}\sum_q(\vec{\nabla}\rho_q)^2 + b_1(\rho\tau\!-\!\vec{j}^{\,2})
- b_1'\sum_q(\rho_q\tau_q\!-\!\vec{j}_q^{\,2})
\nonumber \\
&&
+\tilde{b}_1\Big(\vec{s}\cdot\vec{T} - {\bf J}^2\Big)
+\tilde{b}_1'\sum_q\Big(\vec{s}_q\cdot\vec{T}_q-{\bf J}_q^2 \Big)
\nonumber \\
&&
+\frac{b_3}{3}\rho^{\alpha+2}-\frac{b_3'}{3}\rho^\alpha\sum_q\rho_q^2
-b_4\big[\rho\vec{\nabla}\cdot\vec{J}
+\vec{s}\cdot(\vec{\nabla}\times\vec{j})\big]
\nonumber\end{aligned}$$ $$\begin{aligned}
&&
-b_4'\sum_q\big[\rho_q\vec{\nabla}\cdot\vec{J}_q+\vec{s}_q\cdot (\vec{\nabla}\times\vec{j}_q)\big]
\nonumber\\
&&
+\frac{\tilde{b}_0}{2}\vec{s}^2-\frac{\tilde{b}_0'}{2}\sum_q\vec{s}_q^{\,2}
+\frac{\tilde{b}_2}{2}\sum_{ij}(\nabla_i s_j)^2
\nonumber \\
&&
-\frac{\tilde{b}_2'}{2}\sum_q \sum_{ij}(\nabla_i s_j)_q^2
+\frac{\tilde{b}_3}{3}\rho^\alpha\vec{s}^{\,2}
-\frac{\tilde{b}_3'}{3}\rho^\alpha\sum_q \vec{s}_q^{\,2} \bigg\}\: ,
\nonumber\end{aligned}$$ $$\begin{aligned}
\label{A.4}
&&
\mathcal{E}_{\rm Coul} =
\frac{1}{2} e^2 \int \mathrm{d}\vec{r}
\int \mathrm{d}\vec{r}' \frac{\rho_p(\vec{r}) \rho_p(\vec{r}')}
{|\vec{r} - \vec{r}'|}
\\
&& \qquad \qquad \qquad
- \frac{3}{4}\Big(\frac{3}{\pi} \Big)^{\frac{1}{3}} e^2
\int \mathrm{d}\vec{r} \rho_p(\vec{r})^{\frac{4}{3}} .
\nonumber\end{aligned}$$ The pairing part is given in Eq. (\[Epair\]) of Section 2.2. The functional depends on time-even ($\rho$, $\tau$, ${\bf J}$, $\tilde\rho$) and time-odd ($\vec{j}$, $\vec{s}$, $\vec{T}$) local densities and currents. The total densities are sums of the proton and neutron parts, e.g. $\rho=\rho_n + \rho_p$. In HF+BCS approach, the densities of our main interest, nucleon $\rho$ and pairing $\tilde\rho$, have the form (\[rho\]) and (\[rho\_pair\]). The relation of the Skyrme parameters $b_i$ with the standard ones $t_i,\:x_i$ can be found in [@Re16; @Be03; @Rei92d]
HF+BCS equations {#app-b}
================
The zero-range monopole pairing is defined by Eqs. (\[Epair\])-(\[Gpair\]). It results in the BCS equations: $$[u_i^q]^2=
\frac{1}{2} \: \left\{1 + \frac{e_i^q -
\lambda_q}{\sqrt{\left[e_i^q - \lambda_q\right]^2
+ [\Delta_i^{q}]^2}} \right\},
\label{u_i}$$ $$[v^q_i]^2 =
\frac{1}{2} \: \left\{1 - \frac{e_i^q -
\lambda_{q}}{\sqrt{\left[e_i^q - \lambda_{q}\right]^2
+ [\Delta_i^{q}]^2}} \right\},
\label{v_i}$$ $$\Delta_i^{q}
= - \sum_{j \in q}^{K>0}
f_{j}^q V^{(\rm{pair}, q)}_{i \bar{i} j \bar{j}} \: v_{j}^q u_{j}^q ,
\label{Del_i}$$ $$N_{q}
=
\sum_{i \in q}^{K>0} f_{i}^q
\left\{1 - \frac{e_i^q - \lambda_{q}}{\sqrt{\left[e_i^q -
\lambda_{q}\right]^2 + [\Delta_i^{q}]^2}} \right\} ,
\label{Npair}$$ where $e_i^q$ is the s-p energy, $u_{i}^q, v_{i}^q$ are Bogoliubov coefficients, $\Delta_i^{q}$ is the pairing gap, $\lambda_{q}$ is the chemical potential, $f_{i}^q$ is the cut-off function (\[f-cut\]), $N_p=Z$ and $N_n=N$ are proton and neutron numbers, and $V_{i \bar{i} j \bar{j}}^{(\rm{pair, q})}
=\langle i\bar{i}|V^q_{\rm pair}(\vec{r},\vec{r}')|j\bar{j}\rangle$ is the pairing matrix element.
The total HF+BCS Hamiltonians consists from the local pairing and HF parts: $${\hat h}^q_{\rm HF+BCS}(\rho,{\tilde \rho})=
{\hat h}^q_{\rm pair}(\rho,{\tilde \rho})
+ {\hat h}^q_{\rm HF} (\rho,\tau,{\bf J},\tilde{\rho}) .$$ The pairing part $${\hat h}^q_{\rm pair}(\rho,{\tilde \rho})(\vec{r})=
\frac{\delta \mathcal{E}_{\rm pair}}
{\delta {\tilde \rho_q}(\vec{r})}
=\frac{1}{2}{\tilde \rho_q}(\vec{r})G_q(\vec{r})
\label{hpair}$$ depends on the pairing density $\tilde\rho$ and, in SP case, on total nucleon density $\rho$ (through the factor $G_q(\vec{r})$).
The mean-field part $${\hat h}^q_{\rm HF} (\rho, \tau,{\bf J},\tilde\rho)(\vec{r})=
\frac{\delta \mathcal{E}}{\delta \rho_q(\vec{r})}
+\frac{\delta \mathcal{E}}{\delta \tau_q(\vec{r})}
+\frac{\delta \mathcal{E}}{\delta \mathbf{J}_q(\vec{r})}
\label{hHF}$$ is determined by the time-even densities involved to the Skyrme functional (\[A.3\]). For SP, the Hamiltonian (\[hHF\]) gains dependence on the pairing density $\tilde\rho$.
BCS equations (\[u\_i\])-(\[Npair\]) are coupled to the density-dependent HF equations $$\label{HF}
{\hat h}^q_{\rm HF} (\rho,\tau,{\bf J},\tilde{\rho})(\vec{r})
\psi^q_i(\vec{r}) =e_i^q \psi^q_i(\vec{r})$$ via the pairing $u_{i}^q, v_{i}^q$-factors in the densities, see e.g. Eq. (\[rho\]) for $\rho_q (\vec{r})$. In turn, the BCS equations are affected by the HF mean field through single-particle energies $e_i^q$ and wave functions $\psi_i^q$. In our HF+BCS scheme, BCS equations (\[u\_i\])–(\[Npair\]) are solved at each HF iteration [@Skyax].
QRPA with pp-channel {#app-D}
====================
The QRPA equations read [@Ri80] $$\left(
\begin{array}{cc}
A & B \\
B^* & A^*
\end{array}
\right)
\left(
\begin{array}{c}
X^{(\nu)} \\
Y^{(\nu)}
\end{array}
\right)
=
\hbar E_{\nu} \:
\left(
\begin{array}{cc}
1 & 0 \\
0 & -1
\end{array}
\right)
\left(
\begin{array}{c}
X^{(\nu)} \\
Y^{(\nu)}
\end{array}
\right)
\label{1}$$ where $X^{(\nu)}_{ij}$ and $Y^{(\nu)}_{ij}$ are forward and backward amplitudes of QRPA phonon creation operators $Q^+_{\nu}$ $$Q^+_{\nu} = \frac{1}{2}
\sum_{ij}
\big( X^{(\nu)}_{ij} \: \alpha^+_i \alpha^+_j - Y^{(\nu)}_{ij} \: \alpha_j \: \alpha_i \big) ,
\label{2}$$ $E_{\nu}$ is the phonon energy, $\alpha^+_i (\alpha_j)$ is the quasiparticle creation (annihilation) operator. The RPA matrices $A$ and $B$ have the form $$\begin{aligned}
A_{\omega \omega'} &=& \epsilon_{\omega} \delta_{\omega, \: \omega^{'}}
\nonumber
\\
&+& \sum_{d\:d^{\:'}} \int \!\! \mathrm{d}\vec{r} \! \int \!\! \mathrm{d}\vec{r}' \frac{\delta^2 \mathcal{E}}
{\delta \mathcal{J}^d(\vec{r}) \: \delta \mathcal{J}^{d'}(\vec{r^{'}})} \:
\delta \mathcal{J}^d_{\omega}(\vec{r}) \: \delta \mathcal{J}^{d'}_{\omega'}(\vec{r}')
% \langle \omega |\: \hat{\mathcal{J}}^{(1)}_d(\vec{r}) \:|BCS \rangle^* \:
% \langle \omega^{'} |\: \hat{\mathcal{J}}^{(1)}_{d^{'}}(\vec{r^{'}}) \:|BCS \rangle
\nonumber \\
B_{\omega \omega^{'}} &=& - \:
\sum_{d\:d^{\:'}} \int \!\! \mathrm{d}\vec{r}\! \int \!\! \mathrm{d}\vec{r}' \frac{\delta^2 \mathcal{E}}
{\delta \mathcal{J}^d(\vec{r}) \: \delta \mathcal{J}^{d'}(\vec{r^{'}})} \:
%\nonumber \\
%&& \quad \quad \qquad
\delta \mathcal{J}^d_{\omega}(\vec{r}) \: \delta \mathcal{J}^{d'}_{\omega'}(\vec{r}')
% \langle \omega |\: \hat{\mathcal{J}}^{(1)}_d(\vec{r}) \:|BCS \rangle \:
% \langle \omega^{'} |\: \hat{\mathcal{J}}^{(1)}_{d^{'}}(\vec{r^{'}}) \:|BCS \rangle
\nonumber
\label{9}\end{aligned}$$ where $\omega = ij,\: \bar{i} \bar{j},\: i \bar{j} \in q$ and $\omega' = i'j',\: \bar{i}' \bar{j}',\: i' \bar{j}' \in q'$ ($i > j$); $\delta_{ij,\: i^{'} j^{'}} \equiv \delta_{i i^{'}} \delta_{j j^{'}} -
\delta_{i j^{'}} \delta_{j \:i^{'}}$, $\delta_{i\: \bar{j},\: i^{'} \bar{j^{'}}}
\equiv \delta_{i i^{'}} \delta_{j j^{'}}$, $\delta_{ij,\: i^{'} \bar{j^{'}}} = 0$. Further, $\epsilon_{\omega} \equiv \epsilon_i + \epsilon_j$ is 2qp energy; $\delta \mathcal{J}^d_{\omega}(\vec{r})=\langle \omega |\hat{\mathcal{J}}^d|0\rangle$ is the transition density for the transition between 2qp state $|\omega\rangle$ and BCS vacuum $|0\rangle$, $\hat{\mathcal{J}}^d$ is the density operator in 2qp representation (in terms of $\alpha^+_i\alpha^+_j, \: \alpha_i\alpha_j$); the set $\mathcal{J}^d$ includes the densities and currents $d \in \:\rho,\: \tau,\: {\bf J}, \: \vec{j},
\: \vec{s} , \: \vec{T}, \tilde{\rho}$. For more details see [@Re16].
The pairing functional $\mathcal{E}_{\rm pair}$ (\[Epair\]) leads to additional terms in the residual interaction, which constitute the pp-channel: $$\frac{\delta^2 \mathcal{E}} {\delta \rho(\vec{r}) \: \delta \rho (\vec{r}^{'})},
\:
\frac{\delta^2 \mathcal{E}} {\delta \rho(\vec{r}) \: \delta \tilde{\rho} (\vec{r}^{'})},
\:
\frac{\delta^2 \mathcal{E}} {\delta \tilde{\rho} (\vec{r}) \: \delta \rho (\vec{r}^{'})},
\:
\frac{\delta^2 \mathcal{E}} {\delta \tilde{\rho}(\vec{r}) \: \delta \tilde{\rho} (\vec{r}^{'})} .
\nonumber$$ In SP, all the terms take place while in VP only the last term remains. In the QRPA matrices $A$ and $B$, the pp-channel is presented by the terms: $$\begin{aligned}
&&A^{\rm pair}_{qq'}(i j, i' j') =
-\frac{\eta \: \gamma \: (\gamma -1)}{4\: \rho^{\gamma}_0}
\label{C.1}
\\[-0.2cm]
&\cdot& \sum_{q=n,p} V_q \:
\int \mathrm{d}\vec{r} \: \tilde{\rho}^2_q(\vec{r}) \:\: \rho^{\gamma -2}(\vec{r})
\: \delta \rho_{ij}(\vec{r}) \: \delta \rho_{i'j'}(\vec{r}) ,
\nonumber \\
&&B^{\rm pair}_{qq'}(i j, i' j') = -\: A^{\rm pair}_{qq'}(i j, i' j') ,
\label{C.2}\end{aligned}$$ $$\begin{aligned}
&& \:
A^{\rm pair}_{qq'}(i \bar{j}, i' \bar{j}')
= -\frac{\eta \: \gamma \: (\gamma -1)}{4\: \rho^{\gamma}_0}
\label{C.3} \\
&\cdot& \sum_{q=n,p} V_q \: \int \mathrm{d}\vec{r} \: \tilde{\rho}^2_t(\vec{r}) \:\: \rho^{\gamma -2}(\vec{r})
\delta \rho_{ij}(\vec{r}) \: \delta \rho_{i'j'}(\vec{r})
\nonumber \\
&& +
\frac{1}{2} \: \delta_{q q'} V_q \: \int \mathrm{d}\vec{r}
\Big[1 - \eta \: \Big(\frac{\rho(\vec{r})}
{\rho_0}\Big)^{\gamma} \Big]
\delta \tilde{\rho}_{ij}(\vec{r}) \: \delta \tilde{\rho}_{i'j'}(\vec{r})
\nonumber \\
&& -
\frac{\eta \: \gamma}{2 \rho_0^{\gamma}} \: V_q \: \int \mathrm{d}\vec{r} \:\tilde{\rho}_q(\vec{r})
\: \rho^{\gamma-1}(\vec{r})
\delta \tilde{\rho}_{ij}(\vec{r}) \: \delta \rho_{i'j'}(\vec{r})
\nonumber \\
&& -
\frac{\eta \: \gamma}{2 \rho_0^{\gamma}} \: V_{q^{'}} \: \int \mathrm{d}\vec{r} \:
\tilde{\rho}_{q^{'}}(\vec{r}) \: \rho^{\gamma-1}(\vec{r})
\delta \rho_{ij}(\vec{r}) \: \delta \tilde{\rho}_{i'j'}(\vec{r})
\nonumber\\
&&B^{\rm pair}_{qq'}(i \bar{j}, i' \bar{j}') = -\: A^{\rm pair}_{qq'}(i \bar{j}, i' \bar{j}') ,
\label{C.4}\end{aligned}$$ The contributions (\[C.1\])-(\[C.2\]) and the first term in (\[C.3\]) are zero for both VP ($\eta$=0) and SP (if $\gamma$=1). The last two cross terms in (\[C.3\]) exist only for the density-dependent SP (when $\eta \ne$ 0). Only the second term in (\[C.3\]), including merely pairing transitions densities $\delta \tilde{\rho}$, exists for both VP and SP. Expressions (\[C.1\])-(\[C.4\]) correspond to the formalism derived in [@Te05] for spherical nuclei.
Expressions (\[C.1\])-(\[C.4\]) include contributions to pp-channel from both diagonal ($i=j$) and non-diagonal ($i \ne j$) 2qp configurations. Note that, in the quasiparticle-phonon model (QPM) with the constant pairing [@SSS], only diagonal contributions are taken into account, thus the pp-channel exists only for $K^{\pi}=0^+$ states. In our study for states with $K^{\pi} \ne 0^+$, the selection rules allow only non-diagonal configurations.
V.G. Soloviev, A.V. Sushkov, N.Yu. Shirikova, Phys. Part. Nucl. [**25**]{}, 377 (1994); ibid [**27**]{}, 1643 (1996). P. Ring, P. Schuck, [*The Nuclear Many-Body Problem*]{} (Springer-Verlag, Berlin, 1980). J. Dobaczewski, H. Flocard, J. Treiner, Nucl. Phys. A [**422**]{}, 103 (1984). J. Dobaczewski, W. Nazarewicz, T.R. Werner, J.F. Berger, C.R. Chinn, J. Dechargé, Phys. Rev. C [**53**]{}, 2809 (1996). P.-G. Reinhard, M. Bender, K. Rutz, J.A. Maruhn, Z. Phys. A [**358**]{}, 277 (1997). M. Bender, K. Rutz, P.-G. Reinhard, J.A. Maruhn, Eur. Phys. J. A [**8**]{}, 59 (2000). N. Tajima, Phys. Rev. C [**69**]{} 034305 (2004). J. Terasaki, J. Engel, M. Bender, J. Dobaczewski, W. Nazarewicz, M. Stoitsov, Phys. Rev. C [**71**]{}, 034310 (2005). J. Li, G. Colò, J. Meng, Phys. Rev. C [**78**]{}, 064304 (2008). N. Lo Iudice, A. V. Sushkov, and N. Yu. Shirikova, Phys. Rev. C [**72**]{}, 034303 (2005). A.P. Severyukhin, V.V. Voronov, Nguyen Van Giai, Phys. Rev. C [**77**]{}, 024322 (2008). J. Terasaki, J. Engel, Phys. Rev. C [**82**]{}, 034326 (2010). J. Terasaki, J. Engel, Phys. Rev. C [**84**]{}, 014332 (2011). A. Repko, arXiv:1603.04383v2 \[nucl-th\]. E. Chabanat, P. Bonche, P. Haensel, J. Meyer, R. Schaeffer, Nucl. Phys. A [**635**]{}, 231 (1998). V.O. Nesterenko, V.G. Kartavenko, W. Kleinig, J. Kvasil, A. Repko, R.V. Jolos, P.-G. Reinhard, Phys. Rev. C [**93**]{}, 034301 (2016). N. Paar, D. Vretenar, E. Khan, G. Colò, Rep. Prog. Phys. [**70**]{}, 691 (2007). J. Kvasil, V.O. Nesterenko, W. Kleinig, P.-G. Reinhard, P. Vesely, Phys. Rev. C [**84**]{}, 034303 (2011). A. Repko, P.-G. Reinhard, V.O. Nesterenko, J. Kvasil, Phys. Rev. C [**87**]{}, 024305 (2013). M.N. Harakeh, A. van der Woude, [*Giant Resonances*]{} ( Oxford, Clarendon, 2001). J. Kvasil, V.O. Nesterenko, W. Kleinig, D. Bozik, P.-G. Reinhard, N. Lo Iudice, Eur. Phys. J. A [**49**]{}, 119 (2013). J. Kvasil, V.O. Nesterenko, W. Kleinig, P.-G. Reinhard, Phys. Scr. [**89**]{}, 054023 (2014). V.O. Nesterenko, J. Kvasil, A. Repko, W. Kleinig, P.-G. Reinhard, Phys. Atom. Nucl. [**79**]{}, 842 (2016). T.H.R. Skyrme, Phil. Mag. [**1**]{}, 1043 (1956). D. Vautherin, D.M. Brink, Phys. Rev. C [**5**]{}, 626 (1972). M. Bender, P.-H. Heenen, P.-G. Reinhard, Rev. Mod. Phys. [**75**]{}, 121 (2003). W. Kleinig, V.O. Nesterenko, J. Kvasil, P.-G. Reinhard, P. Vesely, Phys. Rev. C [**78**]{}, 044313 (2008). Evaluated Nuclear Structure Data File, http://www.nndc.bnl.gov/chart/. P.–G. Reinhard, Ann. Phys. (Leipzig) [**504**]{}, 632 (1992). P. Klüpfel, P.–G. Reinhard, T. J. Bürvenich, and J. A. Maruhn, Phys. Rev. C [**79**]{}, 034310 (2009). V.O. Nesterenko, W. Kleinig, J. Kvasil, P. Vesely, P.-G. Reinhard, D.S. Dolci, Phys. Rev. C [**74**]{}, 064306 (2006). P. Fleischer, P. Klupfel, P.-G. Reinhard, and J. A. Maruhn, Phys. Rev. C [**70**]{}, 054321 (2004). G.M. Gurevich, L.E. Lazareva, V.M. Mazur, S.Yu. Merkulov, G.V. Solodukhov, V.A. Tyutin, Nucl. Phys. A [**351**]{}, 257 (1981). J. Kvasil, P. Vesely, V.O. Nesterenko W. Kleinig, P.-G. Reinhard, S. Frauendorf, Int. J. Mod. Phys. E [**18**]{}, 975 (2009). D.H. Youngblood, Y.-W. Lui, H.L. Clark, B. John, Y. Tokimoto, X. Chen, Phys. Rev. C [**69**]{}, 034315 (2004). D. H. Youngblood, Y.-W. Lui, B. John, Y. Tokimoto, H. L. Clark, and X. Chen, Phys. Rev. C [**69**]{}, 054312 (2004). M. Uchida et al, Phys. Rev. C [**69**]{}, 051301(R) (2004). V.O. Nesterenko, A. Repko, P.-G. Reinhard, J. Kvasil, EPJ Web Conf. [**93**]{}, 01020 (2015). G. Alaga, Phys. Rev. [**100**]{}, 432 (1955). P.-G. Reinhard, SKYAX code for axial deformed nuclei.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We develop a theoretical frame for the study of classical and quantum gravitational waves based on the properties of a nonlinear ordinary differential equation for a function $\sigma(\eta)$ of the conformal time $\eta$, called the auxiliary field equation. At the classical level, $\sigma(\eta)$ can be expressed by means of two independent solutions of the ”master equation” to which the perturbed Einstein equations for the gravitational waves can be reduced.
At the quantum level, all the significant physical quantities can be formulated using Bogolubov transformations and the operator quadratic Hamiltonian corresponding to the classical version of a damped parametrically excited oscillator where the varying mass is replaced by the square cosmological scale factor $a^{2}(\eta)$. A quantum approach to the generation of gravitational waves is proposed on the grounds of the previous $\eta-$dependent Hamiltonian. An estimate in terms of $\sigma(\eta)$ and $a(\eta)$ of the destruction of quantum coherence due to the gravitational evolution and an exact expression for the phase of a gravitational wave corresponding to any value of $\eta$ are also obtained. We conclude by discussing a few applications to quasi-de Sitter and standard de Sitter scenarios.
PACS: 98.80.-k Cosmology
PACS: 04.30.-v Gravitational waves
PACS: 03.65.-w Quantum mechanics
author:
- |
A. Geralico$^{1,2,3}$[^1], G. Landolfi$^{1,2}$[^2], G. Ruggeri$^{1,2}$[^3] and G. Soliani$^{1,2}$[^4]\
$^{1}$*Dipartimento di Fisica dell’Università di Lecce, I-73100 Lecce, Italy*\
$^{2}$*INFN, Sezione di Lecce, I-73100 Lecce, Italy*\
$^{3}$*International Center for Relativistic Astrophysics-I.C.R.A.,*\
*Università di Roma ”La Sapienza”, I-00185 Roma, Italy*
title: |
Novel approach to the study of\
quantum effects in the early universe
---
Introduction
============
In a notable paper by Grishchuk and Sidorov \[\[grishchuk\]\], it was shown that relic gravitons can be created from the vacuum quantum fluctuations of the gravitational field during cosmological expansion and can be interpreted as squeezed quantum states of the gravitational field, in analogy to what happens for squeezed states in quantum optics \[\[stoler\]-\[schumaker\]\]. A systematic treatment of the particle creation mechanism is contained in \[\[birrel\]\]. The theory of particle creation is essentially based on the Bogolubov transformations, exploited primary in the formulation of the process of squeezing for the electromagnetic field. An important aspect of the theory of particle creation is the possible loss of coherence in quantum gravitational theories. In a sense, the exploration of decoherence in gravitational theories has been more concerned with the quantum (gravity) effects which can manifest themselves even at scales others than the Planck one. The growing interest in this field relies, among the others, on quite a remarkable mechanism: the amplification of quantum (incidentally vacuum) fluctuations in the metric of a gravitational background (see *e.g.* \[\[grishchuk\],\[grishchuk2\]-\[maggiore\]\]). In the presence of a change of regime under the cosmological evolution, the occupation number of the initial quantum state would get indeed amplified. As long as the change can be considered as adiabatic, the amplification factor approaches one. Nevertheless, in case the change is sudden the amplification mechanism cannot be neglected. In such a case, even the vacuum state transforms into a multiparticle state in the Fock space appropriate to the new regime. The scenario clearly sounds highly attractive under the stronger convintion that, thanks to the great progress we are witnessing in the experimental application of new technologies, the amplification mechanism may provide the possibility to detect quantum effects (*e.g.* relic gravitons) at scales considerably above the Planck one. Because of the semiclassical approximation underlying these studies, the natural formal arena turns out to be that of coherent states. These states are generated by the displacement operator $D\left( \alpha\right) $ (see Section 2). Squeezed states enter in the matter whenever the quadratic operators $\hat{a}^{2}$ and $\hat{a}^{\dagger\,\,2}$ are involved. A squeezed state is generated by the action on a coherent state of the so-called squeeze operator defined in Section 2. As a matter of fact, in a Friedmann-Robertson-Walker (FRW) spacetime the behavior of matter scalar fields as well as of gravitational waves is governed by an equation of the time-dependent type (or time-dependent oscillator, TDO). Thus the problem of both particle creation and metric field fluctuation amplification during the cosmological evolution is reduced to that of solving the quantum TDO problem. The latter problem has been the subject of several studies, mainly in connection with quantum optical arguments. In this paper, the main idea is to make use of the machinery built-up in \[\[profilo\]\] and further developed in \[\[landolfi\]\]. The formalism would enable us to find at every stage information concerning the spectrum of created modes. In the case of a de Sitter gravitational field, the prescription straightforwardly results into full determination of quantities of physical relevance, such as Bogolubov coefficients and the phase of gravitational waves for any value of the conformal time $\eta$. All these quantities are determined exactly. This has been possible by solving a nonlinear ordinary differential equation for an auxiliary field $\sigma(\eta)$ which has been expressed in terms of two independent solutions of the $\eta-$dependent part of the D’Alembert equation for the gravitational perturbation tensor field. Another interesting result achieved in this paper is an estimate in terms of the auxiliary field $\sigma(\eta)$ and the scale factor $a(\eta)$ of the destruction of quantum coherence due to gravitational time-evolution. In our approach, the use of the (nonlinear) auxiliary equation in the linear parametrically excited oscillator equation for $y(\eta)$ (see Eq. (\[tdoy\])) reveals therefore to be quite profitable and makes more strict and compact the formal connection between quantum optics and the theory of gravitational waves. The paper is organized as follows. In Section 2, after a description of some basic properties pertinent to the squeeze operator (\[squeeze\]), we introduce the nonlinear equation (\[eqausiliaria\]) for the auxiliary field $\sigma(t)$ in terms of which the position and momentum operators $Q$ and $P$ turn out to be expressed. The role of the matrix element between squeezed states $\left| \alpha,z\right\rangle $ of the operator $D(\alpha)\,S(z)\,H(t)\,S^{\dagger}(z)\,D^{\dagger}(\alpha)$ is investigated. This matrix element results to be evaluated in terms of the auxiliary field $\sigma(t)$. It is worth noticing that in the expression for the matrix element three energy terms appear, one of them, formula (\[10.2\]), can be interpreted as the energy related to squeezed states which do not preserve the minimum uncertainty. In the theory of gravitational waves, Eq. (\[10.2\]) plays the role of *decoherence* energy of the waves. The Bogolubov transformation is reported whose coefficients are explicitly written in terms of $\sigma$. In Section 3 we discuss classical and quantum aspects of the generation of gravitational waves. An exact formula for the phase of a gravitational wave is obtained. A natural approach to the theory of gravitational waves based on the Kanai-Caldirola oscillator is outlined by means of an operator Hamiltonian expressed in terms of the auxiliary field $\sigma(\eta)$. In Section 4 some applications are displayed. Precisely, we evaluate the decoherence energy in the quasi-de Sitter inflationary model and standard de Sitter spacetime and the role of Bogolubov coefficients in terms of the auxiliary field in the particle creation mechanism is analyzed. Finally, in Section 5 some future perspectives are discussed.
Preliminaries on the squeezed states in generalized oscillators
===============================================================
Let us recall that a squeezed state of a quantum system is defined by \[\[stoler\]\]$$\left| \alpha,z\right\rangle =D(\alpha)\,S(z)\,\left| 0\right\rangle
\label{2.1}$$ where$$D\left( \alpha\right) =e^{\alpha\hat{a}_{0}^{\dagger}-\alpha^{\ast
}\hat{a}_{0}}\label{weyl}$$ is the (unitary) displacement (Weyl) operator, $$S(z)=e^{\frac{1}{2}z\hat{a}_{0}^{\dagger\,\,2}-\frac{1}{2}z^{\ast}\hat{a}_{0}^{2}}\label{squeeze}$$ is the (unitary) *squeeze* operator \[\[stoler\],\[hollerhorst\]\], $\hat{a}_{0}$ and $\hat{a}_{0}^{\dagger}$ stand for $\hat{a}(t)|_{t=t_{0}}$ and $\hat{a}^{\dagger}(t)|_{t=t_{0}}$, respectively, where $\hat{a}(t)$ and $\hat{a}^{\dagger}(t)$ denote the annihilation and creation operators of the system. The complex functions $\alpha(t)$ and $z(t)$ are arbitrary, namely $$\alpha=\left| \alpha\right| \,e^{i\varphi}\;,\qquad z=r\,e^{i\phi
}\;,\label{defr}$$ with $\varphi$, $\phi$ arbitrary (real) $c$-numbers. Notice that for $z=0$, Eq. (\[2.1\]) reproduces the coherent state $\left| \alpha,0\right\rangle
=D(\alpha)\,\,\left| 0\right\rangle $.
The following relations $$\begin{aligned}
b & \doteqdot S^{\dagger}\hat{a}_{0}S=\hat{a}_{0}\cosh r+\hat{a}_{0}^{\dagger}e^{i\phi}\sinh r\label{3.1}\\
b^{\dagger} & \doteqdot S^{\dagger}\hat{a}_{0}^{\dagger}S=\hat{a}_{0}^{\dagger}\cosh r+\hat{a}_{0}e^{-i\phi}\sinh r \label{3.2}$$ hold. This can be readily seen by applying the Baker-Campbell-Hausdorff formula. In other words, the squeeze operator $S(z)$ induces a canonical transformation of the annihilation and creation operators, in the sense that $\left[b,b^{\dagger}\right] =\hat{1}$.
Transformation of the position and momentum variables under the squeeze operator $S(z)$
---------------------------------------------------------------------------------------
From the quantum theory of generalized oscillators, it follows that position and momentum operators $Q$ and $P$ can be expressed by \[\[profilo\],\[landolfi\]\]$$Q=\sqrt{\frac{\hbar}{m}}\,\sigma\text{ }\left( \hat{a}+\hat{a}^{\dagger
}\right) \;,\qquad P=\sqrt{\hbar m}\left( \xi\hat{a}+\xi^{\ast
}\hat{a}^{\dagger}\right) \;, \label{4.2}$$ with $\hat{a}=\hat{a}(t)$, $\hat{a}^{\dagger}=\hat{a}^{\dagger}(t)$ and $$\xi=\frac{-i}{2\sigma}+\left( \dot{\sigma}-\frac{M}{2}\sigma\right) \;,
\label{defxi}$$ where $M=M(t)\equiv\frac{\dot{m}}{m}$, the dot means time derivative, and (the mass) $m=m(t)$ is a given functions of time. The function (c-number) $\sigma(t)$ satisfies the nonlinear ordinary differential equation \[\[ermakov\],\[pinney\]\]$$\ddot{\sigma}+\Omega^{2}\sigma=\frac{1}{4\sigma^{3}}\,\,\,,
\label{eqausiliaria}$$ ($\Omega$ is specified below, see Eq. (\[defOmega2\])), called the auxiliary equation associated with the classical equation of motion $$\ddot{q}+M\dot{q}+\omega^{2}(t)\,q=0\;, \label{eqtdo}$$ and $\omega(t)$ is the time dependent frequency. Via the transformation $$q\rightarrow e^{-\frac{1}{2}\int_{t_{0}}^{t}M(t^{\prime})\,dt^{\prime}}\,y\qquad, \label{qqtilde}$$ Eq. (\[eqtdo\]) can be cast into the equation$$\ddot{y}+\Omega^{2}(t)\,y=0\qquad, \label{tdoy}$$ where$$\Omega^{2}(t)=\frac{1}{4}\left( 4\omega^{2}-2\dot{M}-M^{2}\right) \qquad.
\label{defOmega2}$$ The quantum theory of the generalized oscillator (\[eqtdo\]) can be described by the Hamiltonian operator \[\[lewis\],\[profilo\]\]$$H(t)=\frac{P^{2}}{2m}+\frac{1}{2}m\,\omega^{2}Q^{2} \label{5.5}$$ where the canonical variables $Q,P$ are given by Eq. (\[4.2\]).
Taking account of Eqs. (\[3.1\]) and (\[3.2\]), and choosing $\phi=0$ we obtain[^5]$$\begin{aligned}
S^{\dagger}QS & = e^{r}Q\;,\label{6.1}\\
S^{\dagger}PS &
=2m\sigma\left( \dot{\sigma}-\frac{M}{2}\sigma\right) \sinh r
\, Q+e^{-r}P\;. \label{6.2}$$ The physical meaning of the expression $\dot{\sigma}-\frac{M}{2}\sigma$ will be clarified later. At the present we observe only that whenever the condition $$\dot{\sigma}-\frac{M}{2}\sigma=0 \label{6.3}$$ is fulfilled, then the uncertainty product $\left( \Delta Q\right) \,\left(
\Delta P\right) \,\ $of the variances$$\left( \Delta Q\right) =\sqrt{\left\langle Q^{2}\right\rangle -\left\langle
Q\right\rangle ^{2}}\;,\qquad\left( \Delta P\right) =\sqrt{\left\langle
P^{2}\right\rangle -\left\langle P\right\rangle ^{2}}\;, \label{6.4}$$ attains its minimum, where the expectation value $\left\langle \dots
\right\rangle $ is referred to coherent states \[\[landolfi\]\].
We observe that the operators $S^{\dagger}QS$ and $S^{\dagger}PS$ obey the same commutation relation as $Q$ and $P$, that is$$\left[ S^{\dagger}QS,S^{\dagger}PS\right] =\left[ Q,P\right] =i\hbar\;.$$
However, in contrast to what happens for the operators $b$ and $b^{\dagger}$ (see Eqs. (\[3.1\])-(\[3.2\])), the operators $S^{\dagger}QS$ and $S^{\dagger
}PS$ are not Hermitian conjugate. We point out that the property of Hermitian conjugation is enjoyed by the operators $S^{\dagger}QS$ and $S^{\dagger}PS$ in the case in which the condition (\[6.3\]) is valid.
A possible physical interpretation of the properties (\[6.1\]) and (\[6.2\]) is the following. For a quantum system governed by a Hamiltonian preserving the minimum wave packet (*i.e.*, the condition (\[6.3\]) holds) Eqs. (\[6.1\]) and (\[6.2\]) become $ S^{\dagger}QS =e^{r}Q $ and $ S^{\dagger}PS =e^{-r}P$, respectively. If $\left| \psi\right\rangle $ is the state of the system under consideration, then $\left| \psi^{\prime}\right\rangle =S(r)\,\left|
\psi\right\rangle $ represents the same system squeezed in the space of the position $Q$ by a factor $e^{-r}$ and expanded in the space of the momentum $P$ by the factor $e^{r}$. In fact, we deduce$$\begin{aligned}
e^{-r}\left\langle \psi^{\prime}\left| \,Q\,\right| \psi^{\prime
}\right\rangle & =\left\langle \psi\left| \,Q\,\right| \psi\right\rangle
\quad , \quad
e^{r}\left\langle \psi^{\prime}\left| \,P\,\right| \psi^{\prime
}\right\rangle & =\left\langle \psi\left| \,P\,\right| \psi\right\rangle
\quad.\end{aligned}$$
Now, we shall evaluate a matrix element involving the operator $H(t)$ (see Eq. (\[5.5\])) in the context of squeezing of a quantum system. In doing so, let us consider the following expectation value between squeezed states:$$\begin{gathered}
\left\langle \alpha,z\left| \,D(\alpha)\,S(z)\,H(t)\,S^{\dagger
}(z)\,D^{\dagger}(\alpha)\,\right| \alpha,z\right\rangle =\\
=\left\langle 0\left| \,S^{\dagger}(z)\,D^{\dagger}(\alpha)
D(\alpha)\,S(z)\,H(t)\,S^{\dagger}(z)\,D^{\dagger}(\alpha)D(\alpha)\,S(z)\,\right|
0\right\rangle =\qquad\end{gathered}$$$$\qquad\qquad\qquad=\left\langle 0\left| \,H(t)\,\right| 0\right\rangle
=\frac{1}{2m}\left\langle 0\left| \,P^{2}\,\right| 0\right\rangle +\frac
{1}{2}m\omega^{2}(t)\,\left\langle 0\left| \,Q^{2}\,\right| 0\right\rangle
\quad\quad, \label{9.1}$$ where Eqs. (\[2.1\]) and (\[5.5\]) have been employed.
The expectation values on the right hand side of Eq. (\[9.1\]) can be evaluated from Eq. (\[4.2\]). They read$$\begin{aligned}
\left\langle 0\left| \,P^{2}\,\right| 0\right\rangle & =\hbar\,m\,\left|
\xi\right| ^{2}=\hbar\,m\left[ \frac{1}{4\sigma^{2}}+\left( \frac{M}{2}\sigma-\dot{\sigma}\right) ^{2}\right] \;,\label{9.2}\\
\left\langle 0\left| \,Q^{2}\,\right| 0\right\rangle & =\frac{\hbar}{m}\sigma^{2}\;, \label{9.3}$$ where $\xi$ and $\sigma$ are described by Eqs. (\[defxi\]) and (\[eqausiliaria\]). With the help of Eqs. (\[9.2\]) and (\[9.3\]), Eq. (\[9.1\]) takes the form$$\begin{aligned}
\left\langle \alpha,z\left| \,D(\alpha)\,S(z)\,H(t)\,S^{\dagger
}(z)\,D^{\dagger}(\alpha)\,\right| \alpha,z\right\rangle & =\left\langle
0\left| \,H(t)\,\right| 0\right\rangle =\nonumber\\
& =\left[ \frac{\hbar}{8\sigma^{2}}+\frac{\hbar}{2}\left( \frac{M}{2}\sigma-\dot{\sigma}\right) ^{2}\right] \;+\frac{\hbar}{2}\omega
^{2}(t)\,\sigma^{2}. \label{10.1}$$ The term in the square bracket corresponds to the vacuum expectation value of the kinetic energy of the system, while the last term is related to the vacuum expectation value of the potential energy.
The quantity $$E_{NM}\doteqdot\frac{\hbar}{2}\left( \frac{M}{2}\sigma-\dot{\sigma}\right)
^{2}\label{10.2}$$ (NM=non-minimum) can be interpreted as the energy associated with the squeezed states which do not satisfy the criterium of minimum uncertainty ($E_{NM}\neq0$). When the criterium is verified, then $E_{NM}=0$. In such a case Eq. (\[10.1\]) can be written as $$\begin{aligned}
\left\langle \alpha,z\left| \,D(\alpha)\,S(z)\,H(t)\,S^{\dagger
}(z)\,D^{\dagger}(\alpha)\,\right| \alpha,z\right\rangle & =\left\langle
0\left| \,H(t)\,\right| 0\right\rangle =\nonumber\\
& =\frac{\hbar}{8\sigma^{2}}\;+\frac{\hbar}{2}\omega^{2}(t)\,\sigma
^{2}.\label{10.3}$$ Hence in the minimum uncertainty situation the vacuum expectation values of the kinetic energy and the potential energy turn out to be proportional to $\sigma^{2}$ and $\frac{1}{\sigma^{2}}$, respectively. It is noteworthy that, in general, all the energies appearing in Eq. (\[9.1\]) can be expressed in terms of the auxiliary field $\sigma(t)$ obeying the auxiliary equation (\[eqausiliaria\]). This enhances the convenience of our approach to the study of cosmological quantum effects based on the theory of Eq. (\[eqausiliaria\]), which we are going to develop in Section 3.
We remark that the quantity $\left( \frac{M}{2}\sigma-\dot{\sigma}\right) $ is connected with the expectation value of the operator $\left\{ Q,P\right\}
=QP+PQ$ between vacuum states, *i.e.* $$\left\langle 0\left| QP+PQ\right| 0\right\rangle =2\hbar\sigma\left(
\frac{M}{2}\sigma-\dot{\sigma}\right) \;. \label{11.1}$$ Thus the minimum uncertainty requirement $\frac{M}{2}\sigma=\dot{\sigma}$ implies that the expectation value $\left\langle 0\left| QP+PQ\right|
0\right\rangle $ is vanishing.
A link between Eq. (\[tdoy\]) and the auxiliary equation
--------------------------------------------------------
For later convenience (see Section 3), we shall report a result establishing a relationship involving the solutions of the (linear) equation of motion and the (nonlinear) auxiliary equation $$\ddot{\sigma}+\Omega^{2}(t)\,\sigma=\frac{\kappa}{\sigma^{3}}\qquad,
\label{eqausiliariak}$$ where $\kappa$ is a constant. If $y_{1}$ and $y_{2}$ are two independent solutions of Eq. (\[tdoy\]), then the general solution of the auxiliary equation (\[eqausiliariak\]) can be written as \[\[eliezer\]\] $$\sigma=(Ay_{1}^{2}+By_{2}^{2}+2Cy_{1}y_{2})^{\frac{1}{2}}\qquad,
\label{sigmasqrt}$$ with $A,B,C$ arbitrary constants such that $$AB-C^{2}=\frac{\kappa}{W_{0}^{2}} \label{eqABCW0}$$ where $W_{0}=W_{0}(y_{1},y_{2})=y_{1}\dot{y}_{2}-\dot{y}_{1}y_{2}=const$ is the Wronskian.
It is worth remarking that from the theory of the auxiliary equation (\[eqausiliariak\]) a phase can be given by the real function $\theta(t)$, $$\theta(t)=\int_{t_{0}}^{t}\frac{dt^{\prime}}{\sigma^{2}(t^{\prime})}\qquad.\label{phasetheta}$$ (See \[\[ermakov\],\[pinney\],\[goff\]\]; for some applications: \[\[lewis\],\[profilo\]\].) Here we shall suggest the procedure which can be used to compute the above integral in general cases. To this aim it is convenient to introduce the function$$\psi(t)=\sqrt{A}\,e^{i\alpha}\,y_{1}(t)-\sqrt{B}\,e^{i\beta}\,\,y_{2}(t)\;,$$ where $\alpha,\beta$ are real numbers and $y_{1},y_{2}$ are the two independent solutions appearing in Eq. (\[sigmasqrt\]). We have $$\left| \psi(t)\right| ^{2}=A\,y_{1}^{2}+B\,y_{2}^{2}-2\sqrt{AB}\,\cos
\theta_{0}\,\,y_{1}\,y\,_{2}\;,\label{psit2}$$ where $\theta_{0}=\alpha-\beta$. Comparing Eq. (\[psit2\]) with Eq. (\[sigmasqrt\]) we get $C=-\sqrt{AB}\,\cos\theta_{0}$, so that the condition (\[eqABCW0\]) becomes$$AB\sin^{2}\theta_{0}=\frac{\kappa}{W_{0}^2}\;,\label{ABsinkW0}$$ from which $\kappa\neq0$ whenever $\sin\theta_{0}\neq0$. Hence, the auxiliary field $\sigma$ can be expressed by $$\sigma^2(t)=\,\left| \psi(t)\right| ^{2}\;.$$ To calculate the phase $\theta(t)$ corresponding to the solution (\[sigmasqrt\]) of Eq. (\[eqausiliariak\]), we look for a function $F(t)$ defined by $$F(t)=\frac{1}{2i\sqrt{\kappa }}\ln\frac{\psi(t)}{\psi^{\ast}(t)}\;,\label{F(t)}$$ so that $$\dot{F}=\frac{1}{2i\sqrt{\kappa }}\frac{\dot{\psi}\,\psi^{\ast}-\psi
\dot{\psi}^{\ast}}{\psi\psi^{\ast}}\;.\label{dFt}$$ The numerator in Eq. (\[dFt\]) can be elaborated to give$$\dot{\psi}\,\psi^{\ast}-\psi\dot{\psi}^{\ast}=2i\sqrt{AB}\sin\theta_{0}W_{0}=2i\sqrt{\kappa }\label{psipsistar}$$ where Eq. (\[ABsinkW0\]) has been exploited. Substitution from Eq. (\[psipsistar\]) in Eq. (\[dFt\]) thus yields $$\dot{F}=\frac{1}{\sigma^{2}}\,\,.$$ Then the phase $\theta(t)$ is determined by integrating (\[phasetheta\]), namely$$\theta(t)=\int_{t_{0}}^{t}\frac{dt^{\prime}}{\sigma^{2}(t^{\prime})}=F(t)-F(t_{0})\;,\label{thetaFt}$$ where $F(t)$ is provided by Eq. (\[F(t)\]). We shall recall this general result later.
The Bogolubov coefficients in terms of $\sigma$
-----------------------------------------------
By resorting to the operators$$Q=\sqrt{\frac{\hbar}{2\omega_{0}m_{0}}}\,\left( \hat{a}_{0}+\hat{a}_{0}^{\dagger}\right) \;,\qquad P=-i\sqrt{\frac{\hbar\omega_{0}m_{0}}{2}}\left(
\hat{a}_{0}-\hat{a}_{0}^{\dagger}\right) \;,\label{13.1}$$ where $\hat{a}_{0}=\hat{a}(t_{0})$ is the (mode) annihilation operator in the Schrödinger representation, combining Eqs. (\[13.1\]) and (\[4.2\]) we can derive the Bogolubov transformation$$\hat{a}(t)=\mu(t)\,\,\hat{a}_{0}+\nu(t)\,\,\hat{a}_{0}^{\dagger}\label{13.2}$$ whose coefficients are expressed by $$\mu(t)=\sqrt{\frac{m}{2\omega_{0}m_{0}}}\left( -i\xi^{\ast}+\frac{\omega
_{0}m_{0}}{m}\sigma\right) \;,\quad\nu(t)=\sqrt{\frac{m}{2\omega_{0}m_{0}}}\left( -i\xi^{\ast}-\frac{\omega_{0}m_{0}}{m}\sigma\right) \;,\label{13.3}$$ with $m_{0}=m(t_{0})$, $\omega_{0}=\omega(t_{0})$, and $\xi$ given by Eq. (\[defxi\]). We note that the Bogolubov transformation can be naturally embedded into the relations (\[3.1\]) and (\[3.2\]), where the operators $b$, $b^{\dagger}$ can be identified with $\hat{a}(t)$ and $\hat{a}^{\dagger}(t)$. Equation (\[13.3\]) entails $$\left| \mu\right| ^{2}-\left| \nu\right| ^{2}=1\;. \label{14.1}$$ On the other hand, the uncertainty product can be formulated as follows \[\[landolfi\]\]:$$\begin{aligned}
\left( \Delta Q\right) \,\left( \Delta P\right) & =\frac{\hbar}{2}\sqrt{1+4\sigma^{2}\left( \frac{M}{2}\sigma-\dot{\sigma}\right) ^{2}}=\nonumber\\
& =\frac{\hbar}{2}\,\left| \mu(t)-\nu(t)\right| \,\,\left| \mu
(t)+\nu(t)\right| \geq\frac{\hbar}{2}\;. \label{14.2}$$ The uncertainty formula (\[14.2\]) is closely related to the concept of coherent states for the generalized oscillators. Such coherent states were constructed by Hartley and Ray in 1982 \[\[hartley\]\] taking account of the Lewis-Riesenfield theory \[\[lewis\]\]. These states share all the features of the coherent states of the conventional (time-independent) oscillator except that of the uncertainty formula, in the sense that the product $\left( \Delta
Q\right) \,\left( \Delta P\right) $ turns out to be not minimum. A few years later, Pedrosa showed that the coherent states devised by Hartley and Ray are equivalent to squeezed states \[\[pedrosa\]\].
Classical view and quantum theory generation of gravitational waves via the Kanai-Caldirola oscillator
======================================================================================================
In this section we shall develop a model of propagation of gravitational waves based on the application of the auxiliary equation (\[eqausiliaria\]) for the function $\sigma(\eta)$, in terms of which the Bogolubov coefficients can be built-up. The Bogolubov transformation is a basic concept in the theory of particle creation in external fields. The created particles do exist in squeezed quantum states \[\[grishchuk\]\]. According to \[\[grishchuk\]\], relic gravitons created form zero-point quantum fluctuations during cosmological evolution should now be in strongly squeezed states. In this context the generation of gravitational waves is of fundamental importance.
The theory of generation of gravitational waves in the inflationary universe scenario is based on the action \[11\] $$S=\frac{1}{16\pi G}\int f(R)\,\sqrt{-g}\,d^{4}x\;, \label{SfR}$$ where $f(R)$ is an arbitrary function of the scalar curvature $R$. The theory defined by the above action is conformally equivalent to a pure Einstein theory with scalar-field matter. In linear theory, the gravitational waves decouple from the matter field, so that the main problem is to fix the background model and to desume the relation between the conformal metric.
Starting from Eq. (\[SfR\]), by varying the action with respect to the gravitational perturbation field $h_{j}^{i}$, the equation of motion $$h^{\prime\prime}+2\frac{\tilde{a}^{\prime}}{\tilde{a}}\,h^{\prime}+
(2K-\Delta)\,h=0$$ is obtained, where prime denotes $\frac{d}{d\eta}$, $\tilde{a}(\eta)=\sqrt{\frac{\partial f}{\partial R}}\,
a(\eta)$, $h=h(\eta,\overrightarrow{x})$ is each component of $h_{j}^{i}$, $\Delta $ stands for the Laplace-Beltrami operator, and $K$ means the space curvature. By separating in $h(\eta,\overrightarrow{x})$ the dependence on $\eta$ from the dependence on $\overrightarrow{x}$, we can write $h\sim h_{0}(\eta)\,h_{1}(\overrightarrow{x})$, so that $$\left[ \Delta+(n^{2}-K)\right] \,h_{1}(\overrightarrow{x})=0\;,$$ and $$h_{0}^{\prime\prime}+2\frac{\tilde{a}^{\prime}}{\tilde{a}}
\,h_{0}^{\prime}+(n^{2}+K)\,h_{0}=0\;. \label{eqh0}$$ Equation (\[eqh0\]) can be applied to describe the evolution of gravitational waves in any state of the evolution of the universe, even when resorting to higher derivative theories of gravity \[\[mukhanov\]\]. The elimination of the first derivative in Eq. (\[eqh0\]) leads to the equation (called master equation in \[\[grishchuk\]\]) $$y^{\prime\prime}+[(n^{2}+K)-V(\eta)]\,\,y=0\,\,,
\label{yddotnKV}$$ with $$V(\eta)=\frac{a''}{a} \quad
,\quad y(\eta)=\frac{a(\eta)}{a(\eta_0)}\,h_0(\eta)\,\,.$$ We remark that Eq. (\[eqh0\]) can be regarded as the equation of motion of an oscillator with time-dependent mass $m$ and constant frequency $\bar{\omega}$, $$q^{\prime\prime}+\frac{m^{\prime}}{m}q^{\prime}+
\bar{\omega}^2 q=0\,\,,
\label{qmomega}$$ which is described by the Hamiltonian $$H=\frac{p^{2}}{2m}+\frac{m\bar{\omega}^{2}}{2}q^{2}\;. \label{HKC}$$ The quantum theory of gravitational waves is therefore equivalent to the quantum theory of the Kanai-Caldirola oscillator \[\[kanai\],\[caldirola\]\]. The formal analogy is realized upon the identification: $$m=a^{2}\quad,\quad\bar{\omega}^{2}=n^{2}+K\quad,\quad q\sim h_{0}\;.$$
Exact solution of the parametrically excited oscillator and its associated auxiliary equation
---------------------------------------------------------------------------------------------
We are mainly interested in the period under which the universe accelerates, namely in its *inflationary* stage. Inflation is defined to be a period of accelerating expansion. During such a stage, the universe expands adiabatically and the Friedmann equations can be exploited \[\[mukhanov\]\]. The prototype of the models of inflationary cosmology is based on the de Sitter spacetime, which is a very interesting case concerned with the constant Hubble rate and the scale factor given by $$a(\eta)=-\frac{1}{H_{0}\eta}\quad, \label{adS}$$ where $\eta<0$ and $H_{0}$ denotes the Hubble constant. In the de Sitter case, Eq. (\[yddotnKV\]) takes the form of Eq. (\[tdoy\]), with $\Omega^{2}(\eta)=n^{2}-\frac{2}{\eta^{2}}$. It admits the general solution $$y=\sqrt{-n\eta}\,\left[ \kappa_{1}J_{\frac{3}{2}}(-n\eta)+\kappa_{2}J_{-\frac{3}{2}}(-n\eta)\right] \;, \label{5.1.4}$$ where $J_{\frac{3}{2}}$, $J_{-\frac{3}{2}}$ are Bessel functions of the first kind and the arbitrary constants $\kappa_{1}$ and $\kappa_{2}$ are determined once the initial conditions are imposed. Then from Eq.(\[sigmasqrt\]) we infer that Eq. (\[eqausiliariak\]), which now reads $$\sigma^{\prime\prime}+\left( n^{2}-\frac{2}{\eta^{2}}\right)
\,\sigma =\frac{\kappa}{\sigma^{3}}\; ,\label{5.1.5}$$ is exactly solved by $$\sigma=\sqrt{-n\eta}\left[ AJ_{\frac{3}{2}}^{2}(-n\eta)+BJ_{-\frac{3}{2}}^{2}(-n\eta)+2CJ_{\frac{3}{2}}(-n\eta)\,\,J_{-\frac{3}{2}}(-n\eta)\right]
^{\frac{1}{2}}\;,\label{5.1.6}$$ where constants $A,B,C$ satisfy the condition (\[eqABCW0\]).
### The phase of the auxiliary field $\sigma(t)$
As we have shown in Subsection 2.2, to calculate the phase $\theta$ corresponding to the solution (\[5.1.6\]) of Eq. (\[5.1.5\]) it is convenient to introduce the function$$\psi(-n\eta)=\sqrt{-n\eta}\left[ \sqrt{A}\,e^{i\alpha}\,J_{\frac{3}{2}}(-n\eta)-\sqrt{B}\,e^{i\beta}\,\,J_{-\frac{3}{2}}(-n\eta)\right] \;,
\label{5.1.7}$$ where $\alpha,\beta$ are real numbers. Hence the auxiliary field can be expressed via $\sigma^{2}=\left| \psi\right| ^{2}$ provided that $$AB\sin^{2}\theta_{0}=\frac{\kappa}{W_{0}^2}\;, \label{5.1.9}$$ where $\theta_{0}=\alpha-\beta$ and $W_{0}=W_{0}(\sqrt{-n\eta}J_{\frac{3}{2}}(-n\eta),\sqrt{-n\eta}J_{\frac{3}{2}}(-n\eta)\,)$. Therefore it is an easy matter to see that in the case under consideration the phase $\theta$ determined by integrating Eq. (\[thetaFt\]) is given by ($z=-n\eta$)$$\theta(\eta)=-n\int_{z_{0}}^{z}\frac{dz^{\prime}}{\sigma^{2}(z^{\prime})}=-n\,[F(z)-F(z_{0})]\;, \label{5.2.6}$$ where the function $F(z)$ is defined by $$F(z)=\frac{1}{2i\sqrt{k}}\ln\frac{\psi(z)}{\psi^{\ast}(z)}\; \label{5.2.1}$$ and $\psi$ is given by Eq. (\[5.1.7\]).
On the phase of gravitational waves
-----------------------------------
A deep discussion on the phase of gravitational waves is contained in \[\[kin\]\], where this topic is dwelt on both at the classical and quantum level.
Here we confine ourselves to tackle the problem classically. The study of the phase of gravitational waves by a quantum point of view will be done elsewhere.
For our purpose, first we observe that the Bessel functions $J_{\frac{3}{2}}$ and $J_{-\frac{3}{2}}$ can be explicitly written as follows \[\[gradshteyn\]\]$$J_{\frac{3}{2}}(z)=\sqrt{\frac{2}{\pi z}}\,\left( \frac{\sin z}{z}-\cos
z\right) \quad,\quad J_{-\frac{3}{2}}(z)=-\sqrt{\frac{2}{\pi z}}\,\left(
\sin z+\frac{\cos z}{z}\right) \;.$$ Then, the auxiliary field (\[sigmasqrt\]) can be written as$$\begin{aligned}
\sigma(z) & =\sqrt{\frac{2}{\pi}}\,\left[ A\left( \frac{\sin z}{z}-\cos
z\right) ^{2}+\,\,\,\right. \nonumber\\
& \left. \,\,\,\,\ \,+B\left( \sin z+\frac{\cos z}{z}\right)
^{2}-2C\left( \frac{\sin z}{z}-\cos z\right) \left( \sin z+\frac{\cos z}{z}\right) \right] ^{\frac{1}{2}}\,, \label{sigmaGW}$$ where $y_{1}=\sqrt{z}J_{\frac{3}{2}}$, $y_{2}=\sqrt{z}J_{-\frac{3}{2}}$ and the condition (\[eqABCW0\]) is understood.
Now we shall see that the phase $\theta_{GW}$ of the primordial gravitational waves can be obtained by Eq. (\[sigmaGW\]) under the choice $$A=B\neq0\;,\quad C=0\;, \label{ABCGW}$$ and assuming asymptotically negative values of the conformal time. In doing so, Eqs. (\[5.2.6\]), (\[sigmaGW\]) provide $$\sigma_{in}^{2}\sim\frac{2A}{\pi} \qquad,
\qquad
\theta_{GW}\equiv\theta_{in}=\frac{\pi}{2A}\left( z-z_{0}\right)
\; \; ,
\label{thetaGW}$$ with $$A=\frac{\pi\sqrt{\kappa}}{2n} \; .
\label{AGW}$$ In the case $n^{2}\gg \left| V(\eta)\right| $, the high frequency waves, *e.g.* the solutions of the equation $$y^{\prime\prime}+n^{2}y=0\;,$$ correspond to the following behavior of the gravitational perturbation field $h$: $$h(\eta)=\frac{1}{a}\sin\left( n\eta+\rho\right)$$ ($\rho$ is an arbitrary phase). In an expanding universe, the amplitude $h$ of the waves decreases adiabatically for all $\eta$. The result represented by the calculation of the phase (\[thetaGW\]) of relic gravitational waves suggests one to interpret formula (\[5.2.6\]) as the phase of gravitational waves not only in the case of the primordial cosmological scenario formally corresponding to $\eta\rightarrow-\infty$. Anyway, this subject, which goes beyond the scope of the present paper, deserves further investigation. Here we recall only that the general solution of Eq. (\[tdoy\]), which holds in the case of the de Sitter cosmological model, can also be written as $$y=\frac{c}{\sqrt{\kappa}}\sigma(\eta)\,\cos\left[ \sqrt{\kappa}\theta
(\eta)+\delta\right]$$ where $\theta$ is given by Eq. (\[5.2.6\]), $c$ is a Noether invariant of Eq. (\[tdoy\]), and $\delta$ is an arbitrary constant. Hence for any conformal time $\eta$ in the interval $\left( -\infty,0\right) $, the amplitude $h$ of the gravitational perturbation field can be expressed by $$h\sim\frac{1}{a}\sigma(\eta)\,\sin\left[ \sqrt{\kappa}\theta(\eta
)+const\right] \;.$$
Quantum gravitational waves: Theory in terms of the auxiliary field $\sigma(\eta)$
----------------------------------------------------------------------------------
The approach to the study of gravitational waves we present in this paper is developed starting from the classical Kanai-Caldirola Hamiltonian \[\[kanai\],\[caldirola\]\], Eq. (\[HKC\]). As we have already pointed out, Eq. (\[eqh0\]) can be regarded, in fact, as the equation of motion of an oscillator with time-dependent mass $m=a^{2}$ and constant frequency $\bar{\omega}=\sqrt{n^{2}+K}$ (see Eq. (\[qmomega\])) described by the Hamiltonian (\[HKC\]). So, we make the fundamental identification of the whole temporal part $h_{0}$ of the metric fluctuation amplitude as the basic ”coordinate” variable to quantize as such. (Recall that our procedure for the quantization of gravitational waves is based on the identification of $y=a\,h_{0}$ as the variable to quantize). As a consequence, the quantum theory of gravitational waves turns out to be completely equivalent to the quantum theory of the Kanai-Caldirola oscillator which can be described by the quantum version of the Hamiltonian (\[HKC\]). On the grounds of what we learned in Section 2, the above identification suggests a route which can be successfully pursued whenever we are interested in the characterization of physical effects (quantum decoherence, squeezing, particle production etc.) emerging from the study of inflationary models in the early universe. Section 4 will be devoted to a preliminary exploration of the effectiveness of the idea in the context of expanding universe cosmological models.
Applications
=============
By taking full advantage of the formalism introduced in Section 2, we are in the position to study the dynamical system of the cosmological interest which are described by time-dependent oscillators. In doing so, a key point is the characterization of constants $A,B,C$ in Eq. (\[sigmasqrt\]). It is concerned with the initial (and boundary) conditions. All dynamical aspects under time evolution are enclosed into the function $\sigma$, which obeys the second order nonlinear differential equation (\[eqausiliaria\]). A general condition on $A,B,C$ is provided by Eq. (\[eqABCW0\]). It is not enough, however. Specification of the value and the first time derivative of $\sigma$ at fixed time is thereby needed. Another condition is associated with the requirement that at the initial time $\eta_{i}$ the time-dependent annihilation and creation operators, derived from Eq. (\[4.2\]), go into the standard Dirac-like form, Eq. (\[13.1\]). As for the final condition, it is helpful to reveal that in most cases we want the state at initial time to correspond to a vacuum state. This can be achieved easily under the minimization requirement for $E_{NM}=\frac{\hbar}{2}\left( \frac{M}{2}\sigma-\dot{\sigma}\right) ^{2}$. Indeed, since it provides a measure of the decoherence at the time $\eta$, it has to be vanishing when referring to a vacuum state at the initial time $\eta=\eta_{i}$. Under these circumstances, the whole set of initial conditions for $\sigma$ is given by$$\left\{
\begin{array}
[c]{l}AB-C^{2}=\frac{\kappa}{W_{0}^{2}}\;,\\
\sigma(t_{0})-\left( 4\omega^{2}\right) ^{-1/4}=0\;,\\
\dot{\sigma}(t_{0})-\frac{M(t_{0})}{2}\sigma(t_{0})=0\;.
\end{array}
\right.$$ In a cosmological framework of the FRW type, above system is translated into$$\left\{
\begin{array}
[c]{l}AB-C^{2}=\frac{\kappa}{W_{0}^{2}}\;,\\
\sigma(\eta_{i})=[4(n^{2}+K)]^{-1/4}\;,\\
\sigma^{\prime}\left( \eta_{i}\right) -\frac{a^{\prime}(\eta_{i})}{\sqrt{2 n}a(\eta_{i})}=0\;.
\end{array}
\right. \label{sysABCin}$$
In order to proceed with concrete analysis, it is very customary to resort to the spatially flat inflationary model based on the de Sitter metric. However, a more general and realistic description of the inflation may be provided by a quasi-de Sitter spacetime (see *e.g.* \[\[riotto\]\]). In this case, the Hubble rate is not exactly constant but, rather, it weakly conformal changes with time according to $\tilde{H}^{\prime
}=-\epsilon a^{2}\tilde{H}^{2}$ (that is, $aa^{\prime\prime}=(2-\epsilon
)\,a^{4}\tilde{H}^{2}\,=(2-\epsilon)\,a^{\prime2}$) where $\epsilon$ is a constant parameter. When $\epsilon$ vanishes one gets just the ordinary de Sitter spacetime. For small values of $\epsilon$, a quasi-de Sitter spacetime is associated with the scale factor$$a(\eta)=\frac{-1}{\tilde{H}(1-\epsilon)\eta}\;$$ ($\eta<0$). In the quasi-de Sitter spatially flat scenario, Eq. (\[tdoy\]) reads$$y^{\prime\prime}+\left[ n^{2}-\frac{(2+3\epsilon)}{(1-\epsilon)^{2}\eta^{2}}\right] \,y=0$$ and can be solved in terms of Bessel functions. Precisely, one has the two independent solutions $$y_{1}=\sqrt{-n\eta}J_{\nu}\left( -n\eta\right) \quad,\quad y_{2}=\sqrt{-n\eta}\,Y_{\nu}\left( -n\eta\right) \label{y12ddS}$$ where $\nu=\sqrt{\frac{1}{4}+\frac{(2+3\epsilon)}{(\epsilon-1)^{2}}}$ . The procedure outlined in the previous sections can be applied and we are led to the introduction of the basic function$$\begin{aligned}
\sigma & =\left( Ay_{1}^{2}+By_{2}^{2}+2Cy_{1}y_{2}\right) ^{\frac{1}{2}}=\nonumber\\
& =\sqrt{-n\eta}\left\{ AJ_{\nu}^{2}\left( -n\eta\right) +BY_{\nu}^{2}\left( -n\eta\right) +2CJ_{\nu}\left( -n\eta\right) Y_{\nu}\left(
-n\eta\right) \right\} ^{\frac{1}{2}}$$ where $A,B,C$ are determined by means of the system (\[sysABCin\]), $\eta_{i}$ denoting the conformal time of the beginning of the inflation. Once we are interested in a situation in which the system started very far in the past in a vacuum state, the Bessel function expansions $$\begin{aligned}
J_{\nu}(-n\eta) & \sim\sqrt{-\frac{2}{\pi n\eta}}\left[ \cos(-n\eta
-\frac{\nu}{2}\pi-\frac{\pi}{4})+O\left( \frac{1}{n\eta}\right) \right]
\,\,,\,\,\\
Y_{\nu}(-n\eta) & \sim\sqrt{-\frac{2}{\pi n\eta}}\left[ \sin(-n\eta
-\frac{\nu}{2}\pi-\frac{\pi}{4})+O\left( \frac{1}{n\eta}\right) \right]
\,\,,\end{aligned}$$ for $\nu$ fixed and $n\eta\rightarrow-\infty$ assists us in finding suitable constants $A,B,C$. By taking arbitrary asymptotically negative initial times, the leading terms of Bessel functions $J_{\nu}$, $Y_{\nu}$ give rise to the following behavior for the function: $$\sigma(\eta)=\sqrt{\frac{2}{\pi}}\left\{ A+(B-A)\sin^{2}(-n\eta-\frac{\nu}{2}\pi-\frac{\pi}{4})+C\sin(-2n\eta-\nu\pi-\frac{\pi}{2})+O\left( \frac
{1}{n\eta}\right) \right\} ^{\frac{1}{2}}\,\,.$$ Once the limit $\eta_{i}<<0$ is concerned a natural choice is given by $A=B=\frac{\pi}{4n}$, $C=0$ (recall that we already found this result for the case $\nu=3/2$ associated with the standard de Sitter metric background). So we obtain$$\sigma(\eta)=\sqrt{-\frac{\pi}{4}\eta}\,\left\{ J_{\nu}^{2}\left(
-n\eta\right) +Y_{\nu}^{2}\left( -n\eta\right) \right\} ^{\frac{1}{2}}=\sqrt{-\frac{\pi}{4}\eta}\,\left| H_{\nu}^{1}(-n\eta)\right|
\,\,.\label{sigmaqdS}$$ In the light of our previous results, the decoherence energy $E_{NM}$ at the time $\eta$ of gravitational waves in a quasi-de Sitter model of inflation can be evaluated by inserting (\[sigmaqdS\]) into formula (\[10.2\]). It then results$$E_{NM}=\frac{\hbar}{2}\left[ \sigma^{\prime}-\frac{a^{\prime}(\eta)}{a(\eta
)}\sigma\right] ^{2}=\frac{\hbar}{2}\left[ \sigma^{\prime}+\frac{\sigma
}{(1-\epsilon)\,\eta}\right] ^{2}\,\,$$ where$$\sigma^{\prime}+\frac{\sigma}{(1-\epsilon)\,\eta}=\sqrt{\frac{\pi}{4}}\left\{
\frac{n}{2}\frac{\sqrt{-\eta}}{\left| H_{\nu}^{1}(-n\eta)\right| }\left(
H_{\nu}^{1\ast}H_{\nu+1}+c.c\right) -\left( \nu+\frac{3-\epsilon
}{2(1-\epsilon)}\right) \frac{\left| H_{\nu}^{1}(-n\eta)\right| }{\sqrt{-\eta}}\right\} \,\,.\label{sdotseta}$$ Moreover, since$$\begin{aligned}
\mu(\eta) & =\sqrt{\frac{a^{2}(\eta)}{2n\,a^{2}(\eta_{i})}}\left\{ \left[
\frac{1}{2\sigma}+\frac{n\,a^{2}(\eta_{i})}{a^{2}(\eta)}\sigma(\eta)\right]
-i\left[ \sigma^{\prime}-\frac{a^{\prime}}{a}\sigma\right] \right\}
\;,\quad\\
\nu(\eta) & =\sqrt{\frac{a^{2}(\eta)}{2n\,a^{2}(\eta_{i})}}\left\{ \left[
\frac{1}{2\sigma}-\frac{n\,a^{2}(\eta_{i})}{a^{2}(\eta)}\sigma(\eta)\right]
-i\left[ \sigma^{\prime}-\frac{a^{\prime}}{a}\sigma\right] \right\} \;\end{aligned}$$ at an arbitrary time $\eta$ the Bogolubov coefficients are given by$$\begin{aligned}
\mu(\eta) & =\sqrt{\frac{1}{2n}\,\left( \frac{\,\eta{}_{i}}{\eta{}}\right)
^{\frac{2}{1-\epsilon}}}\left\{ \left[ \frac{1}{2\sigma}+\left(
\frac{\,\eta{}}{\eta_{i}{}}\right) ^{\frac{2}{1-\epsilon}}n\sigma\right]
-i\left[ \sigma^{\prime}+\frac{\sigma}{(1-\epsilon)\,\eta}\right] \right\}
\;,\quad\\
\nu(\eta) & =\sqrt{\frac{1}{2n}\,\left( \frac{\,\eta{}_{i}}{\eta{}}\right)
^{\frac{2}{1-\epsilon}}}\left\{ \left[ \frac{1}{2\sigma}-\left(
\frac{\,\eta{}}{\eta_{i}{}}\right) ^{\frac{2}{1-\epsilon}}n\sigma\right]
\ -i\left[ \sigma^{\prime}+\frac{\sigma}{(1-\epsilon)\,\eta}\right]
\right\} \;\end{aligned}$$ with $\sigma$ and $\sigma^{\prime}+\frac{\sigma}{(1-\epsilon)\,\eta}$ furnished by Eqs. (\[sigmaqdS\]) and (\[sdotseta\]), respectively. Finally, the phase $\theta$ can be evaluated. Due to Eq. (\[thetaFt\]), we get$$\theta\left( \eta\right) =\left. -i \ln\frac
{\psi(\eta)}{\psi^{\ast}(\eta)}\right| _{\eta_{i}}^{\eta}\;,$$ where$$\psi(t)=\sqrt{\frac{\pi}{4n}}\,e^{i\alpha}\left[ \,y_{1}+iy_{2}\right]
=\sqrt{\frac{\pi}{4n}}\,e^{i\alpha}\sqrt{-n\eta}\,H_{\nu}^{1}\,\,.$$ That is, $$\theta\left( \eta\right) =\left. 2 \,\theta_{\nu}^{1}\right| _{\eta_{i}}^{\eta}$$ where $\theta_{\nu}^{1}$ denotes the phase of the Hankel function $H_{\nu}^{1}$.
It is now instructive to focus on a standard de Sitter inflation. In this case $\epsilon=0$, $\nu=\frac{3}{2}$ and $$\sigma=\frac{1}{\sqrt{2n}}\sqrt{1+\frac{1}{n^{2}\eta^{2}}}\qquad.
\label{sigmadssimple}$$ So, the integration of $\frac{1}{\sigma^{2}}$ is straight, *i.e.* $$\int^{\eta}_{\eta_i}
\frac{d\eta'}{\sigma^{2}(\eta')}=
4n^{2}\int\frac{n^{2}\eta^{2}}{1+n^{2}\eta^{2}}d\eta=4n^{2}\left[ \eta-\frac{\tanh^{-1}(n\eta)}{n}\right] \,\,,$$ up to a constant of integration. The above two formulas provide the standard exact (and normalized) solution for the quantum fluctuations of a generic massless scalar field during a de Sitter inflation. Yet, it is interesting to consider a very long inflation by letting the system evolve towards $\eta\rightarrow0$. In this case $E_{NM}$ simply reads$$E_{NM}(\eta)=\frac{\hbar}{2}\left( \dot{\sigma}-\frac{M}{2}\sigma\right)
^{2}=\frac{\hbar}{4}\frac{n}{1+n^{2}\eta^{2}}\,\,.$$ Interestingly, the decoherence energy at $\eta=0$ is finite. Moreover, in the standard de Sitter phase, from Eq. (\[sigmadssimple\]) we obtain the Bogolubov coefficients $$\begin{aligned}
\mu(\eta) & =\frac{1}{2}\frac{\sqrt{n^{2}\eta_{i}^{2}}}{\sqrt{1+n^{2}\eta^{2}}}\left\{ \left[ 1+\frac{\eta^2}{\eta_{i}^{2}}
\right] -\frac{i}{n\eta\,}\right\} \;,\quad\\
\nu(\eta) & =\frac{1}{2}\frac{\sqrt{n^{2}\eta_{i}^{2}}}{\sqrt{1+n^{2}\eta^{2}}}\left\{ \left[ 1-\frac{\eta^2}{\eta_{i}^{2}}
\right] -\frac{i}{n\eta\,}\right\} \;,\end{aligned}$$ ($n^2 \eta_i^2 \gg1$) which in turns imply that $$\left| \nu(\eta)\right| ^{2}=\frac{1}{4}\left( \frac{\eta_{i}}{\eta}-\frac{\eta}{\eta_{i}}\right) ^{2} \label{nu2ds}$$ particles are created out the vacuum at the time $\eta$.
In more refined studies of cosmological effects in the expanding Universe, it turns out to be useful to introduce a cosmological model which allows one to take into account different evolutionary phases of the universe. Once the model has been specified and the equation (\[tdoy\]) solved, one can get an insight into physical effects associated with different cosmological stages. For instance, one can consider a simple cosmological model which includes the inflationary *(i)*, radiation-dominated *(e)* and matter-dominated *(m)* epochs \[\[kolb\]\]. The scale factor has the following dependence on the conformal time: $$\left\{
\begin{array}
[c]{ll}a_{i}(\eta)=-\frac{1}{H_{0}\eta}\,\,, & \qquad\quad\eta_{i}\leq\eta\leq
\eta_{e}<0 \,\, , \\
a_{e}(\eta)=\frac{\left( \eta-2\eta_{e}\right) }{H_{0}\,\eta_{e}^{2}}\,\,, &
\qquad\quad\eta_{e}\leq\eta\leq\eta_{m}\,\,,\\
a_{m}(\eta)=\frac{\left( \eta+\eta_{m}-4\eta_{e}\right) ^{2}}{4H_{0}\eta
_{i}^{2}\,(\eta_{2}-2\eta_{1})}\,\,, & \qquad\quad\eta\geq\eta_{m}\,\, ,
\end{array}
\right. \label{aeta3fasi}$$ where $H_{0}$ denotes the Hubble constant at the inflationary stage, and $\eta_{i}$ represents the beginning of the expansion. In order to determine the $\eta$-dependent amplitude $h_{n}$ for each epoch we have to solve Eq. (\[tdoy\]) with the corresponding varying frequencies, namely $$\Omega_{i}=\left( n^{2}-\frac{2}{\eta^{2}}\right) ^{\frac{1}{2}}\quad,\quad\Omega_{e}=n\quad,\quad\Omega_{m}=\left[ n^{2}-\frac{2}{\left(
\eta+\eta_{m}-4\eta_{e}\right) ^{2}}\right] ^{\frac{1}{2}}\;.
\label{5.7.7.2}$$ Notice that the frequency $\Omega_{e}$ is constant, while $\Omega_{i}$ and $\Omega_{m}$ are varying with the same temporal dependence. The frequency $\Omega_{i}$ characterizes the de Sitter era. The related equation of motion (\[tdoy\]) has already been solved. Taking care about matching data at $\eta_{e}$ and $\eta_{m}$ is needed for the knowledge of the complete form of $\sigma$. The associated $\sigma$’s and their derivatives have to join continuosly at $\eta_{e}$ and $\eta_{m}$, in fact. This step is needed to obtain all the physical information implied in formulas for the Bogolubov coefficients, the decoherence energy, the gravitational phase, and squeezing. Having in mind our previous discussion, employing the model by considering a quasi-de Sitter phase is straight.
In general, the vacuum expectation value of the number operator and the other quantities of the physical interest vary slowly with time if the expansion rate becomes arbitrarily slow. In case the espansion is stopped one should be able to recover time-independent Dirac operators. However, the circumstance does not mean that Bogolubov coefficients trivialize. This is because loss of coherence previously occurred due to the expansion dynamics. A typical situation may be that $\dot{\sigma}$ goes to zero but $\sigma$ does not. $E_{NM}$ goes to zero, indicating that when expansion is stopped the time-dependent gravitational pumping stops as well and there is no further decoherence. If expansion stops from time $\eta_{a}$ to time $\eta_{b}$, then $\forall\eta\in\lbrack\eta_{a},\eta_{b}]$ one gets $$\begin{aligned}
\mu(\eta) & =\sqrt{\frac{a^{2}(\eta_{a})}{2na^{2}(\eta_{i})}}\left[ \frac
{1}{2\sigma(\eta_{a})}+n\frac{a^{2}(\eta_{i})}{a^{2}(\eta_{a})}\,\sigma
(\eta_{a})\right] \;,\quad\\
\nu(\eta) & =\sqrt{\frac{a^{2}(\eta_{a})}{2na^{2}(\eta_{i})}}\left( \frac
{1}{2\sigma(\eta_{a})}-n\frac{a^{2}(\eta_{i})}{a^{2}(\eta_{a})}\,\sigma
(\eta_{a})\right) \;,\end{aligned}$$ which implies$$\left| \mu(\eta)\right| ^{2}=\frac{1}{2}\left[ \frac{a^{2}(\eta_{a})}{4na^{2}(\eta_{i})\,\sigma^{2}(\eta_{a})}+n\frac{a^{2}(\eta_{i})}{a^{2}(\eta_{a})}\,\sigma^{2}(\eta_{a}) +1 \right] \,\,.$$
To get a more clear insight into the results achieved in this section, a few comments are in order. Specifically, our formula for the decoherence energy associated with the dynamical evolution of the gravitational fluctuation modes on a background of the FRW type, is expressed in a very compact form in terms of the auxiliary field $\sigma (\eta)$ and the scale factor $a(\eta)$. On the other hand, the decoherence energy plays an essential role in the relationships for the Bogolubov coefficients. This aspect makes explicit how the energy lost owing to the decoherence effect may be exploited to excite the vacuum state of the model under consideration. Moreover, in our framework this mechanism would be quantified in a compact way by means of the formula $$\label{NU}
|{\nu}(\eta)|^{2}=\frac{a^2(\eta)}{2 n a^2(\eta_i)}
\left[ \left( \frac{1}{2\sigma}-\frac{n a^2(\eta_i)}{a^2(\eta)}
\sigma \right)^{2}
+ \frac{2}{\hbar}E_{NM} \right],$$ where $E_{NM}$ is the decoherence energy $(m=a^{2})$. With respect to other works, in our paper the role of the decoherence energy is made manifest. Furthermore, we observe that, remarkably, the auxiliary field $\sigma $ is nothing but the time-dependent amplitude of the mode solutions to Eq. (\[tdoy\]) for the redshifted gravitational field fluctuations.
Concluding remarks
==================
The main results achieved in this paper have been presented and widely discussed in the Introduction. Therefore we shall conclude by making some final comments concerning challenging perspectives which should be dwelt upon in future developments. The evolution equation of a mode with comoving wavenumber $n$ reduces to the harmonic oscillator equation with time-dependent mass and constant frequency. The approximation behind computations leading to the result actually are applicable only to the infrared region. On general grounds, one therefore expects that predictions for observables may depend sensitively on the physics on the length scales smaller than the Planck one. In order to take into account trans-Planckian physics, it has been recently suggested to make use of effective dispersion relations (see *e.g.* \[\[martin\]\]). The linear dispersion relation is thus replaced by a nonlinear one, $n_{eff}^{2}=a^{2}(\eta)\,F^{2}(n/a)$ , where $F(n/a)$ is an arbitrary function required to behave linearly whenever $n/a$ ($=k$) is below a certain threshold. A time dependent dispersion relation thus enters in the matter. As a consequence, the underlying dynamical model turns out to be that of the harmonic oscillator with both the mass $m$ (=$a^{2}$) and the frequency $\omega$ ($=n_{eff}$) depending on time. In principle, its quantization can still be pursued by resorting to the formalism of Section 2 and it will be studied in detail elsewhere. Nevertheless, a comment is in order. In the light of the discussion in \[\[landolfi\]\], one might wonder, in fact, on whether or not in the cosmological framework the minimum uncertainty criterium can be satisfied under time evolution for some physically reasonable function $F$. It is straightforwardly seen that this is not the case, generally speaking. Indeed, in the cosmological framework the criterium reads as $a^{2}n_{eff}=const$ and implies a purely cubic function $F$, say $F=\alpha_{0}n^{3}/a^{3}$. As a consequence, the uncertainty relation can be minimized only approximately. It is worth noting that this happens in the large wavenumbers limit of a special case of the generalized Corley-Jacobson dispersion relation introduced in \[\[martin\]\] (see Eq. (22) in \[\[martin\]\]).
References {#references .unnumbered}
==========
1. \[grishchuk\]L.P. Grishchuk and Y.V. Sidorov, Phys. Rev. **D42**, 3413 (1990) .
2. \[stoler\]D. Stoler, Phys. Rev. **D1**, 3217 (1970); **D4**, 1925 (1971).
3. \[yuen\]H.P. Yuen, Phys. Rev. **A13**, 2226 (1976) .
4. \[schumaker\] B.L. Schumaker, Phys. Rep. **135**, 317 (1986).
5. \[birrel\]N.D. Birrel and P.C.W. Davies, *Quantum fields in curved space*, Cambridge Univ. Press., Cambridge (1994) .
6. \[grishchuk2\]L.P. Grishchuk, Class. Quantum Grav. **10**, 2449 (1993).
7. L.P. Grishchuk and Y.V. Sidorov, Class. Quant. Grav. **6**, L161 (1989).
8. L.P. Grishchuk and Y.V. Sidorov, Phys. Rev. **D42**, 3413 (1990).
9. L.P. Grishchuk, Class. Quant. Grav. **10**, 2449 (1990).
10. \[bose\]S. Bose and L.P. Grishchuk, Phys. Rev. **D66**, 043529 (2002).
11. \[mukhanov\]V.F. Mukhanov, H.A. Feldman and R.H. Brandenberger, Phys. Rep. **215**, 203 (1992).
12. \[maggiore\]M. Maggiore, Phys. Rep. **331**, 283 (2000).
13. \[profilo\]G. Profilo and G. Soliani, Phys. Rev. **A44**, 2057 (1991) .
14. \[landolfi\]G. Landolfi, G. Ruggeri and G. Soliani, *Amplitude and phase of time dependent Hamiltonian systems under the minimum uncertainty condition*, preprint, quant-ph/0307109 (2003).
15. \[hollerhorst\]J.H. Hollenhorst, Phys. Rev. **D19**, 1669 (1979).
16. \[lewis\]H.R. Lewis,Jr. and W.B. Riesenfeld, J. Math. Phys. **10**, 1458 (1969).
17. \[eliezer\]C.J. Eliezer and A. Gray, SIAM J. Appl. Math. **30**, 463 (1976).
18. \[ermakov\]V.P. Ermakov, Universiteskiya Izvestiya Kiev **20**, 1 (1880).
19. \[pinney\]E. Pinney, Proc. Am. Math. Soc. **1**, 681 (1950).
20. \[goff\]S. Goff and D.F. St. Mary, J. Math. Anal. Appl. **140**, 95 (1989).
21. \[hartley\]J.G. Hartley and J.R. Ray, Phys. Rev. **D25**, 382 (1982) .
22. \[pedrosa\]I.A. Pedrosa, Phys. Rev. **D36**, 1279 (1987).
23. \[kanai\]E. Kanai, Prog. Theor. Phys. **3**, 440 (1948).
24. \[caldirola\]L. Caldirola, Nuovo Cimento **B87**, 241 (1983).
25. \[kin\]Kin-Wang Ng and A.D. Speliotopoulos, Phys. Rev. **D51**, 5636 (1995).
26. \[gradshteyn\]I.S. Gradshteyn and I.M. Ryzhik, *Table of Integrals, Series and Products*, Academic Press, New York (1965).
27. \[martin\]J. Martin and R.H. Brandenberger, Phys.Rev. **D63**, 123501 (2001).
28. \[riotto\]A. Riotto, *Inflation and the theory of cosmological perturbations*, Lectures delivered at the ”ICTP Summer School on Astroparticle Physics and Cosmology”, Trieste, 17 June - 5 July 2002, hep-ph/0210162 (2002).
29. \[kolb\]E.W. Kolb, H.S. Turner, *The early universe*, Addison-Wisley, Reading, 1990.
[^1]: e-mail: [email protected]
[^2]: e-mail: [email protected]
[^3]: e-mail: [email protected]
[^4]: e-mail: [email protected]
[^5]: Stoler \[\[stoler\]\] saw that, generally, states of the type $\left| \alpha,z\right\rangle $ do not describe wave packets relative to the minimum value of the product $\left( \Delta Q\right) \,\left( \Delta
P\right) $, where $\Delta$ means the variance operation (see Eq. (\[6.4\])). The state $\left| \alpha,z\right\rangle $ can describe a wave packet of minimum uncertainty only if $z$ is real ($\phi=0$). In the framework of quantum generalized oscillators, this corresponds to the condition (\[6.3\]).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We use the worldline numerics technique to study a cylindrically symmetric model of magnetic flux tubes in a dense lattice and the non-local Casimir forces acting between regions of magnetic flux. Within a superconductor the magnetic field is constrained within magnetic flux tubes and if the background magnetic field is on the order the quantum critical field strength, $B_k = \frac{m^2}{e} = 4.4 \times 10^{13}$ Gauss, the magnetic field is likely to vary rapidly on the scales where effects are important. In this paper, we construct a cylindrically symmetric toy model of a flux tube lattice in which the non-local influence of on neighbouring flux tubes is taken into account. We compute the effective action densities using the worldline numerics technique. The numerics predict a greater effective energy density in the region of the flux tube, but a smaller energy density in the regions between the flux tubes compared to a locally-constant-field approximation. We also compute the interaction energy between a flux tube and its neighbours as the lattice spacing is reduced from infinity. Because our flux tubes exhibit compact support, this energy is entirely non-local and predicted to be zero in local approximations such as the derivative expansion. This Casimir-Polder energy can take positive or negative values depending on the distance between the flux tubes, and it may cause the flux tubes in neutron stars to form bunches.
In addition to the above results we also discuss two important subtleties of determining the statistical uncertainties within the worldline numerics technique. Firstly, the distributions generated by the worldline ensembles are highly non-Gaussian, and so the standard error in the mean is not a good measure of the statistical uncertainty. Secondly, because the same ensemble of worldlines is used to compute the Wilson loops at different values of $T$ and $x_\mathrm{ cm}$, the uncertainties associated with each computed value of the integrand are strongly correlated. We recommend a form of jackknife analysis which deals with both of these problems.
author:
- Dan Mazur
- 'Jeremy S. Heyl'
bibliography:
- 'references.bib'
title: Casimir Interactions between Magnetic Flux Tubes in a Dense Lattice
---
Introduction
============
In this contribution, we make use of a numerical technique which can be used to compute the effective actions of external field configurations. The technique, called either worldline numerics or the Loop-Cloud Method, was first used by Gies and Langfeld [@Gies:2001zp] and has since been applied to computation of effective actions [@Gies:2001tj; @Langfeld:2002vy; @Gies:2005sb; @Gies:2005ym; @Dunne:2009zz] and Casimir energies [@Moyaerts:2003ts; @Gies:2003cv; @PhysRevLett.96.220401]. More recently, the technique has also been applied to pair production [@2005PhRvD..72f5001G] and the vacuum polarization tensor [@PhysRevD.84.065035]. Worldline numerics is able to compute quantum effective actions in the one-loop approximation to all orders in both the coupling and in the external field, so it is well suited to studying non-perturbative aspects of quantum field theory in strong background fields. Moreover, the technique maintains gauge invariance and Lorentz invariance. The key idea of the technique is that a path integral is approximated as the average of a finite set of $N_l$ representative closed paths (loops) through spacetime. These loops are not mapped to any spacetime lattice, so the theory maintains Lorentz invariance and is distinct from Lattice-based techniques. We use a standard Monte-Carlo procedure to generate loops which have large contributions to the loop average.
We will focus on calculations of the effective action in cylindrically symmetric, extended tubes of magnetic flux using the worldline numerics method. These configurations may be called flux tubes, strings, or vortices, depending on the context. Flux tubes are of interest in astrophysics because they describe magnetic structures near stars and planets, cosmic strings [@vilenkin2000cosmic], and vortices in the superconducting core of neutron stars [@2006pfsb.book..135S; @schmitt2010dense]. Outside of astrophysics, magnetic vortex systems are at the forefront of research in condensed matter physics for the role they play in superconducting systems and in research for their relation to center vortices, a gluonic configuration analogous to magnetic vortices which is believed to be important to quark confinement [@tHooft19781; @2003PrPNP..51....1G]. Currently, we are most interested in the roles played by magnetic flux tubes in neutron star cores.
Our motivation for discussing flux tubes comes from the fact that superconductivity is predicted in the nuclear matter of neutron stars and that some superconducting materials produce a lattice of flux tubes when placed in an external magnetic field. Superconductivity is a macroscopic quantum state of a fluid of fermions that, most notably, allows for the resistanceless conduction of charge. In 1933, Meissner and Ochsenfeld observed that magnetic fields are repelled from superconducting materials [@meisner33]. In 1935, F. and H. London described the Meissner effect in terms of a minimization of the free energy of the superconducting current [@london35]. Then, in 1957, by studying the superconducting electromagnetic equations of motion in cylindrical coordinates, Abrikosov predicted the possible existence of line defects in superconductors which can carry quantized magnetic flux through the superconducting material [@abrikosov57]. A more complete microscopic description of superconducting materials is given by BCS (Bardeen, Cooper, and Schrieffer) theory [@PhysRev.108.1175]. Interested readers may pursue more thorough reviews of superconductivity and superfluidity in neutron stars [@2006pfsb.book..135S; @schmitt2010dense].
Our calculations use the worldline numerics method which is reviewed in detail in section \[sec:qed-effect-act\]. In section \[sec:background\] of this article, we will briefly review the physics of magnetic flux tubes and of nuclear superconductivity in neutron stars to provide context and motivation. We present our models for solitary flux tubes and dense flux tube lattices in section \[sec:calculations\] as well as the details of the calculations. The results of our calculations for scalar and spinor are presented in section \[sec:periodic\_results\]. Our results suggest a small but possibly influential Casimir interaction between flux tubes in a dense lattice that may cause the flux tubes to form bunches. This result and other implications of our calculations are discussed in section \[sec:conclusion\].
QED Effective Action on the Worldline {#sec:qed-effect-act}
=====================================
Worldline numerics is built on the worldline formalism which was initially invented by Feynman [@PhysRev.80.440; @PhysRev.84.108]. Much of the recent interest in this formalism is based on the work of Bern and Kosower, who derived it from the infinite string-tension limit of string theory and demonstrated that it provided an efficient means for computing amplitudes in QCD [@PhysRevLett.66.1669]. For this reason, the worldline formalism is often referred to as ‘string inspired’. However, the formalism can also be obtained straight-forwardly from first-quantized field theory [@1992NuPhB.385..145S], which is the approach we will adopt here. In this formalism the degrees of freedom of the field are represented in terms of one-dimensional path integrals over an ensemble of closed trajectories.
We begin with the QED effective action expressed in the proper-time formalism [@Schwinger:1951], \[eqn:trln\] = -d\^4x \_0\^ e\^[-iTm\^2]{} ( e\^[iT( +e\^0\_)\^2]{} - e\^[iTp\^2]{}). To evaluate $\bra{x}e^{iT(\slashed{p}_\mu +
e\slashed{A}_\mu)^2}\ket{x}$, we recognize that it is simply the propagation amplitude $\braket{x,T}{x,0}$ from ordinary quantum mechanics with $(\slashed{p}_\mu + e\slashed{A}_\mu)^2$ playing the role of the Hamiltonian. We therefore express this factor in its path integral form: e\^[iT(\_+ e\_mu)\^2]{} = x\_() e\^[-\_0\^T d]{} . $\mathcal{N}$ is a normalization constant that we can fix by using our renormalization condition that the fermion determinant should vanish at zero external field: e\^[iTp\^2]{} = x\_p()e\^[-\_0\^T d]{} = e\^[iTp\^2]{} = , We may now write x\_() e\^[-\_0\^T d]{} = , where \[eqn:meandef\] \_x = is the weighted average of the operator $\hat{\mathcal{O}}$ over an ensemble of closed particle loops with a Gaussian velocity distribution.
Finally, combining all of the equations from this section results in the renormalized one-loop effective action for on the worldline: \[eqn:QEDWL\] \^[(1)]{}\[A\_\] = \_0\^ e\^[-m\^2T]{}d\^4x\_ .
Worldline Numerics
------------------
The averages, $\mean{\hat{\mathcal{O}}}$, defined by equation (\[eqn:meandef\]) involve functional integration over every possible closed path through spacetime. The velocities along the paths are drawn from a Gaussian velocity distribution. The prescription of the worldline numerics technique is to compute these averages approximately using a finite set of $N_l$ representative loops on a computer. The worldline average is then approximated as the mean of an operator evaluated along each of the worldlines in the ensemble: \_[i=1]{}\^[N\_l]{} \[x\_i()\].
### Loop Generation {#sec:loopgen}
The velocity distribution for the loops depends on the proper time, $T$. However, generating a separate ensemble of loops for each value of $T$ would be very computationally expensive. This problem is alleviated by generating a single ensemble of loops, $\vec{y}(\tau)$, representing unit proper time, and scaling those loops accordingly for different values of $T$: () = (/T) , \_0\^T d\^2() \_0\^1 dt \^2(t). There is no way to treat the integrals as continuous as we generate our loop ensembles. Instead, we treat the integrals as sums over discrete points along the proper-time interval $[0,T]$. This is fundamentally different from space-time discretization, however. Any point on the worldline loop may exist at any position in space, and $T$ may take on any value. It is important to note this distinction because the worldline method retains Lorentz invariance while lattice techniques, in general, do not.
The challenge of loop cloud generation is in generating a discrete set of points on a unit loop which obeys the prescribed velocity distribution. There are a number of different algorithms for achieving this goal that have been discussed in the literature. Four possible algorithms are compared and contrasted in [@Gies:2003cv]. In this work, we chose an algorithm dubbed “d-loops", which was first described in [@Gies:2005sb]. To generate a “d-loop", the number of points is iteratively doubled, placing the new points in a Gaussian distribution between the existing neighbour points. We quote the algorithm directly:
1. Begin with one arbitrary point $N_0=1$, $y_{N}$.
2. Create an $N_1=2$ loop, i.e., add a point $y_{N/2}$ that is distributed in the heat bath of $y_N$ with $$e^{-\frac{N_1}{4} 2 (y_{N/2} -y_{N})^2}. \label{yn2}$$
3. Iterate this procedure, creating an $N_k=2^k$ points per line (ppl) loop by adding $2^{k-1}$ points $y_{{qN}/{N_k}}$, $q=1,3,\dots, N_k-1$ with distribution $$e^{-\frac{N_k}{4} 2 [y_{qN/N_k} -\frac{1}{2}(y_{(q+1)N/N_k}+
y_{(q-1)N/N_k})]^2}. \label{ynk}$$
4. Terminate the procedure if $N_k$ has reached $N_k=N$ for unit loops with $N$ ppl.
5. For an ensemble with common center of mass, shift each whole loop accordingly.
The above d-loop algorithm was selected since it is simple and about 10% faster than previous algorithms, according to its developers, because it requires fewer algebraic operations. The generation of the loops is largely independent from the main program. Because of this, it was simpler to generate the loops using a Matlab script. This function was used to produce text files containing the worldline data for ensembles of loops. These text files were read into memory at the launch of each calculation. The results of this generation routine can be seen in figure \[fig:worldlineplot\].
When the kernel is called, every thread in every block executes the kernel function with its own unique identifier. Therefore, it is best to generate worldlines in integer multiples of the number of threads per block.
![A single discrete worldline loop shown at several levels of discretization. The loops form fractal patterns and have a strong parallel with Brownian motion. The colour represents the phase of a particle travelling along the loop, and begins at dark blue, progresses in a random walk through yellow, and ends at dark red. The total flux through this particular worldline at $T=1$ and $B=B_k$ is about $0.08 \pi/e$.[]{data-label="fig:worldlineplot"}](images/worldlineplot){width="\linewidth"}
Cylindrical Worldline Numerics
------------------------------
We now consider cylindrically symmetric external magnetic fields. In this case, we may simplify (\[eqn:QEDWL\]), \[eqn:cylEA\] &=& \_0\^\_ d.
### Cylindrical Magnetic Fields {#sec:cylindrical}
We have $\vec{B} =
B(\rho)$ with \[eqn:BWLN\] B() = + if we make the gauge choice that $A_0 =
A_\rho = A_z = 0$.
We begin by considering $\Aphi$ in the form = f\_() so that B\_z()= and the total flux is =F(f\_(L\_)-f\_(0)). It is convenient to express the flux in units of $\frac{2 \pi}{e}$ and define a dimensionless quantity = F.
### Wilson Loop
The quantity inside the angled brackets in equation (\[eqn:QEDWL\]) is a gauge invariant observable called a Wilson loop. We note that the proper time integral provides a natural path ordering for this operator. The Wilson loop expectation value is \[eqn:wilsonloop\] W\_[\_]{} = e\^[ie\_0\^T d(\_ + ()) ]{} e\^[\_0\^T d \_F\_(\_ + ())]{} \_[\_]{} , which we look at as a product between a scalar part ($e^{ie\int_0^T d\tau
\vec{A}(\vec{r}_{\mathrm{ cm}} + \vec{r}(\tau)) \cdot \dot{\vec{r}}}$) and a fermionic part ($\frac{1}{4} \mathrm{ tr} e^{\frac{e}{2}\int_0^T d\tau
\sigma_{\mu \nu}F_{\mu \nu}(\vec{r}_{\mathrm{ cm}} + \vec{r}(\tau))}$).
### Scalar Part
In a magnetic field, the scalar part is related to the flux through the loop, $\Phi_B$, by Stokes theorem: e\^[ie\_0\^T d ]{} &=& e\^[ied]{} = e\^[ie\_ d]{}\
& = & e\^[ie\_ d]{} = e\^[ie\_B]{}. Consequently, this factor accounts for the Aharonov-Bohm phase acquired by particles in the loop.
The loop discretization results in the following approximation of the scalar integral: ()d = \_[i=1]{}\^[N\_]{} \_[\^i]{}\^[\^[i+1]{}]{}()d. Using a linear parameterization of the positions, the line integrals are \_[\^i]{}\^[\^[i+1]{}]{}()d = \_0\^1dt ((t))(\^[i+1]{} - \^i). Using the same gauge choice outlined above ($\vec{A}=A_\phi \hat{\phi}$), we may write ((t)) = f\_(\^2)(-y,x,0), where we have chosen $f_\lambda(\rho^2)$ to depend on $\rho^2$ instead of $\rho$ to simplify some expressions and to avoid taking many costly square roots in the worldline numerics. We then have \_[\^i]{}\^[\^[i+1]{}]{}()d = (x\^iy\^[i+1]{}-y\^i x\^[i+1]{})\_0\^1 dt. The linear interpolation in Cartesian coordinates gives \[eqn:rhoi\] \_i\^2(t) = A\_i + 2B\_it + C\_i t\^2, where A\_i &=& (x\^i)\^2 + (y\^i)\^2\
B\_i&=&x\^i(x\^[i+1]{} - x\^i) + y\^i(y\^[i+1]{}-y\^i)\
C\_i &= &(x\^[i+1]{}-x\^i)\^2 + (y\^[i+1]{}-y\^i)\^2. \[eqn:Ci\]
In performing the integrals along the straight lines connecting each discretized loop point, we are in danger of violating gauge invariance. If these integrals can be performed analytically, than gauge invariance is preserved exactly. However, in general, we wish to compute these integrals numerically. In this case, gauge invariance is no longer guaranteed, but can be preserved to any numerical precision that’s desired.
### Fermion Part
For fermions, the Wilson loop is modified by a factor, W\^ &=& (e\^[ e\_0\^T d\_F\^]{})\
&=& (e\^[\_[x y]{} e\_0\^T dB(x())]{})\
&=&\
&=& , \[eqn:Wfermfpl\] where we have used the relation eB = 2 = 2f’\_(\^2).
This factor represents an additional contribution to the action because of the spin interaction with the magnetic field. Classically, for a particle with a magnetic moment $\vec{\mu}$ travelling through a magnetic field in a time $T$, the action is modified by a term given by \^0\_ = \_0\^T (()) d. The magnetic moment is related to the electron spin $\vec{\mu} = g\left(\frac{e}{2m}\right)\vec{\sigma}$, so we see that the integral in the above quantum fermion factor is very closely related to the classical action associated with transporting a magnetic moment through a magnetic field: \^0\_ = g() \_[x y]{} \_0\^T B\_z(x())d. Qualitatively, we could write W\^ \~.
As a possibly useful aside, we may want to express the integral in terms of $f_\lambda(\rho^2)$ instead of its derivative. We can do this by integrating by parts: \_0\^T df’\_(\^2())&=& \_[i=1]{}\^[N\_]{}\_0\^1 dt f’\_(\^2\_i())\
& = & \_[i=1]{}\^[N\_]{}\
& = & \_[i=1]{}\^[N\_]{}\_0\^1 dt, with $\rho_i^2(t)$ given by equations (\[eqn:rhoi\]) to (\[eqn:Ci\]). The second equality is obtained from integration-by-parts. In the third equality, we use the loop sum to cancel the boundary terms in pairs: \[eqn:Wfermfl\] W\^ = . In most cases, one would use equation (\[eqn:Wfermfpl\]) to compute the fermion factor of the Wilson loop. However, equation (\[eqn:Wfermfl\]) may be useful in cases where $f'_\lambda(\rho^2(\tau))$ is not known or is difficult to compute.
### Renormalization
The field strength renormalization counterterms result from the small $T$ behaviour of the worldline integrand. In the limit where $T$ is very small, the worldline loops are very localized around their center of mass. So, we may approximate their contribution as being that of a constant field with value $\vec{A}(\vec{r}_{\mathrm{ cm}})$. Specifically, we require that the field change slowly on the length scale defined by $\sqrt{T}$. This condition on $T$ can be written T || = | |.
When this limit is satisfied, we may use the exact expressions for the constant field Wilson loops to determine the small $T$ behaviour of the integrands and the corresponding counterterms.
The Wilson loop averages for constant magnetic fields in scalar and fermionic are \_ = eBT and \_ = .
Therefore, the integrand for fermionic in the limit of small $T$ is \[eqn:fermI\] I\_(T) &=&\
&&-+ (eB)\^4 m\^2 T\^2+( -)T\^3 ++O(T\^5). For scalar QED we have \[eqn:scalI\] I\_(T) &=&\
&&-+ ++O(T\^5).
Beyond providing the renormalization conditions, these expansions can be used in the small $T$ regime to avoid a problem with the Wilson loop uncertainties in this region. Consider the uncertainty in the integrand arising from the uncertainty in the Wilson loop: I(T) = W = W. In this case, even though we can compute the Wilson loops for small $T$ precisely, even a small uncertainty is magnified by a divergent factor when computing the integrand for small values of $T$. So, in order to perform the integral, we must replace the small $T$ behaviour of the integrand with the above expansions (\[eqn:fermI\]) and (\[eqn:scalI\]). Our worldline integral then proceeds by analytically computing the integral for the small $T$ expansion up to some small value, $a$, and adding this to the remaining part of the integral [@MoyaertsLaurent:2004]: \_0\^ I(T) dT = \_ + \_. Because this normalization procedure uses the constant field expressions for small values of $T$, this scheme introduces a small systematic uncertainty. To improve on the method outlined here, the derivatives of the background field can be accounted for by using the analytic forms of the heat kernel expansion to perform the renormalization [@Gies:2001tj].
Uncertainty analysis in worldline numerics {#ch:WLError}
------------------------------------------
So far in the worldline numerics literature, the discussions of uncertainty analysis have been unfortunately brief. It has been suggested that the standard deviation of the worldlines provides a good measure of the statistical error in the worldline method [@Gies:2001zp; @Gies:2001tj]. However, the distributions produced by the worldline ensemble are highly non-Gaussian (see figure \[fig:hists\]), and therefore the standard error in the mean is not a good measure of the uncertainties involved. Furthermore, the use of the same worldline ensemble to compute the Wilson loop multiple times in an integral results in strongly correlated uncertainties. Thus, propagating uncertainties through integrals can be computationally expensive due to the complexity of computing correlation coefficients.
The error bars on worldline calculations impact the conclusions that can be drawn from calculations, and also have important implications for the fermion problem, which limits the domain of applicability of the technique (see section \[sec:fermionproblem\]). It is therefore important that the error analysis is done thoughtfully and transparently. The purpose of this section is to contribute a more thorough discussion of uncertainty analysis in the worldline numerics technique to the literature in hopes of avoiding any confusion associated with the above-mentioned subtleties.
There are two sources of uncertainty in the worldline technique: the discretization error in treating each continuous worldline as a set of discrete points, and the statistical error of sampling a finite number of possible worldlines from a distribution. In this section, we discuss each of these sources of uncertainty.
### Estimating the Discretization Uncertainties {#sec:discunc}
The discretization error arising from the integral over $\tau$ in the exponent of each Wilson loop (see equation (\[eqn:wilsonloop\])) is difficult to estimate since any number of loops could be represented by each choice of discrete points. The general strategy is to make this estimation by computing the Wilson loop using several different numbers of points per worldline and observing the convergence properties.
The specific procedure adopted for this work involves dividing each discrete worldline into several worldlines with varying levels of discretization. Since we are using the d-loop method for generating the worldlines (section \[sec:loopgen\]), a $\frac{N_\mathrm{ppl}}{2}$ sub-loop consisting of every other point will be guaranteed to contain the prescribed distribution of velocities.
To look at the convergence for the loop discretization, each worldline is divided into three groups. One group of $\frac{N_\mathrm{ppl}}{2}$ points, and two groups of $\frac{N_\mathrm{ppl}}{4}$. This permits us to compute the average holonomy factors at three levels of discretization: \_[N\_/4]{} = , \_[N\_/2]{} = , and \_[N\_]{} = , where the symbols $\circ$, $\Box$, and $\triangle$ denote the worldline integral, $\int_0^T d\tau A(x_{CM}+x(\tau))\cdot \dot x$, computed using the sub-worldlines depicted in figure \[fig:Division\].
![Diagram illustrating the division of a worldline into three smaller interleaved worldlines[]{data-label="fig:Division"}](images/DiscretizationDivision){width="\linewidth"}
We may put these factors into the equation of a parabola to extrapolate the result to an infinite number of points per line (see figure \[fig:DiscErr\]): \_ \_[N\_]{} - 2 \_[N\_/2]{} +\_[N\_/4]{}. So, we estimate the discretization uncertainty to be \_ |\_[N\_]{} - \_|. Generally, the statistical uncertainties are the limitation in the precision of the worldline numerics technique. Therefore, $N_\mathrm{ppl}$ should be chosen to be large enough that the discretization uncertainties are small relative to the statistical uncertainties.
### Estimating the Statistical Uncertainties
We can gain a great deal of insight into the nature of the statistical uncertainties by examining the specific case of the uniform magnetic field since we know the exact solution in this case. Sections \[sec:notnormal\], \[sec:correlations\], and \[sec:groupingworldlines\] discuss the peculiarities of the statisical uncertainties in the worldline numerics method for the uniform magnetic field.
### The Worldline Ensemble Distribution is not Normal {#sec:notnormal}
A reasonable first instinct for estimating the error bars is to use the standard error in the mean of the collection of individual worldlines: ( W ) = . This approach has been promoted in early papers on worldline numerics [@Gies:2001zp; @Gies:2001tj]. In figure \[fig:resids\], we have plotted the residuals and the corresponding error bars for several values of the proper time parameter, $T$ in black. From this plot, it appears that the error bars are quite large in the sense that we appear to produce residuals which are considerably smaller than would be implied by the sizes of the error bars. This suggests that we have overestimated the size of the uncertainty.
\[fig:resids\]
We can see why this is the case by looking more closely at the distributions produced by the worldline technique. An exact expression for these distributions can be derived in the case of the constant magnetic field [@MoyaertsLaurent:2004]: \[eqn:exactdist\] w(y)&=&\_[n=-]{}\^with f()=. Figure \[fig:hists\] shows histograms of the worldline results along with the expected distributions. These distributions highlight a significant hurdle in assigning error bars to the results of worldline numerics.
Due to their highly non-Gaussian nature, the standard error in the mean is not a good characterization of the distributions that are produced. We should not interpret each individual worldline as an independent measurement of the mean value of these distributions; for large values of $BT$, almost all of our worldlines will produce answers which are far away from the mean of the distribution. This means that the variance of the distribution will be very large, even though our ability to determine the mean of the distribution is relatively precise because of the increasing symmetry about the mean as $T$ becomes large.
### Correlations between Wilson Loops {#sec:correlations}
Typically, numerical integration is performed by replacing the integral with a sum over a finite set of points from the integrand. We will begin the present discussion by considering the uncertainty in adding together two points (labelled $i$ and $j$) in our integral over $T$. Two terms of the sum representing the numerical integral will involve a function of $T$ times the two Wilson loop factors, I = g(T\_i) + g(T\_j) with an uncertainty given by I &=& | |\^2 ()\^2\
& & + | |\^2 ( )\^2\
& & + 2 | |\
& & \_[ij]{} () ()\
&=& g(T\_i)\^2 ()\^2 + g(T\_j)\^2 ()\^2+\
& & 2 | g(T\_i)g(T\_j) | \_[ij]{} () () and the correlation coefficient $\rho_{ij}$ given by \[eqn:corrcoef\] \_[ij]{} = . The final term in the error propagation equation takes into account correlations between the random variables $W(T_i)$ and $W(T_j)$. Often in a Monte Carlo computation, one can treat each evaluation of the integrand as independent, and neglect the uncertainty term involving the correlation coefficient. However, in worldline numerics, the evaluations are related because the same worldline ensemble is reused for each evaluation of the integrand. The correlations are significant (see figure \[fig:corr\]), and this term can’t be neglected. Computing each correlation coefficient takes a time proportional to the square of the number of worldlines. Therefore, it may be computationally expensive to formally propagate uncertainties through an integral.
The point-to-point correlations were originally pointed out by Gies and Langfeld who addressed the problem by updating (but not replacing or regenerating) the loop ensemble in between each evaluation of the Wilson loop average [@Gies:2001zp]. This may be a good way of addressing the problem. However, in the following section, we promote a method which can bypass the difficulties presented by the correlations by treating the worldlines as a collection of worldline groups.
### Grouping Worldlines {#sec:groupingworldlines}
Both of the problems explained in the previous two subsections can be overcome by creating groups of worldline loops within the ensemble. Each group of worldlines then makes a statistically independent measurement of the Wilson loop average for that group. The statistics between the groups of measurements are normally distributed, and so the uncertainty is the standard error in the mean of the ensemble of groups (in contrast to the ensemble of worldlines).
For example, if we divide the $N_l$ worldlines into $N_G$ groups of $N_l/N_G$ worldlines each, we can compute a mean for each group: \_[G\_j]{} = \_[i=1]{}\^[N\_l/N\_G]{}W\_i. Provided each group contains the same number of worldlines, the average of the Wilson loop is unaffected by this grouping: & = & \_[j=1]{}\^[N\_G]{} \_[G\_j]{}\
& = & \_[i=1]{}\^[N\_l]{} W\_i. However, the uncertainty is the standard error in the mean of the groups, = .
Because the worldlines are unrelated to one another, the choice of how to group them to compute a particular Wilson loop average is arbitrary. For example, the simplest choice is to group the loops by the order they were generated, so that a particular group number, $i$, contains worldlines $iN_l/N_G$ through $(i+1)N_l/N_G -1$. Of course, if the same worldline groupings are used to compute different Wilson loop averages, they will still be correlated. We will discuss this problem in a moment.
The basic claim of the worldline technique is that the mean of the worldline distribution approximates the holonomy factor. However, from the distributions in figure \[fig:hists\], we can see that the individual worldlines themselves do not approximate the holonomy factor. So, we should not think of an individual worldline as an estimator of the mean of the distribution. Thus, a resampling technique is required to determine the precision of our statistics. We can think of each group of worldlines as making an independent measurement of the mean of a distribution. As expected, the groups of worldlines produce a more Gaussian-like distribution (see figure \[fig:uncinmean\]), and so the standard error of the groups is a sensible measure of the uncertainty in the Wilson loop value.
We find that the error bars are about one-third as large as those determined from the standard error in the mean of the individual worldlines, and the smaller error bars better characterize the size of the residuals in the constant field case (see figure \[fig:resids2\]). The strategy of using subsets of the available data to determine error bars is called jackknifing. Several previous papers on worldline numerics have mentioned using jackknife analysis to determine the uncertainties, but without an explanation of the motivations or the procedure employed [@2005PhRvD..72f5001G; @PhysRevLett.96.220401; @Dunne:2009zz; @PhysRevD.84.065035].
The grouping of worldlines alone does not address the problem of correlations between different evaluations of the integrands. Figure \[fig:corr\] shows that the uncertainties for groups of worldlines are also correlated between different points of the integrand. However, the worldline grouping does provide a tool for bypassing the problem. One possible strategy is to randomize how worldlines are assigned to groups between each evaluation of the integrand. This produces a considerable reduction in the correlations, as is shown in figure \[fig:corr\]. Then, errors can be propagated through the integrals by neglecting the correlation terms. Another strategy is to separately compute the integrals for each group of worldlines, and then consider the statistics of the final product to determine the error bars. This second strategy is the one adopted for the work presented in this paper. Grouping in this way reduces the amount of data which must be propagated through the integrals by a factor of the group size compared to a delete-1 jackknife scheme, for example. In general, the error bars quoted in the remainder of this paper are obtained by computing the standard error in the mean of groups of worldlines.
### Uncertainties and the Fermion Problem {#sec:fermionproblem}
The fermion problem of worldline numerics is a name given to an enhancement of the uncertainties at large $T$ [@Gies:2001zp; @MoyaertsLaurent:2004]. It should not be confused with the fermion-doubling problem associated with lattice methods. In a constant magnetic field, the scalar portion of the calculation produces a factor of $\frac{BT}{\sinh{(BT)}}$, while the fermion portion of the calculation produces an additional factor $\cosh{(BT)}$. Physically, this contribution arises as a result of the energy required to transport the electron’s magnetic moment around the worldline loop. At large values of $T$, we require subtle cancellation between huge values produced by the fermion portion with tiny values produced by the scalar portion. However, for large $T$, the scalar portion acquires large relative uncertainties which make the computation of large $T$ contributions to the integral very imprecise.
This can be easily understood by examining the worldline distributions shown in figure \[fig:hists\]. Recall that the scalar Wilson loop average for these histograms is given by the flux in the loop, $\Phi_B$: = = . For constant fields, the flux through the worldline loops obeys the distribution function [@MoyaertsLaurent:2004] f(\_B) = . For small values of $T$, the worldline loops are small and the amount of flux through the loop is correspondingly small. Therefore, the flux for small loops is narrowly distributed about $\Phi_B = 0$. Since zero maximizes the Wilson loop ($e^{i0}=1$), this explains the enhancement to the right of the distribution for small values of $T$. As $T$ is increased, the flux through any given worldline becomes very large and the distribution of the flux becomes very broad. For very large $T$, the width of the distribution is many factors of $2\pi/e$. Then, the phase ($e \Phi_B\mod{2\pi}$) is nearly uniformly distributed, and the Wilson loop distribution reproduces the Chebyshev distribution ( the distribution obtained from projecting uniformly distributed points on the unit circle onto the horizontal axis), \_[T]{}w(y) = .
The mean of the Chebyshev distribution is zero due to its symmetry. However, this symmetry is not realized precisely unless we use a very large number of loops. Since the width of the distribution is already 100$\times$ the value of the mean at $T=6$, any numerical asymmetries in the distribution result in very large relative uncertainties of the scalar portion. Because of these uncertainties, the large contribution from the fermion factor are not cancelled precisely.
This problem makes it very difficult to compute the fermionic effective action unless the fields are well localized [@MoyaertsLaurent:2004]. For example, the fermionic factor for non-homogeneous magnetic fields oriented along the z-direction is . For a homogeneous field, this function grows exponentially with $T$ and is cancelled by the exponentially vanishing scalar Wilson loop. For a localized field, the worldline loops are very large for large values of $T$, and they primarily explore regions far from the field. Thus, the fermionic factor grows more slowly in localized fields, and is more easily cancelled by the rapidly vanishing scalar part.
Computing an Effective Action
-----------------------------
The ensemble average in the effective action is simply the sum over the contributions from each worldline loop, divided by the number of loops in the ensemble. Since the computation of each loop is independent of the other loops, the ensemble average may be straightforwardly parallelized by generating separate processes to compute the contribution from each loop. For this parallelization, four Nvidia Tesla C1060 were used through the parallel programming framework. Because can spawn thousands of parallel processing threads with much less computational overhead than an cluster, they excel at handling a very large number of parallel threads, although the clock speed is slower and fewer memory resources are typically available. The architecture has recently been used by another group for computing Casimir forces using worldline numerics [@2011arXiv1110.5936A]. More detailed information about the technical implementation of these calculations on the architecture, including a listing of the source code, can be found in [@2012PhDT........21M].
Once the ensemble average of the Wilson loop has been computed, computing the effective action is a straightforward matter of performing numerical integrals. The effective action density is computed by performing the integration over proper time, $T$. Then, the effective action is computed by performing a spacetime integral over the loop ensemble center of mass. In all cases where a numerical integral was performed, Simpson’s method was used [@burden2001numerical]. Integrals from 0 to $\infty$ were mapped to the interval $[0,1]$ using substitutions of the form $x =
\frac{1}{1+T/T_\mathrm{ max}}$, where $T_\mathrm{ max}$ sets the scale for the peak of the integrand. In the constant field case, for the integral over proper time, we expect $T_\mathrm{ max} \sim 3/(eB)$ for large fields and $T_\mathrm{ max} \sim 1$ for fields of a few times critical or smaller. In section \[ch:WLError\], we presented a detailed discussion of how the statistical and discretization uncertainties can be computed in this technique.
Verification and Validation {#sec:verify}
---------------------------
The worldline numerics software can be validated and verified by making sure that it produces the correct results where the derivative expansion is a good approximation, and that the results are consistent with previous numerical calculations of flux tube effective actions. For this reason, the validation was done primarily with flux tubes with a profile defined by the function f\_(\^2) = . For large values of $\lambda$, this function varies slowly on the Compton wavelength scale, and so the derivative expansion is a good approximation. Also, flux tubes with this profile were studied previously using worldline numerics [@Moyaerts:2003ts; @MoyaertsLaurent:2004].
Among the results presented in [@MoyaertsLaurent:2004] is a comparison of the derivative expansion and worldline numerics for this magnetic field configuration. The result is that the next-to-leading-order term in the derivative expansion is only a small correction to the the leading-order term for $\lambda \gg \lambda_e$, where the derivative expansion is a good approximation. The derivative expansion breaks down before it reaches its formal validity limits at $\lambda \sim \lambda_e$. For this reason, we will simply focus on the leading order derivative expansion, which we call the locally-constant-field () approximation. The effective action of in the approximation is given in cylindrical symmetry by \[eqn:LCFferm\] \^[(1)]{}\_ &=& \_0\^dT \_0\^\_ d\_\
& &{eB(\_)T\
& & - 1 -(eB(\_)T)\^2}.
Figure \[fig:igrandvsT\] shows a comparison between the proper time integrand, , and the approximation result for a flux tube with $\lambda = \lambda_e$ and $\mathcal{F} = 10$. In this case, the approximation is only appropriate far from the center of the flux tube, where the field is not changing very rapidly. In the figure, we can begin to see the deviation from this approximation, which gets more pronounced closer to the center of the flux tube (smaller values of $\rho$).
The effective action density for a slowly varying flux tube is plotted in figure \[fig:vsrhocm\] along with the approximation. In this case, the approximation agrees within the statistics of the worldline numerics.
Astrophysical Background {#sec:background}
========================
Nuclear Superconductivity in Neutron Stars {#subsec:superconductivity}
------------------------------------------
In the dense nuclear matter of a neutron star, it may be possible to have neutron superfluidity, proton superconductivity, and even quark colour superconductivity [@2006pfsb.book..135S; @schmitt2010dense]. The first prediction of neutron-star superfluidity dates back to Migdal in 1959 [@1959NucPh..13..655M]. The arguments which make superfluidity seem likely are based on the temperatures of neutron stars. A short time after their creation, neutron stars are very cold compared to nuclear energy scales. The temperature in the interior may be a few hundred keV. Studies of nuclear matter show that the transition temperature is $T_c \gtrsim 500$ keV [@1989tns..conf..457S]. So, it is expected that the nuclear matter in a neutron star forms condensates of Cooper pairs.
Because of this, some fraction of neutrons in the inner crust of a neutron star are expected to be superfluid. These neutrons make up about a percent of the moment of inertia of the star and are weakly coupled to the nuclear crystal lattice which makes up the remainder of the inner crust. If the neutron vortices in the superfluid component move at nearly the same speed as the nuclear lattice, the vortices can become pinned to the lattice so that the superfluid shares an angular velocity with the crust. This pinning between the fluid and solid crust has observable impacts on the rotational dynamics of the neutron star, for the same reasons that hard-boiling an egg (pinning the yolk to the shell) produces an observable difference in the way it spins.
This picture of a neutron superfluid co-rotating with a solid crust has been used to interpret several types of pulsar timing anomalies. Pulsars are nearly perfect clocks, although they gradually spin-down as they radiate energy. Occasionally, though, pulsars demonstrate deviations from their expected regularity. A glitch is an abrupt increase in the rotation and spin-down rate of a pulsar, followed by a slow relaxation to pre-glitch values over weeks or years. This behaviour is consistent with the neutron superfluid suddenly becoming unpinned from the crust and then dynamically relaxing due to its weak coupling to the crust until it is pinned once again [@1975Natur.256...25A].
The neutron star in Cassiopeia A has been observed to be rapidly cooling [@2010ApJ...719L.167H]. The surface temperature has decreased by about 4% over 10 years. This observation is also strong evidence of superfluidity and superconductivity in neutron stars [@2011PhRvL.106h1101P; @2011MNRAS.412L.108S]. The observed cooling is too fast to be explained by the observed x-ray emissions and standard neutrino cooling. However, the cooling is readily explained by the emission of neutrinos during the formation of neutron Cooper pairs. Based on such a model, the superfluid transition temperature of neutron star matter is $\sim 10^{9}$ K or $\sim 90$ keV.
Further hints regarding superfluidity in neutron stars come from long-term periodic variability in pulsar timing data. For example, variabilities in PSR B1828-11 were initially interpreted as free precession (or wobble) of the star [@2000Natur.406..484S]. If neutron stars can precess, observations could strongly constrain the ratio of the moments of inertia of the crust and the superfluid neutrons. Moreover, the existence of flux tubes (type-II superconductivity) in the crust is generally incompatible with the slow, large amplitude precession suggested by PSR B1828-11 [@2000Natur.406..484S]. The neutron vortices would have to pass through the flux tubes, which should cause a huge dissipation of energy and a dampening of the precession which is not observed [@PhysRevLett.91.101101]. However, recent arguments suggest that the timing variability data is not well explained by free precession and that it more likely suggests that the star is switching between two magnetospheric states [@2010Sci...329..408L]. Nevertheless, other authors suggest not being premature in throwing out the precession hypothesis without further observations [@2012MNRAS.420.2325J].
Magnetic Flux Tubes in Neutron Stars
------------------------------------
If a magnetic field is able to penetrate the proton superfluid on a microscopic level, it must do so by forming a triangular Abrikosov lattice with a single quanta of flux in each flux tube. So, the density of flux tubes is given simply by the average field strength. If the distance between flux tubes is $l_f$, the flux in a circular region within $l_f/2$ of a flux tube is given by F = = 2\_0\^[l\_f/2]{} B dwhere we have introduced a dimensionless measure of flux $\mathcal{F}
= \frac{e}{2\pi}F$. So, the distance between flux tubes is l\_f = . If the magnetic field is the quantum critical field strength, $B_k = \frac{m^2}{e} = 4.4 \times 10^{13}$ Gauss, then the flux tubes are separated by a few Compton electron wavelengths. This is particularly interesting since this is the distance scale associated with non-locality in .
The size of a flux tube profile in laboratory superconductors is on the nanometer or micron scale [@poole2007superconductivity]. Because the flux is fixed, the size of the tube profile determines the strength of the magnetic field within the tube. For laboratory superconductors the field strengths are small compared to the quantum critical field, and the field is slowly varying on the scale of the Compton wavelength. In this case, the quantum corrections to the free energy are known to be much smaller than the classical contribution (see section \[sec:EAisoflux\]). The size scale for the flux tubes in a superconductor is determined by the London penetration depth. In a neutron star, this quantity is estimated to be a small fraction of a Compton wavelength, much smaller than in laboratory superconductors [@lrr-2008-10; @PhysRevLett.91.101101]. In this case, the magnetic field strength at the centre of the tube exceeds the quantum critical field strength and the field varies rapidly, rendering the derivative expansion description of the effective action unreliable.
The Ginzburg-Landau parameter, equation is the ratio of the proton coherence length, $\xi_p \sim 30$ fm, and the London penetration depth of a proton superconductor, $\lambda_p \sim 80$ fm [@PhysRevLett.91.101101]. = \~2 where $\kappa>1/\sqrt{2}$ signals Type-II behavior. We therefore expect that the proton Cooper pairs most likely form a type-II superconductor [@2006pfsb.book..135S]. However, it is possible that physics beyond what is taken into account in the standard picture affects the free energy of a magnetic flux tube. In that case, the interaction between two flux tubes may indeed be attractive in which case the neutron star would be a type-I superconductor.
QED Effective Actions of Flux Tubes {#sec:EAisoflux}
-----------------------------------
Vortices of magnetic flux have very important impacts on the quantum mechanics of electrons. In particular, the phase of the electron’s wavefunction is not unique in such a magnetic field. This is demonstrated by the Aharonov-Bohm effect [@0370-1301-62-1-303; @PhysRev.115.485]. The first calculations of the fermion effective energies of these configurations were for infinitely thin Aharonov-Bohm flux strings [@Gornicki1990271; @1998MPLA...13..379S]. Calculations for thin strings were also performed for cosmic string configurations [@0264-9381-12-5-013]. For these infinitely-thin string magnetic fields, the energy density is singular for small radii. So, it is not possible to define a total energy per unit length. Another approach was to compute the effective action for a finite-radius flux tube where the magnetic flux was confined entirely to the surface of the tube [@PhysRevD.51.810]. This approach results in infinite classical energy densities as well.
Physical flux tube configurations would have a finite radius. The earliest paper to deal with finite radius flux tubes in QED considered the effective action of a step-function profiled flux tube using the Jost function of the related scattering problem [@1999PhRvD..60j5019B]. One of the conclusions from this research was that the quantum correction to the classical energy was relatively small for any value of the flux tube size, for the entire range of applicability of . The techniques from this study were soon generalized to other field profiles including, a delta-function cylindrical shell magnetic field [@PhysRevD.62.085024], and more realistic flux tube configurations such as the Gaussian [@2001PhRvD..64j5011P] and the Nielsen-Olesen vortex [@2003PhRvD..68f5026B]. Flux tube vacuum energies were also analyzed extensively using a spectral method [@2005NuPhB.707..233G; @2006JPhA...39.6799W; @Weigel:2010pf].
The effective actions of flux tubes have been previously analyzed using worldline numerics [@Langfeld:2002vy]. This research investigated isolated flux tubes, but also made use of the loop cloud method’s applicability to situations of low symmetry to investigate pairs of interacting vortices. One conclusion from that investigation was that the fermionic effects resulted in an attractive force between vortices with parallel orientations, and a repulsive force between vortices with anti-parallel orientations. Due to the similarity in scope and technique, the latter mentioned research is the closest to the research presented in this paper.
Calculations {#sec:calculations}
============
In this section, we will further explore the nature of this phenomenon in using a highly parallel implementation of the worldline numerics technique implemented on a heterogeneous and architecture. Specifically, we explore cylindrically symmetric magnetic field profiles for isolated flux tubes and periodic profiles designed to model properties of a triangular lattice. For these calculations, the classical magnetic field configurations are a chosen input to the algorithms and the physical processes that may have created the field configurations do not factor in to the calculations. The worldline numerics algorithm cannot be straight-forwardly applied to spinor calculations in our model lattice because of the well-known fermion problem of worldline numerics (see section \[sec:fermionproblem\]) [@Gies:2001zp; @MoyaertsLaurent:2004]. However, the problem does not affect the scalar QED () calculations. Therefore, we explore the quantum-corrected energies of isolated flux tubes for both scalar and spinor electrons and use this comparison to speculate about the relationship of our cylindrical lattice model and the spinor energies of an Abrikosov lattice of flux tubes that may be found in neutron stars.
Isolated Flux Tubes
-------------------
As discussed in section \[sec:cylindrical\] we will focus on fields with a cylindrical symmetry. In particular to explore isolated magnetic flux tubes, we consider the following profile function (as we used earlier in section \[sec:verify\]): f\_(\^2) = . This gives a magnetic field with a profile \[eqn:isoB\] B\_z(\^2) = . This profile is a smooth flux tube representation that can be evaluated quickly. Moreover, flux tubes with this profile were studied previously using worldline numerics [@Moyaerts:2003ts; @MoyaertsLaurent:2004].
Cylindrical Model of a Flux Tube Lattice
----------------------------------------
![In a type-II superconductor, there are neighbouring flux tubes arranged in a hexagonal Abrikosov lattice which have a non-local impact on the effective action of the central flux tube (left). In our model, we account for the contributions from these neighbouring flux tubes in cylindrical symmetry by including surrounding rings of flux (right).[]{data-label="fig:latticesymmetry"}](images/latticesymmetry){width="\linewidth"}
In a neutron star, we do not have isolated flux tubes. The tubes are likely arranged in a dense lattice with the spacing between tubes on the order of the Compton wavelength, with the size of a flux tube a few percent of the Compton wavelength. Specifically, the maximum size of a flux tube is on the order of the coherence length of the superconductor, which for neutron stars has been estimated to be $\xi
\approx 30$ fm [@PhysRevLett.91.101101]. This situation can be directly computed in the worldline numerics technique. However, this requires us to integrate over two spatial dimensions instead of one. Moreover, it requires the use of more loops to more precisely probe the spatial configurations of the magnetic field. Despite these problems, it is very interesting to consider a dense flux tube lattice. Unlike the isolated flux tube, the wide-tube limit of the configuration doesn’t have zero field, but an average, uniform background field. If this background field is the size of the critical field, there are interesting quantum effects even in the wide-tube limit.
In this section, we build a cylindrically symmetric toy model of a hexagonal flux tube lattice. We focus on one central flux tube and treat the surrounding six flux tubes as a continuous ring with six units of flux at a distance $a$ from the central tube. The next ring will contain twelve units of flux at a distance of $2a$, etc (see figure \[fig:latticesymmetry\]). Because of this condition, the average strength of the field is fixed, and the field becomes uniform in the wide tube limit instead of going to zero. For small values of $\lambda$, we will have non-local contributions from the surrounding rings in addition to the local contributions from the central flux tube. This strategy will result in a simple model relative to the sophisticated flux tube lattice models used in the context of superconducting physics. However, the simplifications are appropriate for a first study of the non-local QED interactions between the regions of flux.
It is difficult to construct a model of this scenario if the flux tubes bleed into one another as they are placed close together. For example, with Gaussian flux tubes or flux tubes with the profile used in the previous section, it is difficult to increase the width of the flux tubes while accounting for the magnetic flux that bleeds out of their regions. Moreover, it is difficult to integrate these schemes to find the profile function $f_\lambda(\rho)$ which is needed to compute the scalar part of the Wilson loop. In order to keep each tube as a distinct entity which stays within its assigned region, we assign a smooth function with compact support to represent each tube. This is most easily done with the bump function, $\Psi(x)$, defined as (x) =
e\^[-1/(1-x\^2)]{} & |x| < 1\
0 &
. This function can be viewed as a rescaled Gaussian. We start by defining the magnetic field outside of the central flux tube. Here, the magnetic field is a constant background field, with the flux ring contributing a bump of width $\lambda$. The height of the bump must go to zero as the width of the flux tube approaches the distance between flux tubes, and should become infinite as the flux tube width goes to zero: B\_z(>) = B\_[bg]{} + A(). with $n\equiv \lfloor\frac{\rho+a/2}{a}\rfloor$.
If we require 6 units of flux in the first outer ring, 12 in the second, and so on (see figure \[fig:latticesymmetry\]), the size of the background field is fixed to $B_{\rm bg} = \frac{6 \mathcal{F}}{ea^2}$. The total flux contribution due to the $\lambda$-dependent terms must be zero:
\_[(n-1/2)a]{}\^[(n+1/2)a]{} A() d= 0
\_[-1]{}\^1 (x+na)(x)dx-Ba\^2n = 0. This fixes the value of the constant $B$ to B=. The numerical constant $q_1$ is defined by q\_1 = \_[-1]{}\^[1]{}(x)dx 0.443991. For a given bump amplitude, $A$, the magnetic field will become negative if $\lambda$ becomes small enough. Therefore, we replace $A$ with its maximum value for which the field is positive if $\lambda >
\lambda_{\rm min}$ for some choice of minimum flux tube size: A=. The choice of $\lambda_{\rm min}$ sets the tube width at which the field between the flux tubes vanishes. If $\lambda < \lambda_{\rm
min}$, the magnetic field between the flux tubes will point in the $-\hat{\vec{z}}$-direction.
Because we are trying to fit a hexagonal peg into a round hole, we must treat the central flux tube differently. For example, the average field inside the central region for a unit of flux, is different than the average field in the exterior region. Therefore, even when $\lambda \rightarrow a$, the field cannot be quite uniform. We consider the field in the central region to be a constant field with a bump centered at $\rho = 0$: B\_z() &=& () +\
& & () (2(-na)/) . The magnetic field profile defined by these equations is shown in figures \[fig:periodicB\] and \[fig:B3d\]. The current density required to created fields with this profile is shown in figure \[fig:current\].
The Classical Action
--------------------
The classical action, $\Gamma^0$, is infinite for this configuration. To obtain a finite result, we must look at the action per unit length in the z-direction, per unit time, and per flux tube region (for $\rho < a/2$). The action for such a region in cylindrical coordinates is given by
= -\_0\^[a/2]{}B\_z()\^2d. After substituting the value of equation (\[eqn:ffbless\]), the magnetic field in the interior region: &=& -\_0\^[a/2]{}\^2 d. After some algebra, we are left with an expression for the classical action, &=& \_0\^[a/2]{}B\_z()\^2d\
& = & -{4 +() }, where $q_3$ is another numerical constant related to integrating the bump function: \[eqn:q3\] q\_3 \_0\^1 x((x))\^2 dx = \_0\^1 x e\^[-]{}dx 0.0187671.
Integrating to Find the Potential Function
------------------------------------------
To compute the Wilson loops, it is generally required to use the vector potential which describes the magnetic field. For us, this means that we must find $f_\lambda(\rho)$ for our magnetic field model. This could always be done numerically, but can be computationally costly since it is evaluated by every discrete point of every worldline in the ensemble. For computations on the device, an increase in the complexity of the kernel often means that less memory resources are available per processing thread, limiting the number of threads that can be computed concurrently. It is therefore preferable to find an analytic expression for this function. From equation (\[eqn:isoB\]), this function is related to the integral of the magnetic field with respect to $\rho^2$. For the inner region, we have f\_(a/2) = 1 + \_[a/2]{}\^’B(’>a/2)d’. We may put the magnetic field, equation (\[eqn:Bzext\]), into this expression to get f\_(>a/2) &=& 1 + (-1) () +\
& & () \_[a/2]{}\^’()d’. The remaining integral is over every bump between $\rho'=a/2$ and $\rho'=\rho$. We express the result as a term which accounts for each completely integrated bump, and an integral over the partial bump if $\rho$ is within a bump: f\_(>a/2) &=& 1 + (-1) ()+\
& & 3n(n-1)() +\
& & ()(2(-na)/), where (x\_0) =
0 & x\_0 -1\
\_[-1]{}\^[x\_0]{}xe\^[-]{}dx + \_[-1]{}\^[x\_0]{}e\^[-]{}dx& |x\_0| < 1\
& x\_0 1
. One of the integrals in $\chi(x_0)$ can be expressed in terms of the exponential integral: \_[-1]{}\^[x\_0]{}xe\^[-]{}dx =. The remaining integral cannot be simplified analytically. To use this integral in our numerical model, it must be computed for each discrete point on each loop for each $\rho_{\rm cm}$ and $T$ value. Therefore, it is worthwhile to consider an approximate expression which models the integral, and can be computed faster than performing a numerical integral each time. To find this approximation, we computed the numerical result at 300 values between $x_0=-1.2$ and $x_0=1.2$. The data was then input into Eureqa Formulize, a symbolic regression program which uses genetic algorithms to find analytic representations of arbitrary data [@formulize]. A similar technique has been used to produce approximate analytic solutions of ODEs [@geneticODEs]. The result is a model of the numerical data points with a maximum error of 0.0001 on the range $|x_0| < 1.0$: \_[-1]{}\^[x\_0]{} e\^[-]{}dx where g(x\_0) = This function evaluates ten times faster than the numerical integral evaluated at the same level of precision with the Gaussian quadrature library functions and with fewer memory registers. Using this approximation introduces a systematic uncertainty which is small compared to that associated with the discretization of the loop integrals, and considerably smaller than the statistical error bars. Using these expressions for the integrals, we can express $f_\lambda(\rho)$ in any region in terms of exponential integrals, exponential, and trigonometric functions with suitable precision. Furthermore, for computation we may express the exponential integral as a continued fraction. The profile function, $f_\lambda(\rho)$, is plotted in figure \[fig:periodicfl\].
Results {#sec:periodic_results}
=======
Comparing Scalar and Fermionic Effective Actions
------------------------------------------------
Because of the fermion problem of worldline numerics [@Gies:2001zp; @MoyaertsLaurent:2004], the 1-loop effective action for the cylindrical flux tube lattice model was not computed for the case of spinor . Performing this calculation in the spinor case would require subtle numerical cancellations between large terms, and represents a significant numerical challenge. Fortunately, the fermion problem does not affect the scalar case. So, we will analyze this model for . However, in this section we will compare the scalar and fermionic effective actions for isolated flux tubes to demonstrate that the behaviour of both theories is qualitatively and numerically similar.
For isolated flux tubes, the decay of the magnetic field for large distances protects the calculations from the fermion problem. Therefore, the effective action can be computed for both scalar and spinor . In figure \[fig:scalfermiratio\], we plot the ratio of the spinor to scalar 1-loop correction term for identical magnetic fields, along with the prediction of the approximation for large values of $\lambda$. The approximation in is given by \[eqn:LCFscal\] \^[(1)]{}\_[scal]{} &=& -\_0\^dT \_0\^\_[cm]{} d\_[cm]{}\
& &{ - 1 +(eB(\_[cm]{})T)\^2}. This can be compared to the spinor approximation, \[eqn:LCFferm2\] \^[(1)]{}\_ &=& \_0\^dT \_0\^\_ d\_\
& &{eB(\_)T\
& & - 1 -(eB(\_)T)\^2}.
There are two important notes to make about figure \[fig:scalfermiratio\]. Firstly, the approximation is only a good approximation for $\lambda \gg \lambda_e$, and isn’t accurate when pushed near its formal validity limits [@MoyaertsLaurent:2004]. The second note is that the statistics of the points computed with worldline numerics are strongly correlated. So, we conclude that the 1-loop correction is larger than the correction for large $\lambda$, and this appears to be reversed for small $\lambda$. However, the large worldline numerics error bars and the invalidity of the approximation near $\lambda = \lambda_e$ prevent us from seeing how this transition happens. Nevertheless, the main conclusion from this figure is that the scalar 1-loop correction reflects the behaviour of the full 1-loop correction to within a factor of about 2 over a wide range of flux tube widths for isolated flux tubes.
Besides using a finite field profile, the fermion problem can also be circumvented by increasing the electron mass. The square of the electron mass sets the scale for the exponential suppression of the large proper time Wilson loops that contribute to the fermion problem. However, if we increase the fermion mass, we are reducing the Compton wavelength of our theory so that the flux tube lattice is no longer dense in terms of the modified Compton wavelength. It is the Compton wavelength of the theory that determines what is meant by ‘dense’. We therefore cannot avoid the fermion problem for dense lattice models by changing the electron mass.
Based on the results presented in figure \[fig:scalfermiratio\], we conclude that the coupling between the electron’s spin and the magnetic field do not have a dramatic effect on the vacuum energy for isolated flux tubes. Therefore, we expect that provides a good model of the underlying vacuum physics near these flux tubes, at least at the level our toy model flux tube lattice.
Flux Tube Lattice
-----------------
The worldline numerics technique computes an effective action density which is then integrated to obtain the effective action. This quantity differs from the Lagrangian in that it is not determined by local operators, but encodes information about the field everywhere through the worldline loops. Like the classical action, the 1-loop term of the effective action per unit length is infinite for a flux tube lattice because the field extends infinitely far. For this reason, we define the effective action to be the action density integrated over the region of a central flux tube $(0<\rho<a/2)$:
&=& -\_0\^[a/2]{}B\_z()\^2d-\
& &\_0\^[a/2]{}\_[cm]{} d\_[cm]{}\_0\^e\^[-m\^2T]{}\
& & { \_[\_[cm]{}]{} - 1 +(eB(\_[cm]{})T)\^2}.
The 1-loop term of the effective action density is plotted in figure \[fig:EAscffpeek\] for the cylindrical flux tube lattice model. The most pronounced feature of this density is that there is a negative contribution from the regions where the field is strong. This contribution has the same sign as the classical term. Therefore, the quantum correction tends to reinforce the classical action. A less pronounced feature is that there is a positive contribution arising from the $\rho_{\rm cm} > \lambda/2$ region, in between the lumps of magnetic field which represent the flux tubes. In this region, the local magnetic field is positive, but small.
To interpret this feature, we consider the relative contributions between the Wilson loops and the counterterm. These terms are shown in figure \[fig:CounterTermDomination\] for the constant field case. For all values of proper time, $T$, the counterterms dominate, giving an overall negative sign. In order for the action density to be positive, there must be a greater contribution from the Wilson loop average than from the counterterm, since this term tends to give a positive contribution to the action. In our flux tube model, this seems to occur in the regions between the flux tubes. In these regions, the local contribution from the counterterm is relatively small because the field is small. However, the contribution from the Wilson loop average is large because the loop cloud is exploring the nearby regions where the field is much larger. The effect is largest where the field is small, but becomes large in a nearby region. We therefore interpret the positive contributions to the 1-loop correction from these regions as a non-local effect. A similar example of such an effect from fields which vary on scales of the Compton wavelength has been observed previously using the worldline numerics technique [@PhysRevD.84.065035].
In figure \[fig:EAscffvsl\], we plot the magnitude of the 1-loop term of the effective action as a function of the flux tube width. As the flux tubes become smaller, there is an amplification of the 1-loop term, just as there is for the classical action. Similarly, for more closely spaced flux tubes, $a$ is smaller, and the 1-loop term increases in magnitude. The ratio of the 1-loop term to the classical term is plotted in figure \[fig:ActRatscff\]. The quantum contribution is greatest for closely spaced, narrow flux tubes, but does not appear to become a significant fraction of the total action.
We observe that the approximation is surprisingly good despite the fact that the magnetic field is varying rapidly on the Compton wavelength scale of the electron. We plot the residuals showing the deviations between the worldline numerics results and the approximation in figure \[fig:EAscffresid\]. To understand this, recall the discussion surrounding figure \[fig:CounterTermDomination\]. The Wilson loop term is sensitive to the average magnetic field through the loop ensemble, $\mean{B}_{\rm e}$. In contrast, the counterterm is sensitive to the magnetic field at the center of mass of the loop, $B_{\rm cm}$. Since these terms carry opposite signs, we can understand the difference from the constant field approximation in terms of a competition between these terms. When $B_{\rm cm} < \mean{B}_{\rm e}$, such as when the center of mass is in a local minimum of the field, there is a reduction of the energy relative to the locally constant field case, with a possibility of the quantum term of the energy density becoming negative. However, when $B_{\rm cm} > \mean{B}_{\rm e}$, such as in a local maximum of the field, there is an amplification of the energy relative to the constant field case. We can put a bound on the difference between the mean field through a loop and the field at the center of mass for small loops (small $T$), | - B\_[cm]{}| |B”(\_0)| T where $|B''(\rho_0)| \ge B''(\rho)$ for all $\rho$ in the loop. This expression is proved the same way as determining the error in numerical integration using the midpoint rectangle rule.
If the field varies rapidly about some mean value on the Compton wavelength scale, the various contributions from local minima and local maxima are averaged out and the mean field approximation provided by the method becomes appropriate. A similar argument applies in the fermion case, where the important quantity is the mean magnetic field along the circumference of the loop. This quantity is also well served by a mean-field approximation when integrating over rapidly varying fields.
Another interesting feature of figure \[fig:EAscffresid\] is that the approximation appears to describe narrower flux tubes better than wider ones, even when the spacing between the flux tubes is held constant. This effect is likely a result of the compact support given to the flux tube profiles. For narrow tubes, we are guaranteed to have many more center of mass points outside the flux tube than inside, giving a smaller energy contribution than for isolated flux tubes without compact support where the distinction between inside and outside is not as abrupt. This also explains why narrow, closely spaced tubes produce a lower energy than is predicted by the approximation.
This argument does not apply to the smooth isolated flux tubes given by equation (\[eqn:isoB\]). For these flux tubes, the only region where there is a large discrepancy between $\mean{B}_{\rm e}$ and $B_{\rm cm}$ is near the center of the flux tube. This is a global maximum of the field, and the only maximum of $|B''(\rho)|$. There are no regions where the average field in the loop ensemble is much stronger than the center of mass magnetic field. So, we expect an amplification of the energy near the flux tube relative to the constant field case. In the flux tubes with compact support, however, there is such a region just outside the flux tube. We can understand the surprisingly close agreement of these results to the approximation in our model in terms of competition between these regions of local minima and maxima of the field (see figure \[fig:EAscffvsconst\]).
Finally, we find that the quantum term remains small compared to the classical action for the range of parameters investigated. This is shown in figure \[fig:ActRatscff\] where we plot the ratio of the term of the action to the classical action. The relative smallness of this correction is consistent with the predictions from homogeneous fields and the derivative expansion, as well as with previous studies on flux tube configurations [@1999PhRvD..60j5019B].
Interaction Energies
--------------------
Using this model, we can study the energy associated with interactions between the flux tubes. Since the flux tubes in our model exhibit compact support, the interaction energy is entirely due to a non-local interaction between nearby flux tubes. Thus, it contrasts with previous research which has investigated the interaction energies between flux tubes which have overlapping fields [@Langfeld:2002vy]. In this case, there is a classical interaction energy ($\propto B_1 B_2$) as well as a quantum correction ($\propto (B_1 + B_2)^4 - B_1^4 - B_2^4$ in the weak-field limit). Even when these field overlap interactions are not present, there are also non-local energies in the vicinity of a flux tube due to the presence of other flux tubes. For example, the energy from nearby flux tubes can interact with a flux tube through the quantum diffusion of the magnetic field. Because of this phenomenon, we expect an interaction energy in the region of the central flux tubes due to the proximity of neighbouring flux tubes, even though no changes are made to the field profile or its derivatives in the region of interest. Since this interaction represents a force due to quantum fluctuations under the influence of external conditions, it is an example of a Casimir force. The Casimir force between two infinitely thin flux tubes in has previously been found to be attractive [@duru1993]. Our model can shed light specifically on this interaction, which is not predicted by local approximations such as the derivative expansion.
Consider a central flux tube with a width $\lambda = 0.5\lambda_e$. When $\lambda_{\rm min} = \lambda$, the magnetic field outside of the flux tube, $B_{\rm bg}$, is zero. Then, if the distance between flux tubes, $a$, is set very large, the energy density will be localized to the central flux tube and there will be no non-local interaction energy due to neighbouring flux tubes. We define the interaction energy, $E_{\rm int}$, as the difference in energy within a distance $a/2$ of the central flux tube between a configuration with a given value of $a$ and a configuration with $a = \infty$. In practice, we use $a =
10,000\lambda_e$ as a suitable stand-in for $a = \infty$: = - +. With this definition, the interaction energy is the energy associated with lowering the distance between flux rings from infinity. This the the analogue in our model of reducing the lattice spacing of the flux tubes.
One complication of this definition is that there is no clear distinction between energy density which ‘belongs’ to the central flux tube and energy density which ‘belongs’ to the neighbouring flux tubes. We continue to use our convention that the total energy for the central flux tube is determined by the integral over the non-local action density in a region within a radius of $a/2$ of the flux tube. As $a$ is taken smaller and smaller, some energy from nearby flux tubes is included within this region, but also, some energy associated with the central flux tube is diffused out of the region. This ambiguity is unavoidable within this model. We can’t numerically compute the energy over all of space and subtract off different contributions, because these energies are infinite.
The interaction energy is plotted in figure \[fig:interaction\]. In this plot, the error bars are 1-sigma error bars that account for the correlations in the means computed for each group of worldlines, \_[E\_[int]{}]{} = , where $\operatorname{Cov}(a,b)$ is the covariance between random variables $a$ and $b$. Recall from figure \[fig:EAscffpeek\] that there is a positive contribution to the effective action, and therefore a negative contribution to the energy from the region between flux tubes. As we reduce $a$, bringing the flux tubes closer together, two considerations become important. Firstly, we are increasing the average field strength meaning there tends to be more flux through the worldline loops which tends to give a negative contribution to the interaction energy. Secondly, we are reducing the spatial volume over which we integrate the energy since we only integrate $\rho$ from $0$ to $a/2$. This effect makes a positive contribution to the interaction energy since we include less and less of the region of negative energy density in our integral.
In figure \[fig:interaction\], there appears to be a landscape with both positive and negative interaction energies at different values of $a$. These appear to be consistent with the interplay between positive and negative contributions described in the previous paragraph. This is consistent with the usual expectation of attractive Casimir forces [@duru1993]. The dominant contribution for the positive energy values is caused by less of the negative energy region contributing as the domain assigned to the flux tube is reduced. However, the point at $a/\lambda_e = 2$ is negative ($-0.3\pm0.1$MeV/$\lambda_e$) indicating that an attractive interaction from nearby flux tubes is dominant.
At $a/\lambda_e = 0.5$, the flux tubes are positioned right next to one another, and the negative contribution from the increase in the mean field appears to be slightly larger than the positive contribution from the loss of the region of negative energy density from the integral. Beyond this, the flux tubes would overlap each other, which approximately corresponds to the critical background field which destroys superconductivity. Based on the above explanation, it appears that the non-local interaction energy between magnetic fields has a strong dependence on the specific profile of the classical magnetic field that was used. This makes it difficult to predict if it will result in attractive or repulsive forces in a more realistic model of a flux tube lattice.
The energy density of a critical strength magnetic field is about $17{\rm GeV}/\lambda_e^3$, so the energy associated with this interaction is relatively small. However, there are no other interactions which affect the energy of moving the flux tubes closer together when they are separated by many coherence lengths. Here, the characteristic distance associated with the interaction, $\lambda_e$, is considerably larger than coherence length or London penetration depth, so the interactions between flux tubes through the order field are heavily suppressed.
Conclusion {#sec:conclusion}
==========
In this paper, we have used the worldline numerics numerical technique to compute the effective action of in non-homogeneous, cylindrically symmetric magnetic fields. The method uses a Monte Carlo generated ensemble of worldline loops to approximate a path integral in the worldline formalism. These worldline loops are generated using a simple algorithm and encode the information about the magnetic field by computing the flux through the loop and the action acquired from transporting a magnetic moment around the loop. This technique preserves Lorentz symmetry exactly and can preserve gauge symmetry up to any required precision.
Computing the quantum effective action for magnetic flux tube configurations is a problem that has generated considerable interest and has been explored through a variety of approaches [@Gornicki1990271; @1998MPLA...13..379S; @0264-9381-12-5-013; @PhysRevD.51.810; @1999PhRvD..60j5019B; @PhysRevD.62.085024; @2001PhRvD..64j5011P; @2003PhRvD..68f5026B; @2005NuPhB.707..233G; @2006JPhA...39.6799W; @Weigel:2010pf; @Langfeld:2002vy] (see section \[sec:EAisoflux\]). Partly, this is because it is a relatively simple problem for analyzing non-homogeneous generalizations of the Heisenberg-Euler action and for exploring limitations of techniques such as the derivative expansion. But, this is also a physically important problem because tubes of magnetic flux are very important for the quantum mechanics of electrons due to the Aharonov-Bohm effect, and they appear in a variety of interesting physical scenarios such as stellar astrophysics, cosmic strings, in superconductor vortices, and quark confinement [@2003PrPNP..51....1G].
In the present context, we are concerned with the role that magnetic flux tubes play in the superconducting nuclear material of compact stars. In this scenario, the effects are particularly interesting because the magnetic flux tubes, if they exist, are confined to tubes which may be only a few percent of the Compton wavelength, $\lambda_{\rm C}$, in radius. Specifically, the flux must be confined to within the London penetration depth of the superconducting material, which for neutron stars has been estimated to be 80 fm $=$ 0.032 $\lambda_e$ [@PhysRevLett.91.101101]. Moreover, the flux tube density is expected to be proportional to the average magnetic field. For a background field near the quantum critical strength, $B_k$, such as in a neutron star, the distance between flux tubes is comparable to a Compton wavelength. This Compton wavelength scale is also the scale at which the non-locality of becomes important and at which powerful local techniques like the derivative expansion are no longer appropriate for computing the effective action.
The free energy associated with these flux tubes is a factor in determining whether the nuclear material of a neutron star is a type-I or type-II superconductor. The free energy of a flux tube is determined by looking at the energies associated with the magnetic field, with the creation of a non-superconducting region in the superconductor, and with interactions between the flux tubes. Flux tubes can only form if it is energetically favourable to do so compared to expelling the field due to the Meissner effect. For a lattice of flux tubes, there is also an energy contribution from the presence of neighbouring flux tubes because of the non-local nature of quantum field theory.
The energy of two flux tubes has been previously computed using worldline numerics methods and for flux tubes with aligned fields, the energy is larger than twice the energy of a single flux tube when the flux tubes are closely spaced [@Langfeld:2002vy]. This result implies that there is a repulsive interaction between the flux tubes due to effects, strengthening the likelihood of the type-II scenario in neutron stars. This interaction energy increases as the flux tubes are placed closer together, and can have a similar magnitude as the correction to the energy when the flux tubes are closely spaced.
We have developed a cylindrically symmetric magnetic field model which reproduces some of the features of a flux tube lattice: for a given central flux tube, there are nearby regions of large magnetic field that interact non-locally, and the large flux tube size limit goes to a large uniform magnetic field instead of to zero field. We have investigated the 1-loop effects from in this model using the worldline numerics technique for various combinations of flux tube size, $\lambda$, and flux tube spacing, $a$.
In contrast to isolated flux tubes, we find that there are some regions where the worldline numerics results are greater than the approximation and other regions where they are less than the approximation. This can be understood by thinking of the difference from the approximation as a competition between the local counterterm and the Wilson loop averages. For magnetic fields that vary on the Compton wavelength scale about some mean field strength, the approximation provides a poor approximation of the energy density, but may provide a good approximation to the total energy density of the field due to it being a good mean field theory approximation to the energy density. The appropriateness of the approximation in this case can be understood as an approximate balance between regions where the field is a local maximum and the magnitude of the quantum correction to the action density is larger than in the constant field case, and regions where the field is a local minimum and the quantum corrections predict a smaller action density than the constant field case. This washing out of the field structure due to non-local effects has also been observed in worldline numerics studies of the vacuum polarization tensor [@PhysRevD.84.065035].
There is a force between nearby magnetic flux tubes due to the quantum diffusion of the energy density. This interaction is non-local and is not predicted by the local derivative expansion. It is an example of a Casimir force (a force resulting from quantum vacuum fluctuations) and it is computed in a very similar way as the Casimir force between conducting bodies in the worldline numerics technique [@Gies:2003cv]. The size of the energy densities involved in this force are small even compared to the 1-loop corrections to the energy densities, which are in turn small compared to the classical magnetic energy density.
Although this interaction energy is small, the interactions between flux tubes in a neutron star due to the order field of the superconductor are suppressed because the distance between the tubes is considerably larger than the coherence length and London penetration depth. Therefore, this force is possibly important for the behaviour of flux tubes in neutron star crusts and interiors. For example, in our lattice model, this force could contribute to a bunching of the worldlines, producing regions where flux tubes are separated by $\sim2\lambda_e$ and other regions which have no flux tubes. Consequently, this force may have important implications for neutron star physics. However, investigating these implications is outside the scope of this paper.
The nature of this interaction energy is expected to depend on the model of the magnetic field profile for the reasons discussed in section \[sec:periodic\_results\]. It is therefore reasonable that forces of either sign, attractive or repulsive, may be possible depending on the particular landscape formed by the magnetic field and the particular definition used of the interaction energy. In a superconductor, the profiles of the magnetic flux tubes are determined by the minimization of the free energy for the interacting system formed by the magnetic and order parameter fields. Therefore, investigating this phenomenon using more realistic models (two-dimensional triangular lattices with field profiles motivated by the physics of superconductors) is an interesting direction for future research. In particular, it would be interesting to determine if certain conditions allowed for a non-local interaction between magnetic flux tubes to be experimentally observable despite the small forces involved.
These conclusions are directly applicable to . However, we have also investigated the relationship between spinor and scalar for isolated flux tubes where the worldline numerics technique can be applied to both cases. We find that both theories have the same qualitative behaviour, and agree within a factor of order unity quantitatively. The arguments and explanations given for the results have strong parallels in the spinor case. The spinor case can also be understood in terms of a competition between the Wilson loop averages and the local counterterm. We therefore speculate that the results from this paper will hold in the spinor case, at least qualitatively. However, addressing the fermion problem so that the spinor case can be studied explicitly for flux tube lattices would be valuable progress in this area of research.
This work was supported by the Natural Sciences and Engineering Research Council of Canada, the Canadian Foundation for Innovation, the British Columbia Knowledge Development Fund. It has made used of the NASA ADS and arXiv.org.
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits